page i

BUSINESS ANALYTICS
Communicating with Numbers
 page ii

The McGraw-Hill Series in Operations and

Decision Sciences
SUPPLY CHAIN MANAGEMENT

Bowersox, Closs, Cooper, and Bowersox

Supply Chain Logistics Management
Fifth Edition

Johnson
Purchasing and Supply Management
Sixteenth Edition

Simchi-Levi, Kaminsky, and Simchi-Levi

Designing and Managing the Supply Chain: Concepts,
Strategies, Case Studies
Third Edition

Stock and Manrodt

Supply Chain Management

PROJECT MANAGEMENT

Larson
Project Management: The Managerial Process
Eighth Edition

SERVICE OPERATIONS MANAGEMENT

Bordoloi, Fitzsimmons, and Fitzsimmons

Service Management: Operations, Strategy, Information
Technology
Ninth Edition

MANAGEMENT SCIENCE

Hillier and Hillier

Introduction to Management Science: A Modeling and Case
Studies Approach with Spreadsheets
Sixth Edition

BUSINESS RESEARCH METHODS

Schindler
Business Research Methods
Thirteenth Edition

BUSINESS FORECASTING

Keating and Wilson

Forecasting and Predictive Analytics
Seventh Edition

BUSINESS SYSTEMS DYNAMICS

Sterman
Business Dynamics: Systems Thinking and Modeling for a
Complex World

OPERATIONS MANAGEMENT

Cachon and Terwiesch

Matching Supply with Demand: An Introduction to Operations
Management
Fourth Edition

Cachon and Terwiesch

Operations Management
Second Edition

Jacobs and Chase

Operations and Supply Chain Management
Sixteenth Edition

Jacobs and Chase

Operations and Supply Chain Management: The Core
Fifth Edition

Schroeder and Goldstein

Operations Management in the Supply Chain: Decisions and
Cases
Eighth Edition

Stevenson
Operations Management
Fourteenth Edition

Swink, Melnyk, and Hartley

Managing Operations Across the Supply Chain
Fourth Edition

BUSINESS STATISTICS

Bowerman, Drougas, Duckworth, Froelich, Hummel, Moninger, and

Schur
Business Statistics and Analytics in Practice
Ninth Edition
Doane and Seward
Applied Statistics in Business and Economics
Sixth Edition

Doane and Seward

Essential Statistics in Business and Economics
Third Edition

Jaggia and Kelly

Business Statistics: Communicating with Numbers
Third Edition

Jaggia and Kelly

Essentials of Business Statistics: Communicating with
Numbers
Second Edition

Lind, Marchal, and Wathen

Basic Statistics for Business and Economics
Ninth Edition

Lind, Marchal, and Wathen

Statistical Techniques in Business and Economics
Eighteenth Edition

McGuckian
Connect Master: Business Statistics
 page iii

BUSINESS ANALYTICS
Communicating with Numbers

Sanjiv Jaggia
California Polytechnic State University

Alison Kelly
Suffolk University

Kevin Lertwachara
California Polytechnic State University
Leida Chen
California Polytechnic State University
 page iv

BUSINESS ANALYTICS

Published by McGraw-Hill Education, 2 Penn Plaza, New York, NY 10121. Copyright ©

2021 by McGraw-Hill Education. All rights reserved. Printed in the United States of America.
No part of this publication may be reproduced or distributed in any form or by any means, or
stored in a database or retrieval system, without the prior written consent of McGraw-Hill
Education, including, but not limited to, in any network or other electronic storage or
transmission, or broadcast for distance learning.

Some ancillaries, including electronic and print components, may not be available to
customers outside the United States.

This book is printed on acid-free paper.

1 2 3 4 5 6 7 8 9 LWI 24 23 22 21 20

ISBN 978-1-260-57601-6
MHID 1-260-57601-9

Cover Image: People head logo Nietz_studio/Shutterstock

All credits appearing on page or at the end of the book are considered to be an extension of
the copyright page.

The Internet addresses listed in the text were accurate at the time of publication. The
inclusion of a website does not indicate an endorsement by the authors or McGraw-Hill
Education, and McGraw-Hill Education does not guarantee the accuracy of the information
presented at these sites.

mheducation.com/highered
 page v

Dedicated to our
families
 page vi

ABOUT THE AUTHORS

Sanjiv Jaggia

Courtesy Sanjiv Jaggia

Sanjiv Jaggia is a professor of economics and finance at Californ

Polytechnic State University in San Luis Obispo. Dr. Jaggia holds
Ph.D. from Indiana University and is a Chartered Financial Analy
(CFA®). He enjoys research in statistics and econometrics applied to
wide range of business disciplines. Dr. Jaggia has published sever
papers in leading academic journals and has co-authored tw
successful textbooks in business statistics. His ability to communicat
in the classroom has been acknowledged by several teaching award
Dr. Jaggia resides in San Luis Obispo with his wife and daughter.
his spare time, he enjoys cooking, hiking, and listening to a wide rang
of music.
Alison Kelly

Courtesy Alison Kelly

Alison Kelly is a professor of economics at Suffolk University

Boston. Dr. Kelly holds a Ph.D. from Boston College and is
Chartered Financial Analyst (CFA®). Dr. Kelly has published in a wid
variety of academic journals and has co-authored two successf
textbooks in business statistics. Her courses in applied statistics an
econometrics are well received by students as well as workin
professionals. Dr. Kelly resides in Hamilton, Massachusetts, with he
husband, daughter, and son. In her spare time, she enjoys exercisin
and gardening.

page vi

Kevin Lertwachara
Teresa Cameron/Frank Gonzales/California Polytechnic State University

Kevin Lertwachara is a professor of information systems at Californ

Polytechnic State University in San Luis Obispo. Dr. Lertwachar
holds a Ph.D. in Operations and Information Management from th
University of Connecticut. Dr. Lertwachara’s research focuses o
technology-based innovation, electronic commerce, health car
informatics, and business analytics and his work has been publishe
in scholarly books and leading academic journals. He teache
business analytics at both the undergraduate and graduate levels an
has received several teaching awards. Dr. Lertwachara resides in th
central coast of California with his wife and three sons. In his spar
time, he coaches his sons’ soccer and futsal teams.

Leida Chen
Courtesy of Leida Chen

Leida Chen is a professor of information systems at Californ

Polytechnic State University in San Luis Obispo. Dr. Chen earned
Ph.D. in Management Information Systems from University o
Memphis. His research and consulting interests are in the areas o
business analytics, technology diffusion, and global informatio
systems. Dr. Chen has published over 50 research articles in leadin
information systems journals, over 30 articles and book chapters
national and international conference proceedings and edited book
and a book on mobile application development. He teaches busines
analytics at both the undergraduate and graduate levels. In his spar
time, Dr. Chen enjoys hiking, painting, and traveling with his wife an
son to interesting places around the world.
 page vii

FROM THE AUTHORS

Data and analytics capabilities have made a leap forward in recen
years and have changed the way businesses make decisions. Th
explosion in the field is partly due to the growing availability of va
amounts of data, improved computational power, and the developmen
of sophisticated algorithms. More than ever, colleges and universitie
need a curriculum that emphasizes business analytics, and companie
need data-savvy professionals who can turn data into insights an
action. To meet these needs, business colleges have been scramblin
to include business analytics courses in their undergraduat
curriculum. At the graduate level, these courses are fast becoming
requirement in MBA programs and in specialized masters an
certificate programs in business analytics.
We wrote Business Analytics: Communicating with Numbers from
the ground up to prepare students to understand, manage, an
visualize the data; apply the appropriate analysis tools; an
communicate the findings and their relevance. Unlike other texts tha
simply repackage traditional statistics and operations research topics
our text seamlessly threads the topics of data wrangling, descriptiv
analytics, predictive analytics, and prescriptive analytics into
cohesive whole.
Experiential learning opportunities have been proven effective
teaching applied and complex subjects such as business analytics.
this text, we provide a holistic analytics process, including dealing wit
real-life data that are not necessarily ’clean’ and/or ’small’. Similarl
we stress the importance of effectively communicating findings b
including features such as a synopsis (a short writing sample) and
sample report (a longer writing sample) in every chapter. Thes
features help students develop skills in articulating the business valu
of analytics by communicating insights gained from a nontechnic
standpoint.
The text’s comprehensive coverage of all relevant busines
analytics topics gives instructors the flexibility to select the topics tha
best align with their course objectives. All chapters are designed to b
modular, making it easy for instructors to include only the ke
chapters, or sections within a chapter, they intend to cover. Fo
example, the text can be adapted for an introductory course coverin
data wrangling (Chapter 2), data visualization (Chapter 3), summa
measures (Chapter 3), statistical theories (Chapters 4 and 5), an
regression analysis (Chapters 6 and 7), or an advanced busines
analytics course focusing on predictive (Chapters 8, 9, 10, 11, and 12
and prescriptive modeling (Chapter 13).

page ix

In this text, we have chosen Excel with Analytic Solver (an Exc
add-in) and R as the software packages due to their accessibility, eas
of use, and powerful features for demonstrating analytics concep
and performing analytics on real-world data. In most chapters, th
instructor can select to cover either Excel or R (or both) based on th
course objectives and the technical background of the students. Thes
software packages can be accessible to students through universi
computer labs, educational licenses, and free, open-source license
For this edition, all examples and exercise problems are solved usin
the latest versions of the software as of writing, namely, Microso
Office Professional 2016, Analytic Solver 2019, and R version 3.5.3
We recommend that the same versions of the software be used
order to replicate the results presented in the text. When new version
of the software packages are released in the future, we plan t
incorporate any substantial changes in future editions of the text o
provide updates online if the differences are relatively minor.

Sanjiv Jagg
Alison Kel
Kevin Lertwachar
Leida Che
 page x

Business Analytics: Communicating

with Numbers
Reviewer Quotes

“I love everything I see! I love the

application demonstrated through the case
study in each chapter, and I like the use of
Excel and Solver, specifically. . . I also love
the case study and examples to help
students apply the material.”

Kristin Pettey, Southwestern College –

Kansas

“I can’t agree with the approach. . . to data

analytics more. I have been looking for a
textbook like this.”

Jahyun Goo, Florida Atlantic University

“End of chapter material is excellent
(“Writing with Big Data”). . .”

Kevin Brown, Asbury University

“I strongly applaud that this text offers a

holistic approach to data analysis and
places the emphasis on communicating
with data. The Big Data and Data Mining
sections seem promising with coverage of
topics of the latest models and methods.”

Hao Chen, UW Platteville

“The TOC includes all the major areas

needed for a foundational level of
knowledge and the added value of
teaching how to communicate with the
information garnered will make a strong
textbook.”

Roman Rabinovich, Boston University

 page xi

Unique Key Features

The pedagogy of Business Analytics includes popular features from
our Business Statistics texts along with new features pertaining to
business analytics. Countless reviewers have added their feedback
and direction to ensure we have built a product that we believe
addresses the needs of the market.

Holistic Approach to Data Analytics

Business analytics is a very broad topic consisting of statistics,
computer science, and management information systems with a
wide variety of applications in business areas including marketing,
HR management, economics, accounting, and finance.
The text offers a holistic approach to business analytics,
combining qualitative reasoning with quantitative tools to identify key
business problems and translate analytics into decisions that
improve business performance.

Integrated Introductory Case

Each chapter opens with a real-life case study that forms the basis
for several examples within the chapter. The questions included in
the examples create a roadmap for mastering the most important
learning outcomes within the chapter. A synopsis of each chapter’s
introductory case is presented once the questions pertaining to the
case have been answered.

page xii

Writing with Big Data

A distinctive feature of Business Analytics is access to select big
data sets with relevance to numerous applications to which students
can relate. In most chapters, we have a designated section where
we use these big data sets to help introduce problems, formulate
possible solutions, and communicate the findings, based on the
concepts introduced in the chapter. Using a sample report, our intent
is to show students how to articulate the business value of analytics
by communicating insights gained from a nontechnical standpoint.
 page xiii

Emphasis on Data Mining

Data mining is one of the most sought-after skills that employers
want college graduates to have. It leverages large data sets and
computer power to build predictive models that support decision
making. In addition to two comprehensive chapters devoted to linear
and logistic regression models, and a chapter on business
forecasting, the text includes four exclusive chapters on data mining.
These include detailed analysis of both supervised and unsupervised
learning, covering relevant topics such as principle component
analysis, k-nearest neighbors, naïve Bayes, classification and
regression trees, ensemble trees, hierarchical and k-means
clustering, and association rules. Each chapter offers relatable real-
world problems, conceptual explanations, and easy-to-follow
computer instructions. There are over 200 exercises in these four
exclusive chapters.

FIGURE 8.5 The cumulative lift chart

Four Chapters on Data Mining

Introduction to Data Mining
Supervised Data Mining: k-Nearest Neighbors and Naïve Bayes
Supervised Data Mining: Decision Trees
Unsupervised Data Mining
TABLE 8.14 Prediction Performance Measures

Performance Measure Model 1 Model 2

RMSE 171.3489 174.1758

ME  11.2530  12.0480

MAD 115.1650 117.9920

MPE −2.05% −2.08%

MAPE 15.51% 15.95%

FIGURE 10.1 A simplified decision tree

page xiv

Computer Software
The text includes hands-on tutorials and problem-solving examples
featuring Microsoft Excel, Analytic Solver (an Excel add-in software
for data mining analysis), as well as R (a powerful software that
merges the convenience of statistical packages with the power of
coding).
Throughout the text, students learn to use the software to solve
real-world problems and to reinforce the concepts discussed in the
chapters. Students will also learn how to visualize and communicate
with data using charts and infographics featured in the software.

page xv

Exercises and McGraw-Hill’s Connect®

Every chapter contains dozens of applied examples from all walks of
life, including business, economics, sports, health, housing, the
environment, polling, psychology, and more.
We also know the importance of ancillaries—like the Instructor’s
Solution Manual (ISM)—and the technology component, specifically
Connect. As we write Business Analytics, we are simultaneously
developing these components with the hope of making them
seamless with the text itself.
We know from experience that these components cannot be
developed in isolation. For example, we review every Connect
exercise as well as evaluate rounding rules and revise tolerance
levels. Given the extremely positive feedback from users of our
Business Statistics texts, we follow the same approach with
Business Analytics.
 page xvi

FOR INSTRUCTORS

You’re in the driver’s seat.

Want to build your own course? No problem. Prefer to use
our turnkey, prebuilt course? Easy. Want to make changes
throughout the semester? Sure. And you’ll save time with
Connect’s auto-grading too.
They’ll thank you for it.
Adaptive study resources like SmartBook® 2.0 help your
students be better prepared in less time. You can transform
your class time from dull definitions to dynamic debates. Find
out more about the powerful personalized learning
experience available in SmartBook 2.0 at
www.mheducation.com/highered/connect/smartbook

Laptop: McGraw-Hill; Woman/dog: George Doyle/Getty Images

Make it simple, make it

affordable.

Connect makes it easy with seamless integration using

any of the major Learning Management Systems—
Blackboard®, Canvas, and D2L, among others—to let you
organize your course in one convenient location. Give your
students access to digital materials at a discount with our
inclusive access program. Ask your McGraw-Hill
representative for more information.

Padlock: Jobalou/Getty Images

Solutions for your challenges.

A product isn’t a solution. Real solutions are affordable,

reliable, and come with training and ongoing support when
you need it and how you want it. Our Customer Experience
Group can also help you troubleshoot tech problems—
although Connect’s 99% uptime means you might not
need to call them. See for yourself at
status.mheducation.com

Checkmark: Jobalou/Getty Images

page xvii
 FOR STUDENTS

Effective, efficient studying.

Connect helps you be more productive with your study time
and get better grades using tools like SmartBook 2.0, which
highlights key concepts and creates a personalized study
plan. Connect sets you up for success, so you walk into class
with confidence and walk out with better grades.

Study anytime, anywhere.

Download the free ReadAnywhere app and access your
online eBook or SmartBook 2.0 assignments when it’s
convenient, even if you’re offline. And since the app
automatically syncs with your eBook and SmartBook 2.0
assignments in Connect, all of your work is available
every time you open it. Find out more at
www.mheducation.com/readanywhere

“I really liked this app—it made it easy to

study when you don't have your textbook
in front of you.”
- Jordan Cunningham,
Eastern Washington University

No surprises.
The Connect Calendar and Reports tools keep you on track
with the work you need to get done and your assignment
scores. Life gets busy; Connect tools help you keep learning
through it all.

Calendar: owattaphotos/Getty Images

Learning for everyone.

McGraw-Hill works directly with Accessibility Services
Departments and faculty to meet the learning needs of all
students. Please contact your Accessibility Services office
and ask them to email accessibility@mheducation.com, or
visit
www.mheducation.com/about/accessibility
for more information.

Top: Jenner Images/Getty Images, Left: Hero Images/Getty Images, Right: Hero Images/Getty Images
 page xviii

Resources for Instructors and

Students
Instructor Library
The Connect Instructor Library is your repository for additional
resources to improve student engagement in and out of class. You
can select and use any asset that enhances your course. The
Connect Instructor Library includes:
Instructor’s Manual
Instructor’s Solutions Manual
Test Bank
Data Sets
PowerPoint Presentations
Digital Image Library

R Package
R is a powerful software that merges the convenience of statistical
packages with the power of coding. It is open source as well as
cross-platform compatible and gives students the flexibility to work
with large data sets using a wide range of analytics techniques. The
software is continuously evolving to include packages that support
new analytical methods. In addition, students can access rich online
resources and tap into the expertise of a worldwide community of R
users. In Appendix C, we introduce some fundamental features of R
and also provide instructions on how to obtain solutions for many
solved examples in the text.
 As with other texts that use R, differences between software
versions are likely to result in minor inconsistencies in analytics
outcomes in algorithm-rich Chapters 9, 10, and 11. In these
chapters, the solved examples and exercise problems are based on
R version 3.5.3 on Microsoft Windows. In order to replicate the
results with newer versions of R, we suggest a line of code in these
chapters that sets the random number generator to the one used on
R version 3.5.3.

Analytic Solver
The Excel-based user interface of Analytic Solver reduces the
learning curve for students allowing them to focus on problem
solving rather than trying to learn a new software package. The
solved examples and exercise problems are based on Analytic
Solver 2019. Newer versions of Analytic Solver will likely produce the
same analysis results but may have a slightly different user interface.
For consistency, we recommend that you use Analytic Solver 2019
with this text.
Analytic Solver can be used with Microsoft Excel for Windows (as
an add-in), or “in the cloud” at AnalyticSolver.com using any device
(PC, Mac, tablet) with a web browser. It offers comprehensive
features for prescriptive analytics (optimization, simulation, decision
analysis) and predictive analytics (forecasting, data mining, text
mining). Its optimization features are upward compatible from the
standard Solver in Excel. Note: Analytic Solver software is not free
with the purchase of this text, but low-cost licenses are available. If
interested in having students get low-cost academic access for class
use, instructors should send an email to support@solver.com to get
their course code and receive student pricing and access information
as well as their own access information.

page xix

Student Resources
Students have access to data files, tutorials, and detailed progress
reporting within Connect. Key textbook resources can also be
accessed through the Additional Student Resources page:
www.mhhe.com/JaggiaBA1e.

McGraw-Hill Customer Care

Contact Information
At McGraw-Hill, we understand that getting the most from new
technology can be challenging. That’s why our services don’t stop
after you purchase our products. You can e-mail our product
specialists 24 hours a day to get product training online. Or you can
search our knowledge bank of frequently asked questions on our
support website.
For customer support, call 800-331-5094 or visit
www.mhhe.com/support. One of our technical support analysts will
be able to assist you in a timely fashion.
 page xx

ACKNOWLEDGMENTS
We would like to thank the following instructors for their feedback
and careful reviews during the development process:
Sung K. Ahn, Washington State University
Triss Ashton, Tarleton State University
Anteneh Ayanso, Brock University
Matthew Bailey, Bucknell University
Palash Beram Saint Louis University
Matthew Brenn, University of Akron
Paul Brooks, Virginia Commonwealth University
Kevin Brown, Asbury University
Kevin Caskey, SUNY New Paltz
Paolo Catasti, Virginia Commonwealth University
Michael Cervetti, University of Memphis
Jimmy Chen, Bucknell University
Jen-Yi Chen, Cleveland State University
Rex Cheung, San Francisco State University
Alan Chow, University of South Alabama,
Matthew Dean, University of Southern Maine
Alisa DiSalvo, St. Mary’s University Twin Cities
Mark Dobeck, Cleveland State University
Joan Donohue, University of South Carolina
Kathy Enget, University at Albany- SUNY
Kathryn Ernstberger, Indiana University Southeast
Robertas Gabrys, University of Southern California
Ridvan Gedik, University of New Haven
Roger Grinde, University of New Hampshire
Thomas Groleau, Carthage College
Babita Gupta, California State University Monterey Bay
Serina Al Haddad, Rollins College
Dan Harpool, University of Arkansas Little Rock
Fady Harfoush, Loyola University Chicago
Paul Holmes, Ashland University
Ping-Hung Hsieh, Oregon State University
Kuang-Chung Hsu, University of Central Oklahoma
Jason Imbrogno, University of Northern Alabama
Marina Johnson, Montclair State University
Jerzy Kamburowski, University of Toledo
Reza Kheirandish, Clayton State University
Esther Klein, St. Francis College
Bharat Kolluri, University of Hartford
Mohammad Merhi, Indiana University
Jacob Miller, Southern Utah University
Sinjini Mitra, California State University Fullerton
Kyle Moninger, Bowling Green State University
Rex Moody, Angelo State University
Ebrahim Mortaz, Pace University
Kristin Pettey, Southwestern College
Daniel Power, University of Northern Iowa
Zbigniew Przasnyski, Loyola Marymount University
Sharma Pillutla, Towson University
Roman Rabinovich, Boston University
Michael Ratajczyk, Saint Mary’s College
R. Christopher L. Riley, Delta State University
Leslie Rush, University of Hawaii West Oahu
Avijit Sarkar, University of Redlands
Dmitriy Shaltayev, Christopher Newport University
Mike Shurden, Lander University
Pearl Steinbuch, Boston University
Alicia Strandberg, Villanova University School of Business
David Taylor, Sacred Heart University
Stan Taylor, California State University Sacramento
Pablo Trejo, Miami Dade College
Nikhil Varaiya, San Diego State University
Timothy Vaughan, University of Wisconsin Eau Claire
Fen Wang, Central Washington University
Kanghyun Yoon, University of Central Oklahoma
John Yu, Saint Joseph’s University
Oliver Yi, San Jose State University
Liang Xu, Slippery Rock University of Pennsylvania
Yong Xu, Ferris State University
Jay Zagorsky, Boston University

page xx

We are grateful for the assistance of the following people wh

helped us develop and review the assessment content and addition
resources:
Anteneh Ayanso, Brock University
John Draper, The Ohio State University
Alex Firsov, Oracle Corporation
Stephanie Henry, Workday, Inc.
Phuong Lam (Rose) Hoang, California Polytechnic State University
Eric Lee Kambestad, California Polytechnic State University
Vincent Mastuntano, Suffolk University
Corey Pang, California Polytechnic State University
Kristen Pettey, Southwestern College
Matthew Spector, California Polytechnic State University
Wendi Sun, Rockland Trust
Zhaowei (Justin) Wang, Citizens Financial Group
Brie Winkles, University of West Alabama

In addition, special thanks to the wonderful and supportive team a

McGraw-Hill: Ryan McAndrews, Product Developer; Thomas Finn
Content Development Editor; Noelle Bathurst, Portfolio Manage
Harper Christopher, Executive Marketing Manager; Sarah Thoma
Market Development Manager; Pat Frederickson, Core Conte
Project Manager; Jamie Koch, Assessment Content Project Manage
Egzon Shaqiri, Designer; Chuck Synovec, Portfolio Director; Natal
King, Marketing Director; Kevin Moran, Director of Digital Content; an
Tim Vertovec, Managing Director.
 page xxi

BRIEF CONTENTS
CHAPTER  1 Introduction to Business Analytics 2
CHAPTER  2 Data Management and Wrangling 30
CHAPTER  3 Data Visualization and Summary Measures 80
CHAPTER  4 Probability and Probability Distributions 138
CHAPTER  5 Statistical Inference 176
CHAPTER  6 Regression Analysis 210
CHAPTER  7 Advanced Regression Analysis 258
CHAPTER  8 Introduction to Data Mining 316
CHAPTER  9 Supervised Data Mining: k-Nearest Neighbors an
Naïve Bayes 366
CHAPTER 10 Supervised Data Mining: Decision Trees 410
CHAPTER 11 Unsupervised Data Mining 474
CHAPTER 12 Forecasting with Time Series Data 518
CHAPTER 13 Introduction to Prescriptive Analytics 560

APPENDIXES
Appendix A  Big Data Sets: Variable Description and Dat
Dictionary 608
Appendix B  Getting Started with Excel and Excel Add-Ins 614
Appendix C  Getting Started with R 621
Appendix D  Statistical Tables 628
Appendix E  Answers to Selected Exercises 632

Index 652
 page xxii

CONTENTS
CHAPTER 1

INTRODUCTION TO BUSINESS ANALYTICS 

1.1 Overview of Business Analytics 4
Important Business Analytics Applications 6

1.2 Types of Data 8
Sample and Population Data 8
Cross-Sectional and Time Series Data 9
Structured and Unstructured Data 10
Big Data 12

1.3 Variables and Scales of Measurement 15

The Measurement Scales 16

1.4 Data Sources and File Formats 20

Fixed-Width Format 21
Delimited Format 22
eXtensible Markup Language 22
HyperText Markup Language 23
JavaScript Object Notation 23

1.5 Writing With Big Data 26

CHAPTER 2
DATA MANAGEMENT AND WRANGLING 30
2.1 Data Management 32
Data Modeling: The Entity-Relationship Diagram 33
Data Retrieval in the Database Environment 35
Data Warehouse and Data Mart 36

2.2 Data Inspection 39

2.3 Data Preparation 45
Handling Missing Values 46
Subsetting 50

2.4 Transforming Numerical Data 57

Binning 58
Mathematical Transformations 62

2.5 Transforming Categorical Data 69

Category Reduction 69
Dummy Variables 72
Category Scores 74

2.6 Writing With Big Data 77

CHAPTER 3

DATA VISUALIZATION AND SUMMARY

MEASURES 80
3.1 Methods to Visualize Categorical and Numerical Variables 
82
Methods to Visualize a Categorical Variable 82
Methods to Visualize a Numerical Variable 86
 Cautionary Comments When Constructing or Interpreting Charts
or Graphs 91

3.2 Methods to Visualize the Relationship Between Two

Variables 96
Methods to Visualize the Relationship between Two Categorical
Variables 96
A Method to Visualize the Relationship between Two Numerical
Variables 101

3.3 Other Data Visualization Methods 105

A Scatterplot with a Categorical Variable 105
A Bubble Plot 107
A Line Chart 108
A Heat Map 109
Options for Advanced Visualizations 112

3.4 Summary Measures 114
Measures of Central Location 114
Measures of Dispersion 119
Measures of Shape 121
Measures of Association 123

3.5 Detecting Outliers 127
A Boxplot 127
z-Scores 129

3.6 Writing With Big Data 133

CHAPTER 4

PROBABILITY AND PROBABILITY

DISTRIBUTIONS 138
4.1 Probability Concepts and Probability Rules 140
Events 140
Assigning Probabilities 141
Rules of Probability 142

4.2 The Total Probability Rule and Bayes’ Theorem 148

The Total Probability Rule and Bayes’ Theorem 148
Extensions of the Total Probability Rule and Bayes’ Theorem 
150

4.3 Random Variables and Discrete Probability Distributions 

153
The Discrete Probability Distribution 154
Summary Measures of a Discrete Random Variable 154

4.4 The Binomial and the Poisson Distributions 156

The Binomial Distribution 156
The Poisson Distribution 159
Using Excel and R to Obtain Binomial and Poisson
Probabilities 161

4.5 The Normal Distribution 164

The Standard Normal Distribution 165
The Transformation of Normal Random Variables 167
Using Excel and R for the Normal Distribution 170

4.6 Writing With Data 172

page xxiv

CHAPTER 5

STATISTICAL INFERENCE 176
5.1 Sampling Distributions 178
 The Sampling Distribution of the Sample Mean 178
The Sampling Distribution of the Sample Proportion 182

5.2 Estimation 185
Confidence Interval for the Population Mean μ 186
Using Excel and R to Construct a Confidence Interval for μ 18
Confidence Interval for the Population Proportion p 190

5.3 Hypothesis Testing 194
Hypothesis Test for the Population Mean μ 197
Using Excel and R to Test μ 201
Hypothesis Test for the Population Proportion p 203

5.4 Writing With Data 207

CHAPTER 6

REGRESSION ANALYSIS 210
6.1 The Linear Regression Model 212
The Components of the Linear Regression Model 212
Estimating a Linear Regression Model with Excel or R 216
Categorical Variables with Multiple Categories 218

6.2 Model Selection 222
The Standard Error of the Estimate, se 223
The Coefficient of Determination, R2 224
The Adjusted R2 225
One Last Note on Goodness-of-Fit Measures 227

6.3 Tests of Significance 229
Test of Joint Significance 229
Test of Individual Significance 231
A Test for a Nonzero Slope Coefficient 233
 Reporting Regression Results 235

6.4 Model Assumptions and Common Violations 241

Required Assumptions of Regression Analysis 241
Common Violation 1: Nonlinear Patterns 243
Common Violation 2: Multicollinearity 244
Common Violation 3: Changing Variability 246
Common Violation 4: Correlated Observations 247
Common Violation 5: Excluded Variables 249
Summary 249
Using Excel and R to Construct Residual Plots 249

6.5 Writing With Big Data 253

CHAPTER 7

ADVANCED REGRESSION ANALYSIS 258

7.1 Regression Models with Interaction Variables 260
The Interaction of Two Dummy Variables 260
The Interaction of a Dummy Variable and a Numerical Variable 
262
The Interaction of Two Numerical Variables 264

7.2 Regression Models for Nonlinear Relationships 272

The Quadratic Regression Model 272
Regression Models with Logarithms 276
The Log-Log Regression Model 277
The Logarithmic Regression Model 278
The Exponential Regression Model 279
Comparing Linear and Log-Transformed Regression Models 
282
Using Excel and R to Compare Linear and Log-Transformed
Regression Models 283
7.3 Linear Probability and Logistic Regression Models 288
The Linear Probability Model 289
The Logistic Regression Model 290
Accuracy of Binary Choice Models 293
Using Analytic Solver and R to Estimate the Logistic Regression
Model 294

7.4 Cross-Validation Methods 300
The Holdout Method 301
Using Analytic Solver and R for the Holdout Method for the
Logistic Regression Model 305
The k-Fold Cross-Validation Method 306

7.5 Writing With Big Data 309

APPENDIX 7.1: The Caret Package in R for the k-fold Cross-

Validation Method 313

CHAPTER 8

INTRODUCTION TO DATA MINING 316

8.1 Data Mining Overview 318
The Data Mining Process 318
Supervised and Unsupervised Data Mining 321

8.2 Similarity Measures 322
Similarity Measures for Numerical Data 323
Similarity Measures for Categorical Data 327

8.3 Performance Evaluation 332
Data Partitioning 332
Oversampling 333
Performance Evaluation in Supervised Data Mining 334
 Performance Evaluation for Classification Models 335
Using Excel to Obtain the Confusion Matrix and Performance
Measures 337
Selecting Cut-off Values 338
Performance Charts for Classification 340
Using Excel to Obtain Performance Charts for Classification 
344
Performance Evaluation for Prediction 345
Using Excel to Obtain Performance Measures for Prediction 
347

8.4 Principal Component Analysis 352

Using Analytic Solver and R to Perform Principal Component
Analysis 355

8.5 Writing With Big Data 364

CHAPTER 9

SUPERVISED DATA MINING: k-NEAREST

NEIGHBORS AND NAÏVE BAYES 366
9.1 Introduction to Supervised Data Mining 368
Comparison of Supervised Data Mining Techniques 369

9.2 The k-Nearest Neighbors Method 370

9.3 The Naïve Bayes Method 390

Transforming Numerical into Categorical Values 398

9.4 Writing With Big Data 404

page xxv
CHAPTER 10

SUPERVISED DATA MINING: DECISION

TREES 410
10.1 Introduction to Classification and Regression Trees
(CART) 412
Classification and Regression Trees (CART) 413

10.2 Classification Trees 415
Using Analytic Solver and R to Develop a Classification Tree 
422

10.3 Regression Trees 440
Using Analytic Solver and R to Develop a Prediction Tree 443

10.4 Ensemble Tree Models 454

Using Analytic Solver and R to Develop Ensemble Classification
Tree Models 456

10.5 Writing With Big Data 469

CHAPTER 11

UNSUPERVISED DATA MINING 474

11.1 Hierarchical Cluster Analysis 476
Hierarchical Cluster Analysis 476
Agglomerative Clustering with Numerical or Categorical
Variables 478
Using Analytic Solver and R to Perform Agglomerative
Clustering 480
Agglomerative Clustering with Mixed Data 485

11.2 k-Means Cluster Analysis 492

 Using Analytic Solver and R to Perform k-Means Clustering 
493

11.3 Association Rule Analysis 502

Using Analytic Solver and R to Perform Association Rule
Analysis 506

11.4 Writing With Big Data 514

CHAPTER 12

FORECASTING WITH TIME SERIES DATA 518

12.1 The Forecasting Process for Time Series 520
Forecasting Methods 521
Model Selection Criteria 521

12.2 Simple Smoothing Techniques 522

The Moving Average Technique 523
The Simple Exponential Smoothing Technique 525
Using Excel for Moving Averages and Exponential Smoothing 
528

12.3 Linear Regression Models for Trend and Seasonality 529

The Linear Trend Model 530
The Linear Trend Model with Seasonality 531
A Note on Causal Models for Forecasting 533

12.4 Nonlinear Regression Models for Trend and Seasonality 

535
The Exponential Trend Model 535
The Polynomial Trend Model 538
The Nonlinear Trend Models with Seasonality 539

12.5 Data Partitioning and Model Selection 544

 Cross-validation of Regression Models with R 547

12.6 Advanced Exponential Smoothing Methods 548

The Holt Exponential Smoothing Method 549
The Holt-Winters Exponential Smoothing Method 552

12.7 Writing With Data 556

CHAPTER 13

INTRODUCTION TO PRESCRIPTIVE
ANALYTICS 560
13.1 Overview of Prescriptive Analytics 562

13.2 Monte Carlo Simulation 562

Modeling Risk and Uncertainty 564
Using Excel and R to Generate Random Observations from a
Discrete Probability Distribution 564
Using Excel and R to Generate Random Observations from a
Continuous Probability Distribution 567
Formulating and Developing a Monte Carlo Simulation 569

13.3 Optimization with Linear Programming 577

Formulating a Linear Programming Model 579
Solving a Linear Programming Problem 581

13.4 Optimization with Integer Programming 596

Capital Budgeting 596
Transportation Problem 599

13.5 Writing With Data 604

APPENDIXES
Appendix A Big Data Sets: Variable Description and Data
Dictionary 608
Appendix B Getting Started with Excel and Excel Add-Ins 614
Appendix C Getting Started with R 621
Appendix D Statistical Tables 628
Appendix E Answers to Selected Exercises 632

Index 652

page xxv
 page 1

BUSINESS ANALYTICS
Communicating with Numbers
 page 2

1 Introduction to Business
Analytics

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 1.1  Explain the importance of business
analytics.
LO 1.2  Explain the various types of data.
LO 1.3  Describe variables and types of
measurement scales.
LO 1.4  Describe different data sources and file
formats.

t every moment, data are being generated at an increasing

A velocity from countless sources in an overwhelming volume. In
just about any contemporary human activity, the analysis of
large amounts of data, under the umbrella of business or data
analytics, helps us make better decisions. Managers, consumers,
bankers, sports enthusiasts, politicians, and medical professionals
are increasingly turning to data to boost a company’s revenue,
deepen customer engagement, find better options on consumer
products, prevent threats and fraud, assess riskiness of loans,
succeed in sports and elections, provide better diagnoses and cures
for diseases, and so on. In the broadest sense, business analytics
involves the methodology of extracting information and knowledge
from data for which an important first step is to understand data and
data types.
In this chapter, we will provide an overview and real-world
examples of business analytics. We will also describe various types
of data and measurement scales of variables that will later help us
choose appropriate techniques for analyzing data. Finally, we will
discuss how we can take advantage of data sources that are publicly
available and review some standard formats that people use to store
and disseminate data.

page 3
©CampPhoto/iStock/Getty Images

INTRODUCTORY CASE
Vacation in Belize
After graduating from a university in southern California, Emily
Hernandez is excited to go on a vacation in Belize with her
friends. There are different airlines that offer flights from
southern California to Belize City. Emily prefers a direct flight
from Los Angeles but is also worried about staying within her
budget. Other, less expensive, options would mean making one
or even two additional stops en route to Belize City. Once
 arriving at Belize City, she plans to take a sea ferry to one of
the resort hotels on Ambergris Caye Island. Purchasing a
vacation package from one of these hotels would be most cost
effective, but Emily wants to make sure that the hotel she
chooses has all the amenities she wants, such as an early
check-in option, a complimentary breakfast, recreational
activities, and sightseeing services. She also wants to make
sure that the hotel is reputable and has good customer
reviews.
Emily starts researching her options for flights and hotels by
searching for deals on the Internet. She has organized and
kept meticulous records of the information she has found so
that she can compare all the options. She would like to use this
information to:
1. Find a flight that is convenient as well as affordable.
2. Choose a reputable hotel that is priced under $200 per night.
A synopsis of this case is provided at the end of Section 1.2.

page 4

1.1 OVERVIEW OF BUSINESS

ANALYTICS
LO 1.1
Explain the importance of business analytics.

Data and analytics capabilities have made a leap forward in recent

years and have changed the way businesses make decisions. In the
broadest sense, business analytics (also referred to as data
analytics) involves the methodology of extracting information and
knowledge from data that improves a company’s bottom-line and
enhances consumer experience. At the core, business analytics
benefits companies by developing better marketing strategies,
deepening customer engagement, enhancing efficiency in
procurement, uncovering ways to reduce expenses, identifying
emerging market trends, mitigating risk and fraud, etc. The explosion
in the field is partly due to the growing availability of vast amounts of
data, improved computational power, and the development of
sophisticated algorithms. More than ever, colleges and universities
need a curriculum that emphasizes business analytics, and
companies need data-savvy professionals who can turn data into
insights and action.
Business analytics is a broad topic, encompassing statistics,
computer science, and information systems with a wide variety of
applications in marketing, human resource management, economics,
finance, health, sports, politics, etc. Unlike data science that is
focused on advanced computer algorithms to develop applications
for end users, business analytics tends to focus more on the analysis
of the available data.
Raw data do not offer much value or insights. With the
advancement of data and analytics capabilities, capturing a large
amount of data or implementing complex statistical models has been
simplified. However, in order to extract value from data, we need to
be able to understand the business context, ask the right questions
from the data, identify appropriate analysis models, and
communicate information into verbal and written language. It is
important to note that numerical results are not very useful unless
they are accompanied with clearly stated actionable business
insights.

BUSINESS ANALYTICS
Business analytics combines qualitative reasoning with
quantitative tools to identify key business problems and
 translate data analysis into decisions that improve business
performance.

There are different types of analytics techniques designed to

extract value from data that can be grouped into three broad
categories: descriptive analytics, predictive analytics, and
prescriptive analytics.
Descriptive analytics refers to gathering, organizing, tabulating,
and visualizing data to summarize ‘what has happened?’
Examples of descriptive analytics include financial reports, public
health statistics, enrollment at universities, student report cards,
and crime rates across regions and time. The descriptive
information can also be presented in a number of formats including
written reports, tables, graphs, and maps.
Predictive analytics refers to using historical data to predict ‘what
could happen in the future?’ Analytical models help identify
associations between variables, and these associations are used
to estimate the likelihood of a specific outcome. Examples of
predictive analytics include identifying customers who are most
likely to respond to specific marketing campaigns, admitted
students who are likely to enroll, credit card transactions that are
likely to be fraudulent, or the incidence of crime at certain regions
and times.
Prescriptive analytics refers to using optimization and simulation
algorithms to provide advice on ‘what should we do?’ It explores
several possible actions and suggests a course of action.
Examples include providing advice on scheduling employees’ work
hours and supply level in order to meet customer demand,
selecting a mix of products to manufacture, choosing an
investment portfolio to meet a financial goal, or targeting marketing
campaigns to specific customer groups on a limited budget.

page 5
 These three categories of analytics serve different functions in
the problem-solving process. Descriptive analytics is often referred
to as business intelligence (BI), which provides organizations and
their users with the ability to access and manipulate data
interactively through reports, dashboards, applications, and
visualization tools. BI uses past data integrated from multiple
sources to inform decision making and identify problems and
solutions. Most BI questions can be solved using complex queries of
the enterprise databases. For example, a typical BI question for an
online music streaming company would be, “During the first quarter
of 2020, how many country songs recommended by the music
service were skipped by a U.S.-based female listener within five
seconds of playing?” The answer to this question can be found by
querying and summarizing historical data and is useful for making
decisions related to the effectiveness of the music recommendation
system of the company.
Predictive and prescriptive analytics, on the other hand, are
commonly considered advanced predictions. They focus on building
predictive and prescriptive models that help organizations
understand what might happen in the future. Advanced prediction
problems are solved using statistics and data mining techniques. For
the online music streaming company example, an advanced
prediction question would be, “What are the key factors that
influence a U.S.-based female listener’s music choice?” The answer
to this question cannot be directly found in the enterprise databases
and requires complex learning and analysis of the relevant historical
data.
The three categories of business analytics can also be viewed
according to the level of sophistication and business values they
offer. For many people, using predictive analytics to predict the
future is more valuable than simply summarizing data and describing
what happened in the past. In addition, predictive techniques tend to
require more complex modeling and analysis tools than most
descriptive techniques. Likewise, using prescriptive techniques to
provide actionable recommendations could be more valuable than
predicting a number of possible outcomes in the future. Turning data-
driven recommendations into action also requires thoughtful
consideration and organizational commitment beyond developing
descriptive and predictive analytical models.
Figure 1.1 categorizes business analytics into three stages of
development based on its value and the level of organizational
commitment to data-driven decision making. Most chapters of this
text can also be grouped within each of the three analytics
categories as shown in Figure 1.1.

FIGURE 1.1 Three stages of business analytics

In addition to the chapters presented in Figure 1.1, Chapters 1

(Introduction to Business Analytics), 4 (Probability and Probability
Distributions), 5 (Statistical Inference), and 8 page 6
(Introduction to Data Mining) cover prerequisite
knowledge to the other topics in the text. Regardless of the analysis
techniques we use, the overall goal of analytics is to improve
business decision making. Essential to this goal is the ability to
communicate the insights and value of data. Throughout the text,
students will not only learn to conduct data analysis but also to tell
an impactful story conveyed in written form to those who may not
know detailed statistical and computational methods.

Important Business Analytics Applications

Data and analytics permeate our daily lives. We are often unaware
of the impact that analytics has on even the most mundane activities,
such as buying clothes, saving for retirement, checking e-mails,
interacting on social media, and watching a TV show. We will now
highlight some of the important applications of business analytics.

Fashion Retailers. Traditional businesses ranging from apparel

companies to food retailers invest heavily in data and analytics
technologies in order to remain competitive and profitable.
Once acclaimed as the company that “dictated how America
dressed” by The New York Times, the Gap, Inc., experienced
disappointing sales in recent years as the fast fashion industry
got increasingly crowded with major international competitors,
such as Zara, H&M, and Uniqlo, entering the U.S. market. To
turn the company around, Art Peck, Gap’s CEO, revolutionized
the company by using insights gained from transactional and
customer databases as well as social media data. Instead of
relying solely on the vision of creative directors, Gap uses
predictive analytics to help determine the next season’s
assortment of products. Peck also encouraged his employees
to use big data to inform a wide range of decision making,
including identifying loyal customers, matching products to
customers, enhancing customer satisfaction, and managing
product rebuys. As the result of these changes, the company is
now able to combine the trends identified in data with real-time
store sales to quickly bring more relevant products to market.
Online Subscription. Netflix, one of the largest content-
streaming companies, transformed the entertainment industry
with its innovative business model. One of the engines that
propelled its success was a sophisticated analytics tool called
CineMatch, a recommendation system for movies and TV
 shows. CineMatch was so crucial to Netflix’s ability to sustain
and grow its customer base that the company launched a one
million dollar competition challenging the contestants to
improve the CineMatch recommendation algorithm. Netflix
provided the contestants with a data set of over 100 million
user ratings given to about 20,000 movies. The winning team,
BellKor’s Pragmatic Chaos, developed an algorithm that
predicted viewers’ movie ratings 10% more accurately than
Netflix’s own algorithm. Nowadays, similar recommendation
systems based on big data and analytics are used pervasively
by companies across industries.
Gaming Industry. Faced with intense competition in the gaming
industry, Harrah’s Entertainment decided to use data analytics
to compete with rivals who are touting opulent resorts and
attractive perks to their customers. By building huge
warehouses of customer data—which casino they visited,
which games and how long they played, where they dined, etc.
—Harrah customized visitors’ experiences based on what it
knew about them and turned “promiscuous” customers into
“monogamous” customers who would only go to Harrah’s.
Analytics also helped drive Harrah’s operating decisions, such
as hotel room pricing and slot machine placement. The
analytics efforts paid off financially. Adjusted earnings per
share increased fivefold within seven years. Improved profits
generated from its data-driven decision making allowed
Harrah’s to invest in facility upgrades and global expansion.
Sports. Sports analytics uses data on players’ statistics and
other sports-related information to drive decision making. Long
before the analytics buzz, the Oakland Athletics page 7
baseball team, referred to as the A’s, used analytics to
build a playoff-bound team. (The A’s approach to the game was
also depicted in the movie Moneyball.) Traditionally,
professional baseball recruiters relied heavily on subjective
measures, such as body type, speed, and projected confidence
for scouting players. Recognizing this shortcoming, the A’s took
 advantage of sabermetrics, which is an application of statistical
analysis to baseball, to assess hidden worth in undervalued
players. The model was built on the premise that a team with a
high on-base percentage was likely to score more runs and,
therefore, win more games. The A’s drafted and traded for
players that fit this system, not necessarily requiring them to fit
the body type and other subjective features that dictated the
decisions of other clubs. Despite being one of the poorest-
funded teams in Major League Baseball, the A’s were able to
pick talented players cheaply, making them one of the most
successful franchises in 2002 and 2003. This success has
since revolutionized all major professional sports, with each
team having its own analytics department or an analytics
expert.
Healthcare. Big data and analytics are also increasingly used in
the medical field. By analyzing 46 billion data points on over
200,000 adult patients, Google’s artificial intelligence system
has developed a model for predicting hospital visit outcomes
and final diagnoses. Google’s system claims to be able to
predict patient death 24 to 48 hours before existing hospital
models can, allowing life-saving procedures to be
administered. Another example of big data in medicine is the
Cancer Moonshot program, sponsored by the U.S. National
Cancer Institute, with an ambitious goal of developing a cure
for cancer. The program provides medical researchers with
access to a vast amount of cancer patient data. By analyzing
biopsy reports, treatment plans, and recovery rates of patients,
researchers are able to study trends in how certain cancer
proteins interact with different treatments and recommend the
most promising treatment plan for the individual patient.

These examples show that analytics is taking a radical approach

to improving organizations’ decision-making abilities. The data-
driven approach is replacing and complementing the traditional
decision-making process that relies heavily on a few experts and
their subjective judgment. This has led to more accurate, often
overlooked, and unexpected findings, which have morphed into
competitive advantages with the help of the right domain expertise.
While the promise of big data analytics is exciting, organizations
cannot neglect the greater responsibilities that come with this newly
acquired power. As we capitalize on the data collected from our
customers and constituents, we must realize that there are
enormous risks involved in using these data, especially in the forms
of data security and privacy. Many stories of data breach and misuse
serve as chilling reminders of these risks. For example, in 2018,
Marriott International disclosed a massive data breach that affected
up to 500 million guests. Private data accessed by the hackers
included names, addresses, phone numbers, e-mail addresses,
passport numbers, and encrypted credit card details. Breaches like
this can lead to identity theft and/or financial fraud. In another
example, Cambridge Analytica, a political data firm hired by
President Donald Trump’s 2016 election campaign, harvested
private information from over 50 million Facebook profiles to create
personality models for voters, which were later used to create digital
ads to influence voter behaviors. Facebook has since tightened its
policies on third-party access to user profiles, but the incident
created enormous backlash against the social media company. As
these cases illustrate, analytics must be conducted in a responsible
and ethical way.
The diverse nature of analytics applications requires deep
understanding of data and data types. In the remaining part of this
chapter, we will describe the various types of data and measurement
scales that help us choose appropriate techniques for analyzing
data. We will also discuss how we can take advantage of data
sources that are publicly available, and review some standard
formats that people use to store and disseminate data.

page 8
1.2 TYPES OF DATA
LO 1.2
Explain the various types of data.

Every day, consumers and businesses use many kinds of data from
various sources to help make decisions. An important first step for
making decisions is to find the right data and prepare it for the
analysis. In general, data are compilations of facts, figures, or other
contents, both numerical and nonnumerical. Data of all types and
formats are generated from multiple sources. Insights from all of
these data can enable businesses to make better decisions, such as
deepening customer engagement, optimizing operations, preventing
threats and fraud, and capitalizing on new sources of revenue. We
often find a large amount of data at our disposal. However, we also
derive insights from relatively small data sets, such as from
consumer focus groups, marketing surveys, or reports from
government agencies.
Data that have been organized, analyzed, and processed in a
meaningful and purposeful way become information. We use a
blend of data, contextual information, experience, and intuition to
derive knowledge that can be applied and put into action in specific
situations.

DATA, INFORMATION, AND KNOWLEDGE

Data are compilations of facts, figures, or other contents, both
numerical and nonnumerical. Information is a set of data that
are organized and processed in a meaningful and purposeful
way. Knowledge is derived from a blend of data, contextual
information, experience, and intuition.
 In the introductory case, Emily is looking for flights from Los
Angeles International Airport (LAX) to Belize City (BZE), as well as
hotels in Belize. An online search on Orbitz.com yields a total of
1,420 combinations of flight schedules. She also finds information on
19 hotels that are within her budget. Figure 1.2 shows a portion of
the airfare and hotel information that Emily finds on the Internet.

FIGURE 1.2 Screen shot of airfare and hotel search results from Orbitz.com
Source: Orbitz.com
 Before we analyze the information that Emily has gathered, it is
important to understand different types of data. In this section, we
focus on the various data categorizations.

Sample and Population Data

There are several ways to categorize data depending on how they
are collected, their format, and specific values they represent. In
most instances, it is not feasible to collect data that comprise the
population of all elements of interest. Therefore, a subset of data—
a sample—is used for the analysis. Traditional statistical techniques
use sample information to draw conclusions about the population.

POPULATION VERSUS SAMPLE

A population consists of all observations or items of interest in
an analysis. A sample is a subset of the population. We
examine sample data to make inferences about the population.

It is generally not feasible to obtain population data due to

prohibitive costs and/or practicality. We rely on sampling because we
are unable to use population data for two main reasons.
Obtaining information on the entire population is expensive.
Suppose we are interested in the average lifespan of a Duracell
AAA battery. It would be incredibly expensive to test each battery.
And moreover, in the end, all batteries would be dead and the
answer to the original question would be useless.
page 9
It is impossible to examine every member of the
population. Consider how the monthly unemployment rate in the
United States is calculated by the Bureau of Labor Statistics (BLS).
Is it reasonable to assume that the the BLS contacts each
individual in the labor force, and asks whether or not he or she is
employed? Given that there are over 160 million individuals in the
 labor force, this task would be impossible to complete in 30 days.
Instead, the BLS conducts a monthly sample survey of about
60,000 households to measure the extent of unemployment in the
United States.
In the introductory case, Emily is working with sample data. The
population would have consisted of all the airlines and hotels, some
of which may not even have shown up in an online search.

Cross-Sectional and Time Series Data

Sample data are generally collected in one of two ways. Cross-
sectional data refer to data collected by recording a characteristic of
many subjects at the same point in time, or without regard to
differences in time. Subjects might include individuals, households,
firms, industries, regions, and countries. The National Basketball
Association (NBA) data presented in Table 1.1 is an example of
cross-sectional data because it displays the team standings at the
end of the 2018–2019 season. The eight teams may not have ended
the season precisely on the same day and time, but the differences
in time are of no relevance in this example. Other examples of cross-
sectional data include the recorded grades of students in a class, the
sale prices of single-family homes sold last month, the current price
of gasoline in different cities in the United States, and the starting
salaries of recent business graduates from the University of
Connecticut.

TABLE 1.1 2018–2019 NBA Eastern Conference

page 10

Team name Wins Losses Winning percentage

Milwaukee Bucks 60 22 0.732

 Team name Wins Losses Winning percentage

Toronto Raptors* 58 24 0.707

Philadephia 76ers 51 31 0.622

Boston Celtics 49 33 0.598

Indiana Pacers 48 34 0.585

Brooklyn Nets 42 40 0.512

Orlando Magic 42 40 0.512

Detroit Pistons 41 41 0.500

*The Toronto Raptors won their first NBA title during the 2018-2019 season.

Time series data refer to data collected over several time

periods focusing on certain groups of people, specific events, or
objects. Time series data can include hourly, daily, weekly, monthly,
quarterly, or annual observations. Examples of time series data
include the hourly body temperature of a patient in a hospital’s
intensive care unit, the daily price of General Electric stock in the first
quarter of 2020, the weekly exchange rate between the U.S. dollar
and the euro over the past six months, the monthly sales of cars at a
dealership in 2020, and the annual population growth rate of India in
the last decade. In these examples, temporal ordering is relevant
and meaningful.
Figure 1.3 shows a plot of the national homeownership rate in the
U.S. from 2000 to 2018. According to the U.S. Census Bureau, the
national homeownership rate in the first quarter of 2016 plummeted
to 63.6% from a high of 69.4% in 2004. An explanation for the
decline in the homeownership rate is the stricter lending practices
caused by the housing market crash in 2007 that precipitated a
banking crisis and deep recession. This decline can also be
attributed to home prices outpacing wages in the sample period.

FIGURE 1.3 Homeownership Rate (in %) in the U.S. from 2000 through 2018
Source: Federal Reserve Bank of St. Louis

Structured and Unstructured Data

When you think of data, the first image that probably pops in your
head is lots of numbers and perhaps some charts and graphs. In
reality, data can come in multiple forms. For example, information
exchange in social networking websites such as Facebook, LinkedIn,
and Twitter also constitute data. In order to better understand the
various forms of data, we make a distinction between structured and
unstructured data.
Generally, structured data reside in a predefined, row-column
format. We use spreadsheet or database applications (see Section
2.1) to enter, store, query, and analyze structured page 11
data. Examples of structured data include numbers,
dates, and groups of words and numbers, typically stored in a
tabular format. Structured data often consist of numerical information
that is objective and is not open to interpretation.
Point-of-sale and financial data are examples of structured data
and are usually designed to capture a business process or
transaction. Examples include the sale of retail products, money
transfer between bank accounts, and the student enrollment in a
university course. When individual consumers buy products from a
retail store, each transaction is captured into a record of structured
data.
Consider the sales invoice shown in Figure 1.4. Whenever a
customer places an order like this, there is a predefined set of data
to be collected, such as the transaction date, shipping address, and
the units of product being purchased. Even though a receipt or an
invoice may not always be presented in rows and columns, the
predefined structure allows businesses and organizations to
translate the data on the document into a row-column format. As we
will see in Example 1.1, the flight information collected for the
introductory case may not fit precisely in a tabular format, but
because of the structured nature of the data, it can easily be
summarized into rows and columns.

FIGURE 1.4 A sample invoice from a retail transaction

 At one time, due to the high cost and performance limitations of
storing and processing data, companies and organizations relied
mostly on structured data, which could be effectively managed with
spreadsheet and database applications. For decades, they relied on
this type of data to run their businesses and operations. Today, with
the advent of the digital age, most experts agree that only about 20%
of all data used in business decisions are structured data. The
remaining 80% are unstructured.
Unlike structured data, unstructured data (or unmodeled data)
do not conform to a predefined, row-column format. They tend to be
textual (e.g., written reports, e-mail messages, doctor’s notes, or
open-ended survey responses) or have multimedia contents (e.g.,
photographs, videos, and audio data). Even though these data may
have some-implied structure (e.g., a report title, e-mail’s subject line,
or a time stamp on a photograph), they are still considered
unstructured as they do not conform to a row-column model required
in most database systems. Social media data such as Twitter,
YouTube, Facebook, and blogs are examples of unstructured data.
 Both structured and unstructured data can be either human-
generated or machine-generated. For structured data, human-
generated data include information on price, income, retail sales,
age, gender, etc., whereas machine-generated data include
information from manufacturing sensors (rotations per minute),
speed cameras (miles per hour), web server logs (number of
visitors), etc. For unstructured data, human-generated data include
texts of internal e-mails, social media data, page 12
presentations, mobile phone conversations, and text
message data, and so on, whereas machine-generated data include
satellite images, meteorological data, surveillance video data, traffic
camera images, and others. In this text, we focus on structured data.
Text analytics and other sophisticated tools to analyze unstructured
data are beyond the scope of this text.

Big Data
Nowadays, businesses and organizations generate and gather more
and more data at an increasing pace. The term big data is a catch-
phrase, meaning a massive volume of both structured and
unstructured data that are extremely difficult to manage, process,
and analyze using traditional data-processing tools. Despite the
challenges, big data present great opportunities to gain knowledge
and business intelligence with potential game-changing impacts on
company revenues, competitive advantage, and organizational
efficiency. More formally, a widely accepted definition of big data is
“high-volume, high-velocity and/or high-variety information assets
that demand cost-effective, innovative forms of information
processing that enable enhanced insight, decision making, and
process automation” (www.gartner.com). The three characteristics
(the three Vs) of big data are:
Volume: An immense amount of data is compiled from a single
source or a wide range of sources, including business
transactions, household and personal devices, manufacturing
equipment, social media, and other online portals.
 Velocity: In addition to volume, data from a variety of sources get
generated at a rapid speed. Managing these data streams can
become a critical issue for many organizations.
Variety: Data also come in all types, forms, and granularity, both
structured and unstructured. These data may include numbers,
text, and figures as well as audio, video, e-mails, and other
multimedia elements.
In addition to the three defining characteristics of big data, we
also need to pay close attention to the veracity of the data and the
business value that they can generate. Veracity refers to the
credibility and quality of data. One must verify the reliability and
accuracy of the data content prior to relying on the data to make
decisions. This becomes increasingly challenging with the rapid
growth of data volume fueled by social media and automatic data
collection. Value derived from big data is perhaps the most important
aspect of any analytics initiative. Having a plethora of data does not
guarantee that useful insights or measurable improvements will be
generated. Organizations must develop a methodical plan for
formulating business questions, curating the right data, and
unlocking the hidden potential in big data.
Big data, however, do not necessarily imply complete
(population) data. Take, for example, the analysis of all Facebook
users. It certainly involves big data, but if we consider all Internet
users in the world, Facebook users are only a very large sample.
There are many Internet users who do not use Facebook, so the
data on Facebook do not represent the population. Even if we define
the population as pertaining to those who use online social media,
Facebook is still one of many social media portals that consumers
use. And because different social media are used for different
purposes, data collected from these sites may very well reflect
different populations of Internet users; this distinction is especially
important from a strategic business standpoint. Therefore, Facebook
data are simply a very large sample.
 In addition, we may choose not to use big data in its entirety even
when they are available. Sometimes it is just inconvenient to analyze
a very large data set as it is computationally burdensome, even with
a modern, high-capacity computer system. Other times, the
additional benefits of working with big data may not justify the
associated costs. In sum, we often choose to work with a small data
set, which in a sense is a sample drawn from big data.

page 13

STRUCTURED, UNSTRUCTURED, AND BIG DATA

Structured data are data that reside in a predefined, row-
column format, while unstructured data do not conform to a
predefined, row-column format. Big data is a term used to
describe a massive volume of both structured and unstructured
data that are extremely difficult to manage, process, and
analyze using traditional data-processing tools. Big data,
however, do not necessary imply complete (population) data.

EXAMPLE 1.1
In the introductory case, Emily is looking for roundtrip flight
schedules offered by different airlines from Los Angeles,
California, to Belize City. Once arriving at Belize City, she plans
to purchase a vacation package from one of the resort hotels
on Ambergris Caye Island. She has compiled information on
flight schedules and hotel options from search results on
Orbitz.com. Her initial search yields 1,420 flight schedules and
19 hotels. In its current format, Emily is overwhelmed by the
amount of data and knows that she needs to refine her search.
She would like to focus on flights that are convenient (short
duration) and relatively inexpensive. For hotels, she would like
an affordable hotel (priced under $200 per night) with good
reviews (above-average review on a 5-point scale). Given
Emily’s preferences, summarize the online information in
tabular form.
SOLUTION:
We first search for roundtrip flight schedules where priority is
given to short flight times and low prices. Even though the flight
information is not in a perfect row-column configuration, the
structured nature of the data allows us to summarize the
information in a tabular format. The four most relevant options
are presented in Panel A of Table 1.2. Given these options, it
seems that the American/Delta choice might be the best for
her. Similarly, once we refine the search to include only hotels
that fall within her budget of under $200 per night and have an
average consumer rating above 4 on a 5-point scale, the
number of hotel options declines from 19 to five. These five
options are presented in Panel B of Table 1.2. In this case, her
choice of hotel is not clear-cut. The hotel with the page 14
highest rating is also the most expensive (X’Tan Ha
– The Waterfront), while the least expensive hotel has the
lowest rating (Isla Bonita Yacht Club) and the rating is based
on only 11 reviews. Emily will now base her final hotel choice
on online reviews. These reviews constitute unstructured data
that do not conform to a row-column format.

TABLE 1.2 Tabular Format of Flight Search and Hotel Results

Panel A. Flight Search Results

Panel B. Hotel Results

Hotel under Average Number of Price per

$200 Rating Reviews night

Isla Bonita Yacht 4.2   11 $130

Club

X’Tan Ha – The 4.6 208 $180

Waterfront

Mata Rocks 4.2 212 $165

Resort

Costa Blu 4.5    8 $165

Blue Tang Inn 4.2   26 $179

SYNOPSIS OF INTRODUCTORY CASE

Emily Hernandez is excited to go on vacation in Belize with her friends. She decides
to search on Orbitz.com for flights that are convenient as well as affordable. She
also wants to find a reputable hotel in Ambergris Caye Island that is priced under
$200 per night. Although she finds 1,420 options for the flight, she focuses on
morning flights with priority given to lower prices and shorter flight times.
David Schulz Photography/Shutterstock

There are four airline options that she finds suitable. With a price of $495.48, the
Delta/American option is the cheapest, but the flight time is over 20 hours each way.
The United/Delta option offers the shortest flight times; however, it comes with a
price tag of $929.91. Emily decides on the American/Delta option, which seems
most reasonable with a price of $632.91, a flight time of under eight hours to Belize,
and only five hours on the return. In regard to hotels, Bonita Yacht Club offers the
cheapest price of $130, but it comes with a relatively low rating of 4.2.
In addition, Emily is concerned about the credibility of the rating as it is based on
only 11 reviews. Although not the cheapest, Emily decides to go with X’Tan Ha – The
Waterfront. It has the highest rating of 4.6 and the price still falls within her budget of
under $200. In these reviews, Emily consistently finds key phrases such as “great
location,” “clean room,” “comfortable bed,” and “helpful staff.” Finally, the pictures
that guests have posted are consistent with the images published by the resort on its
website.

EXERCISES 1.2
Applications
1. A few years ago, it came as a surprise when Apple’s iPhone 4 was found to have
a problem. Users complained of weak reception, and sometimes even dropped
calls, when they cradled the phone in their hands in a particular way. A survey at a
local store found that 2% of iPhone 4 users experienced this reception problem.
a. Describe the relevant population.
b. Is 2% associated with the population or the sample?
 2. Many people regard video games as an obsession for youngsters, but, in fact, the
average age of a video game player is 35 years old. Is the value 35 likely the
actual or the estimated average age of the population? Explain.
3. An accounting professor wants to know the average GPA of the students enrolled
in her class. She looks up information on Blackboard about the students enrolled
in her class and computes the average GPA as 3.29. Describe the relevant
population.
4. Recent college graduates with an engineering degree continue to earn high
salaries. An online search revealed that the average annual salary for an entry-
level position in engineering is $65,000.
a. What is the relevant population?
b. Do you think the average salary of $65,000 is computed from the population?
Explain.
5. Research suggests that depression significantly increases the risk of developing
dementia later in life. Suppose that in a study involving 949 elderly persons, it was
found that 22% of those who had depression went on to develop dementia,
compared to only 17% of those who did not have depression.
a. Describe the relevant population and the sample.
b. Are the numbers 22% and 17% associated with the population or a sample?
6. Go to www.zillow.com and find the sale price of 20 single-family homes sold in
Las Vegas, Nevada, in the last 30 days. Structure the data in a
tabular format and include the sale price, the number of bedrooms,
page 15
the square footage, and the age of the house. Do these data represent cross-
sectional or time series data?
7. Go to www.finance.yahoo.com to get the current stock quote for Home Depot
(ticker symbol = HD). Use the ticker symbol to search for historical prices and
create a table that includes the monthly adjusted close price of Home Depot stock
for the last 12 months. Do these data represent cross-sectional or time series
data?
8. Go to the New York Times website at www.nytimes.com and review the front
page. Would you consider the data on the page to be structured or unstructured?
Explain.
9. Conduct an online search to compare small hybrid vehicles (e.g., Toyota Prius,
Ford Fusion, Chevrolet Volt) on price, fuel economy, and other specifications. Do
you consider the search results structured or unstructured data? Explain.
10. Find Under Armour’s annual revenue from the past 10 years. Are the data
considered structured or unstructured? Explain. Are they cross-sectional or time
series data?
11. Ask 20 of your friends about their online social media usage, specifically whether
or not they use Facebook, Instagram, and Snapchat; how often they use each
social media portal; and their overall satisfaction of each of these portals. Create
a table that presents this information. Are the data considered structured or
unstructured? Are they cross-sectional or time series data?
 12. Ask 20 of your friends whether they live in a dormitory, a rental unit, or other form
of accommodation. Also find out their approximate monthly lodging expenses.
Create a table that uses this information. Are the data considered structured or
unstructured? Are they cross-sectional or time series data?
13. Go to the U.S. Census Bureau website at www.census.gov and search for the
most recent median household income for Alabama, Arizona, California, Florida,
Georgia, Indiana, Iowa, Maine, Massachusetts, Minnesota, Mississippi, New
Mexico, North Dakota, and Washington. Do these data represent cross-sectional
or time series data? Comment on the regional differences in income.

1.3 VARIABLES AND SCALES OF

MEASUREMENT
LO 1.3
Describe variables and types of measurement scales.

For business analytics, we invariably focus on people, firms, or

events with particular characteristics. When a characteristic of
interest differs in kind or degree among various observations
(records), then the characteristic can be termed a variable. Marital
status and income are examples of variables because a person’s
marital status and income vary from person to person. Variables are
further classified as either categorical (qualitative) or numerical
(quantitative). The observations of a categorical variable represent
categories, whereas the observations of a numerical variable
represent meaningful numbers. For example, marital status is a
categorical variable, whereas income is a numerical variable.
For a categorical variable, we use labels or names to identify the
distinguishing characteristic of each observation. For instance, a
university may identify each student’s status as either at the
undergraduate or the graduate level, where the education level is a
categorical variable representing two categories. Categorical
variables can also be defined by more than two categories.
Examples include marital status (single, married, widowed, divorced,
separated), IT firm (hardware, software, cloud), and course grade (A,
B, C, D, F). It is important to note that categories are often converted
into numerical codes for purposes of data processing, which we will
discuss in Chapter 2.
For a numerical variable, we use numbers to identify the
distinguishing characteristic of each observation. Numerical
variables, in turn, are either discrete or continuous. A discrete
variable assumes a countable number of values. Consider the
number of children in a family or the number of points scored in a
basketball game. We may observe values such as 3 children in a
family or 90 points being scored in a basketball game, but we will not
observe 1.3 children or 92.5 scored points. The values that a
discrete variable assumes need not be whole numbers. For
example, the price of a stock for a particular firm is a discrete
variable. The stock price may take on a value of $20.37 or $20.38,
but it cannot take on a value between these two points. A
continuous variable is characterized by uncountable values within
an interval. Weight, height, time, and investment return are all
examples of continuous variables. For example, an unlimited number
of values occur between the weights of 100 and 101 pounds, such
as 100.3, 100.625, 100.8342, and so on. In practice, however,
continuous variables are often measured in discrete values. We may
report a newborn’s weight (a continuous variable) in discrete terms
as 6 pounds 10 ounces and another newborn’s weight in similar
discrete terms as 6 pounds 11 ounces.

page 16

CATEGORICAL AND NUMERICAL VARIABLES

A variable is a general characteristic being observed on a set of
people, objects, or events, where each observation varies in
kind or degree.
The observations of a categorical variable assume names or
labels.
 The observations of a numerical variable assume meaningful
numerical values. A numerical variable can be further
categorized as either discrete or continuous. A discrete
variable assumes a countable number of values, whereas a
continuous variable is characterized by uncountable values.

EXAMPLE 1.2
In the introductory case, Emily has conducted an online search
on airfares and hotels for her planned vacation to Ambergris
Caye Island. She has summarized the information in Table 1.2.
What types of variables are included in the airfare and hotel
data?
SOLUTION:
Airlines and hotels are categorical variables because the
observations—the names—are merely labels. On the other
hand, the roundtrip price, the average rating, the number of
reviews, and the price per night are numerical variables
because the observations are all meaningful numbers. Note
that the roundtrip price, the number of reviews, and the price
per night represent discrete variables because they can only
assume a countable number of values. The average rating is
continuous because it is characterized by uncountable values
within the 0 to 5 interval. The date and time variables are
considered numerical; however, if we were to consider the days
of the week (e.g., Monday, Tuesday, etc.) in the flight
schedules, then the variable days would be considered a
categorical variable.

The Measurement Scales

In order to choose the appropriate techniques for summarizing and
analyzing variables, we need to distinguish between the different
measurement scales. The data for any variable can be classified into
one of four major measurement scales: nominal, ordinal, interval, or
ratio. Nominal and ordinal scales are used for categorical variables,
whereas interval and ratio scales are used for numerical variables.
We discuss these scales in ascending order of sophistication.
The nominal scale represents the least sophisticated level of
measurement. If we are presented with nominal data, all we can do
is categorize or group the data. The values in the data set differ
merely by name or label. Table 1.3 lists the 30 large, publicly-owned
companies that comprise the Dow Jones Industrial Average (DJIA).
The DJIA is a stock market index that shows how these U.S.-based
companies have traded during a standard trading session in the
stock market. Table 1.3 also indicates where stocks of these
companies are traded: on either the National Association of
Securities Dealers Automated Quotations (Nasdaq) or the New York
Stock Exchange (NYSE). These data are classified as nominal scale
because we are simply able to group or categorize them.
Specifically, only five stocks are traded on the Nasdaq, whereas the
remaining 25 are traded on the NYSE.

page 17

TABLE 1.3 Companies of the DJIA and Exchange Where Stock is Traded

Company Exchange Company Exchange

3M (MMM) NYSE Johnson & NYSE

Johnson (JNJ)

American Express NYSE JPMorgan Chase NYSE

(AXP) (JPM)
Company Exchange Company Exchange

Apple (AAPL) Nasdaq McDonald’s (MCD) NYSE

Boeing (BA) NYSE Merck (MRK) NYSE

Caterpillar (CAT) NYSE Microsoft (MFST) Nasdaq

Chevron (CVX) NYSE Nike (NKE) NYSE

Cisco (CSCO) Nasdaq Pfizer (PFE) NYSE

Coca-Cola (KO) NYSE Procter & Gamble NYSE

(PG)

Disney (DIS) NYSE Travelers (TRV) NYSE

DowDupont NYSE United Health NYSE

(DWDP) (UNH)

ExxonMobil NYSE United NYSE

(XOM) Technologies
(UTX)

Goldman Sachs NYSE Verizon (VZ) NYSE

(GS)

Home Depot (HD) NYSE Visa (V) NYSE

IBM (IBM) NYSE Wal-Mart (WMT) NYSE

Intel (INTC) Nasdaq Walgreen (WBA) Nasdaq

Source: https://money.cnn.com/data/dow30/; information retrieved on February 16, 2019

Often, we substitute numbers for the particular categorical

characteristic or trait that we are grouping. For instance, we might
use the number 0 to show that a company’s stock is traded on the
Nasdaq and the number 1 to show that a company’s stock is traded
on the NYSE. One reason why we do this is for ease of exposition;
always referring to the National Association of Securities Dealers
Automated Quotations, or even the Nasdaq, can be awkward and
unwieldy.
Compared to the nominal scale, the ordinal scale reflects a
stronger level of measurement. With ordinal data, we are able to
both categorize and rank the data with respect to some characteristic
or trait. The weakness with ordinal data is that we cannot interpret
the difference between the ranked values because the actual
numbers used are arbitrary. For example, in the introductory case,
Emily looks at hotel reviews where consumers are asked to classify
the service at a particular hotel as excellent (5 stars), very good (4
stars), good (3 stars), fair (2 stars), or poor (1 star). We summarize
the categories and their respective ratings in Table 1.4.

TABLE 1.4 Hotel Survey Categories with Ratings

Category Rating

Excellent 5

Very good 4

Good 3

Fair 2

Poor 1
In Table 1.4, the value attached to excellent (5 stars) is higher than
the value attached to good (3 stars), indicating that the response of
excellent is preferred to good. However, we can easily redefine the
ratings, as we show in Table 1.5.

TABLE 1.5 Hotel Survey Categories with Redefined Ratings

Category Rating

Excellent 100

Very good  80

Good  70

Fair  50

Poor  40

page 18

In Table 1.5, excellent still receives a higher value than good, but
now the difference between the two categories is 30 points (100 –
70), as compared to a difference of 2 points (5 – 3) when we use the
first classification. In other words, differences between categories
are meaningless with ordinal data. (We also should note that we
could reverse the ordering so that, for instance, excellent equals 40
and poor equals 100; this renumbering would not change the nature
of the data.)
As mentioned earlier, observations of a categorical variable are
typically expressed in words but are coded into numbers for
purposes of data processing. When summarizing the results of a
categorical variable, we typically count the number of observations
that fall into each category or calculate the percentage of
observations that fall into each category. However, with a categorical
variable, we are unable to perform meaningful arithmetic operations,
such as addition and subtraction.
With data that is measured on the interval scale, we are able to
categorize and rank the data as well as find meaningful differences
between observations. The Fahrenheit scale for temperatures is an
example of interval-scaled data. Not only is 60 degrees Fahrenheit
hotter than 50 degrees Fahrenheit, the same difference of 10
degrees also exists between 90 and 80 degrees Fahrenheit.
The main drawback of interval-scaled data is that the value of
zero is arbitrarily chosen; the zero point of interval-scaled data does
not reflect a complete absence of what is being measured. No
specific meaning is attached to 0 degrees Fahrenheit other than to
say it is 10 degrees colder than 10 degrees Fahrenheit. With an
arbitrary zero point, meaningful ratios cannot be constructed. For
instance, it is senseless to say that 80 degrees is twice as hot as 40
degrees; in other words, the ratio 80/40 has no meaning.
The ratio scale represents the strongest level of measurement.
The ratio scale has all the characteristics of the interval scale as well
as a true zero point, which allows us to interpret the ratios between
observations. The ratio scale is used in many business applications.
Variables such as sales, profits, and inventory levels are expressed
on the ratio scale. A meaningful zero point allows us to state, for
example, that profits for firm A are double those of firm B. Variables
such as weight, time, and distance are also measured on a ratio
scale because zero is meaningful.
Unlike nominal- and ordinal-scaled variables (categorical
variables), arithmetic operations are valid on interval- and ratio-
scaled variables (numerical variables). In later chapters, we will
calculate summary measures, such as the mean, the median, and
the variance, for numerical variables; we cannot calculate these
measures if the variable is categorical in nature.
 MEASUREMENT SCALES
The observations for any variable can be classified into one of
four major measurement scales: nominal, ordinal, interval, or
ratio.
Nominal: Observations differ merely by name or label.
Ordinal: Observations can be categorized and ranked;
however, differences between the ranked observations are
meaningless.
Interval: Observations can be categorized and ranked, and
differences between observations are meaningful. The main
drawback of the interval scale is that the value of zero is
arbitrarily chosen.
Ratio: Observations have all the characteristics of interval-
scaled data as well as a true zero point; thus, meaningful
ratios can be calculated.
Nominal and ordinal scales are used for categorical variables,
whereas interval and ratio scales are used for numerical
variables.

page 19

EXAMPLE 1.3
FILE
Tween_Survey

The owner of a ski resort two hours outside Boston,

Massachusetts, is interested in serving the needs of the
“tween” population (children aged 8 to 12 years old). He
believes that tween spending power has grown over the past
few years, and he wants their skiing experience to be
memorable so that they want to return. At the end of last year’s
ski season, he asked 20 tweens the following four questions.
Q1. On your car drive to the resort, which music streaming
service was playing?
Q2. On a scale of 1 to 4, rate the quality of the food at the
resort (where 1 is poor, 2 is fair, 3 is good, and 4 is excellent).
Q3. Presently, the main dining area closes at 3:00 pm. What
time do you think it should close?
Q4. How much of your own money did you spend at the
lodge today?
A portion of their responses is shown in Table 1.6. Identify the
scale of measurement for each variable used in the survey.
Given the tween responses, provide suggestions to the owner
for improvement.

TABLE 1.6 Tween Responses to Resort Survey

SOLUTION:

Q1. Responses for music streaming service are nominal

because the observations differ merely in label. Twelve of the
20 tweens, or 60%, listened to Spotify. If the resort wishes to
contact tweens using this means, then it may want to direct
its advertising dollars to this streaming service.
Q2. Food quality responses are on an ordinal scale because
we can both categorize and rank the observations. Eleven of
the 20 tweens, or 55%, felt that the food quality was, at best,
 fair. Perhaps a more extensive survey that focuses solely on
food quality would reveal the reason for their apparent
dissatisfaction.
Q3. Closing time responses are on an interval scale. We can
say that 3:30 pm is 30 minutes later than 3:00 pm, and 6:00
pm is 30 minutes later than 5:30 pm; that is, differences
between observations are meaningful. The closing time
responses, however, have no apparent zero point. We could
arbitrarily define the zero point at 12:00 am, but ratios are still
meaningless. In other words, it makes no sense to form the
ratio 6:00 pm/3:00 pm and conclude that 6:00 pm is twice as
long a time period as 3:00 pm. A review of the closing time
responses shows that the vast majority (19 out of 20) would
like the dining area to remain open later.
Q4. The tweens’ responses with respect to their own money
spent at the resort are on a ratio scale. We can categorize
and rank observations as well as calculate meaningful
differences. Moreover, because there is a natural zero point,
valid ratios can also be calculated. Seventeen of the 20
tweens spent their own money at the lodge. It does appear
that the discretionary spending of this age group is
significant. The owner would be wise to cater to some of their
preferences.

page 20

EXERCISES 1.3
Applications
14. Which of the following variables are categorical and which are numerical? If the
variable is numerical, then specify whether the variable is discrete or continuous.
a. Points scored in a football game.
 b. Racial composition of a high school classroom.
c. Heights of 15-year-olds.
15. Which of the following variables are categorical and which are numerical? If the
variable is numerical, then specify whether the variable is discrete or continuous.
a. Colors of cars in a mall parking lot.
b. Time it takes each student to complete a final exam.
c. The number of patrons who frequent a restaurant.
16. In each of the following scenarios, define the type of measurement scale.
a. A kindergarten teacher marks whether each student is a boy or a girl.
b. A ski resort records the daily temperature during the month of January.
c. A restaurant surveys its customers about the quality of its waiting staff on a
scale of 1 to 4, where 1 is poor and 4 is excellent.
17. In each of the following scenarios, define the type of measurement scale.
a. An investor collects data on the weekly closing price of gold throughout the
year.
b. An analyst assigns a sample of bond issues to one of the following credit
ratings, given in descending order of credit quality (increasing probability of
default): AAA, AA, BBB, BB, CC, D.
c. The dean of the business school at a local university categorizes students by
major (i.e., accounting, finance, marketing, etc.) to help in determining class
offerings in the future.
18. In each of the following scenarios, define the type of measurement scale.
a. A meteorologist records the amount of monthly rainfall over the past year.
b. A sociologist notes the birth year of 50 individuals.
c. An investor monitors the daily stock price of BP following the 2010 oil disaster in
the Gulf of Mexico.
19. FILE Major. A professor records the majors of her 30 students. A portion of the
data is shown in the accompanying table.

Student Major

1 Accounting

2 Management

⋮ ⋮

30 Economics

a. What is the measurement scale of these data?

 b. Summarize the results in tabular form.
c. What information can be extracted from the data?
20. FILE DOW. The accompanying table shows a portion of the 30 companies that
comprise the Dow Jones Industrial Average (DJIA). For each company, the data
set lists the year that it joined the DJIA, its industry, and its stock price (in $) as of
February 15, 2019.

Company Year Industry Price

3M (MMM) 1976 Health Care 208.9

American Express (AXP) 1982 Finance 107.4

⋮ ⋮ ⋮ ⋮

Walgreen (WBA) 2018 Health Care 73.43

a. What is the measurement scale of the Industry variable?

b. What is the measurement scale of the Year variable? What are the strengths of
this type of measurement scale? What are its weaknesses?
c. What is the measurement scale of the Price variable? What are the strengths of
this type of measurement scale?

1.4 DATA SOURCES AND FILE

FORMATS
LO 1.4
Describe different data sources and file formats.

At every moment, data are being generated at an increasing velocity

from countless sources in an overwhelming volume. Many experts
believe that 90% of the data in the world today was created in the
last two years alone. Not surprisingly, businesses continue to
grapple with how to best ingest, understand, and operationalize large
volumes of data.
 We access much of the data in this text by simply using a search
engine like Google. These search engines direct us to data-providing
sites. For instance, searching for economic data leads you to the
Bureau of Economic Analysis (http://bea.gov), the Bureau of Labor
Statistics (http://www.bls.gov), the Federal Reserve Economic Data
(https://research.stlouisfed.org), and the U.S. Census page 21
Bureau (http://www.census.gov). These websites
provide data on inflation, unemployment, gross domestic product
(GDP), and much more, including useful international data. Similarly,
excellent world development indicator data are available at
http://data.worldbank.org.
The Wall Street Journal, The New York Times, USA Today, The
Economist, Business Week, Forbes, and Fortune are all reputable
publications that provide all sorts of data. We would like to point out
that all of these data sources represent only a small portion of
publicly available data. In this text, we have compiled a number of
big data sets, based on online data sources, that are integrated
throughout the text.
Structured data have been used extensively for many years and
are often found in tabular formats. As people work more and more in
collaboration with one another, a need usually arises for an ability to
exchange information between different parties. Formatting data in
an agreed-upon or standardized manner is important for allowing
other people to understand the data contained in a file. There are
many standards for file formats. For example, a text file can be
organized into rows and columns to store in a table. Two common
layouts for simple text files are a fixed-width format and a delimited
format. In addition to text files, we can use a markup language to
provide a structure to data. Three widely used markup languages are
eXtensible Markup Language (XML), HyperText Markup Language
(HTML), and JavaScript Object Notation (JSON). We now provide an
overview of these formats and markup languages.

Fixed-Width Format
In a data file with a fixed-width format (or fixed-length format), each
column starts and ends at the same place in every row. The actual
data are stored as plain text characters in a digital file. Consider the
information in Table 1.7. It shows the first name, telephone number,
and annual salary for three individuals.

TABLE 1.7 Sample Data for Format Illustration

Name Telephone Salary

Rich 419-528-0915 160000

Benjamin 203-991-3608  93000

Eduardo 618-345-1278 187000

The information in Table 1.7 can be organized into a fixed-width

format as shown in Figure 1.5. The first, second, and third columns
of Figure 1.5 are defined to have column widths of 8, 12, and 7
characters, respectively. Every observation or record has the exact
same column widths. The fixed-width file has the simplicity of design
where specific data can be found at the exact same location for
every record. This can help speed up record search when the data
set is very large. Furthermore, because only raw data are stored,
fixed-width files tend to be significantly smaller in size compared to
other data formats such as XML that include data labels and tags.
However, the number of characters of each column (the column
width) needs to be predetermined. In Figure 1.5, the Name column is
predefined to have at most 8 characters; any names with more than
8 characters will be truncated. Finally, at times, the columns seem to
run into each other. For these reasons, other formats are more
popular.

FIGURE 1.5 A fixed-width file format

 page 22

Delimited Format
Another widely used file format to store tabular data is a delimited
format. In Figure 1.6, we show the information in Table 1.7 in a
delimited format, where each piece of data is separated by a
comma.

FIGURE 1.6 A comma-separated value (csv) file format

In a delimited format, a comma is called a delimiter, and the file is

called a comma-delimited or comma-separated value (csv) file.
Sometimes, other characters such as semicolons are used as
delimiters. In a delimited file, each piece of data can contain as many
characters as applicable. For example, unlike the fixed-width file
shown in Figure 1.5, a comma-separated value file does not limit a
person’s name to only eight characters.
Fixed-width and delimited files usually include plain text data that
can be opened in most text editing software such as Microsoft Word
and Notepad in Microsoft Windows, TextEdit in Mac, and online tools
such as Google Docs.
eXtensible Markup Language
The eXtensible Markup Language (XML) is a simple language for
representing structured data. XML is one of the most widely used
formats for sharing structured information between computer
programs, between people, and between computers and people. It
uses markup tags to define the structure of data. Using the
information from Table 1.7, Figure 1.7 shows an example of data
coded in XML format.

FIGURE 1.7 An XML file format

Each piece of data is usually enclosed in a pair of ‘tags’ that

follow specific XML syntax. For example, a telephone number starts
with an opening tag (<Telephone>) and ends with a closing tag
(</Telephone>). The XML code is case-sensitive; therefore,
<Telephone> and <telephone> would indicate two different pieces of
information. The tags in Figure 1.7 are not based on any predefined
standard. The XML language allows each user to page 23
define his or her own tags and document structure,
but XML tag names should generally be self-explanatory. The XML
file format is designed to support readability. This makes the XML file
format especially suitable for transporting data between computer
applications without losing the meanings of the data. However, due
to the additional labels and tags, XML data files tend to be much
larger in size than fixed-width and delimited data files that contain
the same data, making data download and parsing more time-
consuming and computationally intensive.

HyperText Markup Language

Like XML, the HyperText Markup Language (HTML) is a mark-up
language that uses tags to define its data in web pages. The key
distinction between XML and HTML is that XML tells us or computer
applications what the data are, whereas HTML tells the web browser
how to display the data. Using the information from Table 1.7, Figure
1.8 shows an example of data coded in HTML.

FIGURE 1.8 An HTML file format

 Tags such as <table> are used to provide structure for textual
data, such as headings, paragraphs, and tables. In Figure 1.8, the
opening <table> and closing </table> tags indicate the beginning and
completion of a table. Unlike XML where users can define their own
markup tags, HTML tags conform to standards maintained by
organizations such as the World Wide Web Consortium (W3C). Web
browsers such as Google Chrome and Safari are designed to
interpret the HTML code that follows these standards. For example,
the tag <th> is understood by web browsers as a table heading. In
our example, there are three columns and headings (i.e., Name,
Telephone, and Salary). The <tr> tag, on the other hand, defines a
row, and the <td> tag defines each cell within a row. Unlike XML,
HTML is case-insensitive.

JavaScript Object Notation

The JavaScript Object Notation (JSON) has become a popular
alternative to XML in recent years as open data sharing has grown in
popularity. JSON is a standard for transmitting human-readable data
in compact files. Originally a subset of the JavaScript page 24
syntax, JSON is currently a data standard supported by a
wide range of modern programming languages such as C, Java, and
Python. Using the information from Table 1.7, Figure 1.9 shows an
example of data coded in JSON format.

FIGURE 1.9 A JSON file format

The JSON format is self-explanatory just as the XML format is,

but it offers several advantages over the XML format. First, the
JSON format is not as verbose as the XML format, making data files
smaller in size. The difference in size is especially noticeable for very
large data sets. Second, the JSON format supports a wide range of
data types not readily available in the XML format. Finally, parsing
JSON data files is faster and less resource intensive. For these
reasons, the JSON format has become a widely adopted standard
for open data sharing.

DATA FILE FORMATS AND MARKUP LANGUAGES

There are many standards for data file formats. Two common
layouts for simple text files are the fixed-width format and the
delimited format.
With a fixed-width format, each column has a fixed width and
starts and ends at the same place in every row.
With a delimited format, each column is separated by a
delimiter such as a comma. Each column can contain as
many characters as applicable.
Markup languages also provide a structure to data. Three
widely used languages are the eXtensible Markup Language
(XML), the HyperText Markup Language (HTML), and the
JavaScript Object Notation (JSON).
XML is a simple text-based markup language for representing
structured data. It uses user-defined markup tags to specify
the structure of data.
HTML is a simple text-based markup language for displaying
content in web browsers.
JSON is a standard for transmitting human-readable data in
compact files.

page 25

EXERCISES 1.4
Applications
21. A used car salesperson recently sold two Mercedes, three Toyota, six Ford, and
four Hyundai sedans. He wants to record the sales data.
 a. Organize the data into a fixed-width format (eight characters for brand name
and four characters for the number of cars sold).
b. Organize the data in a delimited format.
c. Code the data in XML format.
d. Code the data in HTML table format.
e. Code the data in JSON format.
22. Last year, Oracle hired three finance majors from the local university. Robert
Schneider started with a salary of $56,000, Chun Zhang with $52,000, Sunil
Banerjee with $58,000, and Linda Jones with $60,000. Oracle wants to record the
hiring data.
a. Organize the data into a fixed-width format (10 characters for first name, 10
characters for last name, and six characters for salary).
b. Organize the data in a delimited format.
c. Code the data in XML format.
d. Code the data in HTML table format.
e. Code the data in JSON format.
23. The following table lists the population, in millions, in India and China, the two
most populous countries in the world, for the years 2013 through 2017.

Year India China

2013 1278.56 1357.38

2014 1293.86 1364.27

2015 1309.05 1371.22

2016 1324.17 1378.67

2017 1339.18 1386.40

a. Organize the data into a fixed-width format (four characters for Year, eight
characters for India, and eight characters for China).
b. Organize the data in a delimited format.
24. The following table lists the top five countries in the world in terms of their
happiness index on a 10-point scale, and their corresponding GDP per capita as
reported by the United Nations in 2017.

Country Happiness GDP

Finland 7.769 45670

 Country Happiness GDP

Denmark 7.600 57533

Norway 7.544 75295

Iceland 7.494 73060

Netherlands 7.488 48754

a. Organize the data into a fixed-width format (11 characters for Country, 10
characters for Happiness, and six characters for GDP).
b. Organize the data in a delimited format.
25. The following three students were honored at a local high school for securing
admissions to prestigious universities.

Name University

Bridget Yale

Minori Stanford

Matthew Harvard

a. Organize the data into a fixed-width format (10 characters for Name and 10
characters for University).
b. Organize the data in a delimited format.
c. Code the data in XML format.
d. Code the data in HTML table format.
e. Code the data in JSON format.
26. According to Forbes, Michael Trout of the Los Angeles Angels, with earnings of
$39 million, was the highest-paid player in baseball in 2019. Bryce Harper of the
Philadelphia Phillies ranked second at $36.5 million. The Boston Red Sox pitcher
David Price was baseball’s third-highest-paid player at $32 million.
a. Code the data on the player name, baseball team, and salary in the XML
format.
b. Repeat part a using the HTML table format.
c. Repeat part a using the JSON format.
27. According to U.S. News and World Report, a statistician was the best business
profession in 2019 with 12,600 projected jobs and a median salary of $84,060. A
mathematician was the second-best profession with 900 projected jobs and a
 median salary of $103,010. Interestingly, both of these career paths are related to
data analytics.
a. Code the information on the profession, projected jobs, and median salary in
XML format.
b. Repeat part a using the HTML table format.
c. Repeat part a using the JSON format.

page 26

1.5 WRITING WITH BIG DATA

As mentioned at the beginning of this chapter, data are not very useful unless they
are converted into insightful information that is clearly articulated in written or verbal
language. As such, an important aspect of business analytics is to communicate with
numbers rather than focus on number crunching. In this and subsequent chapters,
we include a sample report based on observations and analysis of data to convey the
information in written form. These reports are intended for a nontechnical audience
who may not be familiar with the details of the statistical and computational methods.
Consider the following case study and accompanying report based on music
popularity data.

Case Study
Since 1940, the Billboard magazine has published a variety of weekly music
popularity charts. Today, the magazine uses a combination of sales volume, airplay
on radio, digital downloads, and online streams in order to determine the chart
rankings. One of the most highly watched Billboard charts is the Hot 100, which
provides a weekly ranking of the top 100 music singles. Each entry on the Hot 100
chart lists the current rank, last week’s rank, highest position, and the number of
weeks the song has been on the chart. Other Billboard charts rank music by genre
such as Pop, Rock, R&B, and Latin or rank the popularity of music albums and
artists.
Maya Alexander is a reporter for her university’s newspaper, The Campus
Gazette. She wants to launch a new Film & Music column that will include
commentary, summarized data, and statistics on music popularity. In an effort to
convince her editor to give her this assignment, Maya researches and evaluates
what The Campus Gazette might be able to use from the Billboard website
(http://www.billboard.com).

Sample Report—Billboard Charts

Music is an integral part of campus life. A new Film & Music column in The
Campus Gazette is likely to be very popular among the campus audience.
The Billboard website publishes weekly data on music popularity charts
ranging from the top 100 singles (Hot 100) to digital sales by music genre. We
can summarize these numerical data for music genres that are most popular
among college students and publish them in the new Film & Music column.
The popularity charts on the Billboard website are coded in the HTML data
format and can be easily imported into a table. For example, the Hot 100
chart shown in Figure 1.10 coded as an HTML table can be readily
downloaded into a table in a text document.

FIGURE 1.10 Billboard’s Hot 100 Chart

Source: Billboard.com

page 27
Our readers may also be interested in following the top music singles by
genre. For example, Table 1.8 is an example of a summary table listing the
top five singles in Country, Pop, and Rock, which is easily compiled from
three different Billboard charts. This summary table can be complemented
with written commentaries based on information provided by the popularity
charts such as the highest chart position and the number of weeks on the
chart for each song.

TABLE 1.8 Top Five Hot Singles by Music Genre

According to the most recent sales and market performance, the Billboard
website also organizes songs and albums into categories such as ‘Gains in
Performance,’ ‘Biggest Gain in Streams,’ and ‘Biggest Gain in Digital Sales.’
These songs and albums have not yet reached the top 5 positions, but their
rankings are quickly moving up on the popularity charts. The Campus
Gazette can establish itself as a music trendsetter on campus by introducing
our readers to these up-and-coming songs and albums in a new Film & Music
column. Figure 1.11 shows an example of an up-and-coming single on the
Hot 100 chart. Similar to the popularity charts, these music categories are
formatted using the HTML standard on the Billboard website and can be
readily imported into a text document. We can create commentaries on
selected songs and albums from these lists to introduce up-and-coming
music to our readers.

FIGURE 1.11 Up-and-coming music on Billboard charts

 Source: Billboard.com

Suggested Case Studies

As you have learned throughout the chapter, data from an endless number of online
sources such as corporations, nonprofit organizations, and government agencies are
available for us to explore and investigate. The following are some suggestions using
online, publicly available data.

Report 1.1. Finland is the happiest country in the world, according to the 2018
Happiness Index Report by the United Nations (http://www.worldhappiness.report). In
fact, several Scandinavian countries have consistently held the top spots among the
156 countries included in the annual Happiness Index Report in the past several
years. Visit the Happiness Index website, explore, and write a report based on the
current data provided on the website.

page 29
Report 1.2. Millions of tourists visit Yosemite National Park in California each year.
Stunning waterfalls, giant redwood trees, and spectacular granite rock formations are
among the main attractions at the iconic park. However, the winding roads leading to
the Yosemite Valley may be closed occasionally due to severe weather conditions.
Visit a weather forecast website such as http://www.weather.com and explore the
weather data around Yosemite Park. Write a report to advise a tourist planning a
visit.

page 28
 page 30

2 Data Management and

Wrangling

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 2.1  Describe the key concepts related to data
management.
LO 2.2  Inspect and explore data.
LO 2.3  Apply data preparation techniques to
handle missing values and to subset data.
LO 2.4  Transform numerical variables.
LO 2.5  Transform categorical variables.

ata wrangling is the process of retrieving, cleansing, integrating,

D transforming, and enriching data to support analytics. It is often
considered one of the most critical and time-consuming steps in
an analytics project. In this chapter, we focus on four key tasks
during the data wrangling process: data management, data
inspection, data preparation, and data transformation.
 We first provide an overview of data management. Although data
management is primarily the responsibility of the information
technology group, understanding the relevant concepts, data
structure, and data retrieval technologies is essential to the success
of analytics professionals. After obtaining relevant data, most
analytics professionals spend a considerable amount of time
inspecting, cleaning, and preparing the data for subsequent analysis.
These tasks often involve counting and sorting relevant variables to
review basic information and potential data quality issues. For data
preparation, we discuss two commonly used techniques: handling
missing values and subsetting data. We then examine two strategies
for handling missing values: omission and imputation. Finally, we
focus on data transformation techniques for both numerical and
categorical variables. For numerical variables, common data
transformation techniques include binning, creating new variables,
and rescaling. For categorical variables, common data
transformation techniques include reducing categories, creating
dummy variables, and creating category scores.

page 31
©CustomersYuliia Mazurkevych/Shutterstoick

INTRODUCTORY CASE
Gaining Insights into Retail Customer Data
Organic Food Superstore is an online grocery store that
specializes in providing organic food products to health-
conscious consumers. The company offers a membership-
based service that ships fresh ingredients for a wide range of
chef-designed meals to its members’ homes. Catherine Hill is a
marketing manager at Organic Food Superstore. She has been
assigned to market the company’s new line of Asian-inspired
 meals. Research has shown that the most likely customers for
healthy ethnic cuisines are college-educated millennials (born
between 1982 and 2000).
In order to spend the company’s marketing dollars
efficiently, Catherine wants to focus on this target demographic
when designing the marketing campaign. With the help of the
information technology (IT) group, Catherine has acquired a
representative sample that includes each customer’s
identification number (CustID), sex (Sex), race (Race),
birthdate (BirthDate), whether the customer has a college
degree (College), household size (HouseholdSize), zip code
(ZipCode), annual income (Income), total spending in 2017
(Spending2017), total spending in 2018 (Spending2018), total
number of orders during the past 24 months (NumOfOrders),
number of days since the last order (DaysSinceLast), the
customer’s rating on the last purchase (Satisfaction), and the
channel through which the customer was originally acquired
(Channel). Table 2.1 shows a portion of the data set.

TABLE 2.1 A Sample of Organic Food Superstore Customers

Catherine will use the Customers data set to:

1. Identify Organic Food Superstore’s college-educated
millennial customers.
2. Compare the profiles of female and male college-educated
millennial customers.
A synopsis of this case is provided at the end of Section 2.3.
 page 32

2.1 DATA MANAGEMENT

LO 2.1
Describe the key concepts related to data management.

Data wrangling is the process of retrieving, cleansing, integrating,

transforming, and enriching data to support subsequent data
analysis. This valuable process focuses on transforming the raw
data into a format that is more appropriate and easier to analyze.
The objectives of data wrangling include improving data quality,
reducing the time and effort required to perform analytics, and
helping reveal the true intelligence in the data.
Data wrangling is an essential part of business analytics. As
mentioned in Chapter 1, the increasing volume and variety of data
compel organizations to spend a large amount of time and resources
in garnering, cleaning, and organizing data before performing any
analysis. As the amount of data grows, the need and difficulties
involving data wrangling increase. In practice, the inability to clean
and organize big data is among the primary barriers preventing
organizations from taking full advantage of business analytics.
Analytics professionals can no longer rely solely on the corporate IT
department for data retrieval and preparation. According to
interviews and expert estimates, analytics professionals spend from
50–80% of their time in the mundane task of collecting and preparing
unruly data, before analytics can be applied (The New York Times,
August 17, 2014).
As such, analytics professionals have to become more self-reliant
and possess the necessary skills for data wrangling as well as
performing data analysis. This practice allows organizations to
address business problems much more quickly and make better-
informed decisions. At the same time, the self-service model
requires analytics professionals to have a broader skill set than just
statistical and data mining techniques.

DATA WRANGLING
Data wrangling is the process of retrieving, cleansing,
integrating, transforming, and enriching data to support
subsequent data analysis.

We first provide an overview of data management. In a very

broad sense, data management is a process that an organization
uses to acquire, organize, store, manipulate, and distribute data.
Organizations today have a plethora of data created and stored
using different, often incompatible, technologies. For the past few
decades, most organizations have adopted the database approach
for storing and managing data. This has tremendously improved the
efficiency and effectiveness of the data management process and
ultimately the quality of data. A database is a collection of data
logically organized to enable easy retrieval, management, and
distribution of data.
The most common type of database used in organizations today
is the relational database. A relational database consists of one or
more logically related data files, often called tables or relations. Each
table is a two-dimensional grid that consists of rows (also called
records or tuples) and columns (also called fields or attributes). A
column (e.g., sex of a customer, price of a product, etc.) contains a
characteristic of a physical object (e.g., products or places), an event
(e.g., business transactions), or a person (e.g., customers, students).
A collection of related fields makes up a record, which represents an
object, event, or person. A software application for defining,
manipulating, and managing data in databases is called a database
management system (DBMS). Popular DBMS packages include
Oracle, IBM DB2, SQL Server, MySQL, and Microsoft Access.

DATA MANAGEMENT
Data management is the process that an organization uses to
acquire, organize, store, manipulate, and distribute data.
The most common type of database (a collection of data) is
the relational database. A relational database consists of one or
more logically related data files, where each data file is a two-
dimensional grid that consists of rows and columns.

page 33

Data Modeling: The Entity-Relationship

Diagram
To understand how and where data can be extracted, one needs to
understand the structure of the data, also known as the data model.
Data modeling is the process of defining the structure of a
database. Relational databases are modeled in a way to offer great
flexibility and ease of data retrieval.
An entity-relationship diagram (ERD) is a graphical
representation used to model the structure of the data. An entity is a
generalized category to represent persons, places, things, or events
about which we want to store data in a database table. A single
occurrence of an entity is called an instance. In most situations, an
instance is represented as a record in a database table. For
example, Claire Johnson is an instance of a CUSTOMER entity, and
organic oatmeal is an instance of a PRODUCT entity. Each entity
has specific characteristics called attributes or fields, which are
represented as columns in a database table. Customers’ last names
and product descriptions are examples of attributes in a database
table.
Entity Relationships
Two entities can have a one-to-one (1:1), one-to-many (1:M), or
many-to-many (M:N) relationship with each other that represents
certain business facts or rules. A 1:1 relationship is less common
than the other two types. In a business setting, we might use a 1:1
relationship to describe a situation where each department can have
only one manager, and each manager can only manage one
department. Recall Organic Food Superstore from the introductory
case; Figure 2.1 shows an ERD for the store that illustrates
examples of 1:M and M:N relationships. The diagram shows three
entities: CUSTOMER, ORDER, and PRODUCT. The relationship
between CUSTOMER and ORDER entities is 1:M because one
customer can place many orders over time, but each order can only
belong to one customer. The relationship between ORDER and
PRODUCT is M:N because an order can contain many products and
the same product can appear in many orders.

FIGURE 2.1 Example of an entity relationship diagram

In Figure 2.1, each entity is represented in a rectangular-shaped

box in which attributes of the entity are listed. For each entity, there
is a special type of attribute called primary key (PK), which is an
attribute that uniquely identifies each instance of the entity. For
example, Customer_ID is the primary key for the CUSTOMER entity
because each customer would have a unique ID number. Because
the primary key attribute uniquely identifies each instance of the
entity, it is often used to create a data structure called an index for
fast data retrieval and searches.
 Some entities (e.g., ORDER) have another special type of
attribute called foreign key (FK). A foreign key is defined as a
primary key of a related entity. Because Customer_ID is the primary
key of the CUSTOMER entity, which shares a relationship with the
ORDER entity, it is considered a foreign key in the ORDER entity. A
pair of the primary and foreign keys is used to establish the 1:M
relationship between two entities. By matching the values in the
Customer_ID fields of the CUSTOMER and ORDER entities, we can
quickly find out which customer placed which order.

page 34

The ERD in Figure 2.1 is not yet complete as it is missing the

ordered products and purchase quantities in the ORDER entity.
Storing these data in the ORDER entity is not appropriate as one
does not know in advance how many products are going to be in
each order, therefore making it impossible to create the correct
number of attributes. To resolve this issue, we simply create an
intermediate entity, ORDER_DETAIL, between the ORDER and
PRODUCT entities, as shown in Figure 2.2. As a result, the M:N
relationship is decomposed into two 1:M relationships. An order has
many detailed line items, but each line item can only belong to one
order. While a product may appear in many orders and order lines,
an order line can only contain one product.

FIGURE 2.2 An expanded entity relationship diagram

In the ORDER_DETAIL entity, two attributes, Order_ID and

Product_ID, together create a unique identifier for each instance. In
this situation, Order_ID and Product_ID are referred to as a
composite primary key, which is a primary key that consists of
more than one attribute. We use a composite primary key when
none of the individual attributes alone can uniquely identify each
instance of the entity. For example, neither Order_ID nor Product_ID
alone can uniquely identify each line item of an order, but a
combination of them can uniquely identify each line item. Because
both Order_ID and Product_ID are primary keys of other entities
related to ORDER_DETAIL, they also serve as foreign keys. By
matching the primary and foreign key pair, the system can quickly
find out which order contains a line item for a particular product.

THE ENTITY-RELATIONSHIP DIAGRAM (ERD)

An entity-relationship diagram (ERD) is a schematic used to
illustrate the structure of the data.
An entity is a generalized category to represent persons,
places, things, or events.
A relationship between entities represents certain business
facts or rules. The types of relationships include one-to-one
(1:1), one-to-many (1:M), and many-to-many (M:N).
A primary key is an attribute that uniquely identifies each
instance of an entity, whereas a foreign key is the primary key
of a related entity. A composite primary key is a primary key
that contains more than one attribute.

The data model represented in an ERD can be converted into

database tables. Based on the ERD in Figure 2.2, Table 2.2 shows
various tables that can be created using Organic Food Superstore
sample data. Can you find out which customer placed an order for
organic sweet potato on October 15, 2018? Did this customer order
any other products? By matching the primary and foreign keys of the
CUSTOMER, ORDER, ORDER_DETAIL, and page 35
PRODUCT tables, we can establish relationships among these
tables. With these relationships, we can extract useful information
from multiple tables. For example, using his customer ID (i.e.,
1531136), we learn that James Anderson was the customer who
placed an order on October 15, 2018, using his mobile phone and
paying for the order with his PayPal account. In addition, we can also
see that he purchased organic sweet potato and organic oatmeal.

TABLE 2.2 Database Tables for Organic Food Superstore

a) CUSTOMER table

b) ORDER table

c) ORDER_DETAIL table

Order_ID Product_ID Quantity

1484001 4378 1
 Order_ID Product_ID Quantity

1482141 4305 1

⋮ ⋮ ⋮

1482141 4330 2

d) PRODUCT table

Data Retrieval in the Database Environment

Once data are stored in a relational database, we can retrieve them
using database queries. The most popular query language used
today is Structured Query Language (SQL). SQL is a language for
manipulating data in a relational database using relatively simple and
intuitive commands. While a comprehensive discussion of SQL is
beyond the scope of this text, we briefly demonstrate how simple
SQL statements can be used to retrieve data from a database.
The basic structure of a SQL statement is relatively simple and
usually consists of three keywords: Select, From, and Where. The
Select keyword is followed by the names of attributes we want to
retrieve. The From keyword specifies the tables from which to
retrieve the data. We usually want to retrieve data based on
selection criteria specified in the Where clause. The page 36
following SQL statement retrieves first and last names of the
customers who live in Los Angles from the CUSTOMER table.

Select Last_Name, First_Name

From CUSTOMER
Where City = “Los Angeles”

While simple queries like the previous one are useful, we often
need to compile data from multiple tables and apply more than one
selection criteria. For example, we may want to retrieve the customer
names, order IDs, and order quantities of organic sweet potato
purchases on October 15, 2018. The following SQL query returns
this information.

Select CUSTOMER.Last_Name, CUSTOMER.First_Name,

ORDER.Order_ID, ORDER_DETAIL.Quantity
From CUSTOMER, ORDER, ORDER_DETAIL, PRODUCT
Where PRODUCT.Product_Name = “Organic Sweet Potato”
and ORDER.Order_Date = “10/15/2018”
and CUSTOMER.Customer_ID = ORDER.Customer_ID
and ORDER.Order_ID = ORDER_DETAIL.Order_ID
and ORDER_DETAIL.Product_ID = PRODUCT.Product_ID

Because the Select clause specifies the attributes to retrieve and

these attributes may come from different tables, we use the
table_name.attribute_name format to tell the DBMS from which table
the attributes are retrieved (e.g., CUSTOMER.Last_Name refers to
the last name attribute in the CUSTOMER table). The From clause
lists all the tables that have the attributes to be retrieved and the
attributes used in the selection criteria. The Where clause lists
multiple conditions and links them using the “and” keyword. The last
three conditions match the values of the primary and foreign key
pairs based on the relationships depicted in Figure 2.2.
 STRUCTURED QUERY LANGUAGE (SQL)
In a relational database, data can be retrieved using Structured
Query Language (SQL) statements, whose basic structure
consists of the Select, From, and Where keywords. SQL
statements specify the attributes, tables, and criteria the
retrieved data must meet.

While the SQL commands in the previous two examples provide

the flexibility for data retrieval, many modern DBMS packages such
as Microsoft Access, Oracle, and SQL Server offer a graphical
interface option where the user constructs a database query by
dragging and dropping the query components (e.g., fields, tables,
conditions, etc.) into a form. The DBMS then translates user
selections into SQL statements to retrieve data.

Data Warehouse and Data Mart

Although the relational database environment provides businesses
with an efficient and effective way to manage data, the proliferation
of isolated databases maintained by different business departments
makes it difficult to analyze data across departments and business
units. This phenomenon has also given rise to data redundancy and
inconsistency. As a result, many organizations have developed an
enterprise data warehouse, which offers an integrated, accurate
“single version of truth” to support decision making.

page 37

An enterprise data warehouse or data warehouse is a central

repository of data from multiple departments within an organization.
One of its primary purposes is to support managerial decision
making, and, therefore, data in a data warehouse are usually
organized around subjects such as sales, customers, or products
that are relevant to business decision making. In order to integrate
data from different databases generated by various business
departments, an extraction, transformation, and load (ETL) process
is undertaken to retrieve, reconcile, and transform data into a
consistent format, and then load the final data into a data
warehouse. A data warehouse provides a historical and
comprehensive view of the entire organization.
As you can imagine, the volume of the data in a data warehouse
can become very large very quickly. While a data warehouse
integrates data across the entire organization, a data mart is a
small-scale data warehouse or a subset of the enterprise data
warehouse that focuses on one particular subject or decision area.
For example, a data mart can be designed to support the marketing
department for analyzing customer behaviors, and it contains only
the data relevant to such analyses.
The structure of a data mart conforms to a multidimensional data
model called a star schema, which is a specialized relational
database model. Figure 2.3 displays the star schema for Organic
Food Superstore. In the star schema, there are two types of tables:
dimension and fact tables. The dimension table describes business
dimensions of interest such as customer, product, location, and time.
The fact table contains facts about the business operation, often in
a quantitative format.

FIGURE 2.3 Star schema of a data mart for Organic Food Superstore
 DATA WAREHOUSE AND DATA MART
A data warehouse is a central repository of data from multiple
departments within an organization to support managerial
decision making. Analytics professionals tend to acquire data
from data marts, which are small-scale data warehouses that
only contain data that are relevant to certain subjects or
decision areas. Data in a data mart are organized using a
multidimensional data model called a star schema, which
includes dimension and fact tables.
 Each of the dimension tables has a 1:M relationship with the fact
table. Hence, the primary keys of the dimension tables are also the
foreign keys in the fact table. At the same time, the combination of
the primary keys of the dimension tables forms the composite
primary key of the fact table. The fact table is usually depicted at the
center surrounded by multiple dimension tables forming the shape of
a star. In reality, it is not uncommon to see multiple fact tables
sharing relationships with a group of dimension tables in a data mart.
One of the key advantages of the star schema is its ability to
“slice and dice” data based on different dimensions. For example, in
the data model shown in Figure 2.3, sales data can be retrieved
based on who the customer is; which product or product category is
involved; the year, quarter, or month of the order; where the
customer lives; or through which channel the order was submitted
simply by matching the primary keys of the dimension tables with the
foreign keys of the fact table.
In recent years, new forms of databases to support big data have
emerged. The most notable is the NoSQL or “Not Only SQL”
database. The NoSQL database is a non-relational database that
supports the storage of a wide range of data types including
structured, semi-structured, and unstructured data. It also offers the
flexibility, performance, and scalability needed to handle extremely
high volumes of data. Analytics professionals will likely see NoSQL
databases implemented alongside relational databases to support
organizations’ data needs in today’s environment.

page 38

EXERCISES 2.1
Applications
1. Which of the following statements correctly describe the data wrangling process?
Select all that apply. Explain if incorrect.
 a. Data wrangling is the process of retrieving, cleansing, integrating, transforming,
and enriching data.
b. Data wrangling is the process of defining and modeling the structure of a
database to represent real-world events.
c. The objectives of data wrangling include improving data quality and reducing
the time and effort required to perform analytics.
d. Data wrangling focuses on transforming the raw data into a format that is more
appropriate and easier to analyze.
2. Which of the following statements about entity-relationship diagrams (ERDs) are
correct? Select all that apply. Explain if incorrect.
a. An entity usually represents persons, places, things, or events about which we
want to store data.
b. A foreign key is an attribute that uniquely identifies each instance of the entity.
c. A composite key is a key that consists of more than one attribute.
d. A relationship between entities represents certain business facts or rules.
3. Which of the following statements correctly identify and describe the key elements
of a relational database? Select all that apply. Explain if incorrect.
a. A table in a relational database is a two-dimensional grid that page 39
contains actual data.
b. A field or a column represents a characteristic of a physical object, an event, or
a person.
c. A relational database includes software tools for advanced data visualization.
d. A tuple or a record in a table represents a physical object, an event, or a
person.
4. Which of the following statements best describes what a foreign key is? Select all
that apply. Explain if incorrect.
a. It is an attribute that uniquely identifies each instance of the entity.
b. It is a primary key that consists of multiple attributes.
c. It is the primary key of a related database table.
d. It is a single occurrence of an entity.
5. Which type of relationship―one-to-one (1:1), one-to-many (1:M), or many-to-
many (M:N)―do the following business rules describe?
a. One manager can supervise multiple employees, and one employee may report
to multiple managers.
b. A business department has multiple employees, but each employee can be
assigned to only one department.
c. A company can have only one CEO, and each CEO can work for only one
company.
d. An academic adviser can work with multiple students, while each student is
assigned to only one adviser.
 e. A golf course offers a membership to many members, and a golfer can
potentially sign up for a membership at multiple golf courses.
f. A soccer team consists of multiple players, while an individual player can play
for only one team at a time.
6. Which of the following statements correctly describe benefits of Structured Query
Language (SQL)? Select all that apply. Explain if incorrect.
a. SQL can be used to manipulate structured, semi-structured, and unstructured
data.
b. SQL commands allow users to select data based on multiple selection criteria.
c. SQL can be used to compile data from multiple tables.
d. SQL commands are relatively simple and intuitive.
7. Which of the following statements about data warehouses and data marts are
correct? Select all that apply. Explain if incorrect.
a. A data warehouse is a subset of the enterprise database that focuses on one
particular subject or decision area.
b. The dimension table describes business dimensions of interest, such as
customer, product, location, and time, while the fact table contains facts about
the business operation.
c. A star schema represents a multidimensional data model.
d. A data warehouse is the central repository of data from multiple departments
within a business enterprise to support managerial decision making.

2.2 DATA INSPECTION

LO 2.2
Inspect and explore data.

Once the raw data are extracted from the database, data
warehouse, or data mart, we usually review and inspect the data set
to assess data quality and relevant information for subsequent
analysis. In addition to visually reviewing data, counting and sorting
are among the very first tasks most data analysts perform to gain a
better understanding and insights into the data. Counting and sorting
data help us verify that the data set is complete or that it may have
missing values, especially for important variables. Sorting data also
allows us to review the range of values for each variable. We can
sort data based on a single variable or multiple variables.
In Example 2.1, we demonstrate how to use counting and sorting
features in Excel and R to inspect and gain insights into the data.
While these features also allow us to detect missing values, we
discuss the treatment of missing values in Section 2.3.
FILE
Gig

EXAMPLE 2.1
BalanceGig is a company that matches independent workers
for short-term engagements with businesses in the
construction, automotive, and high-tech industries. The ‘gig’
employees work only for a short period of time, often on a
particular project or a specific task. A manager at BalanceGig
extracts the employee data from their most recent work
engagement, including the hourly wage page 40
(HourlyWage), the client’s industry (Industry), and
the employee’s job classification (Job). A portion of the Gig
data set is shown in Table 2.3.

TABLE 2.3 Gig Employee Data

EmployeeID HourlyWage Industry Job

1 32.81 Construction Analyst

2 46 Automotive Engineer

⋮ ⋮ ⋮ ⋮
 EmployeeID HourlyWage Industry Job

604 26.09 Construction Other

The manager suspects that data about the gig employees

are sometimes incomplete, perhaps due to the short
engagement and the transient nature of the employees. She
would like to find the number of missing observations for the
HourlyWage, Industry, and Job variables. In addition, she
would like information on the number of employees who (1)
worked in the automotive industry, (2) earned more than $30
per hour, and (3) worked in the automotive industry and earned
more than $30 per hour. Finally, the manager would like to
know the hourly wage of the lowest- and the highest-paid
employees at the company as a whole and the hourly wage of
the lowest- and the highest-paid accountants who worked in
the automotive and the tech industries.
Use counting and sorting functions in Excel and R to find
the relevant information requested by the manager, and then
summarize the results.
Important: Due to different fonts and type settings, copying
and pasting Excel or R functions from this text directly into
Excel or R may cause errors. When such errors occur, you may
need to replace special characters such as quotation marks
and parentheses or delete extra spaces in the functions.
SOLUTION:
Using Excel

a. Open the Gig data file. Note that the employee data are
currently sorted by their employee ID in column A. Scroll to the
end of the data set and note that the last record is in row 605.
With the column heading in row 1, the data set has a total of
604 records.
b. We use two Excel functions, COUNT and COUNTA, to inspect
the number of values in each column. The COUNT function
counts the number of cells that contain numeric values and,
therefore, can only apply to the EmployeeID and HourlyWage
variables. The COUNTA function counts the number of cells
that are not empty and is applicable to all four variables.
Because HourlyWage is a numerical variable, we can enter
either =COUNT(B2:B605) or =COUNTA(B2:B605) in an empty
cell to count the number of values for HourlyWage. We get
604 values, implying that there are no missing values.
Similarly, we enter =COUNTA(C2:C605) and
=COUNTA(D2:D605) in empty cells to count the number of
values for the Industry (column C) and Job (column D)
variables. Because these two variables are non-numerical, we
use COUNTA instead of COUNT. Verify that the number of
records for Industry and Job are 594 and 588, respectively,
indicating that there are 10 and 16 blank or missing values,
respectively, in these two variables.
c. To count the number of employees in each industry, we use
the COUNTIF function. Entering
=COUNTIF(C2:C605,“=Automotive”) in an empty cell will show
that 190 of the 604 employees worked in the automotive
industry. Similarly, entering =COUNTIF(B2:B605,“>30”) in an
empty cell will show that 536 employees earned more than
$30 per hour. Note that the first parameter in the COUNTIF
function is the range of cells to be counted, and the second
parameter specifies the selection criterion. Other logical
operators such as >=, <, <=, and <> (not equal to) can also be
used in the COUNTIF function. page 41
d. To count the number of employees with multiple
selection criteria, we use the COUNTIFS function. For
example, entering =COUNTIFS(C2:C605, “=Automotive”,
B2:B605,“>30”) in an empty cell will show that 181 employees
worked in the automotive industry and earned more than $30
per hour. Additional data ranges and selection criteria can be
 added in corresponding pairs. The >=, <, <=, and <> operators
can also be used in the COUNTIFS function.
e. To sort all employees by their hourly wage, highlight cells A1
through D605. From the menu, click Data > Sort (in the Sort &
Filter group). Make sure that the My data has headers
checkbox is checked. Select HourlyWage for the Sort by
option and choose the Smallest to Largest (or ascending)
order. Click OK.
At the top of the sorted list, verify that there are three
employees with the lowest hourly wage of $24.28. To sort data
in descending order, repeat step e but choose the Largest to
Smallest (or descending) order. Verify that the highest hourly
wage is $51.00.
f. To sort the data based on multiple variables, again highlight
cells A1:D605 and go to Data > Sort. Choose Industry in the
Sort by option and the A to Z (or ascending) order. Click the
Add Level button and choose Job in the Then by option and
the A to Z order. Click the Add Level button again and choose
HourlyWage in the second Then by option and the Smallest to
Largest order. Click OK. We see that the lowest- and the
highest-paid accountants who worked in the automotive
industry made $28.74 and $49.32 per hour, respectively.
Similarly, sorting the data by industry in descending order
(Z to A) and then by job classification and hourly wage in
ascending order reveals that the lowest- and the highest-paid
accountants in the Tech industry made $36.13 and $49.49 per
hour, respectively.
g. To resort the data set to its original order, again highlight cells
A1:D605 and go to Data > Sort. Select each of the Then by
rows and click the Delete Level button. Choose EmployeeID in
the Sort by option and the Smallest to Largest order.
Using R
Before following all R instructions, make sure that you have
read Appendix C (“Getting Started with R”). We assume that
 you have downloaded R and RStudio and that you know how
to import an Excel file. Throughout the text, our goal is to
provide the simplest way to obtain the relevant output. We
denote all function names in boldface and all options within a
function in italics.
a. Import the Gig data file into a data frame (table) and label it
myData. Keep in mind that the R language is case sensitive.
b. We use the dim function in R to count the number of
observations and variables. Verify that the R output shows 604
observations and four variables. Enter:
> dim(myData)
c. Two common functions to display a portion of data are head
and View. The head function displays the first few
observations in the data set, and the View function (case
sensitive) displays a spreadsheet-style data viewer where the
user can scroll through rows and columns. Verify that the first
employee in the data set is an analyst who worked in the
construction industry and made $32.81 per hour. Enter:
> head(myData)
> View(myData)
d. R stores missing values as NA, and we use the is.na function
to identify the observations with missing values. R labels
observations with missing values as “True” and page 42
observations without missing values as “False.” In
order to inspect the Industry variable for missing values, enter:
> is.na(myData$Industry)
e. For a large data set, having to look through all observations is
inconvenient. Alternately, we can use the which function
together with the is.na function to identify “which”
observations contain missing values. The following command
identifies 10 observations by row number as having a missing
value in the Industry variable. Verify that the first observation
with a missing Industry value is in row 24. Enter:
 > which (is.na(myData$Industry))
f. To inspect the 24th observation, we specify row 24 in the
myData data frame. Enter:
> myData[24,]
Note that there are two elements within the square bracket,
separated by a comma. The first element identifies a row
number (also called row index), and the second element after
the comma identifies a column number (also called column
index). Leaving the second element blank will display all
columns. To inspect an observation in row 24 and column 3,
we enter myData[24, 3]. In a small data set, we can also
review the missing values by scrolling to the specific rows and
columns in the data viewer produced by the View function. As
mentioned earlier, the treatment of missing values is
discussed in Section 2.3.
g. To identify and count the number of employees with multiple
selection criteria, we use the which and length functions. In
the following command, we identify which employees worked
in the automotive industry with the which function and count
the number of these employees using the length function. The
double equal sign (==), also called equality operator, is used
to check whether the industry is automotive. In R, text
characters such as ‘Automotive’ are enclosed in quotation
marks. Enter:
> length(which(myData$Industry==‘Automotive’))
We can also use the >, >=, <, <=, and != (not equal to)
operators in the selection criteria. For example, using the
following command, we can determine the number of
employees who earn more than $30 per hour. Enter:
> length(which(myData$HourlyWage > 30))
Note that there are 190 employees in the automotive industry
and there are 536 employees who earn more than $30 per
hour.
h. To count how many employees worked in a particular industry
and earned more than a particular wage, we use the and
operator (&). The following command shows that 181
employees worked in the automotive industry and earned
more than $30 per hour. Enter:
> length(which(myData$Industry==‘Automotive’ &
myData$HourlyWage > 30))
i. We use the order function to sort the observations of a
variable. In order to sort the HourlyWage variable and store
the ordered data set in a new data frame called sortedData1,
enter:
> sortedData1 <- myData[order(myData$HourlyWage),]
> View(sortedData1)
page 43
The View function shows that the lowest and
highest hourly wages are $24.28 and $51.00, respectively. By
default, the sorting is performed in ascending order. To sort in
descending order, enter:
> sortedData1 <- myData[order(myData$HourlyWage,
decreasing = TRUE),]
j. To sort data by multiple variables, we specify the variables in
the order function. The following command sorts the data by
industry, job classification, and hourly wage, all in ascending
order, and stores the ordered data in a data frame called
sortedData2. Enter:
> sortedData2 <- myData[order(myData$Industry,
myData$Job, myData$HourlyWage),]
> View(sortedData2)
The View function shows that the lowest-paid accountant who
worked in the automotive industry made $28.74 per hour.
k. To sort the data by industry and job classification in ascending
order and then by hourly wage in descending order, we insert
a minus sign in front of the hourly wage variable. Verify that
 the highest-paid accountant in the automotive industry made
$49.32 per hour. Enter:
> sortedData3 <- myData[order(myData$Industry,
myData$Job, -myData$HourlyWage),]
> View(sortedData3)
l. The industry and job classification variables are non-
numerical. As a result, to sort the data by industry in
descending order and then by job classification and hourly
wage in ascending order, we use the xtfrm function with the
minus sign in front of the Industry variable. Enter:
> sortedData4 <- myData[order(-xtfrm(myData$Industry),
myData$Job, myData$HourlyWage),]
> View(sortedData4)
The View function reveals that the lowest- and the highest-
paid accountants in the technology industry made $36.13 and
$49.49 per hour, respectively.
m. To sort the data by industry, job, and hourly wage, all in
descending order, we use the decreasing option in the order
function. Verify that the highest-paid sales representative in
the technology industry made $48.87. Enter:
> sortedData5 <- myData[order(myData$Industry,
myData$Job, myData$HourlyWage, decreasing = TRUE),]
> View(sortedData5)
n. To export the sorted data from step m as a comma-separated
value file, we use the write.csv function. Verify that the
exported file is in the default folder (e.g., My Document on
Microsoft Windows). Other data frames in R can be exported
using a similar statement. Enter:
> write.csv(sortedData5,“sortedData5.csv”)
Summary

There are a total of 604 records in the data set. There are no
missing values in the HourlyWage variable. The Industry and
 Job variables have 10 and 16 missing values, respectively.
190 employees worked in the automotive industry, 536
employees earned more than $30 per hour, and 181
employees worked in the automotive industry and earned
more than $30 per hour. page 44
The lowest and the highest hourly wages in the
data set are $24.28 and $51.00, respectively. The three
employees who had the lowest hourly wage of $24.28 all
worked in the construction industry and were hired as
Engineer, Sales Rep, and Accountant, respectively.
Interestingly, the employee with the highest hourly wage of
$51.00 also worked in the construction industry in a job type
classified as Other.
The lowest- and the highest-paid accountants who worked in
the automotive industry made $28.74 and $49.32 per hour,
respectively. In the technology industry, the lowest- and the
highest-paid accountants made $36.13 and $49.49 per hour,
respectively. Note that the lowest hourly wage for an
accountant is considerably higher in the technology industry
compared to the automotive industry ($36.13 > $28.74).

There are many ways to count and sort data to obtain useful
insights. To gain further insights, students are encouraged to
experiment with the Gig data using different combinations of
counting and sorting options than the ones used in Example 2.1.

EXERCISES 2.2
Mechanics
8. FILE Exercise_2.8. The accompanying data set contains two numerical
variables, x1 and x2.
a. For x2, how many of the observations are equal to 2?
 b. Sort x1 and then x2, both in ascending order. After the variables have been
sorted, what is the first observation for x1 and x2?
c. Sort x1 and then x2, both in descending order. After the variables have been
sorted, what is the first observation for x1 and x2?
d. Sort x1 in ascending order and x2 in descending order. After the variables have
been sorted, what is the first observation for x1 and x2?
e. How many missing values are there in x1 and x2?
9. FILE Exercise_2.9. The accompanying data set contains three numerical
variables, x1, x2, and x3.
a. For x1, how many of the observations are greater than 30?
b. Sort x1, x2, and then x3 all in ascending order. After the variables have been
sorted, what is the first observation for x1, x2, and x3?
c. Sort x1 and x2 in descending order and x3 in ascending order. After the
variables have been sorted, what is the first observation for x1, x2, and x3?
d. How many missing values are there in x1, x2, and x3?
10. FILE Exercise_2.10. The accompanying data set contains three numerical
variables, x1, x2, and x3, and one categorical variable, x4.
a. For x4, how many of the observations are less than three?
b. Sort x1, x2, x3, and then x4 all in ascending order. After the variables have
been sorted, what is the first observation for x1, x2, x3, and x4?
c. Sort x1, x2, x3, and then x4 all in descending order. After the variables have
been sorted, what is the first observation for x1, x2, x3, and x4?
d. How many missing values are there in x1, x2, x3, and x4?
e. How many observations are there in each category in x4?

Applications
11. FILE SAT. The following table lists a portion of the average writing and math SAT
scores for the 50 states as well as the District of Columbia, Puerto Rico, and the
U.S. Virgin Islands for the year 2017 as reported by the College Board.

State Writing Math

Alabama 595 571

Alaska 562 544

 State Writing Math

⋮ ⋮ ⋮

Wyoming 633 635

a. Sort the data by writing scores in descending order. Which state has the highest
average writing score? What is the average math score of that state?
b. Sort the data by math scores in ascending order. Which state has the lowest
average math score? What is the average writing score of that state?
c. How many states reported an average math score higher than 600?
d. How many states reported an average writing score lower than 550?
12. FILE Fitness. A social science study conducts a survey of 418 individuals about
how often they exercise, marital status, and annual income. A portion of the
Fitness data is shown in the accompanying table.

page 45
ID Exercise Married Income

1 Always Yes 106299

2 Sometimes Yes 86570

⋮ ⋮ ⋮ ⋮

418 Often No 92690

a. Sort the data by annual income. Of the 10 highest income earners, how many
of them are married and always exercise?
b. Sort the data by marital status and exercise both in descending order. How
many of the individuals who are married and exercise sometimes earn more
than $110,000 per year?
c. How many missing values are there in each variable?
d. How many individuals are married and unmarried?
e. How many married individuals always exercise? How many unmarried
individuals never exercise?
13. FILE Spend. A company conducts a consumer survey with questions about
home ownership (OwnHome: Yes/No), car ownership (OwnCar: Yes/No), annual
household spending on food (Food), and annual household spending on travel
(Travel). A portion of the data is shown in the accompanying table.
 a. Sort the data by home ownership, car ownership, and the travel spending all in
descending order. How much did the first customer on the ordered list spend on
food?
b. Sort the data only by the travel spending amount in descending order. Of the 10
customers who spend the most on traveling, how many of them are
homeowners? How many of them are both homeowners and car owners?
c. How many missing values are there in each variable?
d. How many customers are homeowners?
e. How many customers are homeowners but do not own a car?
14. FILE Demographics. The accompanying table shows a portion of data that
shows an individual’s income (Income in $1,000s), age, sex (F = female, M =
male), and marital status (Married; Y = yes, N = no).

a. Count the number of males and females in the data.

b. What percentages of males and females are married?
c. Of the 10 individuals with the highest income, how many are married males.
d. What are the highest and the lowest incomes of males and females?
e. What are the highest and lowest incomes of married and unmarried males?
15. FILE Admission. College admission is a competitive process where, among
other things, the SAT and high school GPA scores of students are evaluated to
make an admission decision. The accompanying data set contains the admission
decision (Decision; Admit/Deny), SAT score, Female (Yes/No), and high school
GPA (HSGPA) for 1,230 students. A portion of the data is shown in the
accompanying table.
 a. Count the number of male and female students.
b. What percentages of male and female students are admitted?
c. Of the 10 students with the highest HSGPA, how many are males?
d. Of the 10 students with the lowest SAT, how many are females?
e. What are the highest and the lowest SAT scores of admitted male and female
students?

2.3 DATA PREPARATION

LO 2.3
Apply data preparation techniques to handle missing values and to subset data.

Once we have inspected and explored data, we can start the data
preparation process. In this section, we examine two important data
preparation techniques: handling missing values and subsetting
data. As mentioned in Section 2.2, there may be missing values in
the key variables that are crucial for subsequent analysis. Moreover,
most data analysis projects focus only on a portion (subset) of the
data, rather than the entire data set; or sometimes the objective of
the analysis is to compare two subgroups of the data.

page 46

Handling Missing Values

Missing values are a common data quality problem found in data
sets both large and small. This issue can lead to significant reduction
in the number of usable observations and biases in analysis results.
For example, in a data set with 20 variables, if 5% of the values,
spread randomly across the observations and variables, are missing,
then potentially only (1 − 0.05)20 = 0.3585, or 35.85% of the
observations, would be complete and usable. Understanding why the
values are missing is the first step in the treatment of missing values.
Sometimes data are missing because the respondents decline to
provide the information due to its sensitive nature (e.g., race, sexual
orientation, economic status, etc.). In these cases, missing values
are usually not distributed randomly across observations but tend to
concentrate within one or more subgroups. For example, research
has shown that male respondents often skip questions related to
depression in a survey.
In other cases, data values are missing because some of the
items do not apply to every respondent. For instance, patients who
are still in the hospital do not have values in the discharge date
column. Missing values can also be caused by human errors, sloppy
data collection, and equipment failures.
Because missing values are often unavoidable in real life, it is
important to learn how to handle observations with missing values.
There are two common strategies for dealing with missing values:
omission and imputation. The omission strategy, also called
complete-case analysis, recommends that observations with missing
values be excluded from the analysis. This approach is appropriate
when the amount of missing values is small or missing values are
concentrated in a small number of observations. Otherwise, it can
lead to severe loss of data, as illustrated earlier.
The imputation strategy replaces missing values with some
reasonable imputed values. The most commonly used imputation
strategy for numerical variables is the simple mean imputation where
the missing values are replaced with the mean (average) values
across relevant observations. For example, if the annual household
income for an observation is missing, we replace the missing value
with the mean household income across all observations or across a
homogenous group (e.g., households with the same zip code).
 Simple mean imputation is easy to implement and allows
observations with missing values to be included in the analysis
without adding more variability to the data set. However, if a large
number of missing values need to be imputed, simple mean
imputation will likely distort the relationships among variables,
leading to biased results. For example, the total square footage of a
house tends to have a positive relationship with the value of the
house. If a data set contains many missing total square footage
values and these missing values are replaced with mean total
square footage of the rest of the houses in the data set, then the
relationship between total square footage and house value will likely
be distorted. More advanced imputation techniques such as
regression mean imputation that better preserve the relationships
among variables can be used in these cases. Advanced imputation
techniques are beyond the scope of this text.
In the case of categorical variables, the most frequent category is
often used as the imputed value. For instance, if some values of sex
are missing in a data set, we might replace these missing values
with the predominant sex category among the rest of the
observations. For categorical variables, an “Unknown” category may
be created to signify missing values. This approach is especially
useful if the data are missing for a reason; the fact that the values
are missing for these observations may suggest certain patterns and
relationships in the data.
In addition to the omission and imputation strategies, other
approaches may be considered when handling missing values. If the
variable that has many missing values is deemed unimportant or can
be represented using a proxy variable that does not have missing
values, the variable may be excluded from the analysis. Finally,
some analytics techniques such as decision trees (discussed in
Chapter 10) are robust and can be applied to data sets even with the
inclusion of missing values.

page 47
 HANDLING MISSING VALUES
There are two common strategies for dealing with missing
values.
The omission strategy recommends that observations with
missing values be excluded from subsequent analysis.
The imputation strategy recommends that the missing values
be replaced with some reasonable imputed values. For
numerical variables, it is common to use mean imputation. For
categorical variables, it is common to impute the most
predominant category.

In Example 2.2, we demonstrate how to handle missing values

with the omission and imputation strategies using Excel with Analytic
Solver and R.
FILE
Restaurant_Reviews

EXAMPLE 2.2
Sarah Johnson, the manager of a local restaurant, has
conducted a survey to gauge customers’ perception about the
eatery. Each customer rated the restaurant on its ambience,
cleanliness, service, and food using a scale of 1 (lowest) to 7
(highest). Table 2.4 displays a portion of the survey data.

TABLE 2.4 Restaurant Reviews

 Sarah notices that there are a number of missing values in the
survey. Use the Restaurant_Reviews data to first detect the
missing values. Then use both omission and imputation
strategies to handle the missing values.
SOLUTION:
Using Excel with Analytic Solver

a. Open the Restaurant_Reviews data file.

b. We use the Conditional Formatting feature in Excel to detect
missing values. Select the data range B2:D151. Choose
Home > Conditional Formatting > New Rule. . .. Select
Format only cells that contain as the rule type. Select Format
only cells with: Blanks as the rule description. Click on the
Format button. Select the Fill tab and pick the color red. Click
OK. Click OK again. Now the cells with missing values are
highlighted in red.
c. Here, we use the COUNTBLANK function to detect the
number of missing values for each observation. Create a new
column with the column heading “# of Missing Values” in
column F. Enter the formula =COUNTBLANK(B2:E2) in cell
F2. Fill the range F3:F151 with the formula in F2. This formula
shows “0” in the result for the observations without missing
values. For the observations with missing values, the formula
shows the number of variables with missing values. Verify that
observations 2, 13, 26, and 100 each has one missing value,
and observation 134 has two missing values.
d. To count the total number of missing values in the entire data
set, enter the formula =SUM(F2:F151) in cell F152. The result
shows that there is a total of six missing values.
e. To identify variables with missing values, enter the page 48
formula =COUNTBLANK(B2:B151) in cell B152.
Fill the range C152:E152 with the formula in B152. The result
shows that the Ambience (column B), Cleanliness (column C),
and Food (column E) variables each has one missing value,
and the Service variable (column D) has three missing values.
f. We start with the omission strategy to handle missing values.
Choose Data Mining > Transform > Missing Data Handling.
Select data range $B$1:$E$151. Check the box Variable
names in the first row. Select Ambience, Cleanliness, Service,
and Food by holding the Ctrl button. Choose Delete record as
the treatment. Click on the Apply to selected variable(s)
button. Click OK. A new worksheet Imputation is created and
shows a transformed data table with the observations with
missing values deleted. A part of the Imputation worksheet is
displayed in Figure 2.4. The results show that 5 observations
were deleted due to missing values, leaving 145 complete
observations in the final data set. The transformed data set is
ready for data analysis.

FIGURE 2.4 Omission strategy results

g. We now use the simple mean imputation strategy to handle

the missing values. Return to the Restaurant_Reviews
worksheet. Choose Data Mining > Transform > Missing
Data Handling. Select data range $B$1:$E$151. Check the
box Variable names in the first row. Select Ambience,
Cleanliness, Service, and Food by holding the Ctrl button.
Choose Mean as the treatment. Click on the Apply to selected
variable(s) button. Another new worksheet called Imputation1
is created and shows all 150 observations. Figure 2.5 displays
part of the Imputation1 worksheet. The results show that the
mean values were used to replace the missing values and that
all 150 observations are retained. Verify that the service
ratings for observations 2, 100, and 134 are now 5.97, which
 is the mean rating. Again, the transformed data set is ready for
data analysis.

FIGURE 2.5 Imputation strategy results

Using R

a. Import the Restaurant_Reviews data into a data frame (table)

and label it myData.
b. To detect missing values in a data set, we use the is.na
function. Recall from Example 2.1 that missing values are
labelled as NA in R, and the is.na function returns TRUE if a
missing value is detected and FALSE otherwise. Enter:
> is.na(myData)
Figure 2.6 displays a portion of the output, where we have
highlighted the missing values. The service rating for
observation 2 and cleanliness rating for observation 13 are
labeled TRUE because they are missing. The rest of the
values are labeled FALSE because they are not missing.
page 49
FIGURE 2.6 R output for detecting missing values
 To detect missing values, say, in the Service variable, we
enter:
> is.na(myData$Service)
c. If we have a large data set, using the is.na function to detect
missing values can be cumbersome. Alternatively, we can use
the complete.cases function to identify the rows in the data
frame or cases that are complete. Recall that leaving the
second element, after the comma, blank will display all
columns. Enter:
> myData[complete.cases(myData), ]
d. Because our data are mostly complete, listing all the complete
cases may not be convenient. Instead, we can use the not
operator (! character) with the complete.cases function to
identify observations with missing values. The ! character in
front of the complete.cases function identifies individual rows
that are not complete (or cases with missing values). Enter:
> myData[!complete.cases(myData), ]
Figure 2.7 shows observations 2, 13, 26, 100, and 134 have
missing values.

FIGURE 2.7 Observations with missing values

e. To implement the omission strategy, we use the na.omit
function to remove observations with missing values and store
the resulting data set in the omissionData data frame. The
View function displays the updated data. Enter:
> omissionData <- na.omit(myData)
> View(omissionData)
R creates a new data frame, omissionData, that contains 145
complete cases. Verify that there are no missing values in the
new data set. This data set is ready for data analysis.
f. To implement the simple mean imputation strategy, we start
with the original data frame, myData. We then calculate the
average value using the mean function. The option na.rm =
TRUE ignores the missing values when calculating the
average values. In order to compute the average values for
the Ambience and Service variables, enter:
> ambienceMean <- mean(myData$Ambience, na.rm =
TRUE)
> serviceMean <- mean(myData$Service, na.rm = TRUE)
Verify that the means for the Ambience and Service variables
are 4.42953 and 5.965986, respectively.
g. To impute the missing values in the Ambience and page 50
Service variables, we again use the is.na function
to identify the missing values and replace them with the
means calculated in step f. Enter:
> myData$Ambience[is.na(myData$Ambience)] <-
ambienceMean
> myData$Service[is.na(myData$Service)] <- serviceMean
 Students are encouraged to calculate the mean and impute
missing values in the Cleanliness and Food variables and
inspect the resulting data set to make sure that the missing
values have been replaced by the average ratings. The
resulting data set is then ready for data analysis.

Another important data preparation task involves the treatment of

extremely small or large values, referred to as outliers. It is
noteworthy that in the presence of outliers, it is preferred to use the
median instead of the mean to impute missing values. Both Analytic
Solver and R allow easy imputation with the median computed by
using the MEDIAN(data range) function in Excel or the median
function in R. Refer to Chapter 3 for a detailed discussion on outliers
and the median.

Subsetting
The process of extracting portions of a data set that are relevant to
the analysis is called subsetting. It is commonly used to pre-
process the data prior to analysis. For example, a multinational
company has sales data for its global operations, and it creates a
subset of sales data by country and performs analysis accordingly.
For time series data, which are data indexed in time order, we may
choose to create subsets of recent observations and observations
from the distant past in order to analyze them separately. Subsetting
can also be used to eliminate unwanted data such as observations
that contain missing values, low-quality data, or outliers. Sometimes,
subsetting involves excluding variables instead of observations. For
example, we might remove variables that are irrelevant to the
problem, variables that contain redundant information (e.g., property
value and property tax or employee’s age and experience), or
variables with excessive amounts of missing values.
 SUBSETTING
Subsetting is the process of extracting parts of a data set that is
of interest to the analytics professional.

Subsetting can also be performed as part of descriptive analytics

that helps reveal insights in the data. For example, by subsetting
student records into groups with various academic performance
levels, we may be able to identify high-achieving students and
relevant attributes that contribute to their success. Similarly, by
comparing subsets of medical records with different treatment
results, we may identify potential contributing factors of success in a
treatment.
Table 2.5 shows important summary measures from two subsets
of medical records of tuberculosis treatments administered in a
developing country. Here, subsets 1 and 2 represent successful and
unsuccessful treatments, respectively.

TABLE 2.5 Summary Data of Tuberculosis Treatment

Successful Unsuccessful
Summary Measures
Treatments Treatments

% of Male Patients 64.3% 12.7%

Average Education 3.8 2.1

Level

% of Patients with 92.7% 28.1%

Good Incomes

page 51
 Note that the sex, education level (1: lowest; 5: highest), and
income of the patients differ considerably between the two subsets.
Not surprisingly, male patients, especially those with higher
education and income levels, have better success with tuberculosis
treatment than female patients with lower education and income
levels. This simple analysis highlights the importance of contributing
factors in tuberculosis control efforts.
In Example 2.3, we demonstrate how to use subsetting functions
in Excel and R to select or exclude variables and/or observations
from the original data set.
FILE
Customers

EXAMPLE 2.3
In the introductory case, Catherine Hill wants to gain a better
understanding of Organic Food Superstore’s customers who
are college-educated millennials, born between 1982 and
2000. She feels that sex, household size, income, total
spending in 2018, total number of orders in the past 24 months,
and channel through which the customer was acquired are
useful for her to create a profile of these customers. Use Excel
and R to first identify college-educated millenial customers in
the Customers data file. Then, create subsets of female and
male college-educated millenial customers. The synopsis that
follows this example provides a summary of the results.
SOLUTION:
Using Excel

a. Open the Customers data file.

b. We first filter the data set to include only college-educated
millennials. Select the data range A1:N201. From the menu
choose Home > Sort & Filter > Filter. The column headings
(A1:N1) will turn into drop-down boxes.
c. Click on the drop-down box in E1 (College). Uncheck (Select
all), then check the box next to Yes. Click OK. This step
shows only those customers who have a college degree (Yes)
by hiding those who don’t (No) in the data set.
d. Click on the drop-down box in D1 (BirthDate). Select Date
filters > Between. See Figure 2.8. In the Custom AutoFilter
dialog box, enter “1/1/1982” next to the is after or equal to box
or select the date from the calendar object. Select And and
enter “12/31/1999” next to the is before or equal to box or
select the date from the calendar object. Click OK. The data
set now only displays college-educated millennials who were
born between 1982 and 2000.

FIGURE 2.8 Excel’s AutoFilter dialog box

e. Select the entire filtered data that are left in the worksheet.
Copy and paste the filtered data to a new worksheet. Verify
that the new worksheet contains 59 observations of college-
educated millennials. Rename the new worksheet as College-
Educated Millennials.
f. We now exclude the variables that are not relevant page 52
to the current analysis. In the College-Educated
Millennials worksheet, select cell A1 (CustID). From the menu
choose Home > Delete > Delete Sheet Columns to remove
the CustID column. Repeat this step for the Race, BirthDate,
College, ZipCode, Spending2017, DaysSinceLast, and
Satisfaction columns from the data set.
g. To subset the college-educated millennials data by sex, select
column A. From the menu choose Home > Sort & Filter >
Sort A to Z. If prompted, select Expand the selection in the
Sort Warning dialog box and click Sort. The observations are
now sorted by sex in alphabetic order. The female customer
records are followed by male customer records.
h. Create two new worksheets and assign the worksheet names
Female and Male. Copy and paste the female and male
customer records, including the column headings, to the new
Female and Male worksheets, respectively. Table 2.6 shows a
portion of the results.

TABLE 2.6 College-Educated Millennial Customers

a) Female College-Educated Millennials

b) Male College-Educated Millennials

Using R

a. Import the Customers data into a data frame (table) and label
it myData.
b. To select college-educated millennials, we first select all
customers with a college degree. Recall that the double equal
sign (==) is used to check whether the College value is “Yes.”
Enter:
> college <- myData[myData$College==‘Yes’, ]
c. We now use the Birthdate variable to select the millennials
who were born between 1982 and 2000. R usually imports the
date values as text characters, and, therefore, we first need to
convert the BirthDate variable into the date data type using the
as.Date function. The option format = “%m/%d/%Y” indicates
that the BirthDate variable is in the mm/dd/yyyy format. For
example, in order for R to read dates such as 01/13/1990,
enter:
> college$BirthDate <- as.Date(college$BirthDate, format =
“%m/%d/%Y”)
Other common date formats include “%Y-%m-%d”, “%b %d,
%Y”, and “%B %d, %Y” that will read dates specified as 1990-
01-13, Jan 13, 1990, and January 13, 1990, respectively.
d. We also use the as.Date function to specify the page 53
cutoff dates, January 1, 1982, and December 31,
1999, before using them as selection criteria for selecting the
millennials in our data. Enter:
> cutoffdate1 <- as.Date(“01/01/1982”, format = “%m/%d/%Y”)
> cutoffdate2 <- as.Date(“12/31/1999”, format = “%m/%d/%Y”)
> millenials <- college[college$BirthDate >= cutoffdate1 &
college$BirthDate <= cutoffdate2, ]
Verify that the millennials data frame contains 59 college-
educated millennials.
e. To include only the Sex, HouseholdSize, Income,
Spending2018, NumOfOrders, and Channel variables in the
millenials data frame, we specify the column indices of these
variables using the c function. Enter:
> subset1 <- millenials[ , c(2,6,8,10,11,14)]
 Alternately, we can create a new data frame by specifying the
names of the variables to include. Enter:
> subset2 <- millenials[ , c(“Sex”, “HouseholdSize”, “Income”,
“Spending2018”, “NumOfOrders”, “Channel”)]
Note that subset1 and subset2 data are identical.
f. R imports non-numerical variables such as Sex and Channel
as text characters. Before further subsetting and examining
the data, we convert Sex and Channel into categorical
variables (called factors in R) by using the as.factor function.
Enter:
> subset1$Sex <- as.factor(subset1$Sex)
> subset1$Channel <- as.factor(subset1$Channel)
To verify that the Channel variable has been converted into a
factor or a categorical variable, enter:
> is.factor(subset1$Channel)
This command returns TRUE if the variable is a factor, and
FALSE otherwise.
g. To create two subsets of data based on Sex, we use the split
function. Enter:
> sex <- split(subset1, subset1$Sex)
The sex data frame contains two subsets: Female and Male.
We can now access and view the Female and Male subsets.
Enter:
> sex$Female
> sex$Male
> View(sex$Female)
> View(sex$Male)
Verify that there are 21 female college-educated millennials
and 38 male college-educated millennials. Your results should
be similar to Table 2.6.
h. In some situations, we might simply want to subset data based
on data ranges. For example, we use the following statement
to subset data to include observations 1 to 50 and
observations 101 to 200. Enter:
> dataRanges <- myData[c(1:50, 101:200),]

page 54

SYNOPSIS OF INTRODUCTORY CASE

Catherine Hill has been assigned to help market Organic Food Superstore’s new line
of Asian-inspired meals. In order to understand the potential target market for this
product, Catherine subsetted the data that contain a representative sample of the
company’s customers to include only college-educated millennials. She also
partitioned the data set into two subsets based on sex to compare the profiles of
female and male college-educated millennials.

hbpictures/Shutterstock

The differences between the female and male customers have given Catherine
some ideas for the marketing campaign that she is about to design. For example,
the data show that an overwhelming portion of the male customers were acquired
through social media ads, while female customers tend to be enticed by web ads or
referrals. She plans to design and run a series of social media ads about the new
 product line with content that targets male customers. For female customers,
Catherine wants to focus her marketing efforts on web banner ads and the
company’s referral program.
Furthermore, as the male customers seem to place more frequent but smaller
orders than female customers do, Catherine plans to work with her marketing team
to develop some cross-sell and upsell strategies that target male customers. Given
the fact that the company’s male college-educated millennial customers tend to be
high-income earners, Catherine is confident that with the right message and product
offerings, her marketing team will be able to develop strategies for increasing the
total spending of these customers.

EXERCISES 2.3
Mechanics
16. The following table contains three variables and five observations with some
missing values.

x1 x2 x3

248 3.5

124 3.8 55

150 74

196 4.5 32

6.2 63

a. Handle the missing values using the omission strategy. How many observations
remain in the data set and have complete cases?
b. Handle the missing values using the simple mean imputation strategy. How
many missing values are replaced? What are the means of x1, x2, and x3?
17. FILE Exercise_2.17. The accompanying data set contains four variables, x1, x2,
x3, and x4.
a. Subset the data set to include only observations that have a date on or after
May 1, 1975, for x3. How many observations are in the subset data?
b. Split the data set based on the binary values for x4. What are the average
values for x1 for the two subsets?
18. FILE Exercise_2.18. The accompanying data set contains five variables, x1, x2,
x3, x4, and x5.
a. Subset the data set to include only x2, x3, and x4. How many missing values
are there in the three remaining variables?
b. Remove all observations that have “Own” as the value for x2 and have values
lower than 150 for x3. How many observations remain in the data set? What are
the average values for x3 and x4?
19. FILE Exercise_2.19. The accompanying data set contains five variables, x1, x2,
x3, x4, and x5. There are missing values in the data set.
a. Which variables have missing values?
b. Which observations have missing values?
c. How many missing values are in the data set?
d. Handle the missing values using the omission strategy. How many observations
remain in the data set and have complete cases?
20. FILE Exercise_2.20. The accompanying data set contains five variables, x1, x2,
x3, x4, and x5. There are missing values in the data set. Handle the missing
values using the simple mean imputation strategy for numerical variables and the
predominant category strategy for categorical variables.
a. How many missing values are there for each variable?
b. What are the values for imputing missing values in x1, x2, x3, page 55
x4, and x5?
21. FILE Exercise_2.21. The accompanying data set contains five variables, x1, x2,
x3, x4, and x5.
a. Are there missing values for x1? If so, impute the missing values using the
mean value of x1. After imputation, what is the mean of x1?
b. Are there missing values for x2? If so, impute the missing values using the
mean value of x2. After imputation, what is the mean of x2?
c. If there are missing values in x4, impute the missing values using the median
value of x4. (Hint: Use the MEDIAN(data range) function in Excel or the
median function in R.) After imputation, what is the median of x4?
22. FILE Exercise_2.22. The accompanying data set contains five variables, x1, x2,
x3, x4, and x5. There are missing values in the data set.
a. Which variables have missing values?
b. Which observations have missing values?
c. How many missing values are in the data set?
d. Handle the missing values using the omission strategy. How many observations
remain in the data set and have complete cases?
23. FILE Exercise_2.23. The accompanying data set contains four variables, x1, x2,
x3, and x4. There are missing values in the data set.
a. Subset the data set to include only x1, x2, and x3.
b. Which variables have missing values?
c. Which observations have missing values?
d. How many missing values are in the data set?
e. Handle the missing values using the omission strategy. How many observations
remain in the data set and have complete cases?
f. Split the data set based on the categories of x2. How many observations are in
each subset?
24. FILE Exercise_2.24. The accompanying data set contains seven variables, x1,
x2, x3, x4, x5, x6, and x7. There are missing values in the data set.
a. Remove variables x2, x6, and x7 from the data set. Which of the remaining
variables have missing values?
b. Which observations have missing values?
c. How many missing values are in the data set?
d. If there are missing values in x1, replace the missing values with “Unknown.”
How many missing values were replaced?
e. Handle the missing values for numerical variables using the imputation strategy.
If there are missing values in x3, impute the missing values using the mean
value of x3. If there are missing values in x4, impute the missing values using
the median value of x4. If there are missing values in x5, impute the missing
values using the mean value of x5. What are the average values of x3, x4, and
x5 after imputation?
f. Remove observations that have the value “F” for x1 and values lower than
1,020 for x4. How many observations remain in the data set?

Applications
25. FILE Population. The US Census Bureau records the population for the 50
states each year. The accompanying table shows a portion of these data for the
years 2010 to 2018.
 a. Create two subsets of the state population data: one with 2018 population great
than or equal to 5 million and one with 2018 population less than 5 million. How
many observations are in each subset?
b. In the subset of states with 5 million or more people, remove the states with
over 10 million people. How many states were removed?
26. FILE Travel_Plan. Jerry Stevenson is the manager of a travel agency. He wants
to build a model that can predict whether or not a customer will travel within the
next year. He has compiled a data set that contains the following variables:
whether the individual has a college degree (College), whether the individual has
a credit card (CreditCard), annual household spending on food (FoodSpend in $),
annual income (Income in $), and whether the customer has plans to travel within
the next year (TravelPlan, 1 = have travel plans; 0 = do not have travel plans). A
portion of the Travel_Plan data is shown in the accompanying table.

a. Are there any missing values in the data set? If there are, which variables have
missing values? How many missing values are there in the data set?
b. Use the omission strategy to handle missing values. How many observations
are removed due to missing values?
27. FILE Travel_Plan. Refer to the previous exercise for a description of the problem
and data set.
a. Based on his past experience, Jerry knows that whether the individual has a
credit card or not has nothing to do with his or her travel plans and would like to
remove this variable. Remove the variable CreditCard from the data set.
b. In order to better understand his customers with high incomes, page 56
Jerry wants to create a subset of the data that only includes
customers with annual incomes higher than $75,000 and who plan to travel
within the next year. Subset the data to build the list of customers who meet
these criteria. How many observations are in this subset?
c. Return to the original data set. Use the imputation strategy to handle missing
values. If there are missing values for the FoodSpend variable, impute the
missing values using the mean of the variable. If there are missing values for
the Income variable, impute the missing values using the median of the
variable. What are the average values of FoodSpend and Income after
imputation?
28. FILE Football_Players. Denise Lau is an avid football fan and religiously follows
every game of the National Football League. During the 2017 season, she
meticulously keeps a record of how each quarterback has played throughout the
season. Denise is making a presentation at the local NFL fan club about these
 quarterbacks. The accompanying table shows a portion of the data that Denise
has recorded, with the following variables: the player’s name (Player), team’s
name (Team), completed passes (Comp), attempted passes (Att), completion
percentage (Pct), total yards thrown (Yds), average yards per attempt (Avg),
yards thrown per game (Yds/G), number of touch downs (TD), and number of
interceptions (Int).

a. Are there any missing values in the data set? If there are, which variables have
missing values? Which observations have missing values? How many missing
values are there in the data set?
b. Use the omission strategy to handle missing values. How many observations
are removed due to missing values?
29. FILE Football_Players. Refer to the previous exercise for a description of the
data set. Denise feels that, for her presentation, it would remove some biases if
the player names and team names are suppressed. Remove these variables from
the data set.
a. Denise also wants to remove outlier cases where the players have less than
five touchdowns or more than 20 interceptions. Remove these observations
from the data set. How many observations were removed from the data?
b. Return to the original data set. Use the imputation strategy to handle missing
values. If there are missing values for Comp, Att, Pct, Yds, Avg, or Yds/G,
impute the missing values using the mean of the variable. If there are missing
values for TD or Int, impute the missing values using the median of the variable.
What are the average values of Comp, Att, Pct, Yds, Avg, and Yds/G after
imputation?
30. FILE Salaries. Ian Stevens is a human resource analyst working for the city of
Seattle. He is performing a compensation analysis of city employees. The
accompanying data set contains three variables: Department, Job Title, and
Hourly Rate (in $). A few hourly rates are missing in the data.

Department Job Title Hourly Rate

Public Utilities Res&Eval Asst 32.81

Sustainability & Environ Dept StratAdvsr3,Exempt 62.27

 Department Job Title Hourly Rate

⋮ ⋮ ⋮

Public Utilities Capital Prjts Coord, Asst 42.71

a. Split the data set into a number of subsets based on Department. How many
subsets are created?
b. Which subset contains missing values? How many missing values are in that
data set?
c. Use the imputation strategy to replace the missing values with the mean of the
variable. What is the average hourly rate for each subset?
31. FILE Stocks. Investors usually consider a variety of information to make
investment decisions. The accompanying table displays a sample of large publicly
traded corporations and their financial information. Relevant information includes
stock price (Price), dividend as a percentage of share price (Dividend), price to
earnings ratio (PE), earnings per share (EPS), book value, lowest and highest
share prices within the past 52 weeks (52 wk low and 52 wk high), market value
of the company’s shares (Market cap), and earnings before interest, taxes,
depreciation, and amortization (EBITDA in $ billions).

a. As the price to earnings ratio is often considered a better assessment of stock

valuation than stock price or earnings per share alone, the financial analyst
would like to remove Price and EPS from the data set. Remove these variables
from the data set.
b. The financial analyst is most interested in companies with a higher book value
than market cap. Remove all observations that have a lower
book value than market cap. How many observations are left in
page 57
the data set?
32. FILE Stocks. Refer to the previous exercise for a description of the problem and
data set. The financial analyst wants to find out if there are any missing values in
the data set.
a. Are there any missing values in the data set? If there are, which variables have
missing values? Which observations have missing values? How many missing
values are there in the data set?
 b. Use the omission strategy to handle missing values. How many observations
are removed due to missing values?
c. Return to the original data set. If there are missing values for Price, Dividend,
Book Value, 52 wk low, or 52 wk high, replace the missing value with “M,” which
stands for “Missing.” If there are missing values for PE, EPS, Market cap, or
EBITDA, use the imputation strategy to replace the missing values with the
median of the variable. What are the imputed values for the variables with
missing data?
33. FILE Longitudinal_Partial. The accompanying table contains a portion of data
from the National Longitudinal Survey (NLS), which follows over 12,000
individuals in the United States over time. Variables in this analysis include the
following information on individuals: Urban (1 if lives in urban area, 0 otherwise),
Siblings (number of siblings), White (1 if white, 0 otherwise), Christian (1 if
Christian, 0 otherwise), FamilySize, Height, Weight (in pounds), and Income (in
$).

a. Remove Height and Weight from the data set.

b. Subset the data to create two data sets, one that includes observations who live
in an urban area and have incomes over $40,000 and one that includes the rest
of the observations.
34. FILE Longitudinal_Partial. Refer to the previous exercise for a description of
the data set.
a. Are there any missing values in the data set? If there are, which variables have
missing values? Which observations have missing values? How many missing
values are there in the data set?
b. Use the omission strategy to handle missing values. How many observations
are removed due to missing values?
c. Return to the original data set. If there are missing values for Height or Weight,
use the imputation strategy to replace the missing values with the mean of the
variable. If there are missing values for Siblings, FamilySiz, or Income, replace
the missing values with the median of the variable. What are the imputed values
of the variables with missing data?
2.4 TRANSFORMING NUMERICAL
DATA
LO 2.4
Transform numerical variables.

Data transformation is the data conversion process from one

format or structure to another. It is performed to meet the
requirements of statistical and data mining techniques used for the
analysis. Examples of transforming numerical data include
transforming an individual’s date of birth to age, combining height
and weight to create body mass index, calculating percentages, or
converting values to natural logarithms. Sometimes it makes sense
to group a vast range of numerical values into a small number of
“bins.” For example, we might want to arrange numerical ages into
age intervals such as 20 to 39, 40 to 59, and 60 to 80, or convert
dates into seasons such as fall, winter, spring, and summer.

DATA TRANSFORMATION
Data transformation is the data conversion process from one
format or structure to another.

In this section, we describe techniques for transforming numerical

variables into categorical values (or binning) and mathematical
transformations of numerical variables. Section 2.5 explores
techniques for transforming categorical variables.

page 58

Binning
Binning is the process of transforming numerical variables into
categorical variables by grouping the numerical values into a small
number of groups or bins. It is important that the bins are
consecutive and nonoverlapping so that each numerical value falls
into one, and only one, bin. For example, we might want to transform
income values into three groups: below $50,000, between $50,000
and $100,000, and above $100,000. The three income groups can
be labeled as “low,” “medium,” and “high” or “1,” “2,” and “3.” Binning
can be an effective way to reduce noise in the data if we believe that
all observations in the same bin tend to behave the same way. For
example, the transformation of the income values into three groups
makes sense when we are more interested in a person’s earning
power (low, medium, or high) rather than the actual income value.

BINNING
Binning is a common data transformation technique that
converts numerical variables into categorical variables by
grouping the numerical values into a small number of bins.

As noted above, binning reduces the noise in the data often due
to minor observation errors. For example, with binning, outliers in the
data (e.g., individuals with extremely high income, perhaps recorded
incorrectly) will be part of the last bin and, therefore, will not distort
subsequent data analysis. Binning is also useful in categorizing
observations and meeting the categorical data requirements of some
data mining analytics techniques such as naïve Bayes (discussed in
Chapter 9).
In addition to binning numerical values according to user-defined
boundaries, bins are also often created to have equal intervals. For
example, we can create bins that represent an interval of 10 degrees
in temperature or 10 years in age. We can also create bins of equal
counts, where individual bins have the same number of
observations. For example, by binning a class of 200 students into
10 equal-size groups based on their grades, we can find out the
relative academic standing of the students. Students in the bin with
the highest grades represent the top 10% of the class.
In Example 2.4, we demonstrate how to use Analytic Solver and
R to create bins with equal counts, equal intervals, and user-defined
intervals.
FILE
Customers

EXAMPLE 2.4
In order to better understand her customers, Catherine Hill
would like to perform the RFM analysis, a popular marketing
technique used to identify high-value customers. RFM stands
for recency, frequency, and monetary. The RFM ratings can be
created from the DaysSinceLast (recency), NumOfOrders
(frequency), and Spending2018 (monetary) variables.
Following the 80/20 business rule (i.e., 80% of your
business comes from 20% of your best customers), for each of
the three RFM variables, Catherine would like to bin customers
into five equal-size groups, with 20% of the customers included
in each group. Each group is also assigned a score from 1 to 5,
with 5 being the highest. Customers with the RFM rating of 555
are considered the most valuable customers to the company.
In addition to the RFM binning, Catherine would like to bin
the Income variable into five equal intervals. Finally, she would
like to start a tiered membership status where different services
and rewards are offered to customers depending on how much
they spent in 2018. She would like to assign the bronze
membership status to customers who spent less than $250,
silver membership status to those who spent $250 or more but
less than $1,000, and the gold membership status to those who
spent $1,000 or more.
 page 59

Use Analytic Solver and R to bin variables according to

Catherine’s specifications. Summarize the results.
SOLUTION:
Using Analytic Solver

a. Open the Customers data file.

b. To create the recency score, we need to first transform the
variable DaysSinceLast to reverse the order of the data
because the fewer the number of days since the last
purchase, the greater is the recency score. Create a new
column with the column heading DaysSinceLastReverse in
column O. Enter the formula =L2*(−1) in cell O2. Verify that
the value of cell O2 is −101. Fill the range O3:O201 with the
formula in O2.
c. Choose Data Mining > Transform > Transform Continuous
Data > Bin. Select data range $A$1:$O$201. Check the box
Variable names in the first row.
d. Select DaysSinceLastReverse from the variable list. Change
#bins for variable to 5. Choose Equal count for the Bins to be
made with option. Choose Rank of the bin for the Value in the
binned variable is option. Specify both the start and interval
values to be 1. Click on the Apply to Selected Variable button.
Repeat this step for NumOfOrders and Spending2018. Click
on the Finish button.
Two new worksheets, Bin_Output and Bin_Transform, are
created. The Bin_Output worksheet shows the summary
information such as the number of observations in each bin.
The Bin_Transform worksheet displays the scores for each
RFM parameter for the renamed variables,
Binned_DaysSinceLastReverse, Binned_NumOfOrders, and
Binned_Spending2018. Verify that the values of the first
customer for Binned_DaysSinceLastReverse,
 Binned_NumOfOrders, and Binned_Spending2018 are 4, 1,
and 1, respectively.
e. To create the RFM score for each customer, we create a new
column with the column heading RFM in cell S12 in the
Bin_Transform worksheet. Enter =CONCATENATE(R13, N13,
M13) in cell S13. The CONCATENATE function merges
multiple text values into one cell. Recall that cells R13, N13,
and M13 represent the recency, frequency, and monetary
indices, respectively, for the first customer. This formula
creates a 3-digit RFM score for the first customer; verify that
the first customer has an RFM score of 411. Fill the range
S14:S212 with the formula in S13.
f. We now copy and paste the RFM values in the data range
S12:S212 in the Bin_Transform worksheet to the data range
P1:P201 in the Customers data file. To ensure that you paste
the values, and not the formula, use the Paste Special . . .
option and choose Values.
g. We now bin the Income variable into 5 groups with equal
intervals. With the Customers worksheet active, choose Data
Mining > Transform > Transform Continuous Data > Bin.
Select data range $A$1:$P$201. Check the box Variable
names in the first row. Select Income from the variable list.
Change #bins for variable to 5. Choose Equal interval for the
Bins to be made with option. Choose Rank of the bin for the
Value in the binned variable is option. Specify both the start
and interval values to be 1. Click Apply to Selected Variable
and Finish.
h. Two new worksheets, Bin_Output1 and Bin_Transform1, are
created. The Bin_Output1 worksheet shows summary
information such as the ranges of each bin. The
Bin_Transform1 worksheet shows the data set with the new
column, Binned_Income, displaying the customer’s income
category. Verify that the first customer is in the income
category 1. Copy the Binned_Income column in the data
range L12:L212 in the Bin_Transform1 worksheet and paste it
 in the data range Q1:Q201 in the Customers page 60
worksheet using the Paste Special . . . option and
choosing Values.
i. Make sure the Customers worksheet is active. To assign
customers to the tiered membership system that Catherine
designed, we have to first define tiered membership. In order
to describe the range of spending for each membership tier,
enter the values from Table 2.7 in cells S1:T4.
TABLE 2.7 Lookup Table

Column S Column T

Spending Membership

0 Bronze

250 Silver

1000 Gold

j. Add the column heading “Membership_Tier” in cell R1. We

use the VLOOKUP function to bin the spending values into
user-defined categories, or membership tiers in this example.
The VLOOKUP function has four parameters: (1) the lookup
value (a member’s spending in 2018 in column J); (2) a lookup
or reference table (ranges of spending and corresponding
membership tiers in cells S2:T4, without the column
headings); (3) column number in the lookup table that contains
the output (the output, Bronze, Silver, or Gold, is in the second
column in the lookup table); and (4) whether we want to look
up a value within a range (TRUE) or find an exact match
(FALSE).
 Enter =VLOOKUP(J2, $S$2:$T$4, 2, TRUE) in cell R2.
This function assigns the first customer to a membership tier
according to the spending in 2018. Note that the dollar signs in
$S$2:$T$4 ensure that when the formula is copied to another
cell in column R the references to cells S2:T4 remain
unchanged. Fill the range R3:R201 with the formula in R2 to
assign a membership tier to each customer. Verify that the
membership tier for the first customer is Bronze.
Using R

a. Import the Customers data into a data frame and label it

myData.
b. To create the recency score, we first transform the variable
DaysSinceLast to reverse the order of the data because the
fewer the number of days since the last purchase, the greater
is the recency score. Create a new variable called
DaysSinceLastReverse by multiplying DaysSinceLast by −1.
We use the as.numeric function to ensure that the
DaysSinceLastReverse variable is a numerical type. Enter:
> myData$DaysSinceLastReverse <-
as.numeric(myData$DaysSinceLast * -1)
c. We now create five equal-sized bins for
DaysSinceLastReverse (recency), NumOfOrders (frequency),
and Spending2018 (monetary). As the bins represent the
20%, 40%, 60%, 80%, and 100% quantiles, we first use the
quantile function to find the range for each bin and store and
then view the ranges in an object called recencyBins. Enter:
> recencyBins <- quantile(myData$DaysSinceLastReverse,
probs=seq(0, 1, by=0.20))
> recencyBins
Figure 2.9 shows the ranges for the five equal-sized bins.

FIGURE 2.9 Ranges for the recency bins

 page 61
We repeat similar commands for the number of
orders (frequencyBins) and the Spending2018 (monetaryBins)
variables. Enter:
> frequencyBins <- quantile(myData$NumOfOrders,
probs=seq(0, 1, by=0.20))
> monetaryBins <- quantile(myData$Spending2018,
probs=seq(0, 1, by=0.20))
Note that if we were to create 10 equal-sized bins, we would
change the probs option to probs=seq(0, 1, by=0.10).
d. We use the cut function to bin the data. The breaks option of
the cut function specifies the ranges of the bins created in
step c. The labels option assigns a label to each bin. The
right=FALSE option ensures that the intervals are closed on
the left and open on the right. The include.lowest=TRUE
option must be used along with the right=FALSE option so that
the largest value is included in the last bin. Enter:
> myData$Recency <- cut(myData$DaysSinceLastReverse,
breaks=recencyBins,labels=c(“1”, “2”, “3”, “4”, “5”),
include.lowest=TRUE, right=FALSE)
> myData$Frequency <- cut(myData$NumOfOrders,
breaks=frequencyBins, labels=c(“1”, “2”, “3”, “4”, “5”),
include.lowest=TRUE, right=FALSE)
> myData$Monetary <- cut(myData$Spending2018,
breaks=monetaryBins, labels=c(“1”, “2”, “3”, “4”, “5”),
include.lowest=TRUE, right=FALSE)
The prior commands assign numbers 1 to 5 to the equal-sized
bins for the three RFM variables. The RFM indices are stored
in three new variables, Recency, Frequency, and Monetary. If
we were to create and assign labels to 10 equal-sized bins,
 we would use labels=c(“1”, “2”, “3”, “4”, “5”, “6”, “7”,
“8”, “9”, “10”). Other labels can also be assigned to the bins
as appropriate.
e. To create the RFM score, we combine the three RFM indices
using the paste function, and then use the head function to
view the first few observations. Enter:
> myData$RFM <- paste(myData$Recency,
myData$Frequency, myData$Monetary)
> head(myData$RFM)
Verify that the first observation of the RFM variable is 411.
f. We now bin the Income variable into 5 groups with equal
intervals using the cut function. The breaks option specifies 5
bins with equal intervals. We assign numbers 1 (lowest) to 5
(highest) to the bins and then use the head function to view
the first few observations. Enter:
> myData$BinnedIncome <- cut(myData$Income, breaks=5,
labels= c(“1”, “2”, “3”, “4”, “5”), include.lowest=TRUE,
right=FALSE)
> head(myData$BinnedIncome)
Verify that the first observation of the BinnedIncome variable is
1. To create a different number of bins, change the breaks
value (e.g., breaks = 3 will create three bins).
g. We use the levels and cut functions to display the ranges of
the 5 intervals created in step f. Enter:
> levels(cut(myData$Income, breaks=5))
Verify that the first interval is (3.09e+04,5.82e+04] or from
$30,900 up to, and including, $58,200.
h. To find out the number of observations that belong page 62
to each bin, use the table function. Enter:
> table(myData$BinnedIncome)
Verify that the first bin, or the lowest income category, has 67
customers.
i. To create the membership tiers or user-defined bins proposed
by Catherine, we again use the cut function. Recall that
customers who spent less than $250 are assigned to the
Bronze membership, those who spent $250 or more but less
than $1,000 receive the Silver membership, and those who
spent $1,000 or more receive the Gold membership. We use
the breaks option to specify user-defined ranges. The Inf
keyword assigns any values above $1,000 to the last bin. We
then use the head and View functions to view the output in
various formats. Enter:
> myData$MembershipTier <- cut(myData$Spending2018,
breaks = c(0, 250, 1000, Inf),labels = c(“Bronze”, “Silver”,
“Gold”))
> head(myData$MembershipTier)
> View(myData)
Verify that the membership tier for the first customer is Bronze.
Summary
Table 2.8 shows a portion of the Customers data that now
includes variables that have been binned according to
Catherine’s specifications. The first customer has an RFM
score of 411, which suggests that this customer purchased
from Organic Food Superstore quite recently but made very
few purchases and spent very little money during the past year.
Not surprisingly, this customer also has the least desirable
bronze membership. In addition, in terms of income, the
majority of customers belong to the first three bins; only nine
customers have income in the two highest income groups.

TABLE 2.8 Customers Data with Binned Variables

 It is important to note that the binning algorithm is
implemented differently in R and Analytic Solver. In this
example, we obtain an identical result in both software, but in
other cases, R and Analytic Solver may yield different binning
results.

Students are encouraged to experiment with other binning

options with the Customers data. For example, experiment with
binning Income into two categories, High and Low, or redesign the
membership program with four or five tiers. There are many real-life
applications where numerical variables should be converted into
categories, and the number of categories should be determined
appropriately according to the context of the analysis. For example,
to convert students’ exam scores into categories, we might use two
bins (Pass and Fail) or five bins (A, B, C, D, and F grades).

Mathematical Transformations
As discussed earlier, data transformation is an important step in
bringing out the information in the data set, which can then be used
for further data analysis. In addition to binning, another common
approach is to create new variables through mathematical
transformations of existing variables. For example, to page 63
analyze diabetes risk, doctors and dieticians often focus
on body mass index (BMI), which is calculated as weight in
kilograms divided by squared height in meters, rather than focusing
on either weight or height alone. Similarly, in order to analyze trend,
we often transform raw data values into percentages.
Sometimes data on variables such as income, firm size, and
house prices are highly skewed; skewness is discussed in Section
3.4 of Chapter 3. For example, according to a Federal Reserve
report, the richest 1% of families in the U.S. controlled 38.6% of the
country’s wealth in 2016 (CNN, September 27, 2017). The extremely
high (or low) values of skewed variables significantly inflate (or
deflate) the average for the entire data set, making it difficult to
detect meaningful relationships with skewed variables. A popular
mathematical transformation that reduces skewness in data is the
natural logarithm transformation. Another transformation to reduce
data skewness is the square root transformation.
Another common data transformation involves calendar dates.
Statistical software usually stores date values as numbers. For
example, in R, date objects are stored as the number of days since
January 1, 1970, using negative numbers for earlier dates. For
example, January 31, 1970, has a value of 30, and December 15,
1969, has a value of −17. Excel implements a similar approach to
store date values but uses a reference value of 1 for January 1,
1900. Transformation of date values is often performed to help bring
useful information out of the data. A retail company might convert
customers’ birthdates into ages in order to examine the differences
in purchase behaviors across age groups. Similarly, by subtracting
the airplane ticket booking date from the actual travel date, an airline
carrier can identify last-minute travelers, who may behave very
differently from early planners.
Sometimes transforming date values into seasons helps enrich
the data set by creating relevant variables to support subsequent
analyses. For example, by extracting and focusing on the months in
which gym members first joined the health club, we may find that
members who joined during the summer months are more interested
in the aquatic exercise programs, whereas those who joined during
the winter months are more interested in the strength-training
programs. This insight can help fitness clubs adjust their marketing
strategies based on seasonality.
Example 2.5 demonstrates how to use Excel and R to perform
the following mathematical transformations: (1) compute the
percentage difference between two values, (2) perform a logarithm
transformation, and (3) extract information from date values.
FILE
Customers

EXAMPLE 2.5
After a closer review of her customers, Catherine Hill feels that
the difference and the percentage difference between a
customer’s 2017 and 2018 spending may be more useful to
understanding the customer’s spending patterns than the
yearly spending values. Therefore, Catherine wants to
generate two new variables that capture the year-to-year
difference and the percentage difference in spending. She also
notices that the income variable is highly skewed, with most
customers’ incomes falling between $40,000 and $100,000,
with only a few very-high-income earners. She has been
advised to transform the income variable into natural
logarithms, which will reduce the skewness of the data.
Catherine would also like to convert customer birthdates
into ages as of January 1, 2019, for exploring differences in
purchase behaviors of customers across age groups. Finally,
she would like to create a new variable that captures the birth
month of the customers so that seasonal products can be
marketed to these customers during their birth month.
Use Excel and R to transform variables according to
Catherine’s specifications.

page 64
 SOLUTION:
Using Excel

a. Open the Customer data file.

b. Create the column heading SpendingDiff in cell O1. Enter the
formula =J2 - I2 in cell O2. Verify that the resulting value in cell
O2 is −46. Fill the range O3:O201 with the formula in O2.
c. Create the column heading PctSpendingDiff in cell P1. Enter
the formula = (J2 – I2)/I2 in cell P2. In the drop-down menu of
the Home menu tab, change from General to Percentage (%).
Verify that the value in cell P2 is −16.03%. Fill the range
P3:P201 with the formula in P2.
d. Create the column heading IncomeLn in cell R1. The LN
function provides a natural logarithm transformation. Enter the
formula =LN(H2) in cell R2. Verify that the value in cell R2 is
10.8780. Fill the range R3:R201 with the formula in R2.
e. Create the column heading Age in cell S1. The YEARFRAC
function calculates the difference in years between two dates,
and the INT function displays only the integer portion of the
value and discards the decimal places. Enter the formula
=INT(YEARFRAC(D2, “01-01-2019”)) in cell S2. Verify that the
customer’s age as of January 1, 2019, is 32 years. Fill the
range S3:S201 with the formula in S2.
f. Create the column heading BirthMonth in cell T1. The MONTH
function extracts the month element from a date value. Enter
the formula =MONTH(D2) in cell T2. Verify that the value in
cell T2 is 12. Fill the range T3:T201 with the formula in T2.
Table 2.9 shows a portion of the data that includes the five
transformed variables. Other useful Excel functions related to
date values include DAY, YEAR, WEEKDAY, TODAY, and
NOW. The DAY and YEAR functions extract the date and the
year elements, respectively. The WEEKDAY function identifies
the weekday as an integer value from 1 to 7 (1 for Sunday and
7 for Saturday). Finally, =TODAY() and =NOW() return the
values of the current date and current time, respectively.
TABLE 2.9 Five Transformed Variables

Using R
a. Import the Customer data into a data frame (table) and label it myData.
b. We find the spending difference, and then use the head function to view the first
few observations. Enter:
> myData$SpendingDiff <- myData$Spending2018 - myData$Spending2017
> head(myData$SpendingDiff)
Verify that the first observation of the SpendingDiff variable is −46.
c. We create the percentage spending difference and round it to two page 65
decimal places using the round function. We then place the “%”
sign using the paste function and use the head function to view the first few
observations. Enter:
> myData$PctSpendingDiff <- round((myData$SpendingDiff /
myData$Spending2017)*100, digits = 2)
> myData$PctSpendingDiff <- paste(myData$PctSpendingDiff, “%”)
> head(myData$PctSpendingDiff)
Verify that the first observation of the PctSpendingDiff variable is −16.03%.
d. We use the log function for the natural logarithm transformation, and then use the
head function to view the first few observations. Enter:
> myData$IncomeLn <- log(myData$Income)
> head(myData$IncomeLn)
Verify that the first observation of the IncomeLn variable is 10.87805. The
IncomeLn values are slightly different from those in Table 2.9 because Table 2.9
is formatted to show only four decimal places. For the base 10 logarithm
transformation, use the log10 function in place of the log function.
e. To calculate a customer’s age as of January 1, 2019, we first need to convert the
Birthdate variable into the data values and create a new variable for the January
1, 2019, date. Enter:
> myData$BirthDate <- as.Date(myData$BirthDate, format = "%m/%d/%Y")
> endDate <- as.Date("01/01/2019", format = "%m/%d/%Y")
f. We use the difftime function to find out the number of days between the
customer’s birthdate and January 1, 2019. By dividing the difference in days by
365.25, we account for the leap years (by using 365.25 instead of 365) and obtain
 the difference in years. We use the as.numeric function to ensure that the Age
variable has a numerical type. Finally, we use the floor function to remove the
decimal places so that the age of a customer is an integer and the head function
to view the first few observations. Enter:
> myData$Age <- difftime(endDate, myData$BirthDate)/365.25
> myData$Age <- as.numeric(myData$Age)
> myData$Age <- floor(myData$Age)
> head(myData$Age)
Verify that the first customer’s age as of January 1, 2019, is 32 years.
g. We use the months function to extract the month name from the Birthdate
variable, the match function to convert month names (January to December) to
numbers (1 to 12), and the head function to view the first few observations. Enter:
> myData$BirthMonth <- months(myData$BirthDate)
> myData$BirthMonth <- match(myData$BirthMonth, month.name)
> head(myData$BirthMonth)
Verify that the first customer’s birthday is in month 12 (December).
h. We use the View function to display a spreadsheet-style data. The output should
be consistent with Table 2.9. Enter:
> View(myData)

Other useful date-related functions include weekdays and

format. The weekdays function returns the day of the week;
for example > weekdays(as.Date(“2000-12-25”)) returns
“Monday”. The format function returns the specified element
of a date value; for example, > format(as.Date(“2000-12-25”),
“%Y”) returns the year element “2000”. Besides the “%Y”
parameter, “%d” and “%m” specify the date and month
elements. The Sys.Date() and Sys.time() functions return the
current date and time values, respectively.

page 66

Another common transformation for numerical data is rescaling,

which is performed when the variables in a data set are measured
using different scales. For example, annual income measured in
dollars typically ranges from thousands to millions, whereas the
number of children typically contains values in low single digits. The
variability in measurement scales can place undue influence on
larger-scale variables, resulting in inaccurate outcomes. Therefore, it
is commonplace to rescale the data using either standardization or
normalization, especially in data mining techniques; a detailed
discussion on such transformations can be found in Chapter 8.

EXERCISES 2.4
Mechanics
35. FILE Exercise_2.35. The accompanying data contains three variables and six
observations.

x1 x2 x3

248 3.5 78

124 3.8 55

210 1.6 66

150 4.8 74

196 4.5 32

234 6.2 63

a. Bin the values of x1 into two equal-size groups. Label the groups with numbers
1 (lower values) and 2 (higher values). What is the average value of x1 for
group 1? (Hint: Sort the data by group number before calculating the average.)
b. Bin the values of x2 into three equal interval groups. Label the groups with
numbers 1 (lowest values) to 3 (highest values). How many observations are
assigned to group 1?
c. Bin the values of x3 into the following two groups: ≤ 50 and > 50. Label the
groups with numbers 1 (lower values) and 2 (higher values). How many
observations are assigned to group 2?
36. FILE Exercise_2.36. The accompanying data set contains three variables, x1,
x2, and x3.
a. Bin the values of x1 into three equal-size groups. Label the groups with
numbers 1 (lowest values) to 3 (highest values). How many observations are
assigned to group 1?
b. Bin the values of x2 into three equal-interval groups. Label the groups with
numbers 1 (lowest values) to 3 (highest values). How many observations are
assigned to group 2?
c. Bin the values of x3 into the following three groups: < 50,000, between 50,000
and 100,000, and > 100,000. Label the groups with numbers 1 (lowest values)
to 3 (highest values). How many observations are assigned to group 1?
37. FILE Exercise_2.37. The accompanying data set contains three variables, x1,
x2, and x3.
a. Bin the values of x1 into three equal-size groups. Label the groups with “low”
(lowest values), “medium,” and “high” (highest values). How many observations
are assigned to group medium?
b. Bin the values of x2 into three equal-interval groups. Label the groups with “low”
(lowest values), “medium,” and “high” (highest values). How many observations
are assigned to group high?
c. Bin the values of x3 into the following three groups: < 20, between 20 and 30,
and > 30. Label the groups with “low” (lowest values), “medium,” and “high”
(highest values). How many observations are assigned to group low?
38. FILE Exercise_2.38. The accompanying data set contains three variables, x1,
x2, and x3.
a. Bin the values of x1, x2, and x3 into five equal-size groups. Label the groups
with numbers 1 (lowest) to 5 (highest).
b. Combine the group labels of x1, x2, and x3 to create a score like the RFM score
described in Example 2.4. How many observations have the score “431”? How
many observations have the score “222”?
39. The following table contains two variables and five observations.

x1 x2

248 350

124 148

150 130
 x1 x2

196 145

240 180

a. Create a new variable called “Sum” that contains the sum of the values of x1
and x2 for each observation. What is the average value of Sum?
b. Create a new variable called “Difference” that contains the absolute difference
between the values of x1 and x2 for each observation. What is the average
value of Difference?

page 67
40. FILE Exercise_2.40. The accompanying data set contains three
variables, x1, x2, and x3.
a. Create a new variable called “Difference” that contains the difference between
the values of x1 and x2 for each observation (i.e., x2 − x1). What is the average
difference?
b. Create a new variable called “PercentDifference” that contains the percent
difference between the values of x1 and x2 for each observation (i.e., (x2 −
x1)/x1). What is the average percent difference?
c. Create a new variable called “Log” that contains the natural logarithms for x3.
What is the average logarithm value?
41. FILE Exercise_2.41. The accompanying data set contains three variables, x1,
x2, and x3.
a. Create a new variable called “Difference” that contains the difference between
the values of x1 and x2 for each observation (i.e., x2 − x1). What is the average
difference?
b. Create a new variable called “PercentDifference” that contains the percent
difference between the values of x1 and x2 for each observation (i.e., (x2 −
x1)/x1). What is the average percent difference?
c. Create a new variable called “Log” that contains the natural logarithms for x3.
Bin the logarithm values into five equal-interval groups. Label the groups using
numbers 1 (lowest values) to 5 (highest values). How many observations are in
group 2?
42. FILE Exercise_2.42. The accompanying data set contains two variables, Date1
and Date2.
a. Create a new variable called “DifferenceInYear” that contains the difference
between Date1 and Date2 in year for each observation. What is the average
 difference in year? (Hint: Use the YEARFRAC function if you are using Excel to
complete this problem.)
b. Create a new variable called “Month” that contains the month values extracted
from Date1. What is the average month value?
c. Bin the month values in four equal-interval groups. Label the groups using
numbers 1 (lowest values) to four (highest values). Which group has the highest
number of observations?

Application
43. FILE Population. The U.S. Census Bureau records the population for the 50
states each year. The accompanying table shows a portion of these data for the
years 2010 to 2018.

a. Bin the 2017 population values into four equal-size groups. Label the groups
using numbers 1 (lowest values) to 4 (highest values). How many states are
assigned to group 4?
b. Bin the 2018 population values into four equal-interval groups. Label the groups
using numbers 1 (lowest values) to 4 (highest values). How many states are
assigned to group 2? Compare the groups in parts a and b. Which states are in
higher groups for the 2018 population than for the 2017 population?
c. Bin the 2018 population values into the following three groups: < 1,000,000,
between 1,000,000 and 5,000,000, and > 5,000,000. Label the groups using
numbers 1 (lowest values) to 3 (highest values). How many observations are
assigned to group 2?
44. FILE Population. Refer to the previous exercise for a description of the data set.
a. Create a new variable called “Difference” that contains the difference between
the 2018 population and the 2017 population for each state (i.e., 2018
population - 2017 population). What is the average difference?
b. Create a new variable called “PercentDifference” that contains the percent
difference between the 2017 and 2018 population values for the states (i.e.,
(2018 population - 2017 population)/2017 population). What is the average
percent difference?
c. Create a new variable called “Log” that contains the natural logarithms for the
2018 population values for the states. Bin the logarithm values into five equal-
 interval groups. Label the groups using numbers 1 (lowest values) to 5 (highest
values). How many observations are in group 2?
d. Create a new variable called “SquareRoot” that contains the square root of the
2018 population values for the states. Bin the square root values into five equal-
interval groups. Label the groups using numbers 1 (lowest values) to 5 (highest
values). How many observations are in group 2?
e. Compare the groups in parts c and d. Are the groupings the same or different?
45. FILE Credit_Cards. Greg Metcalf works for a national credit card company, and
he is performing a customer value analysis on a subset of credit card customers.
In order to perform the RFM analysis on the customers, Greg has compiled a data
set that contains the dates of the last transaction (LastTransactionDate), total
number of transactions in the past two years (Frequency), and total spending
during the past two years (Spending). A portion of the data set is shown in the
accompanying table.

LastTransactionDate Frequency Spending

5/20/2017 407 41903

8/16/2018 454 35918

⋮ ⋮ ⋮

8/14/2017 49 27918

page 68
a. Greg wants to calculate the number of days between January 1,
2019, and the last transaction date. Create a new variable “DaysSinceLast” that
contains the number of days since the last transaction. (Hint: Use the DATEDIF
function if you are using Excel to complete this problem.) What is the average
number of days since the last purchase for all the customers?
b. Create the RFM scores for each customer. How many customers have an RFM
score of 555? What is their average spending?
c. Create a new variable called “LogSpending” that contains the natural logarithms
for the total spending during the past two years. Bin the logarithm values into
five equal-interval groups. Label the groups using numbers 1 (lowest values) to
5 (highest values). How many observations are in group 2?
d. Create a new variable called “AverageOrderSize” that contains the average
spending per order. This is calculated by dividing total spending (Spending) by
total number of transactions (Frequency) in the past two years. Bin the values
of AverageOrderSize into five equal-interval groups. Label the groups using
numbers 1 (lowest values) to 5 (highest values). How many observations are in
group 2?
e. Compare the groups in parts c and d. Are the groupings the same or different?
46. FILE Game_Players. TurboX is a online video game company that makes three
types of video games: action, role play, and sports. It is interested in
understanding its millennial customers. By combining the data from its customer
database and a customer survey, TurboX compiled a data set that has the
following variables: the player’s satisfaction with the online game purchase
experience (Satisfaction), the enjoyment level of the game played (Enjoyment),
whether the player will recommend the game to others (Recommend), which type
of game the player played (Type), total spending on games last year
(SpendingLastYear), total spending on games this year (SpendingThisYear), and
the date of birth of the player (BirthDate).

a. Bin the total spending on games last year into four equal-size groups. Label the
groups using numbers 1 (lowest values) to 4 (highest values). How many
customers are assigned to group 4?
b. Bin the total spending on games this year into four equal-interval groups. Label
the groups using numbers 1 (lowest values) to 4 (highest values). How many
customers are assigned to group 3?
c. Bin the total spending on games this year into the following three groups: < 250,
between 250 and 500, and > 500. Label the groups using numbers 1 (lowest
values) to 3 (highest values). How many observations are assigned to group 2?
d. Create a new variable called “Difference” that contains the difference between
this year’s and last year’s spending on games for the players (i.e.,
SpendingThisYear - SpendingLastYear). What is the average difference?
e. Create a new variable called “PercentDifference” that contains the percent
difference between this year’s and last year’s spending on games for the
players (i.e., (SpendingThisYear - SpendingLastYear)/SpendingLastYear). What
is the average percent difference?
f. Create a new variable “Age” that contains the players’ ages as of January 1,
2019. What is the average age of the players?
g. Create a new variable “BirthMonth” that contains the players’ birth month
extracted from their dates of birth. Which month is the most frequent birth
month?
47. FILE Engineers. Erin Thomas, an HR manager of an engineering firm, wants to
perform an analysis on the data about the company’s engineers. The variables
included in the data are date of birth (BirthDate), personality type according to the
Myers-Briggs Personality assessment (Personality), annual salary (Salary), level
of the position (Level), and number of professional certificates achieved
(Certificates). The accompanying table shows a portion of the data set.
 a. Create a new variable “Age” that contains the engineers’ ages as of January 1,
2019. What is the average age of the engineers?
b. Bin the age values into three equal-size groups. Label the groups using
numbers 1 (lowest age values) to 3 (highest age values). How many
observations are in group 3?
c. Bin the annual salary values into four equal interval groups. Label the groups
using numbers 1 (lowest salary values) to 4 (highest salary values). How many
engineers are assigned to group 4?
d. Bin the number of professional certificates achieved into the following three
groups: < 2, between 2 and 4, and over 4. Label the groups “Low,” “Medium,”
and “High.” How many engineers are in the “High” group?

page 69
48. FILE Patients. Jerry Stevenson is the manager of a medical
clinic in Scottsdale, AZ. He wants to analyze patient data to identify high-risk
patients for cardiovascular diseases. From medical literature, he learned that the
risk of cardiovascular diseases is influenced by a patient’s age, body mass index
(BMI), amount of exercise, race, and education level. Jerry has compiled a data
set with the following variables for his clinic’s patients: race (Race), education
level (Education), body weight in kilograms (Weight), height in meters (Height),
date of birth (BirthDate), and number of minutes of exercise per week (Exercise).
The accompanying table shows a portion of the data set.

a. Create a new variable called “BMI” that contains the body mass index of the
patients. BMI is calculated as weight in kilograms/(height in meters)2. What is
the average BMI of the patients?
b. Create a new variable “Age” that contains the patients’ ages as of January 1,
2019. What is the average age of the patients?
 c. Bin the patients’ ages into five equal-size groups. Label the groups using
numbers 1 (youngest) to 5 (oldest). How many patients are in group 4?
d. Bin the patients’ total minutes of exercise per week into five equal-size groups.
Label the groups using number 1 (highest values) to 5 (lowest values). How
many patients are in group 5?
e. Bin the patients’ BMI into five equal-size groups. Label the groups using
numbers 1 (lowest values) to 5 (highest values). How many patients are in
group 1?
f. Create a risk score for each patient by concatenating the group numbers
obtained in parts c, d, and e. How many patients are in the risk group of 555?

2.5 TRANSFORMING CATEGORICAL

DATA
LO 2.5
Transform categorical variables.

As discussed in Chapter 1, we use labels or names to identify the

distinguishing characteristics of a categorical variable. For instance,
a firm may identify each customer as either a male or a female.
Here, the sex of a customer is a categorical variable representing
two categories. Categorical variables can also be defined by more
than two categories. Examples include marital status (single,
married, widowed, divorced, separated) and the performance of a
manager (excellent, good, fair, poor). Recall that we use nominal and
ordinal measurement scales to represent categorical variables. In
the above examples, the measurement scales for marital status and
performance of a manager are nominal and ordinal, respectively.
While categorical variables are known to represent less
sophisticated levels of measurement, they are often the most
important variables in the analysis. For example, the sex of a
customer may contain the most useful information on the customer’s
spending behavior. Categorical data do, however, present unique
challenges in data wrangling. As many analysis techniques are
limited in their abilities to handle categorical data directly, steps to
simplify or transform categorical data into numerical formats are
often performed prior to analysis. In this section, we discuss three
common approaches for transforming categorical data: category
reduction, dummy variables, and category scores.

Category Reduction
Sometimes nominal or ordinal variables come with too many
categories. This presents a number of potential problems. First,
variables with too many categories pull down model performance
because, unlike a single parameter of a numerical variable, several
parameters associated with the categories of a categorical variable
must be analyzed. Second, if a variable has some categories that
rarely occur, it is difficult to capture the impact of these categories
accurately. In addition, a relatively small sample may not contain any
observations in certain categories, creating errors when the
analytical model is later applied to a larger data set with observations
in all categories. Third, if one category clearly dominates page 70
in terms of occurrence, the categorical variable will fail to
make a positive impact because modeling success is dependent on
being able to differentiate among the observations.
An effective strategy for dealing with these issues is category
reduction, where we collapse some of the categories to create fewer
nonoverlapping categories. Determining the appropriate number of
categories often depends on the data, context, and disciplinary
norms, but there are a few general guidelines.
The first guideline states that categories with very few
observations may be combined to create the “Other” category. For
example, in a data set that contains the demographic data about
potential customers, if many zip code categories only have a few
observations, it is recommended that an “Other” category be created
for these observations. The rationale behind this approach is that a
critical mass can be created for this “Other” category to help reveal
patterns and relationships in data.
 Another guideline states that categories with a similar impact may
be combined. For example, when studying public transportation
ridership patterns, one tends to find that the ridership levels remain
relatively stable during the weekdays and then change drastically for
the weekends. Therefore, we may combine data from Monday
through Friday into the “Weekdays” category and Saturday and
Sunday into the “Weekends” category to simplify data from seven to
only two categories.
Example 2.6 demonstrates how to use Excel with Analytic Solver
and R for category reduction.
FILE
Customers

EXAMPLE 2.6
After gaining some insights from the Customers data set,
Catherine would like to analyze race. However, in its current
form, the data set would limit her ability to do a meaningful
analysis given the large number of categories of the race
variable; plus some categories have very few observations. As
a result, she needs to perform a series of data transformations
to prepare the data for subsequent analysis. Use Excel with
Analytic Solver and R to create a new category called Other
that represents the two least-frequent categories.
SOLUTION:
Using Excel with Analytic Solver

a. Open the Customers data file.

b. Choose Data Mining > Transform > Transform Categorical
Data > Reduce Categories. See Figure 2.10. In the Reduce
Categories dialog box, select data range $A$1:$N$201. Check
the box First row contains headers. Select Race from the
Category variable drop-down box. Once the variable is
selected, Analytic Solver presents the frequency for each race
category from the most frequent to the least frequent.

FIGURE 2.10 Analytic Solver’s dialog box for Reduce

Categories

In order to combine the two least-frequent categories, we

select the By frequency option and set Limit number of
categories to option to 5. Analytic Solver will automatically
combine the 5th and 6th least-frequent categories into one
category. Click Apply. Click OK. A new worksheet,
Category_Reduction, is created with a new column,
Reduced_Race. Numeric coding is used to represent the
categories, with 1 representing the most frequent category
and 5 representing the least frequent category. If you compare
the list to the original data, you will see that both American
Indian and Pacific Islander have been combined into the new
category 5.
c. Copy the Reduced_Race column in cells F28:F228 and paste
the values (use the Paste Special . . . option and choose
Values) to cells O1:O201 in the Customers worksheet.
d. To retain the category names instead of recoding page 71
them into numbers, we use the IF function. An IF
statement has the following structure: =IF(a logical test
resulting in TRUE or FALSE, result if the logical test is TRUE,
result if the logical test is FALSE). Because there are five
categories in the New_Race variable, we need to use multiple
IF statements, or nested IFs, to change the numbers into
category names. Figure 2.11 provides an illustration of how
the IF function works in this example. In the Customers
worksheet, enter the column heading NewRace in cell P1. In
Cell P2, enter the formula =IF(O2=1, “White”, IF(O2=2,
“Black”, IF(O2=3, “Hispanic”, IF(O2=4, “Asian”, “Other”)))) in
cell P2. Fill the range P3:P201 with the formula in P2. Verify
that the 19th observation represents the first customer in the
Other category.

FIGURE 2.11 Illustration of Excel’s IF function

Using R

a. Import the Customers data into a data frame (table) and label
it myData.
b. First, we inspect the frequency of each Race category to
identify the two least-frequent categories. Enter:
> table(myData$Race)
 The table shows that American Indians and Pacific Islanders
are the two least-frequent categories with only five and three
observations, respectively.
c. We use the ifelse function to create a new variable page 72
called NewRace that uses the Other category to
represent American Indians and Pacific Islanders. Enter:
> myData$NewRace <- ifelse(myData$Race %in%
c(“American Indian”, “Pacific Islander”), “Other”,
myData$Race)
Note that the ifelse function evaluates the values in the Race
variable, and if the value is either American Indian or Pacific
Islander, it replaces it with Other; the original race value is
retained otherwise.
d. We use the table function again to verify that the Other
category has eight observations. Enter:
> table(myData$NewRace)
e. We use the View function to display spreadsheet-style data.
Enter:
> View(myData)
Verify that the 19th customer is the first in the Other category.

Dummy Variables
In many analytical models, such as regression models discussed in
later chapters, categorical variables must first be converted into
numerical variables. For other models, dealing with numerical data is
often easier than categorical data because it avoids the complexities
of the semantics pertaining to each category of the variable. A
dummy variable, also referred to as an indicator or a binary
variable, is commonly used to describe two categories of a variable.
It assumes a value of 1 for one of the categories and 0 for the other
category, referred to as the reference or the benchmark category.
For example, we can define a dummy variable to categorize a
person’s sex using 1 for male and 0 for female, using females as the
reference category. Dummy variables do not suggest any ranking of
the categories and, therefore, without any loss of generality, we can
define 1 for female and 0 for male, using males as the reference
category. All interpretation of the results is made in relation to the
reference category.

A DUMMY VARIABLE
A dummy variable, also referred to as an indicator or a binary
variable, takes on values of 1 or 0 to describe two categories of
a categorical variable.

Oftentimes, a categorical variable is defined by more than two

categories. For example, the mode of transportation used to
commute may be described by three categories: Public
Transportation, Driving Alone, and Car Pooling. Given k categories
of a variable, the general rule is to create k − 1 dummy variables,
using the last category as reference. For the mode-of-transportation
example, we need to define only two dummy variables. Suppose we
define two dummy variables d1 and d2, where d1 equals 1 for Public
Transportation, 0 otherwise, and d2 equals 1 for Driving Alone, 0
otherwise. Here, Car Pooling, the reference category, is indicated
when d1 = d2 = 0. Therefore, adding the third dummy variable for Car
Pooling would create information redundancy; certain analytical
models cannot even be estimated with k dummy variables.
Example 2.7 shows how to create dummy variables using
Analytic Solver and R.
FILE
Customers

page 73
 EXAMPLE 2.7
For the new Asian-inspired meal kits, Catherine feels that
understanding the channels through which customers were
acquired is important to predict customers’ future behaviors. In
order to include the Channel variable in her predictive model,
Catherine needs to convert the Channel categories into dummy
variables. Because web banner ads are probably the most
common marketing tools used by Organic Food Superstore,
she plans to use the Web channel as the reference category
and assess the effects of other channels in relation to the Web
channel. Use Excel with Analytic Solver and R to create the
relevant dummy variables for the Channel variable.
SOLUTION:
Using Excel with Analytic Solver

a. Open the Customers data file.

b. Choose Data Mining > Transform > Transform Categorical
Data > Create Dummies. Select data range $A$1:$N$201
and check the box First row contains headers. Select Channel
from the Variables list. Click on the arrow to send it to the
Variables to be factored box. Click OK. A new worksheet,
Encoding, is created. Four dummy variables,
Channel_Referral, Channel_SM, Channel_TV, and
Channel_Web, have been created.
c. Because we use Web channel as the reference category, we
will copy the dummy variables except Channel_Web to the
Customers worksheet. Copy the dummy variables from cells
Q24:S224 in the Encoding worksheet and paste the values
(use the Paste Special . . . option and choose Values) to cells
O1:Q201 in the Customers worksheet. Verify that the dummy
variable values for the first observation are 0, 1, and 0,
respectively.
Using R
a. Import the Customers data into a data frame (table) and label
it myData.
b. We use the ifelse function to create a dummy variable for
individual categories in the Channel variable. The ifelse
function evaluates the categories in the Channel variable, and,
for example, if the category is Referral, then the function
assigns a 1 to the new Channel_Referral variable, and 0
otherwise.Two other dummy variables, Channel_SM and
Channel_TV, are created similarly. Note that we leave out the
last channel, Web, as it is a reference category. Enter:
> myData$Channel_Referral <- ifelse(myData$Channel ==
“Referral”, 1, 0)
> myData$Channel_SM <- ifelse(myData$Channel == “SM”,
1, 0)
> myData$Channel_TV <- ifelse(myData$Channel == “TV”, 1,
0)
c. We use the View function to display spreadsheet-style data.
Enter:
> View(myData)
Verify that the dummy variable values for the first observation
are 0, 1, and 0, respectively.

There are packages in R, such as dummy, dummies, and

fastDummies, that help create dummy variables. Students are
encouraged to explore and experiment with these other options.

page 74

Category Scores
Finally, another common transformation of categorical variables is to
create category scores. This approach is most appropriate if the data
are ordinal and have natural, ordered categories. For example, in
customer satisfaction surveys, we often use ordinal scales, such as
very dissatisfied, somewhat dissatisfied, neutral, somewhat satisfied,
and very satisfied, to indicate the level of satisfaction. While the
satisfaction variable is categorical, the categories are ordered. In
such cases, we can recode the categories numerically using
numbers 1 through 5, with 1 being very dissatisfied and 5 being very
satisfied. This transformation allows the categorical variable to be
treated as a numerical variable in certain analytical models. With this
transformation, we need not convert a categorical variable into
several dummy variables or to reduce its categories. For an effective
transformation, however, we assume equal increments between the
category scores, which may not be appropriate in certain situations.
Example 2.8 shows how to convert a categorical variable into
category scores using Excel and R.
FILE
Customers

EXAMPLE 2.8
For the new Asian-inspired meal kits, Catherine wants to pay
attention to customer satisfaction. As the customer satisfaction
ratings represent ordinal data, she wants to convert them to
category scores ranging from 1 (Very Dissatisfied) to 5 (Very
Satisfied) to make the variable more readily usable in predictive
models. Use Excel and R to create category scores for the
Satisfaction variable.
SOLUTION:
Using Excel

a. Open the Customers data file.

b. Enter the column heading Satisfaction_Score in cell O1. Enter
the formula =IF(M2=“Very Satisfied”, 5, IF(M2=“Somewhat
Satisfied”, 4, IF(M2=“Neutral”, 3, IF(M2=“Somewhat
 Dissatisfied”, 2, 1)))) in cell O2. Fill the range O3:O201 with
the formula in O2. The scores are now ordered based on the
degree to which the customer is satisfied with Organic Food
Superstore’s service. See Example 2.6 for further information
on the nested IF statement. Verify that the first four
satisfaction scores are 1, 3, 5, and 1, respectively.
Analytic Solver provides an option to create category
scores (Data Mining > Transform > Transform Categorical
Data > Create Category Scores). However, it is not
appropriate for our application because it assigns numerical
scores to the categories based on their order of appearance.
For example, a satisfaction score of Very Dissatisfied is
assigned number 1 as it is the first to appear in the data. If
Neutral was the first observation, then that would have been
assigned number 1.
Using R

a. Import the Customers data into a data frame (table) and label
it myData.
b. We use the ifelse function, in a nested format, to create
category scores for the Satisfaction variable. Enter:
> myData$Satisfaction_Score <- ifelse(myData$Satisfaction
== “Very Dissatisfied”, 1, ifelse(myData$Satisfaction ==
“Somewhat Dissatisfied”, 2, ifelse(myData$Satisfaction ==
“Neutral”, 3, ifelse(myData$Satisfaction == “Somewhat
Satisfied”, 4, 5))))
Note that the ifelse function evaluates the values in the
Satisfaction variable, and if the value is Very Dissatisfied, the
function assigns a 1 to the new Satisfaction_Score page 75
variable. Because it is a nested format, if the value
is not Very Dissatisfied but is Somewhat Dissatisfied, the
function assigns a 2, and so on. If the values in the
Satisfaction variable are none of the first four scores, the
function assigns 5 to the Satisfaction_Score variable.
c. We use the View function to display spreadsheet-style data.
Enter:
> View(myData)
Verify that the first four satisfaction scores are 1, 3, 5, and 1,
respectively.

EXERCISES 2.5
Mechanics
49. The following table has three variables and six observations.

Sex Income Decision

Female 95000 Approve

Male 65000 Approve

Female 55000 Need More Information

Male 72000 Reject

Male 58000 Approve

Male 102000 Approve

a. Convert Sex into dummy variables. Use the most frequent category as the
reference category. Which category is the reference category?
b. Transform Decision into dummy variables. Use the most frequent category as
the reference category. Which category is the reference category?
c. Transform the Decision values into category scores where Approve = 1, Reject
= 2, and Need More Information = 3. How many observations have a category
score of 2?
50. FILE Exercise_2.50. The accompanying data set contains three variables, x1,
x2, and x3.
a. The variable x1 contains 6 categories ranging from “A” to “F.” Reduce the
number of categories to five by combining the two least-frequent categories.
 Name the new category “Other.” How many observations are in the “Other”
category?
b. The variable x2 contains six categories ranging from “A” to “F.” This variable is
ordinal, meaning that the categories are ordered. “A” represents the lowest
level, whereas “F” represents the highest level. Replace the category names
with category scores ranging from 1 (lowest) to 6 (highest). What is the average
category score for x2?
c. The variable x3 contains four unordered categories. To facilitate subsequent
analyses, we need to convert x3 into dummy variables. How many dummy
variables should be created? Create the dummy variables using Category1 as
the reference category.
51. FILE Exercise_2.51. The accompanying data set contains two variables,
Birthdate and LoanDecision.
a. LoanDecision contains three unordered categories. To facilitate subsequent
analyses, we need to convert LoanDecision into dummy variables. How many
dummy variables should be created? Create the dummy variables using “Need
more information” as the reference category.
b. Create a new variable based on LoanDecision. The new variable should have
only two categories: “Approve” and “Not approve.” The “Not approve” category
combines the “Reject” and “Need more information” categories. How many
observations are in the “Not approve” category?
52. FILE Exercise_2.52. The accompanying data set contains two variables, x1 and
x2.
a. The variable x1 contains three categories ranging from “Low” to “High.” Convert
the category names into category scores (i.e., 1 = “Low”, 2 = “Medium”, and 3 =
“High”). How many observations have a category score of 3?
b. Reduce the number of categories in x2 by combining the three least-frequent
categories. Name the new category “Other”. How many observations are in the
“Other” category?
c. Convert the new x2 into dummy variables. How many dummy variables should
be created? Create the dummy variables using the “Other” category as the
reference category.
53. FILE Exercise_2.53. The accompanying data set contains three variables, x1,
x2, and x3.
a. The variable x1 contains 3 categories: S, M, and L. Convert the category names
into category scores (i.e., S = 1, M = 2, and L = 3). How many observations
have a category score of 3?
b. Transform x2 into an appropriate number of dummy variables. How many
dummy variables should be created?
c. The variable x3 contains four categories: A, B, C, and D. Reduce the number of
categories by combining the two least-frequent categories into a new category
 E. How many observations are in the E category?

page 76

Applications
54. FILE Home_Loan. Consider the following portion of data that includes
information about home loan applications. Variables on each application include
the application number (Application), whether the application is conventional or
subsidized by the federal housing administration (LoanType), whether the
property is a single-family or multifamily home (PropertyType), and whether the
application is for a first-time purchase or refinancing (Purpose).

Application LoanType PropertyType Purpose

1 Conventional Single-family Purchase

2 Conventional Multi-family Refinancing

⋮ ⋮ ⋮ ⋮

103 Federal Housing Single-family Purchase

a. Are the variables LoanType, PropertyType, and Purpose nominal or ordinal

data? Why?
b. Which categories are the most frequent categories for LoanType, PropertyType,
and Purpose?
c. To facilitate subsequent analyses, transform LoanType, PropertyType, and
Purpose into dummy variables. Use the most frequent categories of the
variables as the reference categories. Which categories of LoanType,
PropertyType, and Purpose are the reference categories? How many dummy
variables are created in total?
55. FILE Shipment. A manager of a local package delivery store believes that too
many of the packages were damaged or lost. She extracts a sample of 75
packages with the following variables: the package number (Package), the status
of the package (Status: Delivered, Damaged, or Lost), the delivery type (Delivery:
Standard, Express, or Same day), and the size of the package (Size: S, M, L, and
XL). A portion of the data is shown in the accompanying table.

Package Status Delivery Size

1 Delivered Standard XL
 Package Status Delivery Size

2 Damaged Express S

⋮ ⋮ ⋮ ⋮

75 Lost Same day M

a. Transform the Delivery variable into dummy variables. Use the most frequent
category as the reference category. How many dummy variables should be
created? Which category of Delivery is the reference category?
b. Combine the two least frequent categories in the Size variable into a new
category called Other. How many observations are there in the new category?
c. Replace the category names in the Status variable with scores 1 (Lost), 2
(Damaged), or 3 (Delivered). What is the average status score of the 75
packages?
56. FILE Technician. After each thunderstorm, a technician is assigned to do a
check on cellular towers in his or her service area. For each visit, the technician
records in a database the tower number (Tower), the unit model (Model: A or B),
and the extent of the damage to the unit (Damage: None, Minor, Partial, Severe,
and Total for a total loss). A portion of the data is shown in the accompanying
table.

Tower Model Damage

1 A Minor

2 A None

⋮ ⋮ ⋮

98 B Partial

a. Transform the Model variable into dummy variables. How many dummy
variables should be created?
b. Transform the Damage variable into categorical scores ranging from 4 (Total) to
3 (Severe), 2 (Partial), 1 (Minor), and 0 (None). What is the average damage
score of the cell towers?
57. FILE Game_Players. Refer to Exercise 2.46 for a description of the problem and
data set.
a. The variable Satisfaction contains five ordered categories: Very Dissatisfied,
Dissatisfied, Neutral, Satisfied, and Very Satisfied. Replace the category names
 with scores ranging from 1 (Very Dissatisfied) to 5 (Very Satisfied). What is the
average satisfaction score of the players?
b. The variable Enjoyment contains five ordered categories: Very Low, Low,
Neutral, High, and Very High. Replace the category names with scores ranging
from 1 (Very Low) to 5 (Very High). What is the average enjoyment score of the
players?
c. The variable Recommend contains five ordered categories: Definitely Will Not,
Will Not, Neutral, Will, and Definitely Will. Replace the category names with
scores ranging from 1 (Definitely Will Not) to 5 (Definitely Will). What is the
average recommendation score of the players?
d. The variable Type contains three unordered game categories: Action, Role
Play, and Sports. To facilitate subsequent analyses, transform Type into dummy
variables. Use the least frequent category as the reference category. Which
category is the reference category? How many dummy variables should be
created?
58. FILE Engineers. Refer to Exercise 2.47 for a description of the problem and
data set.
a. The variable Personality contains four unordered personality types: Analyst,
Diplomat, Explorer, and Sentinel. To facilitate subsequent analyses, Erin needs
to convert this variable into dummy variables. How many dummy variables
should be created? Create the dummy variables using Analyst
type as the reference category. How many observations are in
page 77
the reference category?
b. The variable Level contains three ordered position levels: Engineer I (lowest),
Engineer II, and Senior Engineer (highest). Replace the level names with
scores ranging from 1 (lowest) to 3 (highest). What is the average level score of
the engineers?
59. FILE Patients. Refer to Exercise 2.48 for a description of the problem and data
set.
a. The variable Race contains five unordered categories: American Indian,
Asian/Pacific Islander, Hispanic, Non-Hispanic Black, and Non-Hispanic White.
Reduce the number of categories to four by combining the two least frequent
categories. Name the new category “Other.” How many observations are in the
“Other” category?
b. Transform the Race variable with the new categories into dummy variables. Use
the most frequent race category in the data as the reference category. Which
category is the reference category? How many dummy variables should be
created?
c. The variable Education contains four ordered categories: HS (lowest
educational attainment level), Some College, College, and Graduate (highest
educational attainment level). Replace the category names with category
scores ranging from 1 (lowest) to 4 (highest). What is the average category
score for Education?
 2.6 WRITING WITH BIG DATA
FILE
TechSales_Reps

Case Study
Cassius Weatherby is a human resources manager at a major technology firm that
produces software and hardware products. He would like to analyze the net promoter
score (NPS) of sales professionals at the company. The NPS measures customer
satisfaction and loyalty by asking customers how likely they are to recommend the
company to others on a scale of 0 (unlikely) to 10 (very likely). This measure is an
especially important indicator for the company’s software business as a large
percentage of the sales leads come from customer referrals. Cassius wants to
identify relevant factors that are linked with the NPS that a sales professional
receives. These insights can help the company make better hiring decisions and
develop a more effective training program.
With the help of the company’s IT group, a data set with over 20,000 records of
sales professionals is extracted from the enterprise data warehouse. The relevant
variables include the product line to which the sales professional is assigned, age,
sex, the number of years with the company, whether the sales professional has a
college degree, personality type based on the Myers-Briggs Personality assessment,
the number of professional certificates acquired, the average score from the 360-
degree annual evaluation, base salary, and the average NPS received. Cassius is
tasked with inspecting and reviewing the data and preparing a report for the
company’s top management team.

Sample Report—Evaluation of Net Promoter

Scores
The net promoter score (NPS) is a key indicator of customer satisfaction and
loyalty. It measures how likely a customer would recommend a product or
company to others. Because our software line for business relies heavily on
customer referrals to generate sales leads, the NPS that our sales
professionals receive is a key indicator of our company’s future success.
dizain/Shutterstock

page 78
From a total of about 20,000 records of sales professionals, we select only
the sales professionals in the software product group and divide them into
two categories: those with an average NPS below nine and those with an
average NPS of nine or ten. When a customer gives a sales professional an
NPS of nine or ten, the customer is considered “enthusiastically loyal,”
meaning that they are very likely to continue purchasing from us and refer
their colleagues to our company. Based on the NPS categorization, we then
divide the sales professionals into two categories: those with zero to three
professional certificates and those with four or more professional certificates.
Table 2.10 shows the results. Of the 12,130 sales professionals in the
software product group, we find that 65.57% have earned less than 4
professional certificates, whereas 34.43% have earned four or more.
However, there appears to be a link between those with 4 or more
professional certificates and NPS values. For those who received an NPS of
nine or ten, we find that 62.60% have earned at least four professional
certificates. Similarly, for those who received an NPS of below nine, we find
that 73.00% earned less than four professional certificates.

TABLE 2.10 Sales Professionals by the Number of Certificates and

NPS Value

Number of Full Sample (n = NPS < 9 (n = NPS ≥ 9 (n =

certificates 12,130) 9,598) 2,532)
 Number of Full Sample (n = NPS < 9 (n = NPS ≥ 9 (n =
certificates 12,130) 9,598) 2,532)

0 to 3 65.57% 73.00% 37.40%

4 or more 34.43% 27.00% 62.60%

Although this might simply suggest that high-achieving employees tend to

be self-motivated to earn professional certificates, we also believe that sales
professionals with sufficient technical knowledge can effectively communicate
and assist their customers in finding technology solutions, which will lead to
increased customer satisfaction and loyalty. Our training and development
program must place a greater emphasis on helping the employees earn
relevant certifications and acquire necessary technical knowledge.
Based on NPS categorization, we then divide the sales professionals into
categories based on personality type. Table 2.11 shows the results. In
addition to professional certification, we find that personality types are linked
with NPS values. Among the four personality types, Diplomats and Explorers
account for 72.69% of all the sales professionals in the software group.
However, when we divide the employees based on the NPS values, these
two personality types account for 91.63% of the group with an average NPS
of nine or ten, whereas they account for only 67.69% for the below nine NPS
group.

TABLE 2.11 Sales Professionals by Personality Type and NPS Value

Myers-Briggs Full Sample (n = NPS < 9 (n = NPS ≥ 9 (n =

Personality Type 12,130) 9,598) 2,532)

Analyst 12.13% 14.47% 3.24%

Diplomat 35.62% 33.07% 45.30%

Explorer 37.07% 34.62% 46.33%

Sentinel 15.19% 17.84% 5.13%

We also examined NPS variations by other variables such as age, sex,

education attainment, sales, and commission but did not find considerable
differences in NPS categorization. Other variables such as salary and the
tenure of the employee with the company are not included in our initial
analysis.
Based on the insights from this analysis, we request that the company
appoint an analytics task force to conduct a more comprehensive analysis of
 sales professionals. We strongly suggest that the analysis focus on
professional certification and personality, among relevant factors for
determining the NPS value. At a minimum, two goals of the task force should
include making recommendations on (1) a redesign of our training and
development program to focus on helping employees acquire relevant
professional certificates and (2) the efficacy of using personality types as part
of the hiring decision.

page 79

Suggested Case Studies

Data wrangling is a crucial step in any data analytics project. The data inspection,
preparation, and transformation techniques discussed in this chapter can be applied
to many data sets. Here are some suggestions using the big data that accompanies
this text.

Report 2.1 FILE CA_Crash. Subset the data set based on the location, day of the
week, type of collision, and lighting condition. Compare these subsets of data to find
interesting patterns. Can you identify any links between crash fatality and the
aforementioned variables? Are there any missing values? Which strategy should you
use to handle the missing values? Because many of the variables are categorical,
you should consider transforming them into dummy variables prior to the analysis.

Report 2.2 FILE House_Price. Subset the data based on variables, such as
number of bedrooms, number of bathrooms, home square footage, lot square
footage, and age of the house. Which variables can be removed when predicting
house prices? Are there any variables that display a skewed distribution? If there are,
perform logarithm transformations for these variables. Would it make sense to
transform some of the numeric values into categorical values using the binning
strategy? Is equal size or equal interval binning strategy more appropriate in these
situations?

Report 2.3 FILE Longitudinal_Survey. Subset the data based on age, sex, or race.
Are there any missing values in the data? Which strategy should you use to handle
the missing values? Consider if any new variables can be created using the existing
variables. Explore the opportunities of transforming numeric variables through
binning and transforming categorical variables by creating dummy variables.

Report 2.4 FILE Mortgage_Approval. Consider whether data distribution skewness

exists in some of the numeric variables, and if it does, determine how to transform
the data into a less skewed distribution. Does it make sense to reduce the number of
categories for some variables? Perform data subsetting and use simple summary
measures such as averages and frequency counts to find out if any differences exist
across subsets.
 page 80

3 Data Visualization and

Summary Measures

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 3.1  Visualize categorical and numerical
variables.
LO 3.2  Construct and interpret a contingency table
and a stacked bar chart.
LO 3.3  Construct and interpret a scatterplot.
LO 3.4  Construct and interpret a scatterplot with a
categorical variable, a bubble plot, a line
chart, and a heat map.
LO 3.5  Calculate and interpret summary measures.
LO 3.6  Use boxplots and z-scores to identify
outliers.

eople often have difficulty processing information provided by data in

its raw form. In this chapter, we present several tabular and graphical
 tools as well as summary measures that help us organize and
P present data. We first summarize a single variable by constructing
a frequency distribution. A frequency distribution is a tabular
method for condensing and summarizing data. For a visual
representation of a frequency distribution, we construct a bar chart for
a categorical variable and a histogram for a numerical variable.
We then examine the relationship between two variables by
constructing a contingency table and a stacked column chart for
categorical variables, and by constructing a scatterplot for numerical
variables. We also discuss additional data visualizations including a
scatterplot with a categorical variable, a bubble plot, a line chart, and
a heat map.
Next, we focus on summary measures. These measures provide
precise, objectively determined values that are easy to calculate,
interpret, and compare with one another. Finally, we examine boxplots
and z-scores, which are two common ways for detecting outliers.
©Tzido Sun/Shutterstock

page 81

INTRODUCTORY CASE
Investment Decision
Dorothy Brennan works as a financial advisor at a large
investment firm. She meets with an inexperienced investor who
has some questions regarding two approaches to mutual fund
investing: growth investing versus value investing. The investor
has heard that growth funds invest in companies whose stock
prices are expected to grow at a faster rate, relative to the
 overall stock market. Value funds, on the other hand, invest in
companies whose stock prices are below their true worth. The
investor has also heard that the main component of investment
return is through capital appreciation in growth funds and
through dividend income in value funds.
The investor shows Dorothy the annual return data for
Fidelity’s Growth Index mutual fund (Growth) and Fidelity’s
Value Index mutual fund (Value). Table 3.1 shows a portion of
the annual returns (in %) for these two mutual funds from 1984
to 2018.
FILE
Growth_Value

TABLE 3.1 Annual Returns (in %) for Growth and Value

Year Growth Value

1984 −4.53 −17.34

1985 36.76 15.67

⋮ ⋮ ⋮

2018 −5.5 −8.59

It is difficult for the investor to draw any conclusions from the

data in its present form. In addition to clarifying the style
differences in growth investing versus value investing, Dorothy
will use the sample information to:
1. Calculate and interpret the typical return for these two mutual
funds.
2. Calculate and interpret the investment risk for these two mutual
funds.
3. Determine which mutual fund provides the greater return
relative to risk.
A synopsis of this case is provided at the end of Section 3.4.

page 82

3.1 METHODS TO VISUALIZE

CATEGORICAL AND NUMERICAL
VARIABLES
LO 3.1
Visualize categorical and numerical variables.

In this section, we present several tabular and graphical tools that

help us organize and present data concerning a single variable. We
examine common ways to summarize both a categorical variable as
well as a numerical variable.

Methods to Visualize a Categorical Variable

Recall from Chapter 1 that a categorical variable consists of
observations that represent labels or names. For example,
participants in a survey are often asked to indicate their gender or
race, or provide ratings of a product. When presented with a
categorical variable, it is often useful to summarize the data with a
frequency distribution and/or a bar chart. We first discuss the
construction of a frequency distribution. Consider the following
example.

A Frequency Distribution for a Categorical Variable

Suppose the Human Resources department of a large technology
company maintains relevant personnel information regarding each
employee’s personality type based on the Myers-Briggs assessment.
The Myers-Briggs assessment breaks down personality types into
four categories:
Analyst: An analyst tends to be open-minded and strong-willed.
He/she likes to work independently and usually approaches things
from a very practical perspective.
Diplomat: A diplomat cares about people and tends to have a lot of
empathy toward others.
Explorer: An explorer tends to be very good at making quick,
rational decisions in difficult situations.
Sentinel: A sentinel likes stability, order, and security, and tends to
be hard working and meticulous.
Table 3.2 shows a portion of the Myers-Briggs assessment results
for 1,000 employees. The Myers-Briggs variable is a categorical
variable that is of nominal scale, and the observations in Table 3.2
merely represent labels. Data presented in this format—that is, in raw
form—are very difficult to interpret. Converting the raw data into a
frequency distribution is often a first step in making the data more
manageable and easier to assess.
FILE
Myers_Briggs

TABLE 3.2 Myers-Briggs Assessment Results

Employee Myers-Briggs Assessment

1 Diplomat

2 Diplomat
 Employee Myers-Briggs Assessment

⋮ ⋮

1000 Explorer

As shown in Table 3.3, the categories of the variable form the first
column of a frequency distribution. We then record the number of
employees that fall into each category. We can readily see from Table
3.3 that the Explorer personality type occurs with the most frequency,
while the Analyst personality type occurs with the least frequency. In
some applications, especially when comparing data sets of differing
sizes, our needs may be better served by focusing on the relative
frequency for each category rather than its frequency. The relative
frequency for each category is calculated by dividing the frequency by
the sample size. We can easily convert relative frequencies into
percentages by multiplying by 100. Table 3.3 shows that 40.4% of the
employees fall into the Explorer personality type.

page 83

TABLE 3.3 Frequency Distribution for the Myers-Briggs Variable

Personality Type Frequency Relative Frequency

Analyst 116 0.116

Diplomat 324 0.324

Explorer 404 0.404

Sentinel 156 0.156

 USING A FREQUENCY DISTRIBUTION TO DISPLAY A
CATEGORICAL VARIABLE
A frequency distribution for a categorical variable groups the
data into categories and records the number of observations that
fall into each category. The relative frequency for each category
equals the proportion of observations in each category.

A Bar Chart
Next, we show a graphical representation of a frequency distribution.
We first construct a vertical bar chart, sometimes referred to as a
column chart. The height of each bar is equal to the frequency or the
relative frequency of the corresponding category. Figure 3.1 shows
the bar chart for the Myers-Briggs variable.

FIGURE 3.1 Bar chart for the Myers-Briggs variable

USING A BAR CHART TO DISPLAY A CATEGORICAL

VARIABLE
 A bar chart depicts the frequency or the relative frequency for
each category of the categorical variable as a series of
horizontal or vertical bars, the lengths of which are proportional
to the values that are to be depicted.

FILE
Transit_Survey

EXAMPLE 3.1
Recently, an urban university conducted a transportation survey
as part of its commitment to reduce its carbon footprint and
comply with the federal Clean Air Act. The survey was
distributed to students, faculty, and staff members in order to
learn the patterns of their daily commute. One of the questions
asked: During a typical school week, how do you commute from
home to school. Possible responses included Drive_Alone,
Public_Transit, Bicycle, Walk, and Other. Six hundred people
responded to the survey. Table 3.4 shows a portion of the survey
results.

TABLE 3.4 Transit_Survey Data

Respondent Mode of Transportation

1 Bicycle

2 Public_Transit

⋮ ⋮

600 Walk
 page 84

Construct a frequency distribution and a bar chart using

Excel and R, and summarize the results.
Important: Due to different fonts and type settings, copying
and pasting Excel or R functions from this text directly into Excel
or R may cause errors. When such errors occur, you may need
to replace special characters such as quotation marks and
parentheses or delete extra spaces in the functions.
SOLUTION:
Using Excel

a. Open the Transit_Survey data file.

b. Enter the column headings Mode of Transportation and
Number of Respondents in cells D1 and E1, respectively. Enter
the column heading Drive_Alone in cell D2. Enter the formula
=COUNTIF(A2:A601, “Drive_Alone”) in cell E2. Enter the
column heading Public_Transit in cell D3. Enter the formula
=COUNTIF(A2:A601, “Public_Transit”) in cell E3. Enter the
column heading Bicycle in cell D4. Enter the formula
=COUNTIF(A2:A601, “Bicycle”) in cell E4. Enter the column
heading Walk in cell D5. Enter the formula
=COUNTIF(A2:A601, “Walk”) in cell E5. Enter the column
heading Other in cell D6. Enter the formula
=COUNTIF(A2:A601, “Other”) in cell E6. Table 3.5 shows the
frequency distribution.
TABLE 3.5 Frequency Distribution for Transit_Survey

Mode of Transportation Number of Respondents

Drive_Alone  57

Public_Transit 273
 Mode of Transportation Number of Respondents

Bicycle 111

Walk 141

Other  18

c. Select cells D2:E6. Choose Insert > Insert Bar Chart. Select
the option on the top left side. (If you are having trouble finding
this option after selecting Insert, look for the horizontal bars
above Charts.) Figure 3.2 shows the bar chart. Note that in this
instance we have constructed a horizontal bar chart. If you wish
to construct a vertical bar chart, then you would choose Insert
> Column Chart.
page 85
FIGURE 3.2 Bar chart for Transit_Survey

d. Formatting (regarding axis titles, gridlines, etc.) can be done by

selecting the ‘+’ sign at the top right of the chart or by selecting
Add Chart Elements from the menu. Check the box next to
 Data Labels in the Chart Elements pop-up box to display
frequencies in the bar chart (or the column chart).
Using R
As mentioned in Chapter 2, before following all R instructions,
make sure that you have read Appendix C (“Getting Started with
R”). We assume that you have downloaded R and RStudio, and
that you know how to import an Excel file. Throughout the text,
our goal is to provide the simplest way to obtain the relevant
output. We denote all function names in boldface and all
options within a function in italics. It is important to keep in mind
that R is case sensitive.
a. Import the Transit_Survey data into a data frame (table) and
label it myData.
b. We use the table function to create a frequency distribution
labeled Frequency. As outlined in Appendix C, we typically
identify a variable within a data frame using the expression
$var, where var denotes the variable name. Here, we need to
enclose the variable name with single quotations because the
variable name, Mode of Transportation, consists of more than
one word. If you retype Frequency, you will see that the
resulting frequency distribution is not very attractive. For this
reason, we use the View function to create a more appealing
frequency distribution. Recall from Chapter 2 that the View
function creates a spreadsheet-style data viewer. Enter:
> Frequency <- table(myData$'Mode of Transportation')
> Frequency
> View(Frequency)
The only difference in the frequency distribution that R
produces compared to the one that appears in Table 3.5 is that
the category names are arranged in alphabetical order.
c. We use the barplot function to construct a bar chart. R offers a
number of options for formatting. Here, we use main to add a
title, xlab to provide a label for the x-axis, horiz to indicate a
 horizontal bar chart, col to define color, xlim to extend the
horizontal axis units from 0 to 300, las = 1 to display the
category names perpendicular to the y-axis, and cex.names to
reduce the font size of the category names so that they are not
truncated. Finally, we use the abline function to insert the y-
axis.
> barplot(Frequency,main=“Bar Chart for Transit Survey”,
xlab=“Number of Respondents”, horiz=TRUE, col=“blue”,
xlim=c(0,300), las=1, cex.names=0.5)
> abline(v=0)
There are a few differences in the bar chart that R produces
compared to the one that appears in Figure 3.2, but these
differences are cosmetic.
Summary
Table 3.5 and Figure 3.2 reveal that the most common
commuting mode at this urban university is public transportation.
Walking and riding a bicycle are the next most common
commuting modes. These results are not surprising for a
university that is located in a city.

page 86

Methods to Visualize a Numerical Variable

With a numerical variable, each observation represents a meaningful
amount or count. The number of patents held by pharmaceutical firms
(count) and household incomes (amount) are examples of numerical
variables. Although different in nature from a categorical variable, we
still use a frequency distribution to summarize a numerical variable.

A Frequency Distribution for a Numerical Variable

When we constructed a frequency distribution for a categorical
variable, the raw data could be categorized in a well-defined way; we
simply counted the number of observations in each category. For a
numerical variable, instead of categories, we construct a series of
intervals (sometimes called classes). We must make certain decisions
about the number of intervals, as well as the width of each interval.
When making these decisions, we consider the following guidelines.
Intervals are mutually exclusive. For example, suppose the first two
intervals of a frequency distribution are defined as 300 < x ≤ 400
and 400 < x ≤ 500, where x is the value of an observation. If x =
400, then it would fall into the first interval. In other words, intervals
do not overlap, and each observation falls into one, and only one,
interval.
The total number of intervals in a frequency distribution usually
ranges from 5 to 20. Smaller data sets tend to have fewer intervals
than larger data sets. Recall that the goal of constructing a
frequency distribution is to summarize the data in a form that
accurately depicts the group as a whole. If we have too many
intervals, then this advantage of the frequency distribution is lost.
Similarly, if the frequency distribution has too few classes, then
considerable accuracy and detail are lost.
Intervals are exhaustive. The total number of intervals covers the
entire sample (or population).
Interval limits are easy to recognize and interpret. For example, the
intervals −10 < x ≤ 0, 0 < x ≤ 10, etc. are preferred to the intervals
−8 < x ≤ 2, 2 < x ≤ 12, etc. Also, as a starting point for
approximating the width of each interval, we often use the formula:
(largest observation – smallest observation)/(number of intervals).

USING A FREQUENCY DISTRIBUTION TO DISPLAY A

NUMERICAL VARIABLE
For a numerical variable, a frequency distribution groups data
into intervals and records the number of observations that falls
into each interval. The relative frequency for each interval equals
the proportion of observations in each interval.
 The Growth variable from the introductory case is a numerical
variable because the data reflect annual returns (in %) for Fidelity’s
Growth Index mutual fund from 1984 to 2018. Here we will create a
frequency distribution with six intervals. The minimum and maximum
observations for the Growth variable are −40.9 and 79.48,
respectively. (Obtaining summary measures will be discussed in
Section 3.4.) Using the approximation formula to find the width of
each interval, we calculate (79.48 − (−40.9))/6 = 20.0633. However,
intervals with a width of 20.0633 would not have limits that are easily
recognizable. For this reason, we will define the lower limit of the first
interval as −50 and have each interval be of width 25 that is, −50 < x
≤ −25, −25 < x ≤ 0, etc., where x is the annual return.
As shown in Table 3.6, the data are more manageable using a
frequency distribution, but some detail is lost because we no longer
see the actual observations. From the frequency distribution, we can
now readily observe that the most likely return for the Growth variable
is between 0% and 25%; there were 17 observations in this interval.
We also note that no observations fall between 50% page 87
and 75%, and only one observation falls between 75%
and 100%; we will see later that this observation is called an outlier.
Summing the values in the frequency column shows that the sample
size (or n) is 35. The relative frequency for each interval is again
calculated by dividing the frequency by the sample size. As before, a
relative frequency can be converted into a percentage by multiplying
by 100. The third column of Table 3.6 shows that 48.57% of the
returns for the Growth variable fell between 0% and 25%.

TABLE 3.6 Frequency Distribution for Growth

Interval (in %) Frequency Relative Frequency

−50 < x ≤ −25 3 0.0857

−25 < x ≤ 0 5 0.1429

 Interval (in %) Frequency Relative Frequency

0 < x ≤ 25 17  0.4857

25 < x ≤ 50 9 0.2571

50 < x ≤ 75 0 0

75 < x ≤ 100 1 0.0286

A Histogram
Next, we show a graphical representation of a frequency distribution.
For numerical data, a histogram is essentially the counterpart to the
vertical bar chart that we use for categorical data.
When constructing a histogram, we typically mark off the interval
limits along the horizontal axis. The height of each bar represents
either the frequency or the relative frequency for each interval. No
gaps appear between the interval limits.

USING A HISTOGRAM TO DISPLAY A NUMERICAL

VARIABLE
A histogram is a series of rectangles where the width and height
of each rectangle represent the interval width and frequency (or
relative frequency) of the respective interval.

Figure 3.3 shows a histogram for the frequency distribution shown

in Table 3.6. The advantage of a visual display is that we can quickly
see where most of the observations tend to cluster, as well as the
spread and shape of the data. From Figure 3.3 we can see that
annual returns for the Growth variable ranged from −50% to 100%.
Annual returns from 0% to 25% were the most likely, whereas annual
returns over 50% were very unlikely.
FIGURE 3.3 Histogram for the annual returns (in %) for Growth

page 88

A histogram also provides information on the shape of the

distribution. In general, the shape of most distributions can be
categorized as either symmetric or skewed. A symmetric distribution
is one that is a mirror image of itself on both sides of its center. That
is, the location of values below the center correspond to those above
the center. As we will see in Chapter 4, the smoothed histogram for
many data sets approximates a bell-shaped curve, which is indicative
of the well-known normal distribution. Figure 3.4(a) shows a
histogram with a symmetric distribution. If the edges were smoothed,
this histogram would look somewhat bell-shaped.

FIGURE 3.4 Histograms with differing shapes

 If the distribution is not symmetric, then it is either positively
skewed or negatively skewed. Figure 3.4(b) shows a histogram with a
positively skewed, or skewed to the right, distribution. The long tail
that extends to the right reflects the presence of a small number of
relatively large values. Figure 3.4 (c) shows a histogram with a
negatively skewed, or skewed to the left, distribution because it has a
long tail extending off to the left. Data with a negatively skewed
distribution have a small number of relatively small values.
The histogram for the Growth variable in Figure 3.3 shows that it is
not a symmetric distribution; however, it is harder to discern whether it
is positively skewed or negatively skewed. It turns out that the
distribution is slightly positively skewed. We are able to make this
conclusion by examining the skewness coefficient for the Growth
variable. As we will see in Section 3.4, the skewness coefficient is a
summary measure that is routinely provided in Excel output.
FILE
Growth_Value

EXAMPLE 3.2
The Value variable from the introductory case shows annual
returns (in %) for Fidelity’s Value mutual fund from 1984 through
 2018. Construct a frequency distribution and a histogram using
Excel and R, and then summarize the results.
SOLUTION:
Before using Excel or R, we need to make some decisions
about the number of intervals, as well as the width of each
interval. For a variable with 35 observations, it would be
reasonable to use five intervals. We then find that the minimum
and the maximum observations for the Value variable are
−46.52 and 44.08, respectively. Using the formula to
approximate the interval width, we calculate (44.08 −(−46.52))/5
= 18.12. It would be perfectly acceptable to construct, for
instance, a frequency distribution with five intervals where each
interval has a width of 20, and the lower limit of the first interval
is −50. However, because one of our objectives is to compare
the Growth returns with the Value returns, we use the same
number of intervals, same width, and same lower limit as we did
when we constructed the frequency distribution for page 89
Growth; that is, we use six intervals, each with a
width of 25, and the first interval has a lower limit of −50.
Using Excel

a. Open the Growth_Value data file.

b. In cell E1 enter the heading Interval Limits, and in cells E2
through E7 enter the upper limit of each interval, so −25, 0, 25,
50, 75, and 100. The reason for these entries will be explained
shortly.
c. From the menu choose Data > Data Analysis > Histogram >
OK. (Note: If you do not see the Data Analysis option under
Data, you must add in the Analysis Toolpak option. From the
menu choose File > Options > Add-Ins and choose Go at the
bottom of the dialog box. Select Analysis Toolpak and then
click OK. If you have installed this option properly, you should
now see Data Analysis under Data.)
d. See Figure 3.5. In the Histogram dialog box, next to Input
Range, select the Value observations. Excel uses the term
“bins” for the interval limits. If we leave the Bin Range box
empty, Excel creates evenly distributed intervals using the
minimum and maximum values of the data as end points. This
approach is rarely satisfactory. In order to construct a histogram
that is more informative, we use the upper limits of each class
as the bin values. Next to Bin Range, we select cells E1:E7 (the
Interval Limits observations). (We check the Labels box
because we have included the names Value and Interval Limits
as part of the selection.) Under Output Options, select Output
Range and enter cell G1 and then select Chart Output. Click
OK.

FIGURE 3.5 Excel’s dialog box for a histogram

Table 3.7 shows the frequency distribution for Value. In the

event that the given interval limits do not include all the data
points, Excel automatically adds another interval labeled “More”
to the resulting frequency distribution and histogram. Because
we observe zero observations in this interval, we delete this
interval for expositional purposes. Also, it is important to note
that Excel defines its intervals by including the value of the
 upper class limit for each interval. For example, if the value −25
appeared in the data, Excel would have accounted for this
observation in the first interval.

TABLE 3.7 Frequency Distribution for Value

Interval (in %) Frequency Relative Frequency

−50 < x ≤ −25 1 0.0286

−25 < x ≤ 0 7 0.2000

0 < x ≤ 25 21  0.6000

25 < x ≤ 50 6 0.1714

50 < x ≤ 75 0 0.0000

75 < x ≤ 100 0 0.0000

In order to calculate the relative frequency for each interval,

we go to cell I2 and enter =H2/35. We then select cell I2 and
drag down to cell I7, and from the menu we choose Home > Fill
> Down. The third column of Table 3.7 shows the relative
frequencies for each interval.

page 90

e. Because Excel leaves spaces between the rectangles in the

histogram, we right-click on any of the rectangles, choose
Format Data Series, change the Gap Width to 0, and then
choose Close. Formatting (regarding axis titles, gridlines, etc.)
can be done by selecting Format > Add Chart Element from
the menu. Figure 3.6 shows the histogram for Value.
 FIGURE 3.6 Histogram for the annual returns (in %) for Value

Note that you can construct a frequency distribution by using

Excel’s COUNTIF function that we used in the last example as
well as in Chapter 2. You can then use the resulting frequency
distribution to plot a histogram. The Histogram option in Excel’s
Data Analysis Toolpak allows us to construct both the frequency
distribution and the histogram with one command.
Using R

a. Import the Growth_Value data into a data frame (table) and

label it myData.
b. We first define the intervals using the seq function. The first
argument in the function is the lower limit of the first interval,
the next argument is the upper limit of the last interval, and the
last argument defines the width of each interval. Enter:
> intervals <- seq(-50, 100, by=25)
c. We then use the cut function with left and right options to
ensure that the intervals are open on the left and closed on the
right; that is, −50 < x ≤ −25, −25 < x ≤ 0, etc. Enter:
> value.cut <- cut(myData$Value, intervals, left=FALSE,
right=TRUE)
d. We use the table function to create a frequency distribution
labeled value.freq. If you retype value.freq, you will see that the
resulting frequency distribution is not very attractive. For these
 reasons, we use the View function to create a more appealing
frequency distribution. Enter:
> value.freq <- table(value.cut)
> value.freq
> View(value.freq)
The frequency distribution that R produces should be
comparable to Table 3.7.
page 91
e. We use the hist function to construct a histogram.
We define the breaks option using the intervals that we defined
in step b. Again, we set the right option equal to TRUE so that
the intervals are right-closed (implying left-opened). As when
constructing a bar chart, R offers a number of options for
formatting. Here, we use main to add a title, xlab to provide a
label for the x-axis, and col to define color. Enter:
> hist(myData$Value, breaks=intervals, right=TRUE,
main=“Histogram for the annual returns (in %) for the Value
Fund”, xlab=“Annual Returns (in %) for Value”, col=“blue”)
The histogram that R produces should be comparable to Figure
3.6.
Summary
From Table 3.7, we see that the range of annual returns for
Value is between −50% and 50%. The range for Value is
narrower than the one for Growth. Similar to Growth, the most
likely return is in the 0% to 25% interval because the most
observations (21) fall in this interval. Unlike Growth, none of the
observations is greater than 50%. From Figure 3.6, we see that
the distribution of Value is not symmetric; it is negatively skewed
with a tail running off to the left.
Cautionary Comments When Constructing or
Interpreting Charts or Graphs
As with many of the analytical methods that we examine throughout
this text, the possibility exists for unintentional, as well as purposeful,
distortions of graphical information. As a careful researcher, you
should follow these basic guidelines:
FILE
Crude_Oil

The simplest graph should be used for a given set of data. Strive for
clarity and avoid unnecessary adornments.
Axes should be clearly marked with the numbers of their respective
scales; each axis should be labeled.
When creating a bar chart or a histogram, each bar/rectangle
should be of the same width. Differing widths create distortions.
The vertical axis should not be given a very high value as an upper
limit. In these instances, the data may appear compressed so that
an increase (or decrease) of the data is not as apparent as it
perhaps should be. For example, Figure 3.7(a) plots the daily price
for a barrel of crude oil for the first quarter of the year. Due to Middle
East unrest, the price of crude oil rose from a low of $83.13 per
barrel to a high of $106.19 per barrel, or approximately 28%
. However, because Figure 3.7(a) uses a high value as

an upper limit on the vertical axis ($325), the rise in price appears
dampened. Figure 3.7(b) shows a vertical axis with an upper limit of
$110; this value better reflects the upper limit observed during this
time period.
page 92
FIGURE 3.7 Misleading vertical axis: unreasonably high upper limit
FILE
Stockprice
The vertical axis should not be stretched so that an increase (or
decrease) of the data appears more pronounced than warranted.
For example, Figure 3.8(a) charts the daily closing stock price of a
large retailer for the week of April 4. It is true that the stock price
declined over the week from a high of $60.15 to a low of $59.46;
this amounts to a $0.69 decrease, or an approximate 1% decline.
However, because the vertical axis is stretched, the drop in stock
price appears more dramatic. Figure 3.8(b) shows a vertical axis
that has not been stretched.

FIGURE 3.8 Misleading scale on vertical axis: stretched scale

EXERCISES 3.1
Applications
1. Fifty pro-football rookies were rated on a scale of 1 to 5, based on performance at a
training camp as well as on past performance. A ranking of 1 indicated a poor
prospect, whereas a ranking of 5 indicated an excellent prospect. The following
frequency distribution was constructed.

Rating Frequency

1  4

2 10

3 14

4 18

5  4

a. How many of the rookies received a rating of 4 or better? How many of the
rookies received a rating of 2 or worse?
b. Construct the relative frequency distribution. What proportion received a rating of
5?
c. Construct a bar chart. Comment on the findings.
2. The following frequency distribution shows the counts of sales of men’s shirts at an
online retailer over the weekend.

Size Frequency

Small  80

Medium 175

Large 210

X-Large 115

page 93
a. Construct the relative frequency distribution. What proportion of
sales were for a medium-sized shirt?
b. Construct a bar chart. Comment on the findings.
3. The following frequency distribution summarizes the counts of purchases by day of
the week for a major domestic retailer.

Day Frequency

Mon 2,504

Tue 2,880

Wed 3,402

Thur 3,566

Fri 4,576

Sat 5,550

Sun 5,022

a. Construct the relative frequency distribution. What proportion of the purchases

occurred on Wednesday?
b. Construct a bar chart using relative frequencies. Comment on the findings.
4. In 2018, the U.S. Census Bureau provided the following frequency distribution for
the number of people (in 1,000s) who live below the poverty level by region.
 Region Number of People

Northeast 6,373

Midwest 7,647

South 16,609 

West 9,069

a. Construct the relative frequency distribution. What proportion of people who live
below the poverty level live in the Midwest?
b. Construct a bar chart. Comment on the findings.
5. A recent poll of 3,057 individuals asked: “What’s the longest vacation you plan to
take this summer?” The following relative frequency distribution summarizes the
results.

Response Relative Frequency

A few days 0.21

A few long weekends 0.18

One week 0.36

Two weeks 0.25

a. Construct the frequency distribution. How many people are going to take a one-
week vacation this summer?
b. Construct a bar chart. Comment on the findings.
6. FILE Dining. A local restaurant is committed to providing its patrons with the best
dining experience possible. On a recent survey, the restaurant asked patrons to
rate the quality of their entrées. The responses ranged from 1 to 5, where 1
indicated a disappointing entrée and 5 indicated an exceptional entrée. A portion of
the 200 responses is as follows:

Response Rating

1 3

2 5
 Response Rating

⋮ ⋮

200 4

a. Construct the frequency distribution that summarizes the results from the survey.
Which rating appeared with the most frequency?
b. Construct a bar chart. Are the patrons generally satisfied with the quality of their
entrées? Explain.
7. FILE Health. Patients at North Shore Family Practice are required to fill out a
questionnaire that gives the doctor an overall idea of each patient’s health. The first
question is: “In general, what is the quality of your health?” The patient chooses
Excellent, Good, Fair, or Poor. A portion of the 150 responses is as follows:

Response Quality

1 Fair

2 Good

⋮ ⋮

150 Good

a. Construct the frequency distribution that summarizes the results from the
questionnaire. What is the most common response to the questionnaire?
b. Construct a bar chart. How would you characterize the health of patients at this
medical practice? Explain.
8. FILE Millennials. A 2014 Religious Landscape Study by the Pew Research
Center found that 35% of Millennials (Americans born between 1981 and 1996)
identified themselves as not religious. A researcher wonders if this finding is
consistent today. She surveys 600 Millennials and asks them to rate their faith.
Possible responses were Strongly Religious, Somewhat Religious, Slightly
Religious, and Not Religious. A portion of the 600 responses is as follows:

Response Faith

1 Slightly Religious

2 Slightly Religious
 Response Faith

⋮ ⋮

600 Somewhat Religious

a. Construct the frequency distribution that summarizes the results from the survey.
What is the most common response to the survey?
b. Construct a bar chart. Do the researcher’s results appear consistent with those
found by the Pew Research Center? Explain.
9. A researcher conducts a mileage economy test involving 80 cars. The frequency
distribution describing average miles per gallon (mpg) appears in the following
table.

Average mpg Frequency

15 ≤ x < 20 15

20 ≤ x < 25 30

25 ≤ x < 30 15

30 ≤ x < 35 10

35 ≤ x < 40  7

40 ≤ x < 45  3

a. Construct the relative frequency distribution. What proportion of the cars got at
least 20 mpg but less than 25 mpg? What proportion of the cars got less than 35
mpg? What proportion of the cars got 35 mpg or more?
b. Construct a histogram. Comment on the shape of the distribution.

page 94
10. Consider the following relative frequency distribution that
summarizes the returns (in %) for 500 small cap stocks.

Return (%) Relative Frequency

−20 ≤ x < −10 0.04

−10 ≤ x < 0 0.25

 Return (%) Relative Frequency

0 ≤ x < 10 0.42

10 ≤ x < 20 0.25

20 ≤ x < 30 0.04

a. Construct the frequency distribution. How many of the stocks had a return of at
least 10% but less than 20%?
b. Construct a histogram. Comment on the shape of the distribution.
11. The manager at a water park constructed the following frequency distribution to
summarize attendance in July and August.

Attendance Frequency

1,000 ≤ x < 1,250  5

1,250 ≤ x < 1,500  6

1,500 ≤ x < 1,750 10

1,750 ≤ x < 2,000 20

2,000 ≤ x < 2,250 15

2,250 ≤ x < 2,500  4

a. Construct the relative frequency distribution. What proportion of the time was
attendance at least 1,750 but less than 2,000? What proportion of the time was
attendance less than 1,750? What proportion of the time was attendance 1,750
or more?
b. Construct a histogram. Comment on the shape of the distribution.
12. Fifty cities provided information on vacancy rates (in %) in local apartments in the
following frequency distribution.

Vacancy Rate (%) Relative Frequency

0≤x<3 0.10

3≤x<6 0.20
 Vacancy Rate (%) Relative Frequency

6≤x<9 0.40

9 ≤ x < 12 0.20

12 ≤ x < 15 0.10

a. Construct the frequency distribution. How many of the cities had a vacancy rate
of at least 6% but less than 9%? How many of the cities had a vacancy rate of at
least 9%?
b. Construct a histogram. Comment on the shape of the distribution.
13. The following relative frequency histogram summarizes the median household
income for the 50 states as reported by the U.S. Census Bureau in 2010.

page 95
a. Is the distribution symmetric? If not, is it positively or negatively
skewed?
b. What percentage of the states had median household income between $45,000
and $55,000?
c. What percentage of the states had median household income between $35,000
and $55,000?
14. The following histogram summarizes Apple Inc.’s monthly stock price for the years
2014 through 2018.
 a. Is the distribution symmetric? If not, is it positively or negatively skewed?
b. Over this five-year period, approximate the minimum monthly stock price and the
maximum monthly stock price.
c. Over this five-year period, which interval had the highest relative frequency?
15. The following histogram summarizes the salaries (in $100,000s) for the 30 highest-
paid portfolio managers at a large investment firm over the past year.

a. Is the distribution symmetric? If not, is it positively or negatively skewed?

b. How many of the portfolio managers earned between $2,000,000 and
$2,400,000?
c. Approximately how many of the portfolio managers earned between $1,200,000
and $2,000,000?
16. FILE Prime. Amazon Prime is a $119-per-year service that gives the company’s
customers free two-day shipping and discounted rates on overnight delivery. Prime
customers also get other perks, such as free e-books. The accompanying table
shows a portion of the annual expenditures (in $) for 100 Prime customers.
 Customer Expenditures

1 1272

2 1089

⋮ ⋮

100 1389

a. Construct the frequency distribution for Expenditures. Use six intervals with
widths of 400 < x ≤ 700; 700 < x ≤ 1,000; etc. How many customers spent
between $701 and $1,000?
b. How many customers spent $1,300 or less? How many customers spent more
than $1,300?
17. FILE Census. The following table lists a portion of median house values (in $) for
the 50 states as reported by the U.S. Census Bureau in 2010.

State House Value

Alabama 117600

Alaska 229100

⋮ ⋮

Wyoming 174000

a. Construct the frequency distribution and the histogram for the median house
values. Use six intervals with widths of 0 < x ≤ 100,000; 100,000 < x ≤ 200,000;
etc. Which interval had the highest frequency? How many of the states had
median house values of $300,000 or less?
b. Is the distribution symmetric? If not, is it positively or negatively skewed?
18. FILE DJIA_2019. The accompanying table shows a portion of the daily price index
for the Dow Jones Industrial Average (DJIA) for the first half of 2019.

Day DJIA

January 1, 2019 23346

January 2, 2019 22686

 Day DJIA

⋮ ⋮

June 28, 2019 26600

a. Construct the frequency distribution and the histogram for the DJIA. Use five
intervals with widths of 22,000 < x ≤ 23,000; 23,000 < x ≤ 24,000; etc. On how
many days during the first half of 2019 was the DJIA more than 26,000?
b. Is the distribution symmetric? If not, is it positively or negatively skewed?
19. FILE Gas_2019. The following table lists a portion of the average price (in $) for a
gallon of gas for the 50 states and the District of Columbia as reported by AAA Gas
Prices on January 2, 2019.

State Price

Alabama 1.94

Alaska 3.06

⋮ ⋮

Wyoming 2.59

a. Construct the frequency distribution and the histogram for the average price of
gas. Use six intervals with widths of 1.70 < x ≤ 2.00; 2.00 < x ≤ 2.30; etc. Which
interval had the highest frequency? How many of the states had average gas
prices greater than $2.60?
b. Is the distribution symmetric? If not, is it positively or negatively skewed?
20. The accompanying figure plots the monthly stock price of a large construction
company from July 2017 through March 2019. The stock has experienced
tremendous growth over this time period, almost tripling in price. Does the figure
reflect this growth? If not, why not?
 21. Annual sales at a small pharmaceutical firm have been rather stagnant over the
most recent five-year period, exhibiting only 1.2% growth over this time frame. A
research analyst prepares the accompanying graph for inclusion in a sales report.

page 96

Does this graph accurately reflect what has happened to sales over the last five
years? If not, why not?

3.2 METHODS TO VISUALIZE THE

RELATIONSHIP BETWEEN TWO
VARIABLES
LO 3.2
Construct and interpret a contingency table and a stacked bar chart.
All of the tabular and graphical tools presented thus far have focused
on describing one variable. However, in many instances we are
interested in the relationship between two variables. People in
virtually every discipline examine how one variable may
systematically influence another variable. Consider, for instance, how
Incomes vary with education.
Sales vary with advertising expenditures.
Stock prices vary with corporate profits.
Crop yields vary with the use of fertilizer.
Cholesterol levels vary with dietary intake.
Employee absences vary with work shift.
We first introduce contingency tables and stacked column charts, two
common tabular and graphical methods that help us summarize the
relationship between two categorical variables. Then we show a
scatterplot, which is a common graphical method that allows us to
determine whether two numerical variables are related in some
systematic way.

Methods to Visualize the Relationship

between Two Categorical Variables
When examining the relationship between two categorical variables, a
contingency table proves very useful. Contingency tables are widely
used in marketing as well as other business applications. Consider
the following example.

A Contingency Table
Recall the Myers_Briggs data set discussed in Section 3.1. If we
expand the data set to include another categorical variable, say one’s
sex, it would allow us to examine the relationship between personality
type and one’s sex. Perhaps we are interested in whether certain
personality types are more prevalent among males versus females.
Table 3.8 shows a portion of the expanded data set, Myers_Briggs2.
FILE
Myers_Briggs2

TABLE 3.8 Expanded Myers-Briggs Assessment Results

Employee Myers-Briggs Assessment Sex

1 Diplomat Female

2 Diplomat Female

⋮ ⋮ ⋮

1000 Explorer Male

page 97

A contingency table that shows the frequencies for personality

type and sex is displayed in Table 3.9. As you can see, 55 female
employees fall into the Analyst personality type whereas 61 male
employees fall into this personality type.

TABLE 3.9 Contingency Table for Expanded Myers-Brigg Example

USING A CONTINGENCY TABLE TODISPLAY TWO

CATEGORICAL VARIABLES
A contingency table shows the frequencies for two categorical
variables, x and y, where each cell represents a mutually
 exclusive combination of the pair of x and y values.

A Stacked Column Chart

The information in a contingency table can be shown graphically
using a stacked column chart. A stacked column chart is an
advanced version of the column chart that we discussed in Section
3.1. It is designed to visualize more than one categorical variable,
plus it allows for the comparison of composition within each category.
Figure 3.9 shows the stacked column chart for personality type
and sex. Each column in the chart represents all the employees of a
personality type, and the two segments in each column represent
female employees and male employees. As the chart shows, no
discernable differences in personality type can be found depending on
whether an employee is female or male.

FIGURE 3.9 A stacked column chart for personality type and sex

In order to illustrate the construction of a contingency table and a

stacked column chart, consider the following example.
FILE
Promotion
 EXAMPLE 3.3
An online retailer recently sent e-mails to customers that
included a promotional discount. The retailer wonders whether
there is any relationship between a customer’s location in the
U.S. (Midwest, Northeast, South, or West) and whether the
customer made a purchase with the discount (yes or no). Table
3.10 shows a portion of the results from 600 e-mail accounts.

TABLE 3.10 Location and Purchase Survey Responses

Email Location Purchase

1 West yes

2 Northeast yes

⋮ ⋮ ⋮

600 South no

page 98

Construct a contingency table and a stacked column chart

using Excel and R, and then summarize the results.
SOLUTION:
Using Excel

a. Open the Promotion data file.

b. Click anywhere on the data (we choose cell A5). From the
menu, select Insert > Pivot Table. Figure 3.10 shows the
Create PivotTable dialog box. Because we clicked on the data
before creating a pivot table, the default option in Select a table
or range should already be populated. We choose to place the
pivot table in the existing worksheet beginning in cell E1. Check
 the box next to the Add this data to the Data Model option.
Then click OK.

FIGURE 3.10 Excel’s Create PivotTable Dialog Box

c. A menu will appear on the right side of the screen called

PivotTable Fields. In the top of this menu you will see all of the
variables in our data set. In the bottom part of the menu, there
is a grid with four fields: Filters, Rows, Columns, and Values;
see Figure 3.11. Drag the Location variable to the Rows field.
Drag the Purchase variable to the Columns field. Drag the
Email variable to the Values field. If the Email variable in the
Values field is not presented as a count (for example, it may be
presented as a sum), you will need to change it. Click the arrow
below the Values field and select Value Field Settings. In the
dialog box, select the Summarize value field by tab, and then,
in the drop-down menu, select Count. Click OK.

FIGURE 3.11 Excel’s Pivot Table Fields

 page 99
The resulting contingency table should be
similar to Table 3.11.
TABLE 3.11 Contingency Table for the Location and Purchase
Example

  Purchase  

Location No Yes Total

Midwest 107 77 184

Northeast 41 102 143

South 24 130 154

West 18 101 119

Total 190 410 600

 Sometimes it is preferable to convert counts to percentages,
as shown in Table 3.12. In order to make this change, go back
to the Value Field Settings dialog box, select the Show values
as tab, and in the drop-down menu select % of Grand Total.
TABLE 3.12 Percent Table for the Location and Purchase
Example

  Purchase  

Location No Yes Total

Midwest 17.83% 12.83% 30.67%

Northeast 6.83% 17.00% 23.83%

South 4.00% 21.67% 25.67%

West 3.00% 16.83% 19.83%

Total 31.67% 68.33% 100.00% 

d. We now illustrate how to create a stacked column chart using

the contingency table. Make sure that the contingency table
shows counts as in Table 3.11. Select the cells E2:G6. Choose
Insert > Insert Column or Bar Chart > Stacked Column.
e. Formatting (regarding axis titles, gridlines, etc.) can be done by
selecting the ‘+’ sign at the top right of the chart or by selecting
Add Chart Elements from the menu. Check the box next to
Data Labels in the Chart Elements pop-up box to display
frequencies in the column chart. The resulting stacked column
chart is shown in Figure 3.12.
page 100
 FIGURE 3.12 A stacked column chart for the Location and
Purchase example

Using R

a. Import the Promotion data into a data frame (table) and label it
myData.
b. In order to create a contingency table, labeled as myTable, we
use the table(row, column) function and specify the row and
column variables. If you retype myTable, you will see a
contingency table that resembles Table 3.11. If we use the
prop.table function, then R returns cell proportions that, when
converted to percentages, are the same as those that appear in
Table 3.12. Enter:
> myTable <- table(myData$Location, myData$Purchase)
> myTable
> prop.table(myTable)
c. To create a stacked column chart similar to Figure 3.12, we
need to first create a contingency table with the Purchase
variable in rows and the Location variable in columns. Enter:
> myNewTable <- table(myData$Purchase, myData$Location)
d. We use the barplot function to construct a column chart. As we
saw when constructing a bar chart and a histogram, R offers a
number of options for formatting. Here we use main to add a
 title; col to define colors for the segments of the columns;
legend to create a legend; xlab and ylab to provide labels for
the x-axis and y-axis, respectively; and ylim to extend the
vertical axis units from 0 to 200. Enter:
> barplot(myNewTable, main=“Location and Purchase”,
col=c('blue','red'), legend=rownames(myNewTable),
xlab='Location', ylab='Count', ylim = c(0,200))
The resulting stacked column chart should look similar to
Figure 3.12.
Summary
Compared to Table 3.10 with just raw data, Table 3.11, Table
3.12, and Figure 3.12 present the results of the location and
purchase example in a much more informative format. We can
readily see that of the 600 e-mail recipients, 410 of them made a
purchase using the promotional discount. With a 68.33%
positive response rate, this marketing strategy seemed
successful. However, there do appear to be some differences
depending on location. Recipients residing in the South and
West were a lot more likely to make a purchase (130 out of 154
and 101 out of 119, respectively) compared to those residing in
the Midwest (77 out of 184). It would be wise for the retailer to
examine if there are other traits that the customers in the South
and West share (age, gender, etc.). That way, in the next
marketing campaign, the e-mails can be even more targeted.

page 101

USING A STACKED COLUMN CHART TO DISPLAY TWO

CATEGORICAL VARIABLES
A stacked column chart is designed to visualize more than one
categorical variable. It allows for the comparison of composition
within each category.
A Method to Visualize the Relationship
between Two Numerical Variables
LO 3.3
Construct and interpret a scatterplot.

When examining the relationship between two numerical variables, a

scatterplot is a simple, yet useful, graphical tool. Each point in a
scatterplot represents a paired observation for the two variables.
When constructing a scatterplot, we generally refer to one of the
variables as x and represent it on the horizontal axis (x-axis) and the
other variable as y and represent it on the vertical axis (y-axis). We
then plot each pairing: (x1, y1), (x2, y2), and so on. Once the data are
plotted, the graph may reveal that
A linear relationship exists between the two variables;
A nonlinear relationship exists between the two variables; or
No relationship exists between the two variables.
For example, Figure 3.13(a) shows points on a scatterplot clustered
together along a line with a negative slope; we infer that the two
variables have a negative linear relationship. Figure 3.13(b) depicts a
positive nonlinear relationship; as x increases, y tends to increase at
an increasing rate. The points in Figure 3.13(c) are scattered with no
apparent pattern; thus, there is no relationship between the two
variables.

FIGURE 3.13 Scatterplots depicting various relationships between two variables

 USING A SCATTERPLOT TO DISPLAY THE RELATIONSHIP
BETWEEN TWO NUMERICAL VARIABLES
A scatterplot is a graphical tool that helps in determining whether
or not two numerical variables are related in some systematic
way. Each point in a scatterplot represents a paired observation
for the two variables.

In order to illustrate a scatterplot, consider the following example.

FILE
Growth_Value

EXAMPLE 3.4
Recall the Growth_Value data set that contains annual returns
for Fidelity’s Growth and Value mutual funds from 1984 to 2018.
Construct a scatterplot of Value against Growth using Excel and
R, and then summarize the results.

page 102
SOLUTION:
Using Excel
a. Open the Growth_Value data file.
b. When constructing a scatterplot, Excel places the variable that
appears in the first column on the x-axis and the variable that
appears in the second column on the y-axis. Because we want
Growth to be on the x-axis and Value on the y-axis, we do not
need to rearrange the order of the columns. We simultaneously
select the observations for the Growth and Value variables and
choose Insert > Insert Scatter or Bubble Chart > Scatter. (If
you are having trouble finding this option, look for the graph
with data points above Charts.) The resulting scatterplot should
be similar to Figure 3.14.

FIGURE 3.14 A scatterplot of Value against Growth

c. Formatting (regarding axis titles, gridlines, etc.) can be done by

selecting Format > Add Chart Element from the menu.
Using R

a. Import the Growth_Value data into a data frame (table) and

label it myData.
b. In order to construct a scatterplot, we use the plot(y~x, . . .)
function. As we saw when constructing a bar chart and a
 histogram, R offers a number of options for formatting. Here,
we use main to add a title; xlab and ylab to provide labels for
the x-axis and the y-axis, respectively; col to define color; and
pch to choose the marker for the data points—in particular, pch
= 16 displays filled circles as data markers. Enter:
> plot(myData$Value~myData$Growth, main=“Scatterplot of
Value against Growth”, xlab=“Growth”, ylab=“Value”,
col=“chocolate”, pch=16)
The resulting scatterplot should be similar to Figure 3.14.
Summary
From Figure 3.14, we can infer that there is a positive
relationship between Value and Growth; that is, as the annual
return for Value increases, the annual return for Growth tends to
increase as well.

page 103

EXERCISES 3.2
Applications
22. FILE Bar. At a local bar in a small Midwestern town, beer and wine are the only
two alcoholic options. The manager conducts a survey on the bar’s customers over
the past weekend. Customers are asked to identify their sex (defined as male or
female) and their drink choice (beer, wine, or soft drink). A portion of the responses
is shown in the accompanying table.

Customer Sex Drink Choice

1 male beer

2 male beer

⋮ ⋮ ⋮

270 female soft drink

 a. Construct a contingency table that cross-classifies the data by Sex and Drink
Choice. How many of the customers were male? How many of the customers
drank wine?
b. Given that a customer is male, what is the likelihood that he drank beer? Given
that a customer is female, what is the likelihood that she drank beer?
c. Construct a stacked column chart. Comment on the findings.
23. FILE Friends. It has been generally believed that it is not feasible for men and
women to be just friends. Others argue that this belief may not be true anymore
because gone are the days when men worked and women stayed at home and the
only way they could get together was for romance (www.npr.org, November 12,
2016). A researcher conducts a survey on 186 students. The students are asked
their sex (male or female) and if it is feasible for men and women to be just friends
(yes or no). A portion of the responses is shown in the accompanying table.

Student Sex Feasible

1 female yes

2 female yes

⋮ ⋮ ⋮

186 male no

a. Construct a contingency table that cross-classifies the data by Sex and Feasible.
How many of the students were female? How many of the students felt that it
was feasible for men and women to be just friends?
b. What is the likelihood that a male student feels that men and women can be just
friends? What is the likelihood that a female student feels that men and women
can be just friends?
c. Construct a stacked column chart. Do male and female students feel the same or
differently about this topic? Explain.
24. FILE Shift. Metalworks, a supplier of fabricated industrial parts, wonders if there is
any connection between when a component is constructed (Shift is equal to 1, 2, or
3) and whether or not it is defective (Defective is equal to Yes if the component is
defective, No otherwise). The supplier collects data on the construction of 300
components. A portion of the data is shown in the accompanying table.

Component Shift Defective

1 1 No

2 1 Yes
 Component Shift Defective

⋮ ⋮ ⋮

300 3 No

a. Construct a contingency table that cross-classifies the data by shift and whether
or not the component is defective. How many components constructed during
Shift 1 were defective? How many components constructed during Shift 2 were
not defective?
b. Given that the component was defective, what is the likelihood that it was
constructed during Shift 2? Given that the component was defective, what is the
likelihood that it was constructed during Shift 3? Does there seem to be any
connection between when a component is constructed and whether or not it is
defective? Explain.
c. Construct a stacked column chart. Are the defect rates consistent across all
shifts? Explain.
25. FILE Athletic. A researcher at a marketing firm examines whether the age of a
consumer matters when buying athletic clothing. Her initial feeling is that Brand A
attracts a younger customer, whereas the more established companies (Brands B
and C) draw an older clientele. For 600 recent purchases of athletic clothing, she
collects data on a customer’s age (Age equals 1 if the customer is under 35, 0
otherwise) and the brand name of the athletic clothing (A, B, or C). A portion of the
data is shown in the accompanying table.

Purchase Age Brand

1 1 A

2 1 A

⋮ ⋮ ⋮

600 0 C

a. Construct a contingency table that cross-classifies the data by Age and Brand.
How many of the purchases were for Brand A? How many of the purchases were
from customers under 35 years old?
b. Given that the purchase was made by a customer under 35 years old, what is the
likelihood that the customer purchased Brand A? Brand B? Brand C? Do the data
seem to support the researcher’s belief? Explain.

page 104
c. Construct a stacked column chart. Does there appear to be a
relationship between the age of the customer and the brand purchased?
26. FILE Study. A report suggests that business majors spend the least amount of
time on course work compared to all other college students (The Washington Post,
January 28, 2017). A provost of a university conducts a survey on 270 students.
Students are asked their major (business or nonbusiness) and if they study hard
(yes or no), where study hard is defined as spending at least 20 hours per week on
course work. A portion of the responses is shown in the accompanying table.

Student Major Study Hard

1 business yes

2 business yes

⋮ ⋮ ⋮

270 nonbusiness no

a. Construct a contingency table that cross-classifies the data by Major and Study
Hard. How many of the students are business majors? How many of the students
study hard?
b. Given that the student is a business major, what is the likelihood that the student
studies hard? Given that the student is a nonbusiness major, what is the
likelihood that the student studies hard? Do the data seem to support the findings
in the report? Explain.
c. Construct a stacked column chart. Comment on the findings.
27. FILE Test_Scores. The accompanying table shows a portion of midterm and final
grades for 32 students. Construct a scatterplot of Final against Midterm. Describe
the relationship.

Final Midterm

86 78

94 97

⋮ ⋮

91 47

28. FILE Life_Obesity. The accompanying table shows a portion of life expectancies
(in years) and obesity rates (in %) for the 50 states and the District of Columbia.
Construct a scatterplot of Life Expectancy against Obesity. Describe the
relationship.
 State Life Expectancy Obesity

Alabama 75.4 36.3

Alaska 78.3 34.2

⋮ ⋮ ⋮

Wyoming 78.3 28.8

29. FILE Consumption. The accompanying table shows a portion of quarterly data for
average U.S. annual consumption (Consumption in $) and disposable income
(Income in $) for the years 2000–2016. Construct a scatterplot of Consumption
against Income. Describe the relationship.

Date Consumption Income

Q1, 2000 28634 31192

Q2, 2000 28837 31438

⋮ ⋮ ⋮

Q4, 2016 35987 39254

30. FILE Return. In order to diversify risk, investors are often encouraged to invest in
assets whose returns have either a negative relationship or no relationship. The
accompanying table shows a portion of the annual return data (in %) on two assets.
Construct a scatterplot of Return B against Return A. In order to diversify risk,
would the investor be wise to include both of these assets in her portfolio? Explain.

Return A Return B

−20 2

−5 0

⋮ ⋮

10 2

31. FILE Healthy_Living. Healthy living has always been an important goal for any
society. Most would agree that a diet that is rich in fruits and vegetables (FV) and
 regular exercise have a positive effect on health, while smoking has a negative
effect on health. The accompanying table shows a portion of the percentage of
these variables observed in various states in the United States.

a. Construct a scatterplot of Health against Exercise. Describe the relationship.

b. Construct a scatterplot of Health against Smoking. Describe the relationship.
32. FILE Car_Price. The accompanying table shows a portion of data consisting of
the price, the age, and the mileage for 20 used sedans.

Price Age Mileage

13590 6 61485

13775 6 54344

⋮ ⋮ ⋮

11988 8 42408

a. Construct a scatterplot of Price against Age. Describe the relationship.

b. Construct a scatterplot of Price against Mileage. Describe the relationship.

page 105

3.3 OTHER DATA VISUALIZATION

METHODS
LO 3.4
Construct and interpret a scatterplot with a categorical variable, a bubble plot, a line chart,
and a heat map.

In Section 3.2, we constructed a scatterplot that was used to visualize

the relationship between two numerical variables. Here, we extend
our discussion of a scatterplot by incorporating a categorical variable.
We also present three additional visualization methods: bubble plots,
line charts, and heat maps.

A Scatterplot with a Categorical Variable

Recall that we use a scatterplot to display the relationship between
two numerical variables. For example, if we plot property value
against square footage, then we anticipate a positive relationship
between these two variables; that is, the bigger the house, the more it
is worth. If we have a third variable in the data set, say the property
type (a single-family home, a condominium, etc.), we can incorporate
this categorical variable within the scatterplot by using different colors
or symbols. This allows us to see if the relationship between property
value and square footage differs across different property types. This
plot is referred to as a scatterplot with a categorical variable.
We illustrate the use of a scatterplot with a categorical variable in
Example 3.5.
FILE
Birth_Life

EXAMPLE 3.5
The Birth_Life data file contains information on the following
variables for 10 countries in 2010: country name (Country
Name), life expectancy (Life Exp in years), birth rate (Birth Rate
in percent), GNI per capita (GNI in $), and level of development
(Development). A portion of the Birth_Life data set is shown in
Table 3.13.
 TABLE 3.13 A Portion of the Birth_Life Data Set

Use Excel and R to construct a scatterplot of birth rate

against life expectancy that also incorporates the Development
variable (categorical). Summarize the results.
SOLUTION:
Using Excel

a. Open the Birth_Life data file.

b. In order to create a scatterplot that incorporates a categorical
variable, the categorical variable should be sorted by category.
Here, the Development variable is already sorted; that is, the
first six countries are developing countries and the remaining
four countries are developed countries. Select Insert > Insert
Scatter or Bubble Chart > Scatter. This creates a placeholder
for the scatterplot in the worksheet.
c. Select Design > Select Data. In the Select Data Source dialog
box, click the Add button. This opens the Edit Series dialog box
for you to select the data for the x-axis and the y-axis of the
scatterplot. Enter “Developing” as the Series name, select cells
B2 through B7 as the Series X values, and select cells C2
through C7 as the Series Y values. Click OK in the Edit Series
dialog box. This plots birth rate against life expectancy for the
six developing countries.
page 106
d. Click Add to open the Edit Series dialog box. Enter
“Developed” as the Series name, select cells B8 through B11
as the Series X values, and select cells C8 through C11 as the
 Series Y values. Click OK in the Edit Series dialog box. This
plots birth rate against life expectancy for the four developed
countries. Click OK in the Select Data Source dialog box.
e. Formatting (regarding axis titles, gridlines, etc.) can be done by
selecting Format > Add Chart Element from the menu. The
resulting scatterplot should be similar to Figure 3.15.

FIGURE 3.15 A scatterplot of birth rate against life expectancy

Using R

a. Import the Birth_Life data into a data frame (table) and label it
myData.
b. To create a scatterplot that incorporates the Development
variable, we use the plot function. Here we use the options
main to add a title; col to define colors for the segments of the
columns; xlab and ylab to provide labels for the x-axis and y-
axis, respectively; pch to define the shape of the marker; and
col to define the color of the marker. The shapes and colors of
the markers are based on the categories of the Development
variable. Enter:
 > plot(myData$'Birth Rate'~myData$'Life Exp',
main=“Scatterplot of Birth Rate against Life Expectancy”, xlab =
“Life Expectancy (in years)”, ylab = “Birth Rate (in %)”, pch=16,
col=ifelse(myData$Development == “Developing”, 20, 26))
c. We add a legend on the right side of the scatterplot using the
legend function. Enter:
> legend(“right”, legend=c(“Developing”, “Developed”), pch=16,
col=c(20, 26))
The resulting scatterplot should be similar to Figure 3.15.
Summary
From Figure 3.15, we see a negative linear relationship between
birth rate and life expectancy. That is, countries with lower birth
rates tend to have higher life expectancies. This relationship
holds true for both developing and developed countries. We also
see that, in general, developed countries have lower birth rates
and higher life expectancies as compared to developing
countries.

page 107

A Bubble Plot
A bubble plot shows the relationship between three numerical
variables. In a bubble plot, the third numerical variable is represented
by the size of the bubble. For instance, a bubble plot may plot a
college student’s study time against screen time and use the size of
the bubble to represent the student’s GPA. This bubble plot would
help us understand the relationships between study time, screen time,
and academic performance.
We illustrate the use of a bubble plot in Example 3.6.
FILE
Birth_Life
 EXAMPLE 3.6
Revisit the Birth_Life data from Example 3.5. Use Excel and R
to construct a bubble plot of birth rate against life expectancy
that uses the GNI variable for the size of the bubbles.
Summarize the results.
SOLUTION:
Using Excel

a. Open the Birth_Life data file.

b. Select Insert > Insert Scatter or Bubble Chart > Bubble. This
creates a placeholder for the plot in the worksheet.
c. Select Design > Select Data. In the Select Data Source dialog
box, click the Add button. This opens the Edit Series dialog box
for you to select the variables for the x-axis, the y-axis, and the
bubble size. Select cells B2 through B11 as the Series X
values, select cells C2 through C11 as the Series Y values, and
select cells D2 through D11 as the Series bubble size. Click OK
in the Edit Series dialog box. Click OK in the Select Data
Source dialog box.
d. Formatting (regarding axis titles, gridlines, etc.) can be done by
selecting Format > Add Chart Element from the menu. The
resulting bubble plot should be similar to Figure 3.16.

FIGURE 3.16 A bubble plot of birth rate, life expectancy, and

GNI
 Using R

a. Import the Birth_Life data into a data frame (table) and label it
myData.
b. We first create an empty plot by using the plot function (with
the appropriate x and y variables) and the option type = “n”,
which means no plotting. Enter:
> plot(myData$‘Birth Rate’~myData$‘Life Exp’, type=“n”)
page 108
c. We then use the symbols function to plot the
bubbles representing the observations. We use the options
circles, inches, and bg to specify the radii, sizes, and color of
the bubbles, respectively. The bubbles are sized based on the
values of the GNI of the countries. As in Example 3.5, we also
use the options main, xlab, and ylab. Enter:
> symbols(myData$‘Birth Rate’~myData$‘Life Exp’,
circles=myData$GNI, inches = 0.5, bg = ‘blue’, main=“A bubble
plot of birth rate, life expectancy, and GNI”, xlab = “Life
Expectancy (in years)”, ylab = “Birth Rate (in %)”)
The resulting bubble plot should be similar to Figure 3.16.
 Summary
From Figure 3.16 we see that a country’s birth rate and its
average life expectancy display a negative linear relationship.
We also see that countries with low birth rates and high life
expectancies have higher GNI per capita, which is indicative of
developed countries.

A Line Chart
A line chart displays a numerical variable as a series of data points
connected by a line. A line chart is especially useful for tracking
changes or trends over time. For example, using a line chart that plots
the sales of Samsung’s smartphones over time, we can easily tell
whether the sales follow an upward, a downward, or a steady trend. It
is also easy for us to identify any major changes that happened in the
past on a line chart. For example, due to a major product recall,
Samsung experienced a significant drop in sales in late 2016. We can
easily identify the event from the line chart because the data point for
that year would dip dramatically.
When multiple lines are plotted in the same chart, we can compare
these observations on one or more dimensions. For example, if we
simultaneously plot the historical sales of Apple’s iPhones alongside
those of Samsung’s smartphones, we would be able to compare the
trends and the rates of change of the two companies. We may even
detect interesting patterns such as whether a drop in the sales of
Samsung’s smartphones coincides with a surge in the sales of
iPhones. In order to illustrate the use of line charts, consider Example
3.7.
FILE
Growth_Value

EXAMPLE 3.7
 Recall the introductory case where data are provided on the
annual returns (in %) for Fidelity’s Growth and Value mutual
funds from 1984 to 2018. Use Excel and R to construct line
charts for Growth and Value. Summarize the results.
SOLUTION:
Using Excel

a. Open the Growth_Value data file.

b. Highlight cells B1 through C36 and then, from the menu, select
Insert > Insert Line or Area Chart > Line.
c. Select Design > Select Data and click the Edit button under
Horizontal (Category) Axis Labels. In the Axis Labels dialog
box, highlight cells A2 through A34 in the Axis label range box.
Click OK. The years appear in the line plot. Due to space
limitations, Excel only displays the even number years along
the horizontal axis.
d. Formatting (regarding axis titles, gridlines, etc.) can be done by
selecting Format > Add Chart Element from the menu. The
resulting line chart should be similar to Figure 3.17.
page 109
FIGURE 3.17 A line chart for the Growth and Value mutual
funds
 Using R

a. Import the Growth_Value data into a data frame (table) and

label it myData.
b. We first use the plot function to create a line chart for the
Growth variable. Here we use the options; col to define the
color of the marker; type to define the type of plot where “l”
means line; and ylim to provide limits on the y-axis. As in
Example 3.5, we also use the options main, xlab, and ylab.
Enter:
> plot(myData$Growth~myData$Year, main=“A line chart for
the Growth and Value mutual funds”, xlab=“Year”, ylab=“Annual
Returns”, col = “blue”, type = “l”, ylim = c(-100, 100) )
c. We then add a red line for the Value mutual fund using the
lines function. Enter:
> lines(myData$Value~myData$Year, col=“red”, type = “l”)
d. We add a legend at the bottom of the chart using the legend
function. The lty =1 option specifies that the legend is based on
the two solid lines. Enter:
> legend(“bottom”, legend=c(“Growth”, “Value”), col=c(“blue”,
“red”), lty=1)
 The resulting line chart should be similar to Figure 3.17.
Summary
The line chart in Figure 3.17 compares the annual returns of
Fidelity’s Growth and Value mutual funds over a period of 35
years. In general, both funds tend to move in the same direction.
However, during the dot-com bubble that occurred roughly
between 1995 and 2000, the technology-heavy Growth fund
shows dramatically higher returns as compared to the Value
fund, which tends to favor traditional companies. The
subsequent burst of the dot-com bubble shows a reversal of that
trend.

A Heat Map
A heat map is an important visualization tool that uses color or color
intensity to display relationships between variables. Heat maps are
especially useful to identify combinations of the categorical variables
that have economic significance. There are a number of ways to
display a heat map, but they all share one thing in common—they use
color to communicate the relationships between the variables that
would be harder to understand by simply inspecting the raw data. For
example, we can use a heat map to show which products page 110
are the best-or worst-selling products at various stores or
show the most-or least-frequently downloaded music genres across
various music streaming platforms. Example 3.8 illustrates the use of
a heat map.
FILE
Bookstores

EXAMPLE 3.8
A national bookstore chain is trying to understand customer
preferences at various store locations. The marketing
 department has acquired a list of 500 of the most recent
transactions from four of its stores. The data set includes the
record number (Record), which one of its four stores sold the
book (BookStore), and the type of book sold (BookType). The
marketing department wants to visualize the data using a heat
map to help it understand customer preferences at different
stores. A portion of the Bookstores data set is shown in Table
3.14.

TABLE 3.14 Transaction Results for Bookstore Example

Record BookStore BookType

1 Store2 Biography

2 Store2 Children book

⋮ ⋮ ⋮

500 Store4 Romance

Use Excel and R to construct a heat map to visualize the sales

of various types of books at different stores. Summarize the
results.
SOLUTION:
Using Excel

a. Open the Bookstores data file.

b. Before we can create a heat map, we need to first construct a
contingency table that summarizes each type of book sold as a
percentage of all books sold at each store. Click anywhere on
the data (we choose cell A5). From the menu, select Insert >
Pivot Table. The option in Select a table or range should
 already be populated with the data range A1:C501. Place the
PivotTable report in the existing worksheet beginning in cell E1.
c. In PivotTable Fields, drag the BookStore and BookType
variables to the Rows and Columns fields, respectively. Drag
Record to the Values field. Click the arrow below Values and
select Value Field Settings. In order to read the Record variable
as a count, click the arrow below the Values tab and select
Value Field Settings. In the dialog box, select the Summarize
value field by tab, and then in the drop-down menu select
Count. In the Show values as tab, we choose % of Row down
menu. Click OK. You should see a contingency table starting in
cell E1.
d. Select the data range F3:K6. From the menu, select Home >
Conditional Formatting > Color Scales > Green − Yellow −
Red Color Scale.
The resulting heat map should be similar to Figure 3.18.
Excel’s color scale works in descending order of magnitude;
that is, for the green-yellow-red color scale, green denotes the
larger values, yellow denotes the median values, and red
denotes the smaller values. We advise you to experiment with
different color scales in Excel to see how the heat map
changes.

FIGURE 3.18 Excel’s heat map for the bookstore example

page 111

Using R

a. Import the Bookstores data into a data frame (table) and label
it myData.
b. We first use the table function to create a contingency table,
labeled myTable, that summarizes the number of each type of
book sold at each store. We then use the rowSums function to
retrieve the total number of books sold at each store and divide
the values in the contingency table by those numbers to get a
modified contingency table that summarizes each type of book
sold as a percentage of all books sold at each store. Enter:
> myTable <-table(myData$BookStore, myData$BookType)
> myTable <- myTable/rowSums(myTable)
c. In order to construct a heat map in R, the data must be
converted into a data matrix, which is a two-dimensional data
structure whose columns must have the same data type and
length. We use the as.matrix function to make this conversion
and label the data matrix as myData.matrix. Enter:
> myData.matrix <- as.matrix(myTable)
d. We use the heatmap function to construct the heat map. We
choose the heat.color pallet with 256 colors using col =
heat.colors(256). By default, the color pallet uses warm colors
(e.g., orange and red) to show smaller values and cool colors
(e.g., yellow and white) to show larger values. We use the scale
= “none” option to ensure that the data are not standardized
when used to plot the heat map. Finally, we use Rowv = NA
and Colv = NA to suppress a tree-like diagram called a
dendrogram. A dendrogram will be discussed later in Chapter
11. Enter:
> heatmap(myData.matrix, col = heat.colors(256), scale =
“none”, Rowv= NA, Colv = NA)
The resulting heat map is shown in Figure 3.19(a).

FIGURE 3.19 R’s heat maps for the bookstore example

e. Alternatively, we can specify our own colors so that less-
popular book genres have cooler colors and the more-popular
genres have warmer colors. Because the most popular book
genre accounts for 25.22% of the book sales at the store, we
create a vector, myBreaks, that contains numbers from 0 (0%)
to 0.30 (30%) with increments of 0.05 (5%). The color assigned
to each number interval is stored in the vector myCol. For
example, if the book genre accounts for 0% to 5% of the book
sales at the store, the color blue will be assigned to that genre;
alternatively, if the book genre accounts for 25% to 30% of the
book sales at the store, a warm red color, coded “red3” will be
assigned. We then assign our own color key using col and
breaks for the new heat map. Enter:
> myBreaks <- c(0, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30)
> myCol <- c(“blue”, “green”, “yellow”, “orange1”, “red1”, “red3”)
> heatmap(myData.matrix, col = myCol, breaks = myBreaks,
scale = “none”, Rowv = NA, Colv = NA)
The resulting heat map is shown in Figure 3.19(b); the colors
may appear different from what you see on your computer
screen.
 page 112

Summary
The heat maps in Figures 3.18 and 3.19 reveal that customers’
book preferences do differ across different store locations. For
example, romance fictions are the most popular books sold at
Store2 but the least popular at Store3. Self-help books are the
least popular books at Store1. The management can use this
information to make decisions about how many copies of each
type of book to stock at each store.

ADDITIONAL GRAPHICAL DISPLAYS

The following graphical displays are additional ways to visualize
relationships between variables that would be harder to
understand by simply inspecting the raw data.
A scatterplot with a categorical variable shows the relationship
between two numerical variables and a categorical variable in
a two-dimensional graph. The categorical variable is typically
represented by different colors for each category.
A bubble plot shows the relationship between three numerical
variables in a two-dimensional graph. The third numerical
variable is represented by the size of the bubble.
A line chart shows a numerical variable as a series of data
points connected by a line. It is especially useful for tracking
changes or trends over time.
A heat map uses color or color intensity to display
relationships between variables. Heat maps are especially
useful for identifying combinations of the categorical variables
that have economic significance.
Options for Advanced Visualizations
In addition to the data visualizations discussed in this chapter,
advanced software tools for constructing sophisticated, aesthetically
pleasing, and interactive visualizations are available. Two popular
advanced visualization tools include the ggplot2 package in R and
Tableau.

page 113

The ggplot2 package is one of the most popular graphic packages

in R for producing elegant, publication-quality graphics. The package
is created based on a set of concepts called the Grammar of
Graphics, which breaks up graphs into data, aesthetics (e.g., color
and size), and geometric (e.g., lines and shapes) components.
Tableau is a stand-alone data visualization software that has
gained tremendous popularity for its attractive output, versatility, and
ease of use. One of the key features of Tableau is its ability to extract
data from a wide range of sources from Excel files on a personal
computer to Oracle databases on a cloud, computing platform. The
product is designed to allow users with little technical experience to
develop attractive visualizations interactively and to drill down the
data for business insights. Users can easily switch between different
visualizations to select the most compelling graphs for their
storytelling. In addition, Tableau allows users to create interactive
dashboards, which can display real-time data, to support business
decision making.
Detailed discussion of the ggplot2 package in R and/or Tableau is
outside of the scope of this text.

EXERCISES 3.3
Applications
33. FILE InternetStocks. A financial analyst wants to compare the performance of the
stocks of two Internet companies, Amazon (AMZN) and Google (GOOG). She
records the average closing prices of the two stocks for the years 2010 through
 2016. A portion of the data is shown in the accompanying table. Construct a line
chart that shows the movements of the two stocks over time using two lines each
with a unique color. Describe the overall trend of price movement for the two
stocks. Which stock shows the greater trajectory of price appreciation?

Year AMZN GOOG

2010 180.00 593.97

2011 173.10 645.90

⋮ ⋮ ⋮

2016 749.86 771.82

34. FILE India_China. It is believed that India will overtake China to become the
world’s most populous nation much sooner than previously thought (CNN, June 19,
2019). The accompanying data file, compiled by the World Bank, contains the
population data, in millions, for India and China from 1960 to 2017. Construct a line
chart that shows the changes in the two countries’ populations over time using two
lines each with a unique color. Describe the overall trend of population growth in
the two countries. Which country shows the faster population growth during the
past 40 years?

Year India China

1960 449.48 667.07

1961 458.49 660.33

⋮ ⋮ ⋮

2017 1339.18 1386.40

35. FILE HighSchool_SAT. The accompanying table shows a portion of the average
SAT math score (Math), the average SAT writing score (Writing), the number of test
takers (Test Taker), and whether the school is a private or public school (Type) for
25 high schools in a major metropolitan area.
 a. Construct a bubble plot that shows the math score on the x-axis, the writing score
on the y-axis, and the number of test takers as the size of the bubble. Do math
score and writing score show a linear, nonlinear, or no relationship? If the
relationship is a linear relationship, is it a positive or negative relationship? Do
math score and the size of the school (using the number of test takers as a
proxy) show a linear, nonlinear, or no relationship?
b. Construct a scatterplot that shows the math score on the x-axis and the writing
score on the y-axis. Use different colors or symbols to show whether the high
school is a private or public school. Describe the relationships between math
score, writing score, and school type. Does the relationship between math score
and writing score hold true for both private and public schools?
36. FILE Car_Price. The accompanying table shows a portion of data consisting of
the selling price, the age, and the mileage for 20 used sedans.

Price Age Mileage

13590 6 61485

13775 6 54344

⋮ ⋮ ⋮

11988 8 42408

page 114
a. Construct a bubble plot that shows price on the x-axis, age on
the y-axis, and mileage as the sizes of the bubbles. Describe the relationships
between price, age, and mileage of these used sedans.
b. Convert Mileage into a categorical variable, Mileage_Category, by assigning all
cars with less than 50,000 miles to the “Low_Mileage” category and the rest to
the “High_Mileage” category. How many cars are in the “High_Mileage”
category?
c. Construct a scatterplot using Price, Age, and Mileage_Category. Use different
colors or symbols to show cars that belong to the different mileage categories.
Describe the relationships between price, age, and mileage of these used
 sedans. Does the relationship between price and age hold true for both mileage
categories?
37. FILE TShirts. A company that sells unisex t-shirts is interested in finding out the
color and size of its best-selling t-shirt. The accompanying data file contains the
size, color, and quantity of t-shirts that were ordered during the last 1,000
transactions. A portion of the data is shown in the accompanying table.

Transaction Quantity Size Color

1 1 XL Purple

2 3 M Blue

⋮ ⋮ ⋮ ⋮

1000 1 S Red

a. Construct a contingency table that shows the total quantity sold for each color
and size combination. How many size M red t-shirts were sold? How many size
XL purple t-shirts were sold?
b. Construct a heat map that displays colors or color intensity based on the total
quantity sold. Which two color and size combinations are the most popular ones?
Which two are the least popular ones?
38. FILE Crime_Analysis. The local police department is performing a crime analysis
to find out which crimes are most likely to occur at which locations. The
accompanying data file contains the types of crimes (CrimeType) that occurred at
various locations (Location) in the city over the past five years. A portion of the data
is shown in the accompanying table.

Record CrimeType Location

1 Narcotics Street

2 Assault Residence

⋮ ⋮ ⋮

10385 Burglary Residence

a. Construct a contingency table that shows the frequencies for CrimeType and
Location combinations. How many of the crimes were for burglary and happened
in a residence?
b. Construct a heat map that displays colors or color intensity based on the
frequencies. Which three crime type and location combinations are the most
 frequent ones?

3.4 SUMMARY MEASURES

LO 3.5
Calculate and interpret summary measures.

In addition to tables and graphs, we can also use numerical

descriptive measures to extract meaningful information from data.
These measures provide precise, objectively determined values that
are easy to calculate, interpret, and compare with one another.
We first calculate several measures of central location, which
attempt to find a typical or central value for the data. We then
examine measures of dispersion and measures of shape. Measures
of dispersion gauge the underlying variability of the data. Measures of
shape reveal whether or not the distribution is symmetric and whether
the tails of the distribution are more or less extreme than the normal
distribution. Finally, we discuss measures of association, which show
whether two numerical variables have a linear relationship.

Measures of Central Location

The term central location relates to the way numerical data tend to
cluster around some middle or central value. Measures of central
location attempt to find a typical or central value that describes the
data. Examples include finding a typical value that describes the
return on an investment, the number of defects in a production
process, the salary of a business graduate, the rental price in a
neighborhood, the number of customers at a local convenience store,
and so on. We discuss the three most widely used measures of
central location: the mean, the median, and the mode. We also
discuss a percentile, which is technically a measure of location
(though not necessarily central location); however, it is also used as a
measure of relative position because it is so easy to interpret.

page 115

The Mean
The arithmetic mean is the primary measure of central location.
Generally, we refer to the arithmetic mean as simply the mean or the
average. In order to calculate the mean of a variable, we simply add
up all the observations and divide by the number of observations. The
only thing that differs between a population mean and a sample mean
is the notation. The population mean is referred to as μ, where μ is the
Greek letter mu (pronounced as “mew”). For observations x1, x2, . . .,

xN, the population mean is calculated as , where N is the number of

observations in the population. The sample mean is referred to as

(pronounced x-bar). For observations x1, x2, . . ., xn, the sample mean

is calculated as , where n is the number of observations in the

sample. We refer to the population mean as a parameter and the

sample mean as a statistic. Because the population mean is
generally unknown, we often use the sample mean to estimate the
population mean.

The Median
The mean is used extensively in data analysis. However, it can give a
misleading description of the center of the distribution in the presence
of extremely small or large observations, also referred to as outliers.
Because the mean can be affected by outliers, we often also calculate
the median as a measure of central location. The median is the
middle value of a data set; that is, an equal number of observations lie
above and below the median. After arranging the data in ascending
order (smallest to largest), we calculate the median as (1) the middle
value if the number of observations is odd or (2) the average of the
two middle values if the number of observations is even.
Many government publications and other data sources publish
both the mean and the median in order to accurately portray a
variable’s typical value. If the mean and the median differ significantly,
then it is likely that the variable contains outliers. For instance, in 2017
the U.S. Census Bureau determined that the median income for
American households was $61,372; however, the mean income was
$86,220. It is well documented that a small number of households in
the United States have income that is considerably higher than the
typical American household income. As a result, these top-earning
households influence the mean by pushing its value significantly
above the value of the median.

The Mode
The mode of a variable is the observation that occurs most frequently.
A variable can have more than one mode, or even no mode. If a
variable has one mode, then we say it is unimodal. If it has two
modes, then it is common to call it bimodal. If two or more modes
exist, then the variable is multimodal. Generally, the mode’s
usefulness as a measure of central location tends to diminish for a
variable with more than three modes. If we want to summarize a
categorical variable, then the mode is the only meaningful measure of
central location.

MEASURES OF CENTRAL LOCATION

The mean is the most commonly used measure of central
location. The population mean is denoted as μ, and the
sample mean is denoted as . One weakness of the mean is
that it is unduly influenced by outliers.
The median is the middle observation of a variable; that is, it
divides the variable in half. The median is especially useful
when outliers are present.
 The mode is the most frequently occurring observation of a
variable. A variable may have no mode or more than one
mode. The mode is the only meaningful measure of central
location for a categorical variable.

page 116
FILE
Growth_Value

EXAMPLE 3.9
Using Excel and R, calculate the mean and the median for the
Growth and the Value variables from the introductory case.
Summarize the results.
SOLUTION:
Using Excel
I. Excel’s Formula Option Excel provides built-in formulas for
virtually every summary measure that we may need. To
illustrate, we follow these steps to calculate the mean and the
median for the Growth variable.
a. Open the Growth_Value data file.
b. Enter =AVERAGE(B2:B36). Verify that the output is 15.1074.
c. To calculate the median, enter =MEDIAN(B2:B36). Verify that
the output is 14.44.
If we want to calculate the mean return for the Value
variable, and because the data occupy cells C2 through C36 on
the spreadsheet, we enter =AVERAGE(C2:C36). Because the
Growth and Value variables each contains 35 unique
observations and no duplicates, it is not practical to compute the
mode. For other variables, we enter =MODE(array) to calculate
the mode, where the notation array specifies the range of cells
to be included in the calculation. When introducing new
functions later in this chapter and other chapters, we will follow
this format. The first and second columns of Table 3.15 show
various descriptive measures and page 117
corresponding function names in Excel. We will
refer back to Table 3.15 on a few occasions in this chapter.

TABLE 3.15 Descriptive Measures and Corresponding Function Names in Excel

and R

Descriptive
Excel R
Measure

Location

Mean =AVERAGE(array) mean(df$var)a

Median =MEDIAN(array) median(df$var)

Mode =MODE(array) NAb

Minimum =MIN(array) min(df$var)

Maximum =MAX(array) max(df$var)

Percentile =PERCENTILE.INC(array, p)c quantile(df$var,

p)c

Multiple NA summary(df)
measures

Dispersion
 Descriptive
Excel R
Measure

Range =MAX(array)-MIN(array) range(df$var)d

Mean =AVEDEV(array) mad(df$var)e

Absolute
Deviation

Sample =VAR.S(array) var(df$var)

Variance

Sample =STDEV.S(array) sd(df$var)

Standard
Deviation

Shape

Skewness =SKEW(array) NA

Kurtosis =KURT(array) NA

Association

Sample =COVARIANCE.S(array1,array2) cov(df)

Covariance

Correlation =CORREL(array1,array2) cor(df)

a The notation df refers to the data frame or file and the notation var refers to the
variable name. The variable name should be specified in single quotations if it
consists of more than one word or if it is a number.
 b NA denotes that a simple function is not readily available.

c The parameter p takes on a value between 0 and 1.

d The range function in R returns the minimum and maximum values, so the range
can be calculated by taking the difference between the two values.
e The mad function calculates the median absolute deviation, rather than the mean
absolute deviation. Alternatively, we can install the ‘lsr’ package in R and use the aad
function, which computes the mean absolute deviation.

II. Excel’s Data Analysis Toolpak Option Another way to obtain

summary measures is to use Excel’s Data Analysis Toolpak
option. One advantage of this option is that it provides numerous
summary measures using a single command. Again, we
illustrate this option using the data from the introductory case.
a. Open the Growth_Value data file.
b. From the menu, choose Data > Data Analysis > Descriptive
Statistics > OK. (Note: As mentioned in Section 3.1, if you do
not see Data Analysis under Data, you must add in the
Analysis Toolpak option.)
c. See Figure 3.20. In the Descriptive Statistics dialog box, click
on the box next to Input Range, then select the data and the
headings for the Growth and Value variables. Select the options
Labels in First Row and Summary Statistics. Select Output
Range and enter cell E1. Then click OK.

FIGURE 3.20 Excel’s Descriptive Statistics dialog box

d. Table 3.16 presents the Excel output. If the output is difficult to
read, highlight the output and choose Home > Format >
Column > Autofit Selection. As noted earlier, Excel provides
numerous summary measures; we have put the measures of
central location in boldface. Note that Excel reports the mode
as #N/A, which means no value appears more than once.

TABLE 3.16 Excel’s Output Using the Data Analysis Toolpak

Growth Value

Mean 15.10742857 Mean 11.44457143

Standard 4.026783239 Standard 3.028998513

Error Error
 Growth Value

Median 14.44 Median 15.09

Mode #N/A Mode #N/A

Standard 23.82277091 Standard 17.91979686

Deviation Deviation

Sample 567.5244138 Sample 321.1191197

Variance Variance

Kurtosis 1.064454138 Kurtosis 1.893804308

Skewness 0.02988603 Skewness −1.007368283

Range 120.38 Range 90.6

Minimum −40.9 Minimum −46.52

Maximum 79.48 Maximum 44.08

Sum 528.76 Sum 400.56

Count 35 Count 35

Using R
Like Excel, R has many built-in formulas or functions. In R, we
denote all function names in boldface and all options within a
function in italics. The first and third columns of Table 3.15 show
various descriptive measures and corresponding function names
in R.
a. Import the Growth_Value data into a data frame (table) and
label it myData.
b. The mean function will return the mean for a specified variable
in a data frame. In order to find the mean for the Growth
variable, enter:
> mean(myData$Growth)
And R returns: 15.10743.
page 118
c. The summary function will return the minimum,
first quartile, median, mean, third quartile, and maximum values
for each variable in a data frame. Enter:
> summary(myData)
Table 3.17 shows the R output using the summary function.
TABLE 3.17 R Output Using the summary Function

Year* Growth Value

Min.   :1984 Min.  :−40.90 Min.   :−46.520

1st Qu. :1992 1st Qu.  : 2.13 1st Qu.  : 1.195

Median :2001 Median : 14.44 Median : 15.090

Mean  :2001 Mean  : 15.11 Mean  : 11.445

3rd Qu. :2010 3rd Qu. : 32.00 3rd Qu. : 22.185

Max.  :2018 Max.   : 79.48 Max.  : 44.080

*Note that in this example, the summary statistics for the

variable Year are not useful.
Summary
 From Tables 3.16 and 3.17 we see that the average return for
the Growth mutual fund is greater than the average return for
the Value mutual fund, 15.11% > 11.445%. Interestingly,
however, the median return for the Value mutual fund is greater
than the median return for the Growth fund, 15.09% > 14.44%.
This example illustrates why it is useful to examine both the
mean and the median when summarizing central location,
especially when outliers may be present. Neither variable has a
mode.

A Percentile
Recall that the median is the middle observation of a variable; that is,
half of the observations fall below this observation and half fall above
it. The median is also called the 50th percentile. In many instances,
we are interested in a percentile other than the 50th percentile. In
general, the pth percentile divides a variable into two parts:
Approximately p percent of the observations are less than the pth
percentile.
Approximately (100 − p) percent of the observations are greater
than the pth percentile.

page 119

A percentile is technically a measure of location; however, it is also

used as a measure of relative position because it is so easy to
interpret. For example, suppose you obtained a raw score of 650 on
the math portion of the SAT, where scores range from 200 to 800. It
may not be readily apparent how you did relative to other students
that took the same test. However, if you know that the raw score
corresponds to the 75th percentile, then you know that approximately
75% of students had scores lower than your score and approximately
25% of students had scores higher than your score.
 Earlier, we found that the median or the 50th percentile for the
Growth mutual fund was 14.44%. When we calculate the 25th, the
50th, and the 75th percentiles for a variable, we have effectively
divided the data into four equal parts, or quarters. Thus, the 25th
percentile is also referred to as the first quartile (Q1), the 50th
percentile is referred to as the second quartile (Q2), and the 75th
percentile is referred to as the third quartile (Q3).
Software packages, like Excel and R, use different algorithms to
calculate percentiles. However, with larger sample sizes, the
differences, if any, tend to be negligible. When calculating the pth
percentile in Excel, we can use the PERCENTILE.INC function. The
corresponding command in R is the quantile function. These
functions are summarized in Table 3.15. However, when we use the
summary function in R, all quartiles are reported.

A PERCENTILE
In general, the pth percentile divides a variable into two parts:
Approximately p percent of the observations are less than the
pth percentile.
Approximately (100 − p) percent of the observations are
greater than the pth percentile.

EXAMPLE 3.10
Using Table 3.17, interpret the first and the third quartiles for the
Growth variable from the introductory case.
SOLUTION:
The first quartile for the Growth variable is 2.13%. Approximately
25% of the returns are less than 2.13%, and approximately 75%
of the returns are greater than 2.13%. The third quartile for the
Growth variable is 32.00%. Approximately 75% of the returns
 are less than 32.00%, and approximately 25% of the returns are
greater than 32.00%.

Measures of Dispersion
While measures of central location reflect the typical or central value
of a variable, they fail to describe the underlying dispersion of the
variable. We now discuss several measures of dispersion that gauge
the variability of a data set. Each measure is a numerical value that
equals zero if all observations are identical and increases as the
observations become more diverse.

The Range
The range is the simplest measure of dispersion; it is the difference
between the maximum and the minimum observations of a variable.
The range is not considered a good measure of dispersion because it
focuses solely on the extreme observations and ignores every other
observation of a variable.

page 120

The Interquartile Range

The interquartile range (IQR) is the difference between the third
quartile and the first quartile, or, equivalently, IQR = Q3 − Q1. We can
think of IQR as the range of the middle 50% of the observations of the
variable. Even though IQR does not depend on the extreme
observations, it does not incorporate all the observations.

The Mean Absolute Deviation

As a starting point, a good measure of dispersion should consider
differences of all observations from the mean (or the median when
outliers are present). However, if we simply average all differences
from the mean, the positives and the negatives will cancel out, even
though they both contribute to dispersion, and the resulting average
will equal zero. The mean absolute deviation (MAD) is an average
of the absolute differences between the observations and the mean.
For sample observations x1, x2, . . ., xn, the sample MAD is calculated

as . If we have access to population data, then we simply

substitute the population mean μ for the sample mean and the
population size N for the sample size n.

The Variance and the Standard Deviation

The variance and the standard deviation are the two most widely
used measures of dispersion. Instead of calculating the average of
the absolute differences from the mean, as in MAD, we calculate the
average of the squared differences from the mean. The squaring of
differences from the mean emphasizes larger differences more than
smaller ones; MAD weighs large and small differences equally.
The formula for the variance differs depending on whether we
have a sample or a population. The sample variance, denoted as s2,

is calculated as ; whereas the population variance, denoted as

the Greek letter sigma squared σ2, is calculated as . Also,

whatever the units of the variable, the variance has squared units. In
order to return to the original units of measurement, we take the
positive square root of s2 or σ2, which gives us either the sample
standard deviation s or the population standard deviation σ.
Table 3.15 shows the function names for various measures of
dispersion in Excel and R. Recall too that Excel’s Descriptive
Statistics option using the Data Analysis Toolpak provides many
summary measures using a single command. For measures of
dispersion, Excel’s Descriptive Statistics option treats the data as a
sample and calculates the sample variance and the sample standard
deviation. These values for the Growth and the Value mutual funds
are shown in Table 3.16.

MEASURES OF DISPERSION
Measures of dispersion can be summarized as follows.
The range is the difference between the maximum and the
minimum observations. The main weakness of the range is
that is ignores all observations except the extremes.
The interquartile range (IQR) is the difference between the
third quartile and the first quartile. The measure does not rely
on the extreme observations; however, it does not incorporate
all observations.
The mean absolute deviation (MAD) is an average of the
absolute differences between the observations and the mean.
The variance is an average of the squared differences
between the observations and the mean. The standard
deviation is the positive square root of the variance.

page 121

An Application Using the Mean and the Standard

Deviation—The Sharpe Ratio
In general, investments with higher returns also carry higher risk.
Investments include financial assets such as stocks, bonds, and
mutual funds. The average return represents an investor’s reward,
whereas variance, or, equivalently, standard deviation, corresponds to
risk.
Nobel Laureate William Sharpe developed what he originally
called the “reward-to-variability” ratio. However, academics and
finance professionals prefer to call it the Sharpe ratio. The Sharpe
ratio is often calculated as , where is the mean return for

investment i, Rf is the mean return for a risk-free asset such as a

Treasury bill (T-bill), and si is the standard deviation for investment i.
The numerator of the Sharpe ratio measures the extra reward that
investors receive for the added risk taken—this difference is often
called excess return. The higher the Sharpe ratio, the better the
investment compensates its investors for risk.

EXAMPLE 3.11
Use the information in Table 3.18 to
a. Compare the risk of investing in Growth versus Value using the
standard deviation.
b. Calculate and interpret the Sharpe ratios for Growth and Value.
Assume that the return on a 1-year T-bill is 2%.
SOLUTION:

a. Because the standard deviation of Growth is greater than the

standard deviation of Value, 23.8228 > 17.9198, Growth is
considered riskier than Value.
b. Given that the return on a 1-year T-bill is 2%, Rf = 2. Plugging
in the values of the relevant means and standard deviations
into the Sharpe ratio yields:

We had earlier shown that the Growth mutual fund had a higher
return, which is good, along with a higher standard deviation,
 which is bad. We can use the Sharpe ratio to make a valid
comparison between the funds. The Growth mutual fund
provides a higher Sharpe ratio than the Value mutual fund
(0.5502 > 0.5270); therefore, the Growth mutual fund offered
more reward per unit of risk.

Measures of Shape
In this section, we examine measures of shape, namely, the
skewness coefficient and the kurtosis coefficient.

The Skewness Coefficient

Recall from Section 3.1 that a symmetric distribution is one that is a
mirror image of itself on both sides of its center. The skewness
coefficient measures the degree to which a distribution is not
symmetric about its mean. A common way to calculate the skewness

coefficient is When we use Excel’s Descriptive

Statistics option in its Analysis Toolpak, this is the formula that is

applied. R also calculates the skewness coefficient, but it requires the
use of a package and the output differs slightly depending on the
package that is used.

page 122

A symmetric distribution has a skewness coefficient of zero. Thus,

the normal distribution has a skewness coefficient of zero. A positively
skewed distribution has a positive skewness coefficient, whereas a
negatively skewed distribution has a negative skewness coefficient. If
two return distributions are the same except for skewness (for
example, they have the same mean and standard deviation),
investors would prefer the distribution with positive skewness because
that distribution implies a greater probability of extremely large gains.
The Kurtosis Coefficient
The kurtosis coefficient is a summary measure that tells us whether
the tails of the distribution are more or less extreme than the normal
distribution. A distribution that has tails that are more extreme than
the normal distribution is leptokurtic (lepto from the Greek word for
slender). A return distribution is often leptokurtic, which means that its
tails are longer than the normal distribution—implying the existence of
outliers. If a return distribution is in fact leptokurtic, but we assume
that it is normally distributed in statistical models, then we will
underestimate the likelihood of very bad or very good returns. A
platykurtic (platy from the Greek word for broad) distribution is one
that has shorter tails, or tails that are less extreme, than the normal
distribution.
The kurtosis coefficient of a normal distribution is 3.0. A common

way to calculate the kurtosis coefficient is Many

statistical packages report excess kurtosis, which is calculated as

the kurtosis coefficient minus 3. Thus, a normal distribution has
excess kurtosis equal to zero. Excel calculates excess kurtosis when
we use the Descriptive Statistics option in its Analysis Toolpak. R also
calculates the kurtosis coefficient, but again, it requires the use of a
package and the output varies depending on the package that is
used.

MEASURES OF SHAPE
Measures of shape can be summarized as follows.
The skewness coefficient measures the degree to which a
distribution is not symmetric about its mean. A symmetric
distribution has a skewness coefficient of 0. A positively
(negatively) skewed distribution has a positive (negative)
skewness coefficient.
The kurtosis coefficient measures whether the tails of a
distribution are more or less extreme than the normal
 distribution. Because the normal distribution has a kurtosis
coefficient of 3, it is common to calculate the excess kurtosis of
a distribution as the kurtosis coefficient minus 3.

EXAMPLE 3.12
Interpret the skewness and the kurtosis coefficients in Table
3.16 for the Growth and Value variables from the introductory
case.

SOLUTION: The skewness coefficient and the (excess) kurtosis

coefficient for Growth are 0.0299 and 1.06, respectively. These
values imply that the return distribution for Growth is positively
skewed, and the distribution has longer tails than the normal
distribution. With a skewness coefficient of −1.0074 and a
(excess) kurtosis coefficient of 1.8938, the return distribution for
Value is negatively skewed, and it too has longer tails than the
normal distribution.

page 123

Measures of Association
In Section 3.2 we used a scatterplot to visually assess whether two
numerical variables had some type of systematic relationship. Here,
we present two numerical measures of association that quantify the
direction and strength of the linear relationship between two variables,
x and y. It is important to point out that these measures are not
appropriate when the underlying relationship between the variables is
nonlinear.

The Covariance
An objective numerical measure that reveals the direction of the linear
relationship between two variables is called the covariance. Like
variance, the formula for the covariance depends on whether we have
a sample or a population. The sample covariance, denoted as sxy is

calculated as . The population covariance, denoted as σxy is

calculated as .

The value of the covariance can be negative, positive, or 0.

If the covariance is negative, then x and y have a negative linear
relationship.
If the covariance is positive, then x and y have a positive linear
relationship.
If the covariance is zero, then x and y have no linear relationship.
The covariance is difficult to interpret because it is sensitive to the
units of measurement. That is, the covariance between two variables
might be 100 and the covariance between two other variables might
be 100,000, yet all we can conclude is that both sets of variables have
a positive linear relationship. We cannot comment on the strength of
the relationships.

The Correlation Coefficient

An easier measure to interpret is the correlation coefficient. It
describes both the direction and the strength of the linear relationship
between x and y. The sample correlation coefficient, denoted as rxy, is

calculated as , where sxy denotes the sample covariance and sx

and sy denote the sample standard deviations of the variables x and

y, respectively. The population correlation coefficient, denoted as ρxy

(the Greek letter rho), is calculated as , where σxy denotes the

population covariance and σx and σy denote the population standard
deviations. The correlation coefficient is unit-free because the units in
the numerator cancel with those in the denominator. The value of the
correlation coefficient falls between −1 and 1.
If the correlation coefficient equals −1, then x and y have a perfect
negative linear relationship.
If the correlation coefficient equals 0, then x and y are not linearly
related.
If the correlation coefficient equals 1, then x and y have a perfect
positive linear relationship.
Other values for the correlation coefficient must be interpreted with
reference to −1, 0, or 1. For instance, a correlation coefficient equal to
−0.80 indicates a strong negative relationship, whereas a correlation
coefficient equal to 0.12 indicates a weak positive relationship. Table
3.15 shows the function names for the covariance and the correlation
coefficient in Excel and R.

MEASURES OF ASSOCIATION
Measures of association can be summarized as follows.
The covariance between two variables x and y indicates
whether they have a negative linear relationship, a positive
linear relationship, or no linear relationship.
The correlation coefficient between two variables x and y
indicates the direction and the strength of the linear
relationship.

page 124
FILE
Growth_Value
 EXAMPLE 3.13
Using Excel and R, calculate the correlation coefficient between
the Growth and the Value variables from the introductory case.
Then summarize the results.
SOLUTION:
Using Excel

a. Open the Growth_Value data file.

b. Note that the data for the Growth variable are in cells B2
through B36 (array1) and the data for the Value variable are in
cells C2 through C36 (array2). We enter =CORREL(B2:B36,
C2:C36), and Excel returns 0.6572.
Using R

a. Import the Growth_Value data into a data frame (table) and

label it myData.
b. The cor function will return a matrix that lists the correlation
coefficient for each pairing of variables in the data frame. Enter:
> cor(myData)
And R returns:

Year Growth Value

Year 1.00000000 -0.08071914 -0.07909256
Growth -0.08071914 1.00000000 0.65719566
Value -0.07909256 0.65719566 1.00000000

We are interested in the correlation coefficient between Growth

and Value, which appears twice in this matrix (see boldface
values). We also see the value 1 along the diagonal of the
matrix, which measures the correlation between each variable
and itself. The correlation coefficient between Year and Growth
 and the correlation coefficient between Year and Value are
meaningless in this application.
Summary
The correlation coefficient between Growth and Value of 0.6572
indicates that the variables have a moderate positive linear
relationship. In other words, on average, when one fund’s return
is above its mean, the other fund’s return tends to be above its
mean, and vice versa. In order to diversify risk, an investor is
often advised to invest in assets (such as stocks, bonds, and
mutual funds) whose returns are not strongly correlated. If asset
returns do not have a strong positive correlation, then if one
investment does poorly, the other may still do well.

SYNOPSIS OF INTRODUCTORY CASE

Growth investing and value investing are two fundamental styles in stock and mutual
fund investing. Proponents of growth investing believe that companies that are
growing faster than their peers are trendsetters and will be able to maintain their
superior growth. By investing in the stocks of these companies, they expect their
investment to grow at a rate faster than the overall stock market. By comparison,
value investors focus on the stocks of companies that are trading at a discount
relative to the overall market or a specific sector. Investors of value stocks believe that
these stocks are undervalued and that their price will increase once their true value is
recognized by other investors. The debate between growth and
value investing is age-old, and which style dominates depends on
page 125
the sample period used for the analysis.
 Ingram Publishing/Getty Images

An analysis of annual return data for Fidelity’s Growth and Value mutual funds for
the years 1984 through 2018 provides important information for an investor trying to
determine whether to invest in a growth mutual fund, a value mutual fund, or both
types of mutual funds. Over this period, the mean return for the Growth fund of
15.11% is greater than the mean return for the Value fund of 11.44%. While the mean
return typically represents the reward of investing, it does not incorporate the risk of
investing. Standard deviation tends to be the most common measure of risk with
financial data. Because the standard deviation for the Growth fund (23.82%) is greater
than the standard deviation for the Value fund (17.92%), the Growth fund is likelier to
have returns farther above and below its mean. Finally, given a risk-free rate of 2%,
the Sharpe ratio for the Growth fund is 0.5502, compared to that for the Value fund of
0.5270, indicating that the Growth fund provides more reward per unit of risk.
Assuming that the behavior of these returns will continue, the investor will favor
investing in Growth over Value. A commonly used disclaimer, however, states that
past performance is no guarantee of future results. Because the two styles often
complement each other, it might be advisable for the investor to add diversity to his
portfolio by using them together.

EXERCISES 3.4
Applications
39. FILE Corporations. Monthly stock prices (in $) for Corporation A and Corporation
B are collected for five years. A portion of the data is shown in the accompanying
table.

Date A B
 Date A B

1/1/2014 47.95 65.59

2/1/2014 52.95 65.61

⋮ ⋮ ⋮

12/1/2018 66.09 92.64

a. Calculate the mean and the standard deviation for each corporation’s stock price.
b. Which corporation had the higher average stock price over the time period?
Which corporation’s stock price had greater dispersion as measured by the
standard deviation?
40. FILE HD_Lowe’s. Annual revenues (in $ millions) for Home Depot and Lowe’s
Corporation are collected for 13 years. A portion of the data is shown in the
accompanying table.

Year HD Lowe’s

2006 77019 43243

2007 79022 46927

⋮ ⋮ ⋮

2018 100904 68619

a. For each company, calculate the mean and the median revenues for this time
period. Which company had higher average revenues?
b. For each company, calculate the variance and the standard deviation for this time
period. Which company’s revenues had more dispersion as measured by the
standard deviation?
41. FILE Prime. The accompanying table shows a portion of the annual expenditures
(in $) for 100 Amazon Prime customers.

Customer Expenditures

1 1272

2 1089
 Customer Expenditures

⋮ ⋮

100 1389

a. Find the mean and the median of annual expenditures.

b. Calculate and interpret the first quartile and the third quartile for annual
expenditures. [Note: If you are using Excel, use the PERCENTILE.INC function
to calculate a percentile.]
42. FILE Gas_2019. The following table lists a portion of the average price (in $) for a
gallon of gas for the 50 states and the District of Columbia as reported by AAA Gas
Prices on January 2, 2019.

State Price

Alaska 3.06

Alabama 1.94

⋮ ⋮

Wyoming 2.59

a. Find the mean and the median for gas price.

b. Calculate and interpret the first quartile and the third quartile for gas price. [Note:
If you are using Excel, use the PERCENTILE.INC function to calculate a
percentile.]
c. Calculate the sample variance and the sample standard deviation.

page 126
43. FILE Rent. The following table shows a portion of the monthly
rent and square footage for 40 rentals in a large college town.

House Rent Sqft

1 645 500

2 675 648

⋮ ⋮ ⋮

40 2400 2700
 a. Calculate the mean and the standard deviation for monthly rent.
b. Calculate the mean and the standard deviation for square footage.
44. Refer to the previous exercise for a description of the data.
a. The skewness and (excess) kurtosis coefficients for monthly rent are 1.0198 and
0.4790, respectively. Interpret these values.
b. The skewness and (excess) kurtosis coefficients for square footage are 3.0573
and 12.3484, respectively. Interpret these values.
45. FILE Highway. Many environmental groups and politicians are suggesting a return
to the federal 55-mile-per-hour (mph) speed limit on America’s highways. They
argue that not only will a lower national speed limit reduce greenhouse emissions, it
will also increase traffic safety. A researcher believes that a lower speed limit will
not increase traffic safety because he feels that traffic safety is based on the
variability of the speeds with which people are driving, rather than the average
speed. The researcher gathers the speeds of 40 cars from a highway with a speed
limit of 55 mph (Highway 1) and the speeds of 40 cars from a highway with a speed
limit of 65 mph (Highway 2). A portion of the data is shown in the accompanying
table.

Car Highway 1 Highway 2

1 60 70

2 55 65

⋮ ⋮ ⋮

40 52 65

a. Calculate the mean and the median for each highway.

b. Calculate the standard deviation for each highway.
c. Do the data support the researcher’s belief? Explain.
46. FILE Firms. Monthly stock prices (in $) for Firm A and Firm B are collected for five
years. A portion of the data is shown in the accompanying table.

Date A B

1/1/2014 63.85 75.56

2/1/2014 66.04 78.68

⋮ ⋮ ⋮

12/1/2018 89.98 126.38

 a. Calculate the mean, the variance, and the standard deviation for each firm’s
stock price.
b. Which firm had the higher average stock price over the time period?
c. Which firm’s stock price had greater dispersion as measured by the standard
deviation?
47. FILE Country. The accompanying table shows a portion of the annual returns (in
%) for a mutual fund focusing on investments in Latin America and a mutual fund
focusing on investments in Canada from 1994 through 2018.

Year Latin America Canada

1994 −23.17 −11.98

1995 −16.46 19.39

⋮ ⋮ ⋮

2018 −10.37 −14.29

a. Which fund had the higher reward over this time period? Explain.
b. Which fund was riskier over this time period? Explain.
c. Given a risk-free rate of 2%, which fund has the higher Sharpe ratio? What does
this ratio imply?
48. Refer to the previous exercise for a description of the data.
a. The skewness and (excess) kurtosis coefficients for the Latin America fund are
0.3215 and −0.7026, respectively. Interpret these values.
b. The skewness and (excess) kurtosis coefficients for the Canada fund are
−0.2531 and 0.0118, respectively. Interpret these values.
49. FILE Tech_Energy. The accompanying table shows a portion of the annual
returns (in %) for a technology mutual fund and an energy mutual fund from 1982
through 2018.

Year Technology Energy

1982 56.32 −12.16

1983 52.47 20.27

⋮ ⋮ ⋮

2018 −8.79 −24.92

 a. Which fund had the higher average return?
b. Which fund was riskier over this time period?
c. Given a risk-free rate of 2%, which fund had the higher Sharpe ratio? What does
this ratio imply?
50. FILE Test_Scores. The accompanying data file shows midterm and final grades
for 32 students. Calculate and interpret the correlation coefficient.
51. FILE Life_Obesity. The accompanying data file shows life expectancies (in years)
and obesity rates (in %) for the 50 states and the District of Columbia. Calculate
and interpret the correlation coefficient.

page 127
52. FILE Happiness_Age. Many attempts have been made to relate
happiness with various factors. A 2018 study in the Journal of Happiness relates
happiness with age and finds that holding everything else constant, people are
least happy when they are in their mid-40s. The accompanying data file shows a
respondent’s age and his/her perception of well-being on a scale from 0 to 100.
a. Calculate and interpret the correlation coefficient between age and happiness.
b. Construct a scatterplot to point out a flaw with the correlation analysis in part a.

3.5 DETECTING OUTLIERS

LO 3.6
Use boxplots and z-scores to identify outliers

Extremely large or small observations for a variable are referred to as

outliers, which can unduly influence summary statistics, such as the
mean or the standard deviation. In a small sample, the impact of
outliers is particularly pronounced. Sometimes, outliers may just be
due to random variations, in which case the relevant observations
should remain in the data set. Alternatively, outliers may indicate bad
data due to incorrectly recorded observations or incorrectly included
observations in the data set. In such cases, the relevant observations
should be corrected or simply deleted from the data set. However,
there are no universally agreed upon methods for treating outliers.
Various approaches are used across different disciplines and
applications.
 In any event, it is important to be able to identify potential outliers
so that one can take corrective actions, if needed. In this section, we
first construct a boxplot, which is an effective tool for identifying
outliers. A series of boxplots are also useful when comparing similar
information for a variable gathered at another place or time. Another
method for detecting outliers is to use z-scores.

A Boxplot
A common way to quickly summarize a variable is to use a five-
number summary. A five-number summary shows the minimum value,
the quartiles (Q1, Q2, and Q3), and the maximum value of the
variable. A boxplot, also referred to as a box-and-whisker plot, is a
convenient way to graphically display the five-number summary of a
variable. In general, a boxplot is constructed as follows:
Plot the five-number summary values in ascending order on the
horizontal axis.
Draw a box encompassing the first and third quartiles.
Draw a dashed vertical line in the box at the median.
Calculate the interquartile range (IQR). Recall that IQR = Q3 − Q1.
Draw a line (“whisker”) that extends from Q1 to the minimum value
that is not farther than 1.5 × IQR from Q1. Similarly, draw a line that
extends from Q3 to the maximum value that is not farther than 1.5 ×
IQR from Q3.
Use an asterisk (or other symbol) to indicate observations that are
farther than 1.5 × IQR from the box. These observations are
considered outliers.
Consider the boxplot in Figure 3.21. The left whisker extends from Q1
to the minimum value (Min) because Min is not farther than 1.5 × IQR
from Q1. The right whisker, on the other hand, does not extend from
Q3 to the maximum value because there is an observation that is
farther than 1.5 × IQR from Q3. The asterisk indicates that this
observation is considered an outlier.
FIGURE 3.21 An example of a boxplot

A boxplot is also used to informally gauge the shape of the

distribution. Symmetry is implied if the median is in the center of the
box and the left and right whiskers are equidistant from their
respective quartiles. If the median is left of center and the right
whisker is longer than the left whisker, then the distribution is
positively skewed. Similarly, if the median is right of page 128
center and the left whisker is longer than the right
whisker, then the distribution is negatively skewed. If outliers exist, we
need to include them when comparing the lengths of the left and right
whiskers.
From Figure 3.21, we note that the median is located to the left of
center and that an outlier exists on the right side. Here the right
whisker is longer than the left whisker, and if the outlier is included,
then the right whisker becomes even longer. This indicates that the
underlying distribution is positively skewed.
Unfortunately, Excel does not provide a simple and straightforward
way to construct a boxplot; however, R does. Consider the following
example.
FILE
Growth_Value

EXAMPLE 3.14
Use R to construct a boxplot for the Growth and Value variables
from the introductory case. Interpret the results.
SOLUTION:
Using R
a. Import the Growth_Value data into a data frame (table) and
label it myData.
b. We use the boxplot function. For options within the function,
we use main to provide a title, xlab to label the x-axis, names to
label each variable, horizontal to construct a horizontal boxplot
(as opposed to a vertical boxplot), and col to give color to the
IQR portion. Enter:
> boxplot(myData$Growth, myData$Value, main= “Boxplots for
Growth and Value”, xlab=“Annual Returns, 1984-2018 (in
percent)”, names =c(“Growth”,“Value”), horizontal = TRUE,
col=“gold”)
Figure 3.22 shows the output that R returns.

FIGURE 3.22 Boxplots for Growth and Value

page 129
c. To treat outliers, we use the out parameter in the
boxplot function to extract and store their values in a new data
frame (table). Enter:
> outliersGrowth <- boxplot(myData$Growth)$out
 > outliersValue <- boxplot(myData$Value)$out
Verify that there is one outlier (79.48) for the Growth variable
and one outlier (−46.52) for the Value variable.
d. One approach to treat outliers in R is to replace them with NAs,
which represent missing values. The %in% operator is for
value matching, and we use it here to find the outliers in the
Growth and Value variables. We then use the ifelse function to
replace the outliers with NAs and store updated data in two
new variables, newGrowth and newValue. Enter:
> myData$newGrowth <- ifelse(myData$Growth %in%
outliersGrowth, NA, myData$Growth)
> myData$newValue <- ifelse(myData$Value %in%
outliersValue, NA, myData$Value)
Verify that the newGrowth observation for year 1999 and the
newValue observation for year 2008 are now NAs. Note: With
outliers replaced with NAs, we have the option to implement the
omission and the imputation strategies for treatment of missing
values described in Section 2.4.
e. Once the outliers are replaced with NAs, we can recalculate
summary measures. We use the summary function to compare
the means of the original variables with outliers and the new
variables without outliers. Enter:
> summary(myData)
Verify that the means of the newGrowth and newValue
variables are 13.210 and 13.149, respectively, and that they are
different from the means of the original Growth and Value
variables.
Summary
The median returns for the two mutual funds are indicated by
the bold and wider vertical lines in Figure 3.22. As we already
found, the median returns for the two funds are similar (15.09%
for Value versus 14.44% for Growth). However, Value has an
 outlier on the left-hand side, as indicated by the circle, while
Growth has an outlier on the right-hand side.
For Value, the outlier on the left-hand side coupled with a
median that falls to the right of center in the IQR box suggests
that this distribution is negatively skewed. This is consistent with
the negative skewness coefficient that was calculated for this
variable in Section 3.4.
On the other hand, Growth has an outlier on the right-hand
side with a median that falls to the left of center in the IQR box.
The distribution of Growth is positively skewed. Again, this is
consistent with the positive skewness coefficient that was
calculated for this variable in Section 3.4.

z-Scores
The mean and the standard deviation are the most extensively used
measures of central location and dispersion, respectively. Unlike the
mean, it is not easy to interpret the standard deviation intuitively. All
we can say is that a low value for the standard deviation indicates that
the observations are close to the mean, while a high value for the
standard deviation indicates that the observations are more dispersed
from the mean.
We will first use the empirical rule to make precise statements
regarding the percentage of observations that fall within a specified
number of standard deviations from the mean. We page 130
then compute a z-score that measures the relative
location of an observation and indicates whether it is an outlier.

The Empirical Rule

If the observations are drawn from a relatively symmetric and bell-
shaped distribution—perhaps determined by an inspection of its
histogram—then we can make precise statements about the
percentage of observations that fall within certain intervals. Symmetry
and bell-shaped are characteristics of the normal distribution, a topic
that we discuss in more detail in Chapter 4. The normal distribution is
often used as an approximation for many real-world applications.
Given a sample mean , a sample standard deviation s, and a
relatively symmetric and bell-shaped distribution, the empirical rule
states that

Approximately 68% of all observations fall in the interval ,

Approximately 95% of all observations fall in the interval , and

Almost all observations fall in the interval .

Figure 3.23 illustrates the empirical rule.

FIGURE 3.23 Graphical representation of the empirical rule

EXAMPLE 3.15
 A large lecture class has 280 students. The professor has
announced that the mean score on an exam is 74 with a
standard deviation of 8. The distribution of scores is bell-shaped.
a. Approximately how many students scored within 58 and 90?
b. Approximately how many students scored more than 90?
SOLUTION:

a. The score 58 is two standard deviations below the mean

, while the score 90 is two standard

deviations above the mean . For a bell-

shaped distribution, the empirical rule states that approximately

95% of the observations fall within two standard deviations of
the mean. Therefore, about 95% of 280 students, or 0.95 × 280
= 266 students, scored within 58 and 90.
b. We know that the score 90 is two standard deviations above
the mean. Because approximately 95% of the observations fall
within two standard deviations of the mean, we can infer that
5% of the observations fall outside the interval. Therefore,
given the symmetry of the distribution, about half of 5%, or
2.5%, of 280 students scored above 90. Equivalently, about
seven students (0.025 × 280), scored above 90 on the exam. If
the professor uses a cutoff score above 90 for an A, then only
seven students in the class are expected to get an A.

page 131

Calculating z-Scores
It is often instructive to use the mean and the standard deviation to
find the relative location of an observation. Suppose a student gets a
score of 90 on her accounting exam and a score of 90 on her
marketing exam. While the student’s scores are identical in both
classes, her relative position in these classes may be quite different.
What if the mean score was different in the classes? Even with the
same mean scores, what if the standard deviation was different in the
classes? Both the mean and the standard deviation are needed to
find the relative position of this student in both classes.
We use the z-score to find the relative position of an observation
by dividing the difference of the observation from the mean by the
standard deviation, or, equivalently, . A z-score is a unitless

measure. It measures the distance of an observation from the mean

in terms of standard deviations. For example, a z-score of two implies
that the observation is two standard deviations above the mean.
Similarly, a z-score of −1.5 implies that the observation is 1.5 standard
deviations below the mean.
Converting observations into z-scores is also called standardizing
the observations. Standardization is a common technique used in
data analytics when dealing with variables measured using different
scales. We will revisit this topic in Chapter 8 to discuss the use of
standardization techniques in data mining.

EXAMPLE 3.16
The mean and the standard deviation of scores on an
accounting exam are 74 and eight, respectively. The mean and
the standard deviation of scores on a marketing exam are 78
and 10, respectively. Find the z-scores for a student who scores
90 in both classes.

SOLUTION: The z-score in the accounting class is .

Similarly, the z-score in the marketing class is .

Therefore, the student has fared relatively better in accounting

 because she is two standard deviations above the mean, as
compared to marketing, where she is only 1.2 standard
deviations above the mean.

If the distribution of a variable is relatively symmetric and bell-

shaped, we can also use z-scores to detect outliers. Because almost
all observations fall within three standard deviations of the mean, it is
common to treat an observation as an outlier if its z-score is more
than 3 or less than −3. Such observations must be reviewed to
determine if they should remain in the data set.

EXAMPLE 3.17
Table 3.18 shows the minimum and maximum observations as
well as the means and standard deviations for the Growth and
Value variables from the introductory case. Calculate the z-
scores for the minimum and the maximum observations for each
variable. Are the results consistent with the boxplots constructed
in Figure 3.22? Explain.

TABLE 3.18 Summary Statistics for the Growth and Value Variables (in %)

page 132

For Growth, the z-scores for the minimum and the

SOLUTION:
maximum observations are:
For Value, the z-scores for the minimum and the maximum
observations are:

For Growth, because the absolute value of both z-scores is less

than 3, it would suggest that there are no outliers for the Growth
variable. However, the boxplot in Figure 3.22 indicates that there
is an outlier. How do we resolve this apparent inconsistency?
Remember that z-scores are reliable indicators of outliers when
the distribution is relatively bell-shaped and symmetric. Because
the boxplot indicates that the Growth variable is positively
skewed, we are better served identifying outliers in this case
with a boxplot. For Value, the minimum observation is identified
as an outlier in the boxplot as well as by its z-score with a value
of −3.2347.

IDENTIFYING OUTLIERS
A boxplot is a convenient way to graphically display the five-
number summary of a variable. If outliers are present, then
they are indicated as asterisks (or another symbol) that are
farther than 1.5 × IQR from the box.
 The z-score measures the relative position of an observation
within a distribution and is calculated as . If the

distribution of a variable is relatively symmetric and bell-

shaped, then it is common to treat an observation as an outlier
if its z-score is more than 3 or less than −3.

EXERCISES 3.5
Applications
53. Consider the following boxplot.

a. Does the boxplot indicate possible outliers in the data?

b. Comment on the skewness of the underlying distribution.
54. Consider the following boxplot.

a. Does the boxplot indicate possible outliers in the data?

b. Comment on the skewness of the underlying distribution.
55. Consider the following five-point summary for a variable that was obtained using
200 observations.

page 133
a. Interpret Q1 and Q3.
b. Calculate the interquartile range. Determine whether any outliers exist.
 c. Is the distribution symmetric? If not, comment on its skewness.
56. Consider the following five-point summary for a variable that was obtained using
500 observations.

a. Interpret Q1 and Q3.

b. Calculate the interquartile range. Determine whether any outliers exist.
c. Is the distribution symmetric? If not, comment on its skewness.
57. A sample of the salaries of assistant professors on the business faculty at a local
university revealed a mean income of $100,000 with a standard deviation of
$10,000. Assume that salaries follow a bell-shaped distribution. Use the empirical
rule to answer the following questions.
a. Approximately what percentage of the salaries fall between $90,000 and
$110,000?
b. Approximately what percentage of the salaries fall between $80,000 and
$120,000?
c. Approximately what percentage of the salaries are greater than $120,000?
58. It is often assumed that IQ scores follow a bell-shaped distribution with a mean of
100 and a standard deviation of 16. Use the empirical rule to answer the following
questions.
a. Approximately what percentage of scores are between 84 and 116?
b. Approximately what percentage of scores are less than 68?
c. Approximately what percentage of scores are more than 116?
59. The historical returns on a portfolio had an average return of 8% and a standard
deviation of 6%. Assume returns on the portfolio follow a bell-shaped distribution.
Use the empirical rule to answer the following questions.
a. Approximately what percentage of returns were between 2% and 14%?
b. Approximately what percentage of returns were greater than 14%?
c. Approximately what percentage of returns were below −4%?
60. On average, an American professional football game lasts about three hours, even
though the ball is actually in play only 11 minutes. Assume that game times follow a
bell-shaped distribution and that the standard deviation is 0.4 hour. Use the
empirical rule to answer the following questions.
a. Approximately what percentage of games last between 2.2 hours and 3.8 hours?
b. Approximately what percentage of games last longer than 3.4 hours?
c. Approximately what percentage of games last less than 2.2 hours?
61. FILE Prime. The accompanying data file shows the annual expenditures
(Expenditures in $) for 100 Amazon Prime customers.
 a. Construct a boxplot for the Expenditures variable. Does the boxplot suggest that
outliers exist?
b. Use z-scores to determine if there are any outliers for the Expenditures variable.
Are your results consistent with part a? Explain why or why not.
62. FILE Debt. The accompanying data file shows the average monthly debt
payments (Debt in $) for residents of 26 metropolitan areas.
a. Construct a boxplot for the Debt variable. Does the boxplot suggest that outliers
exist?
b. Use z-scores to determine if there are any outliers for the Debt variable. Are your
results consistent with part a? Explain why or why not.
63. FILE Gas_2019. The accompanying data file shows the average price of gas
(Price in $ per gallon) for the 50 states during January 2019.
a. Construct a boxplot for the Price variable. Does the boxplot suggest that outliers
exist?
b. Use z-scores to determine if there are any outliers for the Price variable. Are your
results consistent with part a? Explain why or why not.
c. Use the information from the boxplot to replace outliers with NAs. What is the
mean of the average price of gas without the outliers?
64. FILE Tech_Energy. The accompanying data file shows the annual return (in %)
for Fidelity’s Technology and Energy mutual funds from 1982 through 2018.
a. Construct a boxplot for the Technology variable. Does the boxplot suggest that
outliers exist?
b. Use z-scores to determine if there are any outliers for the Technology variable.
Are your results consistent with part a? Explain why or why not.
c. Construct a boxplot for the Energy variable. Does the boxplot suggest that
outliers exist?
d. Use z-scores to determine if there are any outliers for the Energy variable. Are
your results consistent with part c? Explain why or why not.
e. Use the information from the two boxplots to replace outliers with NAs. What are
the new averages for the Technology and Energy variables?

3.6 WRITING WITH BIG DATA

When confronted with a very large data set, a necessary first step for any analysis is to
convert the raw data into a more meaningful form. Tabular and graphical methods as
well as summary measures prove very useful. Consider the following big data case.
FILE
House_Price

page 134

Case Study
An investor currently owns real estate in the college town of Blacksburg, Virginia—
home to the Virginia Tech Hokies. He would like to expand his holdings by purchasing
similar rental property in either Athens, Georgia, or Chapel Hill, North Carolina. As a
preliminary step, he would like information on house prices in these two areas. He is
interested in properties that have at least two bedrooms and that are listed for less
than $1,000,000. The following report will summarize previous sales that have
satisfied these criteria.

Sample Report – Investing in College Town

Real Estate
There are a number of reasons why you might consider investing in a rental
property near a university. First, there’s a large pool of renters, including
students, faculty, and staff. Second, because many universities are unable to
house their students beyond freshman year, students offer a steady stream of
rental demand. Finally, university towns tend to be filled with restaurants,
shopping, and nightlife. All of these factors can make it easier for you to market
your property.

kali9/Getty Images

The following report examines house prices in Athens, Georgia—home to

the University of Georgia Bulldogs—and Chapel Hill, North Carolina—home to
the University of North Carolina Tar Heels. The sample consists of 293 house
sales in Athens and 351 house sales in Chapel Hill. In addition, all houses in
the sample had at least two bedrooms and sold for less than $1,000,000. Table
3.19 provides the most relevant summary measures for the analysis.

TABLE 3.19 Summary Measures for House Prices (in $) in Athens and
Chapel Hill

Summary Measure Athens, GA Chapel Hill, NC

Mean 219,671 429,152

Median 177,500 395,000

Minimum 41,125 105,000

Maximum 910,000 950,000

Standard deviation 147,648 186,762

Coefficient of variation 0.67 0.44

Number of houses 293 351

The average house price in Athens is $219,671, as opposed to $429,152 in

Chapel Hill, a difference of almost $210,000. In Athens, the median house price
is $177,500, suggesting that half of the house prices are below this value and
half are above this value. The corresponding value in Chapel Hill is $395,000.
The difference in medians between these two cities is close to $218,000. In
both cities, the median is quite a bit less than the mean, which implies that
outliers, some extremely high house prices in this case, are likely present.
While the mean and the median represent where house prices tend to
cluster, they do not relay information about the variability in house prices.
Generally, standard deviation is used as a measure of variability. The standard
deviation for house prices in Chapel Hill is greater than the standard deviation
for house prices in Athens ($186,762 > $147,648), suggesting that, compared
to Athens, house prices in Chapel Hill are more dispersed from the mean.
 page 135

Table 3.20 shows the relative frequency distribution for the house prices in
both cities. The relative frequency distribution reinforces the findings from the
summary measures, that is, house prices in Athens are more affordable than
houses in Chapel Hill. In Athens, 51% of the prices fell in the range $100,000
to $200,000, but only 9% of the prices in Chapel Hill fell in this range.
Moreover, 91% of houses in Athens versus only 51% of houses in Chapel Hill
sold for less than $400,000.

TABLE 3.20 Frequency Distribution for House Prices in Athens and

Chapel Hill

Athens, Chapel Hill,

Interval (in $)
GA NC

   0 < x ≤ 100,000 0.11 0.00

100,000 < x ≤ 200,000 0.51 0.09

200,000 < x ≤ 300,000 0.19 0.17

300,000 < x ≤ 400,000 0.10 0.25

400,000 < x ≤ 500,000 0.03 0.17

500,000 < x ≤ 600,000 0.03 0.14

600,000 < x ≤ 700,000 0.02 0.07

700,000 < x ≤ 800,000 0.00 0.07

800,000 < x ≤ 900,000 0.01 0.03

 Athens, Chapel Hill,
Interval (in $)
GA NC

900,000 < x ≤ 0.00 0.01

1,000,000

Finally, Figure 3.24 shows the boxplots of house prices for each city. The
boxplots reveal two more major points with respect to house prices in these
two cities:

FIGURE 3.24 Boxplots of house prices in Athens, Georgia, and Chapel

Hill, North Carolina

In each boxplot, the median is off-center within the box,

being located to the left of center.
In each boxplot, there are outliers on the right side.
However, there are far fewer outliers in the Chapel Hill
distribution as compared to the Athens distribution.
These two observations suggest that both distributions are positively skewed.
This implies that the bulk of the house prices falls in the lower end of the
distribution, and there are relatively few very-high-priced houses.
This report summarizes house prices in Athens, Georgia, and Chapel Hill,
North Carolina. On average, houses in Chapel Hill are almost twice as
expensive as those in Athens. Moreover, if outliers are removed from the
analysis, house prices in Athens are less variable than house prices in Chapel
 Hill. However, before any investor purchases property in either city, many other
factors should be considered, such as the strength of the rental market and
average income in each city.

page 136

Suggested Case Studies

Tabular and graphical methods as well as summary measures prove very useful when
analyzing the big data that accompanies this text. Here are some suggestions.

Report 3.1 FILE House_Price. Perform a similar analysis to the one conducted in this
section, but choose two other college towns.

Report 3.2 FILE College_Admissions. Use tabular and graphical methods as well as
summary measures to examine the SAT scores of those students who were admitted
to the School of Business & Economics versus those students who were admitted to
the School of Arts & Letters.

Report 3.3 FILE Longitudinal_Survey. Use tabular and graphical methods as well as
summary measures to examine the weight of a respondent depending on whether or
not the respondent is an outgoing or a shy individual.

Report 3.4 FILE TechSales_Reps. Use tabular and graphical methods as well as
summary measures to examine the salaries of sales representatives depending on
their personality types and gender in the software and hardware groups.

page 137
 page 138

4 Probability and Probability

Distributions

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 4.1  Describe probability concepts and the
rules of probability.
LO 4.2  Apply the total probability rule and Bayes’
theorem.
LO 4.3  Describe a discrete random variable and
its probability distribution.
LO 4.4  Calculate probabilities for binomial and
Poisson distributions.
LO 4.5  Describe the normal distribution and
calculate its associated probabilities.

very day we make choices about issues in the presence of

E uncertainty. By figuring out the chances of various events, we
are better prepared to make the more desirable choices. For
example, given the weather forecast, we determine whether we
should wear a jacket or carry an umbrella. Similarly, retailers tweak
their sales force in anticipation of an increase or decrease in
shoppers, and the Federal Reserve adjusts interest rates based on
its anticipation of growth and inflation. This chapter presents the
essential probability tools needed to frame and address many real-
world issues involving uncertainty.
We extend our discussion about probability by introducing the
concept of a random variable. A random variable summarizes the
results of an experiment in terms of numerical values and can be
classified as discrete or continuous depending on the range of
values that it assumes. A discrete random variable assumes a
countable number of distinct values, whereas a continuous random
variable is characterized by uncountable values. Finally, we discuss
two widely used discrete probability distributions, namely the
binomial and the Poisson distributions. We also discuss the normal
distribution, which is the most extensively used continuous
probability distribution and is the cornerstone of statistical inference.

page 139
©Seastock/Shutterstock

INTRODUCTORY CASE
Linking Support for Legalizing Marijuana
with Age Group
Support for marijuana legalization in the United States has
grown remarkably over the past few decades. In 1969, when
the question was first presented, only 12% of Americans were
in favor of its legalization. This support increased to over 25%
by the late 1970s. While support was stagnant from 1981 to
 1997, the turn of the century brought a renewed interest in its
legalization, with the percentage of Americans in favor
exceeding 30% by 2000 and 40% by 2009.
Alexis Lewis works for a drug policy institute that focuses on
science, health, and human rights. She is analyzing the
demographic breakdown of marijuana supporters. Using results
from a Pew Research Center survey conducted from August
23-September 2, 2016, she has found that support for
marijuana legalization varies considerably depending on a
person’s age group. Alexis compiles information on support
based on age group, as shown in Table 4.1.

TABLE 4.1 Percentage Support for Legalizing Marijuana by Age Group

Age Group Support

Millennial (18–35) 71%

Generation X (36–51) 57 

Baby Boomer (52–70) 56 

Silent (71–88) 33 

Alexis understands that an important factor determining the

fate of marijuana legalization concerns each age group’s ability
to sway the vote. For adults eligible to vote, she breaks down
each age group’s voting power. The Millennial, Generation X,
Baby Boomer, and Silent generations account for 31%, 25%,
31%, and 13% of the voting population, respectively. She
would like to use the data to accomplish the following tasks.
1. Calculate and interpret relevant conditional, unconditional,
and joint probabilities.
 2. Calculate and interpret the probability of all Americans who
support the legalization of marijuana.
A synopsis of this case is provided in Section 4.2.

page 140

4.1 PROBABILITY CONCEPTS AND

PROBABILITY RULES
LO 4.1
Describe probability concepts and the rules of probability.

We are better prepared to deal with uncertainty if we know the

probabilities that describe which events are likely and which are
unlikely. A probability is a numerical value that measures the
likelihood that an event occurs. This value is between zero and one,
where a value of zero indicates impossible events and a value of one
indicates definite events. In order to define an event and assign the
appropriate probability to it, it is useful to first establish some
terminology and impose some structure on the situation.
An experiment is a process that leads to one of several possible
outcomes. The diversity of the outcomes of an experiment is due to
the uncertainty of the real world. When changing a branding strategy,
it is uncertain if the profits will go up or down. You can think of this as
an experiment because the actual outcome will be determined only
over time. Other examples of an experiment include the extent of
fraudulent credit card transactions, the sale of a new product, and
the letter grade in a course.
A sample space, denoted by S, of an experiment contains all
possible outcomes of the experiment. For example, suppose the
sample space representing the letter grade in a course is given by S
= {A, B, C, D, F}. The sample space for an experiment need not be
unique. For example, in the above experiment, we can also define
the sample space with just P (pass) and F (fail) outcomes; that is, S
= {P, F}. Note that if the teacher also gives out an I (incomplete)
grade, then neither of the sample spaces defined above is valid
because they do not contain all possible outcomes of the
experiment.

Events
An event is a subset of the sample space. A simple event consists of
just one of the possible outcomes of an experiment. Getting an A in
a course is an example of a simple event. An event may also contain
several outcomes of an experiment. For example, we can define an
event as getting a passing grade in a course; this event is formed by
the subset of outcomes A, B, C, and D.
Events are considered exhaustive if they include all outcomes in
the sample space. In the grade-distribution example, the events of
getting grades A and B are not exhaustive events because they do
not include many feasible grades in the sample space. However, the
events P and F, defined as “pass” and “fail,” respectively, are
exhaustive.
Another important probability concept concerns mutually
exclusive events. For two mutually exclusive events, the occurrence
of one event precludes the occurrence of the other. Going back to
the grade-distribution example, while the events of getting grades A
and B are not exhaustive, they are mutually exclusive because you
cannot possibly get an A as well as a B in the same course. Grades
P and F, on the other hand, are both mutually exclusive and
exhaustive.

EXPERIMENTS AND EVENTS

An experiment is a process that leads to one of several
possible outcomes. A sample space, denoted S, of an
 experiment contains all possible outcomes of the experiment.
An event is any subset of outcomes of the experiment. It is
called a simple event if it contains a single outcome.
Events are exhaustive if all possible outcomes of an
experiment belong to the events. Events are mutually
exclusive if they do not share any common outcome of an
experiment.

For any experiment, we can define events based on one or more

outcomes of the experiment and also combine events to form new
events. The union of two events, denoted A ∪ B, is page 141
the event consisting of all outcomes in A or B. A
useful way to illustrate these concepts is through the use of a Venn
diagram. Figure 4.1 shows a Venn diagram where the rectangle
represents the sample space S and the two circles represent events
A and B. The union A ∪ B is the portion in the Venn diagram that is
included in either A or B.

FIGURE 4.1 The union of two events, A ∪ B

The intersection of two events, denoted A ∩ B, is the event

consisting of all outcomes in A and B. Figure 4.2 depicts the
intersection of two events A and B. The intersection A ∩ B is the
portion in the Venn diagram that is included in both A and B.

FIGURE 4.2 The intersection of two events, A ∩ B

The complement of event A, denoted Ac, is the event consisting of

all outcomes in the sample space S that are not in A.

COMBINING EVENTS
The union of two events, denoted A ∪ B, is the event
consisting of all outcomes in A or B.
The intersection of two events, denoted A ∩ B, is the event
consisting of all outcomes in A and B.
The complement of event A, denoted Ac, is the event
consisting of all outcomes in the sample space S that are not
in A.

Assigning Probabilities
Now that we have described a valid sample space and the various
ways in which we can define events from that sample space, we are
ready to assign probabilities. When we arrive at a probability, we
generally are able to categorize the probability as a subjective
probability, an empirical probability, or a classical probability.
Regardless of the method used, there are two defining properties of
probability.

page 142

THE TWO DEFINING PROPERTIES OF PROBABILITY

1. The probability of any event A is a value between 0 and 1; that

is, 0 ≤ P(A) ≤ 1.
2. The sum of the probabilities of any list of mutually exclusive
and exhaustive events equals 1.

The subjective probability is based on an individual’s personal

judgment or experience. For example, a manager may instinctively
feel that 14% of consumers will respond positively to the firm’s social
media campaign. This personal assessment of probability does not
explicitly reference any data. Subjective probabilities differ from
person to person and may contain a high degree of personal bias.
In many instances, we calculate probabilities by referencing data
based on the observed outcomes of an experiment. The empirical
probability of an event is the observed relative frequency with which
an event occurs. If over a ten year period in a particular city, 152
restaurants out of 180 opened and closed within one year, and
closure within a year is our benchmark for failure, we can determine
the empirical probability of restaurant failures for that city as 152/180
= 0.84. It is important to note that the empirical probability is reliable
if the experiment is run a very large number of times. In the above
example, it may be misleading to report a failure rate of 0.40 based
on 2 out of 5 restaurants opening and closing within one year.
In a more narrow range of well-defined problems, we can
sometimes deduce probabilities by reasoning about the problem.
The resulting probability is a classical probability. Classical
probabilities are often used in games of chance. They are based on
the assumption that all outcomes of an experiment are equally likely.
Therefore, the classical probability of an event is computed as the
number of outcomes belonging to the event divided by the total
number of outcomes. For example, we can calculate the probability
of rolling a two with a fair six-sided die as 1/6 = 0.1667.

CATEGORIZING PROBABILITIES
A subjective probability is calculated by drawing on personal
and subjective judgment.
An empirical probability is calculated as a relative frequency
of occurrence.
A classical probability is based on logical analysis rather than
on observation or personal judgment.
Since empirical and classical probabilities generally do not vary
from person to person, they are often grouped as objective
probabilities.

Rules of Probability
We will now present various rules used to combine probabilities of
events.
The complement rule follows from one of the defining properties
of probability: The sum of probabilities assigned to simple events in a
sample space must equal one. Therefore, for an event A and its
complement Ac, we get P(A) + P(Ac) = 1. Rearranging this equation,
we obtain the complement rule.

THE COMPLEMENT RULE

The complement of an event A is derived as P(Ac) = 1 − P(A).
The complement rule is quite straightforward and rather simple, but it
is widely used and powerful.

page 143

EXAMPLE 4.1
A manager at Moksha Yoga Center believes that 37% of
female and 30% of male open house attendees she contacts
will purchase a membership.
a. What is the probability that a randomly selected female
contacted by the manager will not purchase a membership?
b. What is the probability that a randomly selected male
contacted by the manager will not purchase a membership?
SOLUTION:

a. Let’s define A as the event that a randomly selected female

contacted by the manager will purchase a membership; thus,
P(A) = 0.37. In this problem, we are interested in the
complement of A. So P(Ac) = 1 − P(A) = 1 − 0.37 = 0.63.
b. Similarly, we define B as the event that a randomly selected
male contacted by the manager will purchase a membership,
so P(B) = 0.30. Thus, P(Bc) = 1 − P(B) = 1 − 0.30 = 0.70.

The addition rule allows us to find the probability of the union of

two events. Suppose we want to find the probability that either A
occurs or B occurs, so in probability terms, P(A ∪ B). Recall from
Figure 4.1 that the union, A ∪ B, is the portion in the Venn diagram
that is included in event A or event B, whereas the intersection, A ∩
B, is the portion in the Venn diagram that is included in both A and B.
 If we try to obtain P(A ∪ B) by simply summing P(A) with P(B),
then we overstate the probability because we double-count the
probability of the intersection of A and B, P(A ∩ B). It is common to
refer to P(A ∩ B) as the joint probability of events A and B. When
implementing the addition rule, we sum P(A) and P(B) and then
subtract P(A ∩ B) from this sum.

THE ADDITION RULE

The probability that event A or event B occurs is derived as

EXAMPLE 4.2
Anthony feels that he has a 75% chance of getting an A in
Statistics and a 55% chance of getting an A in Managerial
Economics. He also believes he has a 40% chance of getting
an A in both classes.
a. What is the probability that he gets an A in at least one of
these courses?
b. What is the probability that he does not get an A in either of
these courses?
SOLUTION:

a. Let P(AS) correspond to the probability of getting an A in

Statistics and P(AM) correspond to the probability of getting an
A in Managerial Economics. Thus, P(AS) = 0.75 and P(AM) =
0.55. In addition, the joint probability that Anthony gets an A in
both classes, P(AS ∩ AM) = 0.40. In order to find the
 probability that he receives an A in at least one of these
courses, we calculate:

page 144
b. The probability that he does not receive an A in
either of these two courses is actually the complement of the
union of the two events; that is, P((AS ∪ AM)c). We calculated
the union in part a, so using the complement rule we have

An alternative expression that correctly captures the required

probability is A common mistake is to

calculate the probability as P((AS ∩ AM)c) = 1 − P(AS ∩ AM) =
1 − 0.40 = 0.60, which simply indicates that there is a 60%
chance that Anthony will not get an A in both courses. This is
clearly not the required probability that Anthony does not get
an A in either course.

Note that for mutually exclusive events A and B, the joint probability
is zero; that is, P(A ∩ B) = 0. We need not concern ourselves with
double-counting, and, therefore, the probability of the union is simply
the sum of the two probabilities.
In business applications, the probability of interest is often a
conditional probability. Examples include the probability that a
customer will make an online purchase conditional on receiving an e-
mail with a discount offer; the probability of making a six-figure salary
conditional on getting an MBA; and the probability that sales will
improve conditional on the firm launching a new innovative product.
Let’s use an example to illustrate the concept of conditional
probability. Suppose the probability that a recent business college
graduate finds a suitable job is 0.80. The probability of finding a
suitable job is 0.90 if the recent business college graduate has prior
work experience. Here, the probability of an event is conditional on
the occurrence of another event. If A represents “finding a job” and B
represents “prior work experience,” then P(A) = 0.80 and the
conditional probability is denoted as P(A ∣ B) = 0.90. In this example,
the probability of finding a suitable job increases from 0.80 to 0.90
when conditioned on prior work experience. In general, the
conditional probability, P(A ∣ B), is greater than the unconditional
probability, P(A), if B exerts a positive influence on A. Similarly, P(A
∣ B) is less than P(A) when B exerts a negative influence on A.
Finally, if B exerts no influence on A, then P(A ∣ B) equals P(A). It is
common to refer to “unconditional probability” simply as “probability.”
We rely on the Venn diagram in Figure 4.2 to explain the
conditional probability. Because P(A ∣ B) represents the probability of
A conditional on B (B has occurred), the original sample space S
reduces to B. The conditional probability P(A ∣ B) is based on the
portion of A that is included in B. It is derived as the ratio of the
probability of the intersection of A and B to the probability of B.

CONDITIONAL PROBABILITY
The probability that A occurs given that B has occurred is
derived as

EXAMPLE 4.3
 Economic globalization is defined as the integration of national
economies into the international economy through trade,
foreign direct investment, capital flows, migration, and the
spread of technology. Although globalization is generally
viewed favorably, it also increases the vulnerability of a country
to economic conditions of other countries. An economist
predicts a 60% chance that country A will perform poorly and a
25% chance that country B will perform poorly. There is also a
16% chance that both countries will perform poorly.
a. What is the probability that country A performs poorly given
that country B performs poorly?
page 145
b. What is the probability that country B performs
poorly given that country A performs poorly?
c. Interpret your findings.
SOLUTION:
We first write down the available information in probability
terms. Defining A as “country A performing poorly” and B as
“country B performing poorly,” we have the following
information: P(A) = 0.60, P(B) = 0.25, and P(A ∩ B) = 0.16.

a.

b.

c. It appears that globalization has definitely made these

countries vulnerable to the economic woes of the other
country. The probability that country A performs poorly
increases from 60% to 64% when country B has performed
poorly. Similarly, the probability that country B performs poorly
 increases from 25% to 27% when conditioned on country A
performing poorly.

In some situations, we are interested in finding the joint

probability P(A ∩ B). Using the conditional probability formula

we can easily derive P(A ∩ B) = P(A ∣ B)P(B). Because

we calculate the product of two probabilities to find P(A ∩ B), we

refer to it as the multiplication rule for probabilities.

THE MULTIPLICATION RULE

The joint probability of events A and B is derived as P(A ∩ B) =
P(A ∣ B)P(B).

EXAMPLE 4.4
A manager believes that 14% of consumers will respond
positively to the firm’s social media campaign. Also, 24% of
those who respond positively will become loyal customers. Find
the probability that the next recipient of their social media
campaign will react positively and will become a loyal
customer.
SOLUTION:
Let the event R represent a consumer who responds positively
to a social media campaign and the event L represent a loyal
customer. Therefore, P(R) = 0.14 and P(L ∣ R) = 0.24. We
calculate the probability that the next recipient of a social media
campaign will react positively and become a loyal customer as
P(R ∩ L) = P(L ∣ R)P(R) = 0.24 × 0.14 = 0.0336.
 Of particular interest to researchers is whether or not two events
influence one another. Two events are independent if the
occurrence of one event does not affect the probability of the
occurrence of the other event. Similarly, events are considered
dependent if the occurrence of one is related to the probability of the
occurrence of the other. We generally test for the independence of
two events by comparing the conditional probability of one event, for
instance P(A ∣ B), to the probability, P(A). If these two probabilities
are the same, we say that the two events, A and B, are independent;
if the probabilities differ, the two events are dependent.

page 146

INDEPENDENT VERSUS DEPENDENT EVENTS

Two events, A and B, are independent if P(A ∣ B) = P(A) or,
equivalently, P(A ∩ B) = P(A ∣ B)P(B) = P(A)P(B). Otherwise,
the events are dependent.

EXAMPLE 4.5
Suppose that for a given year there is a 2% chance that your
desktop computer will crash and a 6% chance that your laptop
computer will crash. Moreover, there is a 0.12% chance that
both computers will crash. Is the reliability of the two computers
independent of each other?
SOLUTION:
Let event D represent the outcome that your desktop crashes
and event L represent the outcome that your laptop crashes.
Therefore, P(D) = 0.02, P(L) = 0.06, and P(D ∩ L) = 0.0012.
The reliability of the two computers is independent because
In other words, if your laptop crashes, it does not alter the
probability that your desktop also crashes. Equivalently, we
show that the events are independent because P(D ∩ L) =
P(D)P(L) = 0.0012.

EXERCISES 4.1
Applications
1. Consider the following scenarios to determine if the mentioned combination of
attributes represents a union or an intersection.
a. A marketing firm is looking for a candidate with a business degree and at least
five years of work experience.
b. A family has decided to purchase Toyota or Honda.
2. You apply for a position at two firms. Let event A represent the outcome of getting
an offer from the first firm and event B represent the outcome of getting an offer
from the second firm.
a. Explain why events A and B are not exhaustive.
b. Explain why events A and B are not mutually exclusive.
3. The probabilities that stock A will rise in price is 0.40 and that stock B will rise in
price is 0.60. Further, if stock B rises in price, the probability that stock A will also
rise in price is 0.50.
a. What is the probability that at least one of the stocks will rise in price?
b. Are events A and B mutually exclusive? Explain.
c. Are events A and B independent? Explain.
4. Fraud detection has become an indispensable tool for banks and credit card
companies to combat fraudulent credit card transactions. A fraud detection firm
raises an alarm on 5% of all transactions and on 80% of fraudulent transactions.
What is the probability that the transaction is fraudulent if the firm does not raise
an alarm? Assume that 1% of all transactions are fraudulent.
5. Dr. Miriam Johnson has been teaching accounting for over 20 years. From her
experience, she knows that 60% of her students do homework regularly.
 Moreover, 95% of the students who do their homework regularly pass the course.
She also knows that 85% of her students pass the course.
a. What is the probability that a student will do homework regularly and also pass
the course?
b. What is the probability that a student will neither do homework regularly nor
pass the course?
c. Are the events “pass the course” and “do homework regularly” mutually
exclusive? Explain.
d. Are the events “pass the course” and “do homework regularly” independent?
Explain.
6. Mike Danes has been delayed in going to the annual sales event at one of his
favorite apparel stores. His friend has just texted him that there are only 20 shirts
left, of which eight are in size M, 10 in size L, and two in size XL. Also nine of the
shirts are white, five are blue, and the remaining are of mixed
colors. Mike is interested in getting a white or a blue shirt in size
page 147
L. Define the events A = Getting a white or a blue shirt and B = Getting a shirt in
size L.
a. Find P(A), P(Ac), and P(B).
b. Are the events A and B mutually exclusive? Explain.
c. Would you describe Mike’s preference by the events A ∪ B or A ∩ B?
7. An analyst estimates that the probability of default on a seven-year AA-rated bond
is 0.06, while that on a seven-year A-rated bond is 0.13. The probability that they
will both default is 0.04.
a. What is the probability that at least one of the bonds defaults?
b. What is the probability that neither the seven-year AA-rated bond nor the
seven-year A-rated bond defaults?
c. Given that the seven-year AA-rated bond defaults, what is the probability that
the seven-year A-rated bond also defaults?
8. A manufacturing firm just received a shipment of 20 assembly parts, of slightly
varied sizes, from a vendor. The manager knows that there are only 15 parts in
the shipment that would be suitable. He examines these parts one at a time.
a. Find the probability that the first part is suitable.
b. If the first part is suitable, find the probability that the second part is also
suitable.
c. If the first part is suitable, find the probability that the second part is not suitable.
9. Apple products have become a household name in America, with the average
household owning 2.6 Apple products (CNBC, October 10, 2017). Suppose that in
the Midwest, the likelihood of owning an Apple product is 61% for households with
kids and 48% for households without kids. Suppose there are 1,200 households
in a representative community, of which 820 are with kids and the rest are without
kids.
 a. Are the events “household with kids” and “household without kids” mutually
exclusive and exhaustive? Explain.
b. What is the probability that a household is without kids?
c. What is the probability that a household is with kids and owns an Apple
product?
d. What is the probability that a household is without kids and does not own an
Apple product?
10. As part of the 2010 financial overhaul, bank regulators are renewing efforts to
require Wall Street executives to cut back on bonuses (The Wall Street Journal,
March 5, 2019). It is known that 10 out of 15 members of the board of directors of
a company are in favor of the bonus. Suppose two members were randomly
selected by the media.
a. What is the probability that both of them were in favor of the bonus?
b. What is the probability that neither of them was in favor of the bonus?
11. Christine Wong has asked Dave and Mike to help her move into a new apartment
on Sunday morning. She has asked them both, in case one of them does not
show up. From past experience, Christine knows that there is a 40% chance that
Dave will not show up and a 30% chance that Mike will not show up. Dave and
Mike do not know each other and their decisions can be assumed to be
independent.
a. What is the probability that both Dave and Mike will show up?
b. What is the probability that at least one of them will show up?
c. What is the probability that neither Dave nor Mike will show up?
12. It is reported that 85% of Asian, 78% of white, 70% of Hispanic, and 38% of black
children have two parents at home. Suppose there are 500 students in a
representative school, of which 280 are white, 50 are Asian, 100 are Hispanic,
and 70 are black.
a. Are the events “Asian” and “black” mutually exclusive and exhaustive? Explain.
b. What is the probability that a child is not white?
c. What is the probability that a child is white and has both parents at home?
d. What is the probability that a child is Asian and does not have both parents at
home?
13. Surgery for a painful, common back condition often results in significantly reduced
back pain and better physical function than treatment with drugs and physical
therapy. A researcher followed 803 patients, of whom 398 ended up getting
surgery. After two years, of those who had surgery, 63% said they had a major
improvement in their condition, compared with 29% among those who received
nonsurgical treatment.
a. What is the probability that a patient had surgery? What is the probability that a
patient did not have surgery?
b. What is the probability that a patient had surgery and experienced a major
improvement in his or her condition?
 c. What is the probability that a patient received nonsurgical treatment and
experienced a major improvement in his or her condition?
14. Despite the availability of several modes of transportation, including metro and
ride-booking services, most people in the Washington, D.C., area continue to
drive their own cars to get around (The Washington Post, June 6, 2019).
According to a survey, 62% of area adults use their own cars daily. Suppose only
38% of the area’s adults under 35 use their own cars daily. It is known that 43% of
the area’s adults are under 35.
a. What is the probability that a Washington, D.C., area adult is under 35 and uses
his/her own car daily?
b. If a Washington, D.C., area adult uses his/her own car daily, what is the
probability that he/she is under 35?

page 148

4.2 THE TOTAL PROBABILITY RULE

AND BAYES’ THEOREM
LO 4.2
Apply the total probability rule and Bayes’ theorem.

In this section, we present two important rules in probability theory:

the total probability rule and Bayes’ theorem. The total probability
rule is a useful tool for breaking the computation of a probability into
distinct cases. Bayes’ theorem uses this rule to update the
probability of an event that has been affected by a new piece of
evidence.

The Total Probability Rule and Bayes’

Theorem
Often in business the probability of an event is not readily apparent
from the given information. The total probability rule expresses the
probability of an event in terms of joint or conditional probabilities.
Let P(A) denote the probability of an event of interest. We can
express P(A) as the sum of probabilities of the intersections of A with
some mutually exclusive and exhaustive events corresponding to an
experiment. For instance, consider event B and its complement Bc.
Figure 4.3 shows the sample space partitioned into these two
mutually exclusive and exhaustive events. The circle, representing
event A, consists entirely of its intersections with B and Bc.
According to the total probability rule, P(A) equals the sum of P(A ∩
B) and P(A ∩ Bc).

FIGURE 4.3 The total probability rule: P(A) = P(A ∩ B) + P(A ∩ Bc)

Oftentimes the joint probabilities needed to compute the total

probability are not explicitly specified. Therefore, we use the
multiplication rule to derive these probabilities from the conditional
probabilities as P(A ∩ B) = P(A ∣ B)P(B) and P(A ∩ Bc) = P(A ∣
Bc)P(Bc).
The total probability rule is also needed to derive Bayes’ theorem.
Bayes’ theorem is a procedure for updating probabilities based on
new information. The original probability is an unconditional
probability called a prior probability, in the sense that it reflects only
what we know now before the arrival of any new information. On the
basis of new information, we update the prior probability to arrive at
a conditional probability called a posterior probability.
Suppose we know that 99% of the individuals who take a lie
detector test tell the truth. Therefore, the prior probability of telling
the truth is 0.99. Suppose an individual takes the lie detector test
and the results indicate that the individual lied. Bayes’ theorem
updates a prior probability to compute a posterior probability, which
in the above example is essentially a conditional probability based
on the information that the lie detector has detected a lie.
Let P(B) denote the prior probability and P(B ∣ A) the posterior
probability. Note that the posterior probability is conditional on event
A, representing new information. Recall the conditional probability
formula from Section 4.1:

page 149

In some instances, we may have to evaluate P(B ∣ A), but we do not

have explicit information on P(A ∩ B) or P(A). However, given
information on P(B), P(A ∣ B), and P(A ∣ Bc), we can use the total
probability rule to find P(B ∣ A), as shown in the following definition
box.

THE TOTAL PROBABILITY RULE AND BAYES’ THEOREM

The posterior probability P(B ∣ A) can be found using the
information on the prior probability P(B) along with the
conditional probabilities P(A ∣ B) and P(A ∣ Bc) as
 or equivalently,

In the above formula, we have used Bayes’ theorem to update the

prior probability P(B) to the posterior probability P(B ∣ A).
Equivalently, we can use Bayes’ theorem to update the prior
probability P(A) to derive the posterior probability P(A ∣ B) by
interchanging the events A and B in the above formula.

EXAMPLE 4.6
In a lie-detector test, an individual is asked to answer a series
of questions while connected to a polygraph (lie detector). This
instrument measures and records several physiological
responses of the individual on the basis that false answers will
produce distinctive measurements. Assume that 99% of the
individuals who go in for a polygraph test tell the truth. These
tests are considered to be 95% reliable. In other words, there is
a 95% chance that the test will detect a lie if an individual
actually lies. Let there also be a 0.5% chance that the test
erroneously detects a lie even when the individual is telling the
truth. An individual has just taken a polygraph test and the test
has detected a lie. What is the probability that the individual
was actually telling the truth?
SOLUTION:
First we define some events and their associated probabilities.
Let D and T correspond to the events that the polygraph
detects a lie and that an individual is telling the truth,
respectively. We are given that P(T ) = 0.99, implying that P(T
c) = 1 − 0.99 = 0.01. In addition, we formulate P(D ∣ T c) = 0.95

and P(D ∣ T ) = 0.005. We need to find P(T ∣ D) when we are

not explicitly given P(D ∩ T ) and P(D). We can use Bayes’
theorem to find

Although we can use this formula to solve the problem directly,

we use Table 4.2 to help solve the problem systematically.

TABLE 4.2 Computing Posterior Probabilities for Example 4.6

Prior Conditional Joint Posterior

Probability Probability Probability Probability

P(T ) = P(D ∣ T ) = P(D ∩ T ) = P(T ∣ D) =

0.99 0.005 0.00495 0.3426

P(T c) = P(D ∣ T c) = P(D ∩ T c) = P(T c ∣ D) =

0.01 0.95 0.00950 0.6574

P(T ) + P(T P(D) = P(T ∣ D) + P(T

c) = 1 0.01445 c ∣ D) = 1

page 150
 The first column presents prior probabilities and the second
column shows related conditional probabilities. We first
compute the denominator of Bayes’ theorem by using the total
probability rule, P(D) = P(D ∩ T ) + P(D ∩ T c). Joint
probabilities are calculated as products of conditional
probabilities with their corresponding prior probabilities. For
instance, in Table 4.2, in order to obtain P(D ∩ T ), we multiply
P(D ∣ T ) with P(T ), which yields P(D ∩ T ) = 0.005 × 0.99 =
0.00495. Similarly, we find P(D ∩ T c) = 0.95 × 0.01 = 0.00950.
Thus, according to the total probability rule, P(D) = 0.00495 +

0.00950 = 0.01445. Finally,

The prior probability of an individual telling the truth is 0.99.

However, given the new information that the polygraph
detected the individual telling a lie, the posterior probability of
this individual telling the truth is now revised downward to
0.3426.

Extensions of the Total Probability Rule and

Bayes’ Theorem
So far we have used the total probability rule as well as Bayes’
theorem based on two mutually exclusive and exhaustive events,
namely, B and Bc. We can easily extend the analysis to include n
mutually exclusive and exhaustive events, B1, B2, . . . , Bn.

EXTENSIONS OF THE TOTAL PROBABILITY RULE AND

BAYES’ THEOREM
If B1, B2, . . . Bn represent n mutually exclusive and exhaustive
events, then Bayes’ theorem, for any i = 1, 2, . . . , n, extends to
 or equivalently,

EXAMPLE 4.7
Scott Myers is a security analyst for a telecommunications firm
called Webtalk. Although he is optimistic about the firm’s future,
he is concerned that its stock price will be considerably
affected by the condition of credit flow in the economy. He
believes that the probability is 0.20 that credit flow will improve
significantly, 0.50 that it will improve only marginally, and 0.30
that it will not improve at all. He also estimates that the
probability that the stock price of Webtalk will go up is 0.90 with
significant improvement in credit flow in the economy, 0.40 with
marginal improvement in credit flow in the economy, and 0.10
with no improvement in credit flow in the economy.
a. Based on Scott’s estimates, what is the probability that the
stock price of Webtalk goes up?
b. If we know that the stock price of Webtalk has gone up, what
is the probability that credit flow in the economy has improved
significantly?
SOLUTION:
As always, we first define the relevant events and their
associated probabilities. Let S, M, and N denote significant,
 marginal, and no improvement in credit flow, respectively. Then
P(S) = 0.20, P(M) = 0.50, and P(N) = 0.30. In addition, if we
allow G to denote an increase in stock price, we formulate P(G
∣ S) = 0.90, P(G ∣ M) = 0.40, and P(G ∣ N) = 0.10. We need to
calculate P(G) in part a and P(S ∣ G) in part b. Table 4.3 aids in
assigning probabilities.

page 151

TABLE 4.3 Computing Posterior Probabilities for Example 4.7

Prior Conditional Joint Posterior

Probability Probability Probability Probability

P(S) = 0.20 P(G ∣ S) = P(G ∩ S) = P(S ∣ G) =

0.90 0.18 0.4390

P(M) = 0.50 P(G ∣ M) = P(G ∩ M) = P(M ∣ G) =

0.40 0.20 0.4878

P(N) = 0.30 P(G ∣ N) = P(G ∩ N) = P(N ∣ G) =

0.10 0.03 0.0732

P(S) + P(M) P(G) = 0.41 P(S ∣ G) + P(M ∣

+ P(N) = 1 G) + P(N ∣ G) =
1

a. In order to calculate P(G), we use the total probability rule,

P(G) = P(G ∩ S) + P(G ∩ M) + P(G ∩ N). The joint
probabilities are calculated as products of conditional
probabilities with their corresponding prior probabilities. For
instance, in Table 4.3, P(G ∩ S) = P(G ∣ S)P(S) = 0.90 × 0.20
= 0.18. Therefore, the probability that the stock price of
Webtalk goes up equals P(G) = 0.18 + 0.20 + 0.03 = 0.41.
b. According to Bayes’ theorem, .

We use the total probability rule in the denominator to find

P(G) = 0.18 + 0.20 + 0.03 = 0.41. Therefore,
Note that the prior probability of a

significant improvement in credit flow is revised upward from

0.20 to a posterior probability of 0.4390.

EXAMPLE 4.8
We can now answer the questions posed by Alexis Lewis in the
introductory case of this chapter. Use the information gathered
by Alexis to
a. Formulate relevant conditional, unconditional, and joint
probabilities.
b. Calculate the probability of all Americans who support the
legalization of marijuana.
SOLUTION:

a. For ease of interpretation, let M, G, B, and S denote

“Millennial,” “Generation X,” “Baby Boomer,” and “Silent”
generations, respectively. Based on data from the study, the
following probability statements can be formulated with
respect to the relative size of each generation’s voting bloc:
P(M) = 0.31, P(G) = 0.25, P(B) = 0.31, P(S) = 0.13.
Now let L denote “support for legalizing marijuana.” Again,
based on data from the study, conditional probabilities can be
specified as P(L ∣ M) = 0.71, P(L ∣ G) = 0.57, P(L ∣ B) = 0.56,
and P(L ∣ S) = 0.33.
 Therefore, the probability that a randomly selected adult
supports legal marijuana and is in the Millennial generation is
determined as P(L ∩ M) = 0.71 × 0.31 = 0.2201. Similarly, P(L
∩ G) = 0.1425, P(L ∩ B) = 0.1736, and P(L ∩ S) = 0.0429.
b. By combining all generations, we deduce the total probability
of support for legalizing marijuana as P(L) = P(L ∩ M) + P(L ∩
G) + P(L ∩ B) + P(L ∩ S) = 0.2201 + 0.1425 + 0.1736 +
0.0429 = 0.5791.

page 152

SYNOPSIS OF INTRODUCTORY CASE

Driven by growing public support, the legalization of marijuana in America has been
moving at a breakneck speed in recent years. Changing demographics can help
explain how the tide has turned in marijuana’s favor, especially with Millennials, who
are on the verge of becoming the nation’s largest living generation. A recent Pew
Research Study provides interesting data regarding support for marijuana
legalization. Two factors seem to drive support for the issue: generation (or age
group) and the relative size of a generation’s voting bloc. Based on data from the
study, Millennials and Baby Boomers have the most voting power, each comprising
31% of the voting population; the Generation X and Silent generations represent
25% and 13% of the voting population, respectively. Given other conditional
probabilities, we determine that a staggering 57.91% of Americans supported the
legalization of marijuana.
To put it in perspective, suppose that there are 1,000 randomly selected adult
attendees at a conference. The results imply that there would be about 310
Millennial, 250 Generation X, 310 Baby Boomer, and 130 Silent attendees. Further,
the supporters of marijuana legalization would include about 220 Millennial, 143
Generation X, 174 Baby Boomer, and 43 Silent attendees.
Millennials are now as large a political force as Baby Boomers. In general,
Millennials tend to be liberal on social issues such as gay rights, immigration, and
marijuana. This shift in population has not gone unnoticed by political parties, which
all hope to court the more than 75 million of these eligible young voters.

EXERCISES 4 2
 EXERCISES 4.2
Applications
15. Christine has always been weak in mathematics. Based on her performance prior
to the final exam in Calculus, there is a 40% chance that she will fail the course if
she does not have a tutor. With a tutor, her probability of failing decreases to 10%.
There is only a 50% chance that she will find a tutor at such short notice.
a. What is the probability that Christine fails the course?
b. Christine ends up failing the course. What is the probability that she had found
a tutor?
16. An analyst expects that 20% of all publicly traded companies will experience a
decline in earnings next year. The analyst has developed a ratio to help forecast
this decline. If the company is headed for a decline, there is a 70% chance that
this ratio will be negative. If the company is not headed for a decline, there is a
15% chance that the ratio will be negative. The analyst randomly selects a
company and its ratio is negative. What is the posterior probability that the
company will experience a decline?
17. The State Police are trying to crack down on speeding on a particular portion of
the Massachusetts Turnpike. To aid in this pursuit, they have purchased a new
radar gun that promises greater consistency and reliability. Specifically, the gun
advertises ± one-mile-per-hour accuracy 98% of the time; that is, there is a 0.98
probability that the gun will detect a speeder, if the driver is actually speeding.
Assume there is a 1% chance that the gun erroneously detects a speeder even
when the driver is below the speed limit. Suppose that 95% of the drivers drive
below the speed limit on this stretch of the Massachusetts Turnpike.
a. What is the probability that the gun detects speeding and the driver was
speeding?
b. What is the probability that the gun detects speeding and the driver was not
speeding?
c. Suppose the police stop a driver because the gun detects speeding. What is the
probability that the driver was actually driving below the speed limit?
18. According to data from the National Health and Nutrition Examination Survey,
33% of white, 49.6% of black, 43% of Hispanic, and 8.9% of Asian women are
obese. In a representative town, 48% of women are white, 19% are black, 26%
are Hispanic, and the remaining 7% are Asian.
a. Find the probability that a randomly selected woman in this town is obese.
b. Given that a woman is obese, what is the probability that she is white?
c. Given that a woman is obese, what is the probability that she is black?
d. Given that a woman is obese, what is the probability that she is Asian?
19. An analyst thinks that next year there is a 20% chance that the world economy
will be good, a 50% chance that it will be neutral, and a 30% chance that it will be
poor. She also predicts probabilities that the performance of a start-up firm,
 Creative Ideas, will be good, neutral, or poor for each of the economic states of
the world economy. The following table presents probabilities for three states of
the world economy and the corresponding conditional probabilities for Creative
Ideas.

page 153
Probability
State of Performance Conditional
of
the World of Creative Probability of
Economic
Economy Ideas Creative Ideas
State

Good 0.20 Good 0.60

Neutral 0.30
Poor 0.10

Neutral 0.50 Good 0.40

Neutral 0.30
Poor 0.30

Poor 0.30 Good 0.20

Neutral 0.30
Poor 0.50

a. What is the probability that the performance of the world economy will be
neutral and that of Creative Ideas will be poor?
b. What is the probability that the performance of Creative Ideas will be poor?
c. The performance of Creative Ideas was poor. What is the probability that the
performance of the world economy had also been poor?
20. A crucial game of the Los Angeles Lakers basketball team depends on the health
of its key player. According to his doctor’s report, there is a 40% chance that he
will be fully fit to play, a 30% chance that he will be somewhat fit to play, and a
30% chance that he will not be able to play at all. The coach has estimated the
chances of winning at 80% if the player is fully fit, 60% if he is somewhat fit, and
40% if he is unable to play.
a. What is the probability that the Lakers will win the game?
b. You have just heard that the Lakers won the game. What is the probability that
the key player had been fully fit to play in the game?
21. A 2015 national survey by the Washington Post–Kaiser Family Foundation finds
that there is a big sex divide between Americans when identifying as feminist or
strong feminist. The results of the survey are shown in the following table. In
addition, per the 2010 U.S. Census Current Population Survey, 50.8% of the
American population is female and 49.2% is male.
 Sex Feminist or Strong Feminist

Female 66%

Male 41%

a. Calculate the probability that a randomly selected American adult is a female

who also identifies as feminist or strong feminist.
b. Calculate the probability that a randomly selected American adult is a male who
also identifies as feminist or strong feminist.
c. What percentage of American adults identifies as feminist or strong feminist?
d. If a randomly selected American adult identifies as feminist or strong feminist,
what is the probability that this adult is a female?
22. There is a growing public support for marijuana law reform, with polls showing
more than half the country is in favor of some form of marijuana legalization.
However, opinions on marijuana are divided starkly along political party lines. The
results of a recent Pew Research Center survey are shown in the following table.
In addition, assume that 27% of Americans identify as Republicans, 30% as
Democrats, and 43% as independents.

Political Party Support

Republican 41%

Democrat 66%

Independent 63%

a. Calculate the probability that a randomly selected American adult supports

marijuana legalization and is a Republican.
b. Calculate the probability that a randomly selected American adult supports
marijuana legalization and is a Democrat.
c. Calculate the probability that a randomly selected American adult supports
marijuana legalization and is an Independent.
d. What percentage of American adults supports marijuana legalization?
e. If a randomly selected American adult supports marijuana legalization, what is
the probability that this adult is a Republican?
4.3 RANDOM VARIABLES AND
DISCRETE PROBABILITY
DISTRIBUTIONS
LO 4.3
Describe a discrete random variable and its probability distribution.

We often have to make important decisions in the face of uncertainty.

For example, a car dealership has to determine the number of cars
to hold on its lot when the actual demand for cars is unknown.
Similarly, an investor has to select a portfolio when the actual
outcomes of investment returns are not known. This uncertainty is
captured by what we call a random variable. A random variable
summarizes outcomes of an experiment with numerical values.

page 154

We generally use the uppercase letter X to denote a random

variable and the lowercase letter x to denote the value that X may
assume. A discrete random variable assumes a countable number
of distinct values such as x1, x2, x3, and so on. A continuous
random variable, on the other hand, is characterized by
uncountable values within an interval. Unlike the case of a discrete
random variable, we cannot describe the possible values of a
continuous random variable X with a list x1, x2, . . . because the
value (x1 + x2)/2, not in the list, might also be possible.

DISCRETE VERSUS CONTINUOUS RANDOM VARIABLES

A random variable is a function that assigns numerical values
to the outcomes of an experiment. A discrete random variable
assumes a countable number of distinct values. A continuous
 random variable, on the other hand, is characterized by
uncountable values in an interval.

Examples of discrete random variables include the number of sales

people who hit their target for the quarter, the number of employees
leaving a firm, or the number of firms filing for bankruptcy in a given
month. Similarly, the return on a mutual fund, time to completion of a
task, or the volume of beer sold as 16 ounces are examples of
continuous random variables.

The Discrete Probability Distribution

Every discrete random variable is associated with a probability
distribution that provides the probability that the random variable X
assumes a particular value x, or equivalently, P(X = x). Random
variables can also be defined in terms of their cumulative
distribution function, or, equivalently, P(X ≤ x). All discrete
probability distributions satisfy the following two properties.

TWO KEY PROPERTIES OF DISCRETE PROBABILITY

DISTRIBUTIONS
The probability of each value x is a value between 0 and 1, or,
equivalently, 0 ≤ P(X = x) ≤ 1.
The sum of the probabilities equals 1. In other words, ΣP(X =
xi) = 1, where the sum extends over all values x of X.

Summary Measures of a Discrete Random

Variable
The analysis of probability distributions is useful because it allows us
to calculate various probabilities associated with the different values
that the random variable assumes. In addition, it allows us to
calculate summary measures for a random variable. These summary
measures include the mean, the variance, and the standard
deviation.
One of the most important probabilistic concepts in statistics is
that of the expected value, also referred to as the mean. The
expected value of the discrete random variable X, denoted by E(X),
or simply μ, is a weighted average of all possible values of X. The
variance and the standard deviation of a random variable are
measures of variability.

SUMMARY MEASURES OF A DISCRETE RANDOM

VARIABLE
Consider a discrete random variable X with values x1, x2, x3, . .
. , which occur with probabilities P(X = xi).

The expected value of X, denoted as E(X) or μ, is calculated as

page 155

The variance of X, denoted as Var(X) or σ2, is calculated as

The standard deviation of X, denoted as SD(X) or σ, is

calculated as
 EXAMPLE 4.9
Brad Williams is the owner of a large car dealership in Chicago.
Brad decides to construct an incentive compensation program
that equitably and consistently compensates employees on the
basis of their performance. He offers an annual bonus of
$10,000 for superior performance, $6,000 for good
performance, $3,000 for fair performance, and $0 for poor
performance. Based on prior records, he expects an employee
to perform at superior, good, fair, and poor performance levels
with probabilities 0.15, 0.25, 0.40, and 0.20, respectively.
a. Calculate the expected value of the annual bonus amount.
b. Calculate the variance and the standard deviation of the
annual bonus amount.
c. What is the total annual amount that Brad can expect to pay in
bonuses if he has 25 employees?
SOLUTION:

a. Let the random variable X denote the bonus amount (in

$1,000s) for an employee. The first and second columns of
Table 4.4 represent the probability distribution of X. The
calculations for the mean are provided in the third column. We
weigh each outcome by its respective probability, xiP(X = xi),
and then sum these weighted values. Thus, as shown at the
bottom of the third column, E(X) = μ = ΣxiP(X = xi) = 4.2, or
$4,200.
TABLE 4.4 Calculations for Example 4.9

xi P(X = xi) xiP(X = xi) (xi − μ)2P(X = xi)

 xi P(X = xi) xiP(X = xi) (xi − μ)2P(X = xi)

10 0.15 10 × 0.15 = 1.5 (10 − 4.2)2 × 0.15 = 5.05

6 0.25 6 × 0.25 = 1.5 (6 − 4.2)2 × 0.25 = 0.81

3 0.40 3 × 0.40 = 1.2 (3 − 4.2)2 × 0.40 = 0.58

0 0.20 0 × 0.20 = 0   (0 − 4.2)2 × 0.20 = 3.53

Total = 4.2 Total = 9.97

b. The last column of Table 4.4 shows the calculation for the
variance. We first calculate each xi’s squared difference from
the mean (xi − μ)2, weigh each value by the appropriate
probability, (xi − μ)2P(X = xi), and then sum these weighted
squared differences. Thus, as shown at the bottom of the last
column, Var(X) = σ2 = Σ(xi − μ)2P(X = xi) = 9.97, or 9.97 (in
($1,000s)2). The standard deviation is the positive square root
of the variance,

c. In part a we found that the expected bonus of an employee is

$4,200. Because Brad has 25 employees, he can expect to
pay $4,200 × 25 = $105,000 in bonuses.

page 156

EXERCISES 4.3
 Applications
23. Fifty percent of the customers who go to Auto Center for tires buy four tires and
30% buy two tires. Moreover, 18% buy fewer than two tires, with 5% buying none.
a. What is the probability that a customer buys three tires?
b. Construct a cumulative probability distribution for the number of tires bought.
24. Jane Wormley is a professor of management at a university. She expects to be
able to use her grant money to fund up to two students for research assistance.
While she realizes that there is a 5% chance that she may not be able to fund any
student, there is an 80% chance that she will be able to fund two students.
a. What is the probability that Jane will fund one student?
b. Construct a cumulative probability distribution of the random variable defined as
the number of students that Jane will be able to fund.
25. A marketing firm is considering making up to three new hires. Given its specific
needs, the management feels that there is a 60% chance of hiring at least two
candidates. There is only a 5% chance that it will not make any hires and a 10%
chance that it will make all three hires.
a. What is the probability that the firm will make at least one hire?
b. Find the expected value and the standard deviation of the number of hires.
26. An appliance store sells additional warranties on its refrigerators. Twenty percent
of the buyers buy the limited warranty for $100 and 5% buy the extended warranty
for $200. What is the expected revenue for the store from the warranty if it sells
120 refrigerators?
27. Organizers of an outdoor summer concert in Toronto are concerned about the
weather conditions on the day of the concert. They will make a profit of $25,000
on a clear day and $10,000 on a cloudy day. They will take a loss of $5,000 if it
rains. The weather channel has predicted a 60% chance of rain on the day of the
concert. Calculate the expected profit from the concert if the likelihood is 10% that
it will be sunny and 30% that it will be cloudy.
28. The manager of a publishing company plans to give a $20,000 bonus to the top
15%, $10,000 to the next 30%, and $5,000 to the next 10% of sales
representatives. If the publishing company has a total of 200 sales
representatives, what is the expected bonus that the company will pay?
29. You are considering buying insurance for your new laptop computer, which you
have recently bought for $1,500. The insurance premium for three years is $80.
Over the three-year period there is an 8% chance that your laptop computer will
require work worth $400, a 3% chance that it will require work worth $800, and a
2% chance that it will completely break down with a scrap value of $100. Should
you buy the insurance?
30. An investor considers investing $10,000 in the stock market. He believes that the
probability is 0.30 that the economy will improve, 0.40 that it will stay the same,
and 0.30 that it will deteriorate. Further, if the economy improves, he expects his
investment to grow to $15,000, but it can also go down to $8,000 if the economy
 deteriorates. If the economy stays the same, his investment will stay at $10,000.
What is the expected value of his investment?

4.4 THE BINOMIAL AND THE

POISSON DISTRIBUTIONS
LO 4.4
Calculate probabilities for binomial and Poisson distributions.

Different types of experiments generate different discrete probability

distributions. A simple, yet widely used, distribution is called the
discrete uniform distribution. This distribution has a finite number
of values where each value is equally likely. Suppose we know that
the weekly production at a facility is between 31 and 40 units and
that production follows a discrete uniform distribution. Here, the
random variable can only assume one out of ten equally likely values
between 31 and 40. For example, the probability that weekly
production is 34 is 1/10. Similarly, the probability that weekly
production is less than 34 is 3/10.
In this section, we focus on the binomial and the Poisson
probability distributions.

The Binomial Distribution

Before we can discuss the binomial distribution, we first must ensure
that the experiment satisfies the conditions of a Bernoulli process,
which is a particular type of experiment described below.

page 157

A BERNOULLI PROCESS
 A Bernoulli process consists of a series of n independent and
identical trials of an experiment such that on each trial:
There are only two possible outcomes, conventionally labeled
success and failure; and
The probabilities of success and failure remain the same from
trial to trial.

We use p to denote the probability of success, and, therefore, 1 − p

is the probability of failure.
A binomial random variable x is defined as the number of
successes achieved in the n trials of a Bernoulli process. The
possible values of x include 0, 1, . . . , n. Many experiments fit the
conditions of a Bernoulli process. For instance:
A customer defaults or does not default on a loan.
A consumer reacts positively or negatively to a social media
campaign.
A drug is either effective or ineffective.
A college graduate applies or does not apply to graduate school.
Our goal is to attach probabilities to various outcomes of a Bernoulli
process. The result is a binomial distribution.

THE BINOMIAL DISTRIBUTION

For a binomial random variable X, the probability of x
successes in n Bernoulli trials is

for x = 0, 1, 2, . . . , n. By definition, 0! = 1.
The formula consists of two parts, which we explain with a scenario
where historically 85% of the customers in a store make a purchase.
Suppose we want to compute the probability that exactly one of the
three customers in the store will make a purchase.

The first term, , tells us how many sequences with x

successes and n − x failures are possible in n trials. We refer to

the first term as the binomial coefficient, which is really the familiar
combination formula used to find the number of ways to choose x
objects from a total of n objects, where the order in which the x
objects are listed does not matter. For instance, in order to
calculate the number of sequences that contain exactly one
success in three trials of a Bernoulli process, we substitute x = 1
and n = 3 into the formula and calculate

So there are three sequences having

exactly 1 success.
The second part of the equation, px(1 − p)n − x, represents the
probability of any particular sequence with x successes and n − x
failures. For example, we can obtain the probability of one success
in three trials as 0.85 × 0.15 × 0.15 = (0.85)1 × (0.15)2 = 0.0191. In
other words, each sequence consisting of 1 success in 3 trials has
a 1.91% chance of occurring.
We obtain the overall probability of getting 1 success in 3 trials as
P(X = 1) = 3 × 0.0191 = 0.0573.

page 158

Moreover, we could use the formulas shown in Section 4.3 to

calculate the expected value and the variance for any binomial
random variable. For the binomial distribution, these formulas
simplify to E(X) = np and Var(X) = np(1 − p).

EXAMPLE 4.10
In the United States, about 30% of adults have four-year
college degrees (US Census, July 31, 2018). Suppose five
adults are randomly selected.
a. What is the probability that none of the adults has a college
degree?
b. What is the probability that no more than two of the adults
have a college degree?
c. What is the probability that at least two of the adults have a
college degree?
d. Calculate the expected number of adults with a college
degree.
SOLUTION: First, this problem satisfies the conditions for a
Bernoulli process with a random selection of five adults, n = 5.
Here, an adult either has a college degree, with probability p =
0.30, or does not have a college degree, with probability 1 − p
= 1 − 0.30 = 0.70.
a. In order to find the probability that none of the adults has a
college degree, we let x = 0 and find

In other words, from a random sample of five adults, there is a

16.81% chance that none of the adults has a college degree.
b. We find the probability that no more than two adults have a
college degree as
 We have already found P(X = 0) from part a. So we now
compute P(X = 1) and P(X = 2):

Next we sum the three relevant probabilities and obtain P(X ≤

2) = 0.1681 + 0.3602 + 0.3087 = 0.8370. From a random
sample of five adults, there is an 83.7% likelihood that no
more than two of them will have a college degree.
c. We find the probability that at least two adults have a college
degree as

A simpler method uses one of the key properties of a

probability distribution, which states that the sum of the
probabilities over all values of X equals 1. Therefore, P(X ≥ 2)
can be written as 1 − [P(X = 0) + P(X = 1)]. We have already
calculated P(X = 0) and P(X = 1) from parts a and b, so

page 159
From a random sample of five adults, there is a
47.17% likelihood that at least two adults will have a college
 degree.
d. We calculate the expected number of adults with a college
degree as

The Poisson Distribution

Another important discrete probability distribution is the Poisson
distribution. It is particularly useful in problems that deal with finding
the number of occurrences of a certain event over time or space,
where space refers to area or region. For simplicity, we call these
occurrences “successes.” Before we can discuss the Poisson
distribution, we first must ensure that our experiment satisfies the
conditions of a Poisson process.

A POISSON PROCESS
An experiment satisfies a Poisson process if
The number of successes within a specified time or space
interval equals any integer between zero and infinity.
The number of successes counted in nonoverlapping intervals
are independent.
The probability of success in any interval is the same for all
intervals of equal size and is proportional to the size of the
interval.

For a Poisson process, we define the number of successes achieved

in a specified time or space interval as a Poisson random variable.
Like the Bernoulli process, many experiments fit the conditions of a
Poisson process. For instance:
The number of customers who use a new banking app in a day.
The number of spam e-mails received in a month.
The number of defects in a 50-yard roll of fabric.
The number of bacteria in a specified culture.
We use the following formula for calculating probabilities associated
with a Poisson random variable.

THE POISSON DISTRIBUTION

For a Poisson random variable X, the probability of x successes
over a given interval of time or space is

for x = 0, 1, 2, . . . , where μ is the mean number of successes

and e ≈ 2.718 is the base of the natural logarithm.

If X is a Poisson random variable, then Var(X) = E(X) = μ.

page 160

EXAMPLE 4.11
Anne is concerned about staffing needs at the Starbucks that
she manages. She believes that the typical Starbucks
customer averages 18 visits to the store over a 30-day month.
a. How many visits should Anne expect in a 5-day period from a
typical Starbucks customer?
b. What is the probability that a customer visits the chain five
times in a 5-day period?
c. What is the probability that a customer visits the chain no
more than two times in a 5-day period?
d. What is the probability that a customer visits the chain at least
three times in a 5-day period?
SOLUTION: In applications of the Poisson distribution, we first
determine the mean number of successes in the relevant time
or space interval. We use the Poisson process condition that
the probability that success occurs in any interval is the same
for all intervals of equal size and is proportional to the size of
the interval. Here, the relevant mean will be based on the rate
of 18 visits over a 30-day month.
a. Given the rate of 18 visits over a 30-day month, we can write
the mean for the 30-day period as μ30 = 18. For this problem,
we compute the proportional mean for a five-day period as μ5

= 3 because In other words, on average, a typical

Starbucks customer visits the store three times over a five-day

period.
b. In order to find the probability that a customer visits the chain
five times in a five-day period, we calculate

c. For the probability that a customer visits the chain no more

than two times in a five-day period, we find P(X ≤ 2) = P(X = 0)
+ P(X = 1) + P(X = 2). We calculate the individual probabilities,
and then find the sum:
 Thus, P(X ≤ 2) = 0.0498 + 0.1494 + 0.2241 = 0.4233. There is
approximately a 42% chance that a customer visits the chain
no more than two times in a five-day period.
d. We find P(X ≥ 3) as 1 − [P(X = 0) + P(X = 1) + P(X = 2)].
Based on the probabilities in part c, we have P(X ≥ 3) = 1 −
[0.0498 + 0.1494 + 0.2241] = 1 − 0.4233 = 0.5767. Thus,
there is about a 58% chance that a customer will frequent the
chain at least three times in a five-day period.

page 161

Using Excel and R to Obtain Binomial and

Poisson Probabilities
As you may have noticed, it is somewhat tedious and cumbersome
to solve binomial and Poisson distribution problems using the
formulas. This issue becomes even more pronounced when X
assumes a wide range of values. In Examples 4.12 and 4.13, we will
show how to solve binomial and Poisson distribution problems with
Excel and R.

EXAMPLE 4.12
 People turn to social media to stay in touch with friends and
family members, connect with old friends, catch the news, look
for employment, and be entertained. According to a recent
survey, 68% of all U.S. adults are Facebook users. Consider a
sample of 100 randomly selected American adults.
a. What is the probability that exactly 70 American adults are
Facebook users?
b. What is the probability that no more than 70 American adults
are Facebook users?
c. What is the probability that at least 70 American adults are
Facebook users?
SOLUTION:
We let X denote the number of American adults who are
Facebook users. We also know that p = 0.68 and n = 100.
Using Excel
We use Excel’s BINOM.DIST function to calculate binomial
probabilities. In order to find P(X = x), we enter
=BINOM.DIST(x, n, p, 0) where x is the number of successes,
n is the number of trials, and p is the probability of success. If
we enter a “1” for the last argument in the function, then Excel
returns P(X ≤ x).
a. In order to find the probability that exactly 70 American adults
are Facebook users, P(X = 70), we enter =BINOM.DIST(70,
100, 0.68, 0) and Excel returns 0.0791.
b. In order to find the probability that no more than 70 American
adults are Facebook users, P(X ≤ 70), we enter
=BINOM.DIST(70, 100, 0.68, 1) and Excel returns 0.7007.
c. In order to find the probability that at least 70 American adults
are Facebook users, P(X ≥ 70) = 1 − P(X ≤ 69), we enter =1-
BINOM.DIST(69, 100, 0.68, 1) and Excel returns 0.3784.
Using R
 We use R’s dbinom and pbinom functions to calculate
binomial probabilities. In order to calculate P(X = x), we enter
dbinom(x, n, p) where x is the number of successes, n is the
number of trials, and p is the probability of success. In order to
calculate P(X ≤ x), we enter pbinom(x, n, p).
a. In order to find P(X = 70), we enter:
> dbinom(70, 100, 0.68)
And R returns: 0.07907911.
b. In order to find P(X ≤ 70), we enter:
> pbinom(70, 100, 0.68)
And R returns: 0.7006736.
c. In order to find P(X ≥ 70) = 1 − P(X ≤ 69), we enter:
> 1 − pbinom(69, 100, 0.68)
And R returns: 0.3784055.

page 162

EXAMPLE 4.13
The sales volume of craft beer continues to grow, amounting to
24% of the total beer market in the U.S. (USA Today, April 2,
2019). It has been estimated that 1.5 craft breweries open
every day. Assume this number represents an average that
remains constant over time.
a. What is the probability that no more than 10 craft breweries
open every week?
b. What is the probability that exactly 10 craft breweries open
every week?
 SOLUTION:We let X denote the number of craft breweries that
open every week and compute the weekly mean, μ = 1.5 × 7 =
10.5.
Using Excel
We use Excel’s POISSON.DIST function to calculate Poisson
probabilities. In order to find P(X = x), we enter
=POISSON.DIST(x, μ, 0) where x is the number of successes
over some interval and μ is the mean over this interval. If we
enter a “1” for the last argument in the function, then Excel
returns P(X ≤ x).
a. In order to find the probability that no more than 10 craft
breweries open every week, P(X ≤ 10), we enter
=POISSON.DIST(10, 10.5, 1) and Excel returns 0.5207.
b. In order to find the probability that exactly 10 craft breweries
open every week, P(X = 10), we enter =POISSON.DIST(10,
10.5, 0) and Excel returns 0.1236.
Using R
We use R’s dpois and ppois functions to calculate Poisson
probabilities. In order to calculate P(X = x), we enter dpois(x, μ)
where x is the number of successes over some interval and μ
is the mean over this interval. In order to calculate P(X ≤ x), we
enter ppois(x, μ).
a. In order to find P(X ≤ 10), we enter:
> ppois(10, 10.5)
And R returns: 0.5207381.
b. In order to find P(X = 10), we enter:
> dpois(10, 10.5)
And R returns: 0.1236055.

EXERCISES 4.4
 Applications
31. At a local community college, 40% of students who enter the college as freshmen
go on to graduate. Ten freshmen are randomly selected.
a. What is the probability that none of them graduates from the local community
college?
b. What is the probability that at most nine will graduate from the local community
college?
c. What is the expected number that will graduate?
32. As of 2018, 30% of Americans have confidence in U.S. banks, which is still below
the pre-recession level of 41% reported in June 2007 (www.gallup.com, June 28,
2018).
a. What is the probability that fewer than half of four Americans in 2018 have
confidence in U.S. banks?
b. What would have been the corresponding probability in 2007?
33. Approximately 45% of Baby Boomers—those born between 1946 and 1964—are
still in the workforce (www.pewresearch.org, May 11, 2015). Six Baby Boomers
are selected at random.
a. What is the probability that exactly one of the Baby Boomers is still in the
workforce?

page 163
b. What is the probability that at least five of the Baby Boomers
are still in the workforce?
c. What is the probability that less than two of the Baby Boomers are still in the
workforce?
d. What is the probability that more than the expected number of the Baby
Boomers are still in the workforce?
34. According to the Census Bureau projections, Hispanics will comprise 28.6% of the
total population in 2060 (CNN, March 6, 2019). For comparison, they comprised
just 18.1% of the total population in 2018.
a. What is the expected value and the standard deviation of Hispanics in a random
sample of 5,000 people in 2018?
b. What is the corresponding expected value and the standard deviation projected
for 2060?
35. The arduous task of combing and tying up long hair and a desire to assimilate has
led to approximately 25% of Sikh youths giving up on wearing turbans.
a. What is the probability that exactly two in a random sample of five Sikh youths
wear a turban?
b. What is the probability that two or more in a random sample of five Sikh youths
wear a turban?
c. What is the probability that more than the expected number of Sikh youths wear
a turban in a random sample of five Sikh youths?
 d. What is the probability that more than the expected number of Sikh youths wear
a turban in a random sample of 10 Sikh youths?
36. Researchers from leading universities have shown that divorces are contagious
(Chicago Tribune, August 16, 2018). A split-up between immediate friends
increases a person’s own chances of getting divorced from 36% to 63%, an
increase of 75%.
a. Compute the probability that more than half of four randomly selected
marriages will end in divorce if the couple’s immediate friends have split up.
b. Redo part a if it is known that none of the couple’s immediate friends has split
up.
37. Sixty percent of a firm’s employees are men. Suppose four of the firm’s
employees are randomly selected.
a. What is more likely, finding three men and one woman or two men and two
women?
b. Do you obtain the same answer as in part a if 70% of the firm’s employees had
been men?
38. The principal of an architecture firm tells her client that there is at least a 50%
chance of having an acceptable design by the end of the week. She knows that
there is only a 25% chance that any one designer would be able to do so by the
end of the week.
a. Would she be correct in her statement to the client if she asks two of her
designers to work on the design, independently?
b. If not, what if she asks three of her designers to work on the design,
independently?
39. Suppose 40% of recent college graduates plan on pursuing a graduate degree.
Fifteen recent college graduates are randomly selected.
a. What is the probability that no more than four of the college graduates plan to
pursue a graduate degree?
b. What is the probability that exactly seven of the college graduates plan to
pursue a graduate degree?
c. What is the probability that at least six but no more than nine of the college
graduates plan to pursue a graduate degree?
40. A manager at 24/7 Fitness Center is strategic about contacting open house
attendees. With her strategy, she believes that 40% of the attendees she contacts
will purchase a club membership. Suppose she contacts 20 open house
attendees.
a. What is the probability that exactly 10 of the attendees will purchase a club
membership?
b. What is the probability that no more than 10 of the attendees will purchase a
club membership?
c. What is the probability that at least 15 of the attendees will purchase a club
membership?
41. Fraud detection has become an indispensable tool for banks and credit card
companies to combat fraudulent credit card transactions. A fraud detection firm
has detected some form of fraudulent activities in 1.31%, and serious fraudulent
activities in 0.87%, of transactions. Assume that fraudulent transactions remain
stable.
a. What is the probability that fewer than 2 out of 100 transactions are fraudulent?
b. What is the probability that fewer than 2 out of 100 transactions are seriously
fraudulent?
42. New Age Solar installs solar panels for residential homes. Because of the
company’s personalized approach, it averages three home installations daily.
a. What is the probability that New Age Solar installs solar panels in at most four
homes in a day?
b. What is the probability that New Age Solar installs solar panels in at least three
homes in a day?
43. On average, there are 12 potholes per mile on a particular stretch of the state
highway. Suppose the potholes are distributed evenly on the highway.
a. Find the probability of finding fewer than two potholes in a quarter-mile stretch
of the highway.
b. Find the probability of finding more than one pothole in a quarter-mile stretch of
the highway.
44. A tollbooth operator has observed that cars arrive randomly at an average rate of
360 cars per hour.
a. Find the probability that two cars arrive during a specified one-minute period.

page 164
b. Find the probability that at least two cars arrive during a
specified one-minute period.
c. Find the probability that 40 cars arrive between 10:00 am and 10:10 am.
45. A textile manufacturing process finds that on average, two flaws occur per every
50 yards of material produced.
a. What is the probability of exactly two flaws in a 50-yard piece of material?
b. What is the probability of no more than two flaws in a 50-yard piece of material?
c. What is the probability of no flaws in a 25-yard piece of material?
46. Motorists arrive at a Gulf gas station at the rate of two per minute during morning
hours.
a. What is the probability that more than two motorists will arrive at the Gulf gas
station during a one-minute interval in the morning?
b. What is the probability that exactly six motorists will arrive at the Gulf gas
station during a five-minute interval in the morning?
c. How many motorists can an employee expect in her three-hour morning shift?
47. Airline travelers should be ready to be more flexible as airlines once again cancel
thousands of flights this summer. The Coalition for Airline Passengers Rights,
 Health, and Safety averages 400 calls a day to help stranded travelers deal with
airlines. Suppose the hotline is staffed for 16 hours a day.
a. Calculate the average number of calls in a one-hour interval, 30-minute interval,
and 15-minute interval.
b. What is the probability of exactly six calls in a 15-minute interval?
c. What is the probability of no calls in a 15-minute interval?
d. What is the probability of at least two calls in a 15-minute interval?
48. According to the Centers for Disease Control and Prevention (CDC), the aging of
the U.S. population is translating into many more visits to doctors’ offices and
hospitals. It is estimated that an average person makes four visits a year to
doctors’ offices and hospitals.
a. What are the mean and the standard deviation of an average person’s number
of monthly visits to doctors’ offices and hospitals?
b. What is the probability that an average person does not make any monthly
visits to doctors’ offices and hospitals?
c. What is the probability that an average person makes at least one monthly visit
to doctors’ offices and hospitals?
49. Last year, there were 24,584 age-discrimination claims filed with the Equal
Employment Opportunity Commission. Assume there were 260 working days in
the fiscal year for which a worker could file a claim.
a. Calculate the average number of claims filed on a working day.
b. What is the probability that exactly 100 claims were filed on a working day?
c. What is the probability that no more than 100 claims were filed on a working
day?
50. American adults are watching significantly less television than they did in previous
decades. In 2016, Nielsen reported that American adults are watching an average
of five hours and four minutes, or 304 minutes, of television per day.
a. Find the probability that an average American adult watches more than 320
minutes of television per day.
b. Find the probability that an average American adult watches more than 2,200
minutes of television per week.

4.5 THE NORMAL DISTRIBUTION

LO 4.5
Describe the normal distribution and calculate its associated probabilities.
Recall that a discrete random variable X assumes a countable
number of distinct values such as x1, x2, x3, and so on. A continuous
random variable, on the other hand, is characterized by uncountable
values because it can take on any value within an interval.
Furthermore, we can calculate the probability that a discrete random
variable X assumes a particular value x; that is, P(X = x). For
instance, for a binomial random variable, we can calculate the
probability of exactly one success in n trials; that is, P(X = 1). We
cannot make this calculation with a continuous random variable.
The probability that a continuous random variable assumes a
particular value x is zero; that is, P(X = x) = 0. This occurs because
we cannot assign a nonzero probability to each of the uncountable
values and still have the probabilities sum to one. Therefore, for a
continuous random variable, P(a ≤ X ≤ b) = P(a < X < b) = P(a ≤ X <
b) = P(a < X ≤ b) because P(X = a) and P(X = b) are both zero.

page 165

A continuous random variable is completely described by its

probability density function, denoted by f(x). The graph of f(x)
approximates the relative frequency polygon for the population. The
probability that the variable assumes a value within an interval, say
P(a ≤ X ≤ b), is defined as the area under f(x) between points a and
b. Moreover, the entire area under f(x) over all values of x must
equal one; this is equivalent to the fact that, for discrete random
variables, the probabilities add up to one.
A simple probability distribution for a continuous random variable
is called the continuous uniform distribution. This distribution is
appropriate when the underlying random variable has an equally
likely chance of assuming a value within a specified range. Suppose
you are informed that your new refrigerator will be delivered between
2:00 pm and 3:00 pm. Let the random variable X denote the delivery
time of your refrigerator. This variable is bounded below by 2:00 pm
and above by 3:00 pm for a total range of 60 minutes. Here, the
probability of delivery between 2:00 pm and 2:30 pm equals 0.50
(=30/60), as does the probability of delivery between 2:30 pm and
3:00 pm. Similarly, the probability of delivery in any 15-minute
interval equals 0.25 (=15/60), and so on. Other examples of random
variables that follow the continuous uniform distribution include the
scheduled flight time between cities and the waiting time for a
campus bus. Any specified range for each of these random variables
can be assumed to be equally probable.
In this section, we focus on the most widely used continuous
probability distribution: the normal distribution. It is the familiar bell-
shaped distribution, also referred to as the Gaussian distribution.
One reason for its extensive use is that it closely approximates the
probability distribution for a wide range of random variables of
interest. Examples of random variables that closely follow a normal
distribution include salaries of employees, cumulative debt of college
graduates, and advertising expenditure of firms. Another important
function of the normal distribution is that it serves as the cornerstone
of statistical inference, a topic discussed in detail in Chapter 5.
A graph depicting the normal probability density function f(x) is
often referred to as the normal curve or the bell curve, symmetric
around the mean. We generally use the cumulative distribution
function P(X ≤ x) to compute probabilities for a normally distributed
random variable, where P(X ≤ x) is simply the area under the normal
curve up to the value x. Fortunately, we do not necessarily need the
knowledge of integral calculus to compute probabilities for the
normal distribution. Instead, we rely on a table to find probabilities.
We can also compute probabilities with Excel, R, and other statistical
packages. The specifics of how to use the table are explained next.

The Standard Normal Distribution

The standard normal distribution is a special case of the normal
distribution with a mean equal to zero and a standard deviation (or
variance) equal to one. Using the letter Z to denote a random
variable with the standard normal distribution, we have E(Z) = 0 and
SD(Z) = 1. As usual, we use the lowercase letter z to denote the
value that the standard normal variable Z may assume.
We will first show how to compute probabilities related to the
standard normal distribution. Later, we will show that any normal
distribution is equivalent to the standard normal distribution when the
unit of measurement is changed to measure standard deviations
from the mean.
Virtually all introductory statistics texts include the z table, that
provides areas (probabilities) under the z curve. However, the format
of these tables is sometimes different. In this text, the z table
provides cumulative probabilities P(Z ≤ z); this table appears on two
pages in Appendix D and is labeled Table D.1. The left-hand page
provides cumulative probabilities for z values less than or equal to
zero. The right-hand page shows cumulative probabilities for z
values greater than or equal to zero. Given the symmetry of the
normal distribution and the fact that the area under the entire curve
is one, other probabilities can be easily computed.

page 166

THE STANDARD NORMAL DISTRIBUTION

The standard normal random variable Z is a normal random
variable with E(Z) = 0 and SD(Z) = 1. The z table lists
cumulative probabilities P(Z ≤ z) against positive and negative z
values.

Figure 4.4 represents the standard normal probability density

function (z distribution). Because the random variable Z is symmetric
around its mean of 0, P(Z < 0) = P(Z > 0) = 0.5. As is the case with
all continuous random variables, we can also write the probabilities
as P(Z ≤ 0) = P(Z ≥ 0) = 0.5.

FIGURE 4.4 Standard normal probability density function

As mentioned earlier, the z table provides cumulative probabilities
P(Z ≤ z) for a given z. Consider, for example, a cumulative
probability P(Z ≤ 1.52). Because z = 1.52 is positive, we can look up
this probability from the right-hand page of the z table in Appendix D;
Table 4.5 shows a portion of the table.

TABLE 4.5 Portion of the Right-Hand Page of the z Table

z 0.00 0.01 0.02

0.0 0.5000 0.5040 ↓

0.1 0.5398 0.5438 ↓

⋮ ⋮ ⋮ ⋮

1.5 → → 0.9357

The first column of the table, denoted as the z column, shows values
of z up to the tenth decimal point, while the first row of the table,
denoted as the z row, shows hundredths values. Thus, for z = 1.52,
we match 1.5 on the z column with 0.02 on the z row to find a
corresponding probability of 0.9357. The arrows in Table 4.5 indicate
that P(Z ≤ 1.52) = 0.9357. Note that the area to the right of 1.52 can
be computed as P(Z > 1.52) = 1 − P(Z ≤ 1.52) = 1 − 0.9357 =
0.0643.
Suppose we want to find P(Z ≤ −1.96). Because z is a negative
value, we can look up this probability from the left-hand page of the z
table in Appendix D to find P(Z ≤ −1.96) = 0.0250. As before, the
area to the right of −1.96 can be computed as P(Z > −1.96) = 1 −
P(Z ≤ −1.96) = 1 − 0.0250 = 0.9750.
So far we have computed probabilities for given z values. Now
we will evaluate z values for given cumulative probabilities;
noncumulative probabilities can be evaluated using symmetry.
Suppose we need to find z given P(Z ≤ z) = 0.6808.
Because the probability is already in a cumulative format—that is,
P(Z ≤ z) = 0.6808—we simply look up 0.6808 from the body of the
table (right-hand side) to find the corresponding z value from the
row/column of z. Table 4.6 shows the relevant portion of the z table.
Therefore, z = 0.47.

page 167

TABLE 4.6 Portion of the z Table

Now suppose we need to find z given P(Z ≤ z) = 0.0643. Here, z

must be negative because the probability to its left is less than 0.50.
We look up the cumulative probability 0.0643 from the left-hand page
of the z table in Appendix D to find z = −1.52. Finally, if the given
cumulative probability is not shown in the table, we find the
approximate z value. For example, given P(Z ≤ z) = 0.90, we
approximate the value as z = 1.28.
The Transformation of Normal Random
Variables
Any normal random variable can be transformed into the standard
normal random variable to derive the relevant probabilities. In other
words, any normally distributed random variable X with mean μ and
standard deviation σ can be transformed (standardized) into the
standard normal variable Z with mean equal to zero and standard
deviation equal to one. We transform X into Z by subtracting from X
its mean and dividing by its standard deviation; this is referred to as
the standard transformation.

THE STANDARD TRANSFORMATION: CONVERTING X INTO

Z
Any normally distributed random variable X with mean μ and
standard deviation σ can be transformed into the standard
normal random variable Z as

Therefore, any value x has a corresponding value z given by

We are now in a position to solve any normal distribution problem by

first transforming it to the z distribution.
 EXAMPLE 4.14
Scores on a management aptitude exam are normally
distributed with a mean of 72 and a standard deviation of 8.
a. What is the probability that a randomly selected manager will
score above 60?
b. What is the probability that a randomly selected manager will
score between 68 and 84?
SOLUTION:
Let X represent scores with μ = 72 and σ = 8. We will use the
standard transformation to solve these problems.

a. The probability that a manager scores above 60 is P(X > 60).

Figure 4.5 shows the probability as the shaded area to the
right of 60. We derive

Because P(Z > −1.5) = 1 − P(Z ≤ −1.5), we look up −1.50 in

the z table (left-hand side) to get this probability as 1 − 0.0668
= 0.9332.
page 168
FIGURE 4.5 Finding P(X > 60)
b. When solving for the probability that a manager scores
between 68 and 84, we find P(68 ≤ X ≤ 84). The shaded area
in Figure 4.6 shows this probability. We derive
We compute

this probability using the z table as P(Z ≤ 1.5) − P(Z < −0.5) =
0.9332 − 0.3085 = 0.6247.

FIGURE 4.6 Finding P(68 ≤ X ≤ 84)

So far we have used the standard transformation to compute
probabilities for given x values. We can use the inverse
transformation, x = μ + zσ, to compute x values for given
probabilities.

THE INVERSE TRANSFORMATION: CONVERTING Z INTO X

The standard normal variable Z can be transformed to the
normally distributed random variable X with mean μ and
standard deviation σ as X = μ + Zσ.

Therefore, any value z has a corresponding value x given by x

= μ + zσ.

page 169

EXAMPLE 4.15
Scores on a management aptitude examination are normally
distributed with a mean of 72 and a standard deviation of 8.
a. What is the lowest score that will place a manager in the top
10% (90th percentile) of the distribution?
b. What is the highest score that will place a manager in the
bottom 25% (25th percentile) of the distribution?
SOLUTION: Let X represent scores on a management aptitude
examination with μ = 72 and σ = 8. We will use the inverse
transformation x = μ + zσ to solve these problems.
a. The 90th percentile is a numerical value x such that P(X < x) =
0.90. We look up 0.90 (or the closest value to 0.90) in the z
table (right-hand side) to get z = 1.28 and use the inverse
 transformation to find x = 72 + 1.28(8) = 82.24. Therefore, a
score of 82.24 or higher will place a manager in the top 10%
of the distribution (see Figure 4.7).

FIGURE 4.7 Finding x given P(X < x) = 0.90

b. The 25th percentile is a numerical value x such that P(X < x) =

0.25. Using the z table (left-hand side), we find the
corresponding z value that satisfies P(Z < z) = 0.25 as −0.67.
We then solve x = 72 − 0.67(8) = 66.64. Therefore, a score of
66.64 or lower will place a manager in the bottom 25% of the
distribution (see Figure 4.8).

FIGURE 4.8 Finding x given P(X < x) = 0.25

 page 170

Using Excel and R for the Normal

Distribution
Example 4.16 illustrates the use of Excel and R functions with
respect to the normal distribution.

EXAMPLE 4.16
The Vanguard Balanced Index Fund seeks to maintain an
allocation of 60% to stocks and 40% to bonds. With low fees
and a consistent investment approach, this fund ranks fourth
out of 792 funds that allocate 50% to 70% to stocks (US News,
March 2017). Based on historical data, the expected return and
standard deviation of this fund are estimated as 7.49% and
6.41%, respectively. Assume that the fund returns are stable
and are normally distributed.
a. What is the probability that the fund will generate a return
between 5% and 10%?
b. What is the lowest return of the fund that will place it in the top
10% (90th percentile) of the distribution?
SOLUTION: We let X denote the return on the Vanguard
Balanced fund. We know that X is normally distributed with μ =
7.49 and σ = 6.41.
As mentioned in Chapters 2 and 3, due to different fonts
and type settings, copying and pasting Excel or R functions
from this text directly into Excel or R may cause errors. When
such errors occur, you may need to replace special characters
such as quotation marks and parentheses or delete extra
spaces in the functions.
Using Excel
We use Excel’s NORM.DIST and NORM.INV functions to solve
problems pertaining to the normal distribution. In order to find
P(X ≤ x), we enter =NORM.DIST(x, μ, σ, 1) where x is the
value for which we want to evaluate the cumulative probability,
μ is the mean of the distribution, and σ is the standard
deviation of the distribution. (If we enter “0” for the last
argument in the function, then Excel returns the height of the
normal distribution at the point x. This feature is useful if we
want to plot the normal distribution.) If we want to find a
particular x value for a given cumulative probability
(cumulprob), then we enter =NORM.INV(cumulprob, μ, σ).
a. In order to find the probability of a return between 5% and
10%, P(5 ≤ X ≤ 10), we enter =NORM.DIST(10, 7.49, 6.41, 1)
− NORM.DIST(5, 7.49, 6.41, 1). Excel returns 0.3035.
b. In order to find the lowest return that will place it in the top
10% (90th percentile) of the distribution, P(X > x) = 0.10, we
enter =NORM.INV(0.90, 7.49, 6.41). Excel returns 15.7047.
Using R
 We use R’s pnorm and qnorm functions to solve problems
associated with the normal distribution. In order to find P(X ≤
x), we enter pnorm(x, μ, σ, lower.tail = TRUE) where x is the
value for which we want to evaluate the cumulative probability,
μ is the mean of the distribution, and σ is the standard
deviation of the distribution. If we enter “lower.tail=FALSE” for
the last argument in the function, then R returns P(X > x). If we
want to find a particular x value for a given cumulative
probability (cumulprob), then we enter qnorm(cumulprob, μ, σ).
a. In order to find P(5 ≤ X ≤ 10), we enter:
> pnorm(10, 7.49, 6.41, lower.tail=TRUE) − pnorm(5, 7.49,
6.41, lower.tail=TRUE)
And R returns: 0.3034746.
page 171
b. In order to solve for x to satisfy P(X > x) = 0.10,
we enter:
> qnorm(0.90, 7.49, 6.41)
And R returns: 15.70475.

EXERCISES 4.5
Applications
51. The historical returns on a balanced portfolio have had an average return of 8%
and a standard deviation of 12%. Assume that returns on this portfolio follow a
normal distribution.
a. What percentage of returns were greater than 20%?
b. What percentage of returns were below −16%?
52. A professional basketball team averages 105 points per game with a standard
deviation of 10 points. Assume points per game follow the normal distribution.
a. What is the probability that a game’s score is between 85 and 125 points?
b. What is the probability that a game’s score is more than 125 points? If there are
82 games in a regular season, in how many games will the team score more
than 125 points?
53. The average rent in a city is $1,500 per month with a standard deviation of $250.
Assume rent follows the normal distribution.
a. What percentage of rents are between $1,250 and $1,750?
b. What percentage of rents are less than $1,250?
c. What percentage of rents are greater than $2,000?
54. Suppose that the miles-per-gallon (mpg) rating of passenger cars is a normally
distributed random variable with a mean and a standard deviation of 33.8 mpg
and 3.5 mpg, respectively.
a. What is the probability that a randomly selected passenger car gets at least 40
mpg?
b. What is the probability that a randomly selected passenger car gets between 30
and 35 mpg?
c. An automobile manufacturer wants to build a new passenger car with an mpg
rating that improves upon 99% of existing cars. What is the minimum mpg that
would achieve this goal?
55. Business interns from a small private college earn an average annual salary of
$43,000. Let the salary be normally distributed with a standard deviation of
$18,000.
a. What percentage of business interns make between $40,000 and $50,000?
b. What percentage of business interns make more than $80,000?
56. A financial advisor informs a client that the expected return on a portfolio is 8%
with a standard deviation of 12%. There is a 15% chance that the return would be
above 16%. If the advisor is right about her assessment, is it reasonable to
assume that the underlying return distribution is normal?
57. According to a company’s website, the top 25% of the candidates who take the
entrance test will be called for an interview. You have just been called for an
interview. The reported mean and standard deviation of the test scores are 68 and
8, respectively. What is the possible range for your test score if you assume that
the scores are normally distributed?
58. The time required to assemble an electronic component is normally distributed
with a mean and a standard deviation of 16 minutes and 4 minutes, respectively.
a. Find the probability that a randomly picked assembly takes between 10 and 20
minutes.
b. It is unusual for the assembly time to be above 24 minutes or below 6 minutes.
What proportion of assembly times falls in these unusual categories?
59. According to the Mortgage Banking Association, loans that are 90 days or more
past due are considered seriously delinquent (housingwire.com, May 14, 2019). It
has been reported that the rate of seriously delinquent loans has an average of
9.1%. Let the rate of seriously delinquent loans follow a normal distribution with a
standard deviation of 0.80%.
a. What is the probability that the rate of seriously delinquent loans is above 8%?
 b. What is the probability that the rate of seriously delinquent loans is between
9.5% and 10.5%?
60. The manager of a night club in Boston stated that 95% of the customers are
between the ages of 22 and 28 years. If the age of customers is normally
distributed with a mean of 25 years, calculate its standard deviation.
61. An estimated 1.8 million students take on student loans to pay ever-rising tuition
and room and board. It is also known that the average cumulative debt of recent
college graduates is about $22,500. Let the cumulative debt among recent college
graduates be normally distributed with a standard deviation of $7,000.
Approximately how many recent college graduates have accumulated student
loans of more than $30,000?
62. On average, an American professional football game lasts about three hours,
even though the ball is actually in play only 11 minutes (SB Nation, April 1, 2019).
Assume that game times are normally distributed with a standard deviation of 0.4
hour.
a. Find the probability that a game lasts less than 2.5 hours.
b. Find the probability that a game lasts either less than 2.5 hours or more than
3.5 hours.

page 172
c. Find the maximum value for the game time that will place it in
the bottom 1% of the distribution.
63. A young investment manager tells his client that the probability of making a
positive return with his suggested portfolio is 90%. If it is known that returns are
normally distributed with a mean of 5.6%, what is the risk, measured by standard
deviation, that this investment manager assumes in his calculation?
64. You are considering the risk-return profile of two mutual funds for investment. The
relatively risky fund promises an expected return of 8% with a standard deviation
of 14%. The relatively less risky fund promises an expected return and standard
deviation of 4% and 5%, respectively. Assume that the returns are approximately
normally distributed.
a. Which mutual fund will you pick if your objective is to minimize the probability of
earning a negative return?
b. Which mutual fund will you pick if your objective is to maximize the probability of
earning a return above 8%?
65. A new car battery is sold with a two-year warranty whereby the owner gets the
battery replaced free of cost if it breaks down during the warranty period. Suppose
an auto store makes a net profit of $20 on batteries that stay trouble-free during
the warranty period; it makes a net loss of $10 on batteries that break down. The
life of batteries is known to be normally distributed with a mean and a standard
deviation of 40 and 16 months, respectively.
a. What is the probability that a battery will break down during the warranty
period?
b. What is the expected profit of the auto store on a battery?
 c. What is the expected monthly profit on batteries if the auto store sells an
average of 500 batteries a month?

4.6 WRITING WITH DATA

Case Study

Jasper White/Image Source

Professor Lang is a professor of economics at Salem State University. She has been
teaching a course in Principles of Economics for over 25 years. Professor Lang has
never graded on a curve because she believes that relative grading may unduly
penalize or benefit a student in an unusually strong or weak class. She always uses
an absolute scale for making grades, as shown in the two left columns of Table 4.7.
TABLE 4.7 Grading Scales with Absolute Grading versus Relative Grading

Absolute Grading Relative Grading

Grade Score Grade Probability

A 92 and above A 0.10

B 78 up to 92 B 0.35

C 64 up to 78 C 0.40

D 58 up to 64 D 0.10

F Below 58 F 0.05

A colleague of Professor Lang’s has convinced her to move to relative grading

because it corrects for unanticipated problems. Professor Lang decides to
experiment with grading based on the relative scale as shown in the two right
columns of Table 4.7. Using this relative grading scheme, the top 10% of students
will get A’s, the next 35% B’s, and so on. Based on her years of teaching experience,
Professor Lang believes that the scores in her course follow a normal distribution
with a mean of 78.6 and a standard deviation of 12.4.
Professor Lang wants to use the above information to
1. Calculate probabilities based on the absolute scale. Compare these probabilities to
the relative scale.
2. Calculate the range of scores for various grades based on the relative scale.
Compare these ranges to the absolute scale.
3. Determine which grading scale makes it harder to get higher grades.

page 173

Sample Report—Absolute Grading versus

Relative Grading
Many teachers would confess that grading is one of the most difficult tasks of
their profession. Two common grading systems used in higher education are
relative and absolute. Relative grading systems are norm-referenced or
curve-based, in which a grade is based on the student’s relative position in
class. Absolute grading systems, on the other hand, are criterion-referenced,
in which a grade is related to the student’s absolute performance in class. In
short, with absolute grading, the student’s score is compared to a
predetermined scale, whereas with relative grading, the score is compared to
the scores of other students in the class.
Let X represent a grade in Professor Lang’s class, which is normally
distributed with a mean of 78.6 and a standard deviation of 12.4. This
information is used to derive the grade probabilities based on the absolute
scale. For instance, the probability of receiving an A is derived as P(X ≥ 92) =
P(Z ≥ 1.08) = 0.14. Other probabilities, derived similarly, are presented in
Table 4.8.

TABLE 4.8 Probabilities Based on Absolute Scale and Relative Scale

Probability Based on Probability Based

Grade
Absolute Scale on Relative Scale

A 0.14 0.10

B 0.38 0.35

C 0.36 0.40

D 0.07 0.10

F 0.05 0.05

The second column of Table 4.8 shows that 14% of students are expected
to receive A’s, 38% B’s, and so on. Although these numbers are generally
consistent with the relative scale restated in the third column of Table 4.8, it
appears that the relative scale makes it harder for students to get higher
grades. For instance, 14% get A’s with the absolute scale compared to only
10% with the relative scale.
Alternatively, we can compare the two grading methods on the basis of the
range of scores for various grades. The second column of Table 4.9 restates
the range of scores based on absolute grading. In order to obtain the range of
scores based on relative grading, it is once again necessary to apply
concepts from the normal distribution. For instance, the minimum score
required to earn an A with relative grading is derived by solving for x in P(X ≥
x) = 0.10. Because P(X ≥ x) = 0.10 is equivalent to P(Z ≥ z) = 0.10, it follows
that z = 1.28. Inserting the proper values of the mean, the standard deviation,
and z into x = μ + zσ yields a value of x equal to 94.47. Ranges for other
grades, derived similarly, are presented in the third column of Table 4.9.

TABLE 4.9 Range of Scores with Absolute Grading versus Relative

Grading

Range of Scores Range of Scores

Grade Based on Absolute Based on Relative
Grading Grading

A 92 and above 94.47 and above

B 78 up to 92 80.21 up to
94.47

C 64 up to 78 65.70 up to
80.21

D 58 up to 64 58.20 up to
65.70

F Below 58 Below 58.20

Once again comparing the results in Table 4.9, the use of the relative scale
makes it harder for students to get higher grades in Professor Lang’s courses.
For instance, in order to receive an A with relative grading, a student must
have a score of at least 94.47 versus a score of at least 92 with absolute
grading. Both absolute and relative grading methods have their merits and
teachers often make the decision on the basis of their teaching philosophy.
However, if Professor Lang wants to keep the grades consistent with her
earlier absolute scale, she should base her relative scale on the probabilities
computed in the second column of Table 4.8.
 page 174

Suggested Case Studies

Report 4.1. Enacted in 1998, the Children’s Online Privacy Protection Act requires
firms to obtain parental consent before tracking the information and the online
movement of children; however, the act applies to those children ages 12 and under.
Teenagers are often unaware of the consequences of sharing their lives online. Data
reapers create huge libraries of digital profiles and sell these profiles to advertisers,
who use it to detect trends and micro-target their ads back to teens. For example, a
teen searching online for ways to lose weight could become enticed by an ad for
dietary supplements, fed into his/her network by tracking cookies. As a preliminary
step in gauging the magnitude of teen usage of social networking sites, a researcher
surveys 200 teen girls and 200 teen boys. Of teen girls, 166 use social networking
sites; of teen boys, 156 use social networking sites.
In a report, use the sample information to
Determine the probability that a teen uses social networking sites.
Determine the probability that a teen girl uses a social networking site.
New legislation would ban Internet companies from sending targeted advertising to
children under 16 and give these children and their parents the ability to delete
their digital footprint and profile with an “eraser button”. Given the probabilities that
you calculated with respect to teen usage of social networking sites, do you think
that this legislation is necessary? Explain.

Report 4.2. Senior executives at Skyhigh Construction, Inc., participate in a pick-

your-salary plan. They choose salaries in a range between $125,000 and $150,000.
By choosing a lower salary, an executive has an opportunity to make a larger bonus.
If Skyhigh does not generate an operating profit during the year, then no bonuses are
paid. Skyhigh has just hired two new senior executives, Allen Grossman and Felicia
Arroyo. Each must decide whether to choose Option 1: a base pay of $125,000 with
a possibility of a large bonus or Option 2: a base pay of $150,000 with a possibility of
a bonus, but the bonus would be one-half of the bonus under Option 1.
Grossman, 44 years old, is married with two young children. He bought his home
at the height of the market and has a rather large monthly mortgage payment.
Arroyo, 32 years old, just completed her MBA at a prestigious Ivy League university.
She is single and has no student loans due to a timely inheritance upon entering
graduate school. Arroyo just moved to the area, so she has decided to rent an
apartment for at least one year. Given their personal profiles, inherent perceptions of
risk, and subjective views of the economy, Grossman and Arroyo construct their
individual probability distributions with respect to bonus outcomes shown in Table
4.10.

TABLE 4.10 Grossman’s and Arroyo’s Probability Distributions

 Probability

Bonus (in $) Grossman Arroyo

0 0.35 0.20

50,000 0.45 0.25

100,000 0.10 0.35

150,000 0.10 0.20

In a report, use the sample information to

Compute expected values to evaluate payment plans for Grossman and Arroyo.
Help Grossman and Arroyo decide whether to choose Option 1 or Option 2 for
his/her compensation package.

Report 4.3. Akiko Hamaguchi is the manager of a small sushi restaurant called Little
Ginza in Phoenix, Arizona. As part of her job, Akiko has to purchase salmon every
day for the restaurant. For the sake of freshness, it is important that she buys the
right amount of salmon daily. Buying too much may result in
wastage and buying too little may disappoint some customers on
page 175
high-demand days.
Akiko has estimated that the daily consumption of salmon is normally distributed
with a mean of 12 pounds and a standard deviation of 3.2 pounds. She has always
bought 20 pounds of salmon every day. Lately, she has been criticized by the owners
because this amount of salmon was too often resulting in wastage. As part of cost
cutting, Akiko is considering a new strategy. She will buy salmon that is sufficient to
meet the daily demand of customers on 90% of the days.
In a report, help Akiko use the above information to
Calculate the probability that the demand for salmon at Little Ginza is above 20
pounds.
Calculate the probability that the demand for salmon at Little Ginza is below 15
pounds.
Determine the amount of salmon that should be bought daily so that the restaurant
meets demand on 90% of the days.
 page 176

5 Statistical Inference

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 5.1  Describe the sampling distribution of the
sample mean.
LO 5.2  Describe the sampling distribution of the
sample proportion.
LO 5.3  Construct a confidence interval for the
population mean.
LO 5.4  Construct a confidence interval for the
population proportion.
LO 5.5  Conduct a hypothesis test for the
population mean.
LO 5.6  Conduct a hypothesis test for the
population proportion.

n the last chapter, we had information on population parameters,

such as the population proportion or the population mean, for the
 analysis of discrete and continuous random variables. In many
I instances, we do not have information on the parameters, so we
make statistical inferences on the basis of sample statistics. In this
chapter, we first discuss how to evaluate the properties of sample
statistics. In particular, we examine the probability distributions of the
sample mean and the sample proportion based on simple random
sampling.
We then turn our attention to estimation and hypothesis testing,
which are the two basic methodologies of statistical inference. In
order to estimate an unknown population parameter, we develop and
interpret a confidence interval. We use a hypothesis test to challenge
the status quo, or some belief about an underlying population
parameter. We focus on confidence intervals and hypothesis tests
for the population mean and the population proportion.

page 177
©Ken Seet/Corbis Images/SuperStock

INTRODUCTORY CASE
Undergraduate Study Habits
Are today’s college students studying hard or hardly studying?
A study asserts that, over the past six decades, the number of
hours that the average college student studies each week has
been steadily dropping (The Wall Street Journal, April 10,
2019). In 1961, students invested 24 hours per week in their
academic pursuits, whereas today’s students study an average
of 14 hours per week.
Susan Knight is a dean at a large university in California.
She wonders if the study trend is reflective of the students at
 her university. She randomly selects 35 students and asks their
average study time per week (in hours). A portion of the
responses is shown in Table 5.1.
FILE
Study_Hours

TABLE 5.1 Number of Hours Spent Studying

   25   

   19   

   ⋮   

   16   

Summary measures: hours and s = 7.2155 hours.

Susan wants to use the sample information to

1. Determine if the mean study time of students at her university
is below the 1961 national average of 24 hours per week.
2. Determine if the mean study time of students at her university
differs from today’s national average of 14 hours per week.
A synopsis of this case is provided at the end of Section 5.3.

page 178

5.1 SAMPLING DISTRIBUTIONS

In many applications, we are interested in the characteristics of a
population. For instance, a ride sharing company is interested in the
average income (population mean) in a large city. Similarly, a banker
is interested in the default probability (population proportion) of
mortgage holders. Recall that the population mean and the
population proportion are parameters that describe a numerical
variable and a categorical variable, respectively. Because it is
extremely difficult, if not impossible, to analyze the entire population,
we generally make inferences about the characteristics of the
population on the basis of a random sample drawn from the
population.
It is important to note that there is only one population, but many
possible samples of a given size can be drawn from the population.
Therefore, a population parameter is a constant, even though its
value may be unknown. On the other hand, a statistic, such as the
sample mean or the sample proportion, is a variable whose value
depends on the particular sample that is randomly drawn from the
population.

PARAMETER VERSUS STATISTIC

A parameter is a constant, although its value may be unknown.
A statistic is a variable whose value depends on the chosen
random sample.

Consider the income in a large city as the variable of interest. If

you decide to make inferences about the population mean income
on the basis of a random draw of 38 residents, then the sample
mean is the relevant statistic. Note that the value of will change if

you choose a different random sample of 38 residents. In other

words, is a variable whose value depends on the chosen random
sample. The sample mean is commonly referred to as the
estimator, or the point estimator, of the population mean.

In the income example, the sample mean is the estimator of the

mean income of residents in the large city. If the average derived
from a specific sample is $54,000, then is the estimate of

the population mean. Similarly, if the variable of interest is the default

probability of mortgage holders, then the sample proportion of
defaults, denoted by from a random sample of 80 mortgage

holders is the estimator of the population proportion. If 10 out of 80

mortgage holders in a given sample default, then is

the estimate of the population proportion.

ESTIMATOR AND ESTIMATE

When a statistic is used to estimate a parameter, it is referred
to as an estimator. A particular value of the estimator is called
an estimate.

The Sampling Distribution of the Sample

Mean
We first focus on the probability distribution of the sample mean

which is also referred to as the sampling distribution of Because

is a variable, its sampling distribution is simply the probability

distribution derived from all possible samples of a given size from the
population. Consider, for example, a mean derived from a sample of
n observations. Another mean can similarly be derived from a
different sample of n observations. If we repeat this process a very
large number of times, then the frequency distribution of the sample
means can be thought of as its sampling distribution. In particular,
we will discuss the expected value and the standard deviation of the
sample mean. We will also study the conditions under which the
sampling distribution of the sample mean is normally distributed.

page 179

LO 5.1
Describe the sampling distribution of the sample mean.

Let the random variable X represent a certain characteristic of a

population under study, with an expected value, E(X) = μ, and a
variance, Var(X) = σ2. For example, X could represent the income of
city residents or the return on an investment. We can think of μ and
σ2 as the mean and the variance of an individual observation drawn
randomly from the population of interest, or simply as the population
mean and the population variance. Let the sample mean be based

on a random sample of n observations from this population.

The expected value of is the same as the expected value of the

individual observation—that is, In other words, if we

were to sample repeatedly from a given population, the average

value of the sample means will equal the average value of all
individual observations in the population, or, simply, the population
mean. This is an important property of an estimator, called
unbiasedness, that holds irrespective of whether the sample mean is
based on a small or a large sample. An estimator is unbiased if its
expected value equals the population parameter of interest.
 The variance of is equal to In other words, if we were

to sample repeatedly from a given population, the variance of the

sample mean will equal the variance of all individual observations in
the population, divided by the sample size, n. Note that is

smaller than the variance of X, which is equal to Var(X) = σ2. This is

an intuitive result, suggesting that the variability between sample
means is less than the variability between observations. Because
each sample is likely to contain both high and low observations, the
highs and lows cancel one another, making the variance of smaller

than the variance of X. As usual, the standard deviation of is

calculated as the positive square root of the variance. However, in
order to distinguish the variability between samples from the
variability between individual observations, we refer to the standard
deviation of as the standard error of the sample mean, computed

as

THE EXPECTED VALUE AND THE STANDARD ERROR OF

THE SAMPLE MEAN

The expected value of the sample mean equals the

population mean; that is, The standard error of the

sample mean equals the population standard deviation divided

by the square root of the sample size; that is,
 EXAMPLE 5.1
The chefs at a local pizza chain in Cambria, California, strive to
maintain the suggested size of their 16-inch pizzas. Despite
their best efforts, they are unable to make every pizza exactly
16 inches in diameter. The manager has determined that the
size of the pizzas is normally distributed with a mean of 16
inches and a standard deviation of 0.8 inch.
a. What are the expected value and the standard error of the
sample mean derived from a random sample of two pizzas?
b. What are the expected value and the standard error of the
sample mean derived from a random sample of four pizzas?
c. Compare the expected value and the standard error of the
sample mean with those of an individual pizza.

SOLUTION: We know that the population mean μ = 16 and the

population standard deviation σ = 0.8. We use and

to calculate the following results.

a. With the sample size

b. With the sample size

page 180
c. The expected value of the sample mean for both
sample sizes is identical to the expected value of the
individual pizza. However, the standard error of the sample
mean with n = 4 is lower than the one with n = 2. For both
sample sizes, the standard error of the sample mean is lower
 than the standard deviation of the individual pizza. This result
confirms that averaging reduces variability.

Sampling from a Normally Distributed Population

An important feature of the sampling distribution of the sample mean
is that, irrespective of the sample size n, is normally distributed if

the population X from which the sample is drawn is normally

distributed. In other words, if X is normally distributed with expected
value μ and standard deviation σ, then is also normally distributed

with expected value μ and standard error Moreover, if is

normally distributed, then it can be transformed into a standard

normal variable Z by subtracting from its mean, and then dividing

by its standard error.

SAMPLING FROM NORMALLY DISTRIBUTED POPULATION

For any sample size n, the sampling distribution of is normal

if the population X from which the sample is drawn is normally

distributed.

EXAMPLE 5.2
Use the information in Example 5.1 to answer the following
questions:
a. What is the probability that a randomly selected pizza is less
than 15.5 inches?
b. What is the probability that the average of two randomly
selected pizzas is less than 15.5 inches?
c. What is the probability that the average of four randomly
selected pizzas is less than 15.5 inches?
d. Comment on the computed probabilities.

SOLUTION: Because the population is normally distributed, the

sampling distribution of the sample mean is also normal. Figure
5.1 depicts the shapes of the three distributions based on the
population mean μ = 16 and the population standard deviation
σ = 0.8.

FIGURE 5.1 Normal distribution of the sample mean

Note that when the sample size n = 1, the sample mean is

the same as the individual observation x.
a. We use the standard transformation to derive
There is a 26.43%

chance that an individual pizza is less than 15.5 inches.

page 181
b. Here we use the standard transformation to derive

In a random sample of two

pizzas, there is an 18.94% chance that the average size is

less than 15.5 inches.

c. Again we find but now n = 4. Therefore,

In a random sample

of four pizzas, there is a 10.56% chance that the average size

is less than 15.5 inches.
d. The probability that the average size is under 15.5 inches, for
four randomly selected pizzas, is less than half of that for an
individual pizza. This is due to the fact that while X and have

the same expected value of 16, the variance of is less than

that of X.

The Central Limit Theorem

For making statistical inferences, it is essential that the sampling
distribution of is normally distributed. So far we have only

considered the case where is normally distributed because the

population X from which the sample is drawn is normally distributed.
What if the underlying population is not normally distributed?
Here we present the central limit theorem (CLT), which perhaps
is the most remarkable result of probability theory. The CLT states
that the sum or the average of a large number of independent
observations from the same underlying distribution has an
approximate normal distribution. The approximation steadily
improves as the number of observations increases. In other words,
irrespective of whether or not the population X is normally
distributed, the sample mean computed from a random sample of
size n will be approximately normally distributed as long as n is
sufficiently large. How large a sample is necessary? Practitioners
often use the normal distribution approximation when n ≥ 30.

THE CENTRAL LIMIT THEOREM FOR THE SAMPLE MEAN

For any population X with expected value μ and standard
deviation σ, the sampling distribution of will be approximately

normal if the sample size n is sufficiently large. As a general

guideline, the normal distribution approximation is justified
when n ≥ 30.

EXAMPLE 5.3
For the month of May, a coffee chain advertised a Happy Hour
between the hours of 3 pm and 5 pm when customers could
enjoy a half-price iced coffee. The manager of one of the
chains wants to determine if the Happy Hour has had a
lingering effect on the amount of money customers now spend
on iced coffee.
Before the marketing campaign, customers spent an
average of $4.18 on iced coffee with a standard deviation of
$0.84. Based on 50 customers sampled after the marketing
campaign, the average amount spent is $4.26. If the coffee
chain chose not to pursue the marketing campaign, how likely
is it that customers will spend an average of $4.26 or more on
iced coffee?

SOLUTION: If the coffee chain did not pursue the marketing

campaign, spending on iced coffee would still have mean μ =
4.18 and standard deviation σ = 0.84. We need to calculate the
probability that the sample mean is at least 4.26, or

The population from which the sample is drawn is not known to

be normally distributed. However, because n ≥ 30, from the
central limit theorem we know that is approximately normally

distributed. Therefore, as shown in Figure 5.2,

It is quite

plausible (probability = 0.2514) that in a sample of page 182

50 customers, the sample mean is $4.26 or more
even if the coffee chain did not pursue the marketing
campaign.

FIGURE 5.2 Finding

The Sampling Distribution of the Sample
Proportion
LO 5.2
Describe the sampling distribution of the sample proportion.

Our discussion thus far has focused on the population mean, but in
many business applications, we are concerned with the population
proportion. For instance, a banker is interested in the default
probability of mortgage holders. Or an online retailer cares about the
proportion of customers who make a purchase after receiving a
promotional e-mail. In these examples, the parameter of interest is
the population proportion p.
As in the case of the population mean, we almost always make
inferences about the population proportion on the basis of sample
data. Here, the relevant statistic (estimator) is the sample proportion,
a particular value (estimate) is denoted by Because is a

variable, we need to discuss its sampling distribution.

It turns out that the sampling distribution of is closely related to

the binomial distribution. Recall from Chapter 4 that the binomial

distribution describes the number of successes X in n trials of a
Bernoulli process where p is the probability of success; thus, is

the number of successes X divided by the sample size n. It can be

shown that the expected value and the variance of the sampling
distribution of are and respectively. Note that
because , it implies that is an unbiased estimator of p.

Analogous to our discussion regarding the standard error of the

sample mean, we refer to the standard deviation of the sample
proportion as the standard error of the sample proportion; that is,

THE EXPECTED VALUE AND THE STANDARD ERROR OF

THE SAMPLE PROPORTION

The expected value of the sample proportion is equal to the

population proportion; that is, The standard error of the

sample proportion equals se

In this text, we make statistical inferences about the population

proportion only when the sampling distribution of is approximately

normal. From the earlier discussion of the central limit theorem

(CLT), we can conclude that is approximately normally distributed

when the sample size is sufficiently large. In general, the normal

distribution approximation is justified when np ≥ 5 and n(1 − p) ≥ 5.
Moreover, if is normally distributed, then it can be transformed into

a standard normal variable Z by subtracting from its expected

value, and then dividing by its standard error.

 page 183

THE CENTRAL LIMIT THEOREM FOR THE SAMPLE

PROPORTION

For any population proportion p, the sampling distribution of is

approximately normal if the sample size n is sufficiently large.

As a general guideline, the normal distribution approximation is
justified when np ≥ 5 and n(1 − p) ≥ 5.

EXAMPLE 5.4
A study found that 55% of British firms experienced a cyber-
attack in the past year (BBC, April 23, 2019).
a. What are the expected value and the standard error of the
sample proportion derived from a random sample of 100
firms?
b. In a random sample of 100 firms, what is the probability that
the sample proportion is greater than 0.57?
SOLUTION:

a. Given that p = 0.55 and n = 100, the expected value and the

standard error of

b. With n = 100, the normal approximation for the sample

proportion is justified for p = 0.55. As shown in Figure 5.3, we

find that
 FIGURE 5.3 Finding

With a chance of 34.46%, it is quite plausible that the

proportion of British firms who experience a cyber-attack is at
least 0.57.

EXERCISES 5.1
Applications
1. According to a survey, high school girls average 100 text messages daily. Assume
the population standard deviation is 20 text messages. Suppose a random
sample of 50 high school girls is taken.
a. What is the probability that the sample mean is more than 105?
b. What is the probability that the sample mean is less than 95?
c. What is the probability that the sample mean is between 95 and 105?
2. Beer bottles are filled so that they contain an average of 330 ml of beer in each
bottle. Suppose that the amount of beer in a bottle is normally distributed with a
standard deviation of four ml.
a. What is the probability that a randomly selected bottle will have less than 325
ml of beer?

page 184
 b. What is the probability that a randomly selected six-pack of beer will have a
mean amount less than 325 ml?
c. What is the probability that a randomly selected 12-pack of beer will have a
mean amount less than 325 ml?
d. Comment on the sample size and the corresponding probabilities.
3. The weight of people in a small town in Missouri is known to be normally
distributed with a mean of 180 pounds and a standard deviation of 28 pounds. On
a raft that takes people across the river, a sign states, “Maximum capacity 3,200
pounds or 16 persons.” What is the probability that a random sample of 16
persons will exceed the weight limit of 3,200 pounds?
4. Despite its nutritional value, seafood is only a tiny part of the American diet, with
the average American eating just 16 pounds of seafood per year. Janice and Nina
both work in the seafood industry and they decide to create their own random
samples and document the average seafood diet in their sample. Let the standard
deviation of the American seafood diet be seven pounds.
a. Janice samples 42 Americans and finds an average seafood consumption of 18
pounds. How likely is it to get an average of 18 pounds or more if she had a
representative sample?
b. Nina samples 90 Americans and finds an average seafood consumption of 17.5
pounds. How likely is it to get an average of 17.5 pounds or more if she had a
representative sample?
c. Which of the two women is likely to have used a more representative sample?
Explain.
5. A small hair salon in Denver, Colorado, averages about 30 customers on
weekdays with a standard deviation of six. It is safe to assume that the underlying
distribution is normal. In an attempt to increase the number of weekday
customers, the manager offers a $2 discount on five consecutive weekdays. She
reports that her strategy has worked because the sample mean of customers
during this five-weekday period jumps to 35.
a. How unusual would it be to get a sample average of 35 or more customers if
the manager had not offered the discount?
b. Do you feel confident that the manager’s discount strategy has worked?
Explain.
6. Suppose that the typical college student graduates with $28,650 in debt. Let debt
among recent college graduates be normally distributed with a standard deviation
of $7,000.
a. What is the probability that the average debt of four recent college graduates is
more than $25,000?
b. What is the probability that the average debt of four recent college graduates is
more than $30,000?
7. Forty families gathered for a fund-raising event. Suppose the individual
contribution for each family is normally distributed with a mean and a standard
deviation of $115 and $35, respectively. The organizers would call this event a
 success if the total contributions exceed $5,000. What is the probability that this
fund-raising event is a success?
8. A doctor is getting sued for malpractice by four of her former patients. It is
believed that the amount that each patient will sue her for is normally distributed
with a mean of $800,000 and a standard deviation of $250,000.
a. What is the probability that a given patient sues the doctor for more than
$1,000,000?
b. If the four patients sue the doctor independently, what is the probability that the
total amount they sue for is over $4,000,000?
9. Suppose that the miles-per-gallon (mpg) rating of passenger cars is normally
distributed with a mean and a standard deviation of 33.8 and 3.5 mpg,
respectively.
a. What is the probability that a randomly selected passenger car gets more than
35 mpg?
b. What is the probability that the average mpg of four randomly selected
passenger cars is more than 35 mpg?
c. If four passenger cars are randomly selected, what is the probability that all of
the passenger cars get more than 35 mpg?
10. Suppose that IQ scores are normally distributed with a mean of 100 and a
standard deviation of 16.
a. What is the probability that a randomly selected person will have an IQ score of
less than 90?
b. What is the probability that the average IQ score of four randomly selected
people is less than 90?
c. If four people are randomly selected, what is the probability that all of them
have an IQ score of less than 90?
11. A 2019 Pew Research study finds that the number of undocumented immigrants
living in the United States has dropped to the level it was in 2004. While its share
has declined, California still accounts for approximately 23% of the nation’s
estimated 10.5 million undocumented immigrants.
a. In a sample of 50 undocumented immigrants, what is the probability that more
than 20% live in California?
b. In a sample of 200 undocumented immigrants, what is the probability that more
than 20% live in California?
c. Comment on the reason for the difference between the computed probabilities
in parts a and b.
12. Suppose that a study finds that 33% of teenagers text while driving. The study
was based on 100 teen drivers.
a. Discuss the sampling distribution of the sample proportion.
b. What is the probability that the sample proportion is less than 0.30?
c. What is the probability that the sample proportion is within ±0.02 of the
population proportion?
 page 185
13. A car manufacturer is concerned about poor customer
satisfaction at one of its dealerships. The management decides to evaluate the
satisfaction surveys of its next 40 customers. The dealer will be fined if the
number of customers who report favorably is between 22 and 26. The dealership
will be dissolved if fewer than 22 customers report favorably. It is known that 70%
of the dealer’s customers report favorably on satisfaction surveys.
a. What is the probability that the dealer will be fined?
b. What is the probability that the dealership will be dissolved?
14. At a new exhibit in the Museum of Science, people are asked to choose between
50 or 100 random draws from a machine. The machine is known to have 60 green
balls and 40 red balls. After each draw, the color of the ball is noted and the ball is
put back for the next draw. You win a prize if more than 70% of the draws result in
a green ball. Would you choose 50 or 100 draws for the game? Explain.
15. Suppose that one in six smartphone users have fallen prey to cyber-attack.
a. Discuss the sampling distribution of the sample proportion based on a sample
of 200 smartphone users. Is it appropriate to use the normal distribution
approximation for the sample proportion?
b. What is the probability that more than 20% of smartphone users in the sample
have fallen prey to cyber-attack?

5.2 ESTIMATION
Given sample data, we use the sample statistics to make inferences
about the unknown population parameters, such as the population
mean and the population proportion. Two basic methodologies
emerge from the inferential branch of statistics: estimation and
hypothesis testing. Although the sample statistics are based on a
portion of the population, they contain useful information to estimate
the population parameters and to conduct tests regarding the
population parameters. In this section, we focus on estimation.
As mentioned earlier, when a statistic is used to estimate a
parameter, it is referred to as a point estimator, or simply an
estimator. A particular value of the estimator is called a point
estimate or an estimate. Recall that the sample mean is the
estimator of the population mean μ, and the sample proportion is
the estimator of the population proportion p.
Suppose in a sample of 25 ultra-green cars, we find that the
mean miles per gallon (mpg) of the cars is mpg; similarly,
suppose we calculate the proportion of these cars that get an mpg
greater than 100 as Thus, the estimate for the mean mpg of

all ultra-green cars is 96.52 mpg, and the estimate for the proportion
of all ultra-green cars with mpg greater than 100 is 0.28. It is
important to note that these estimates are based on a sample of 25
cars and, therefore, are likely to vary between samples. Often it is
more informative to provide a range of values—an interval—rather
than a single point estimate for the unknown population parameter.
This range of values is called a confidence interval, also referred to
as an interval estimate, for the population parameter.

CONFIDENCE INTERVAL
A confidence interval, or interval estimate, provides a range of
values that, with a certain level of confidence, contains the
population parameter of interest.

In order to construct a confidence interval for the population

mean μ or the population proportion p, it is essential that the
sampling distributions of and follow, or approximately follow, a

normal distribution. Other methods that do not require the normality

condition are not discussed in this text. Recall from Section 5.1 that
follows a normal distribution when the underlying population is

normally distributed; this result holds irrespective of the sample size

n. If the underlying population is not normally distributed, then by the
central limit theorem, the sampling distribution of will be
approximately normal if the sample size is sufficiently large—that is,
when n ≥ 30. Similarly, the sampling distribution of page 186

is approximately normal if the sample size is sufficiently large—that

is, when np ≥ 5 and n(1 − p) ≥ 5.
The main ingredient for developing a confidence interval is the
sampling distribution of the underlying statistic. The sampling
distribution of for example, describes how the sample mean varies

between samples. Recall that the variability between samples is

measured by the standard error of If the standard error is small, it
implies that the sample means are not only close to one another,
they are also close to the unknown population mean μ.
A confidence interval is generally associated with a margin of
error that accounts for the standard error of the estimator and the
desired confidence level of the interval. The symmetry of the
sampling distributions of both the population mean and the
population proportion allow us to construct a confidence interval by
adding and subtracting the same margin of error to the point
estimate.

GENERAL FORMAT OF THE CONFIDENCE INTERVAL FOR

μ AND p
The confidence interval for the population mean and the
population proportion is constructed as
 An analogy to a simple weather example is instructive. If you feel
that the outside temperature is about 50 degrees, then perhaps you
can, with a certain level of confidence, suggest that the actual
temperature is between 40 and 60 degrees. In this example, 50
degrees is analogous to a point estimate of the actual temperature,
and 10 degrees is the margin of error that is added to and subtracted
from this point estimate.

Confidence Interval for the Population Mean

μ
LO 5.3
Construct a confidence interval for the population mean.

Before constructing a confidence interval for the population mean μ,

we first need to introduce a new distribution. Let’s see why this is the
case.

Recall that if is normally distributed, then any value can be

transformed to its corresponding value z as In reality,

the population standard deviation σ is rarely known. With σ

unknown, the standard error of , given by can be conveniently

estimated by where s denotes the sample standard deviation.

For convenience, we denote this estimate of the standard error of

also by
 Another standardized statistic, which uses the estimator S in

place of σ, is computed as The random variable T follows the

Student’s t distribution, more commonly known as the t

distribution.
The t distribution is actually a family of distributions, which are
similar to the z distribution in that they are all bell-shaped and
symmetric around 0. However, all t distributions have slightly broader
tails than the z distribution. Each t distribution is identified by the
degrees of freedom, or simply, df. The degrees of freedom determine
the extent of the broadness of the tails of the distribution; the fewer
the degrees of freedom, the broader the tails. Because the t
distribution is defined by the degrees of freedom, it is common to
refer to it as the tdf distribution.

THE tdf DISTRIBUTION

If a random sample of size n is taken from a normal population

with a finite variance, then the statistic follows the tdf

distribution, where df denotes degrees of freedom.

Locating tdf Values and Probabilities

Table 5.2 lists tdf values for selected upper-tail probabilities and
degrees of freedom df. Table D.2 of Appendix D provides a more
complete table. Because the tdf distribution is a family of page 187
distributions identified by the df parameter, the t table is
not as comprehensive as the z table. It only lists probabilities
corresponding to a limited number of values. Also, unlike the
cumulative probabilities in the z table, the t table provides the
probabilities in the upper tail of the distribution.
TABLE 5.2 Portion of the t Table

We use the notation tα,df to denote a value such that the area in
the upper tail equals α for a given df. In other words, for a random
variable Tdf, the notation tα,df represents a value such that P(Tdf ≥
tα,df ) = α. Similarly, tα/2,df represents a value such that P(Tdf ≥ tα/2,df)
= α/2. Figure 5.4 illustrates the notation.

FIGURE 5.4 Graphical depiction of P(Tdf ≥ tα,df) = α

When determining the value tα,df, we need two pieces of

information: (a) the sample size n or the degrees of freedom df, and
(b) α. For instance, suppose we want to find the value tα,df with α =
0.05 and df = 10; that is, t0.05,10. Using Table 5.2, we look at the first
column labeled df and find the row 10. We then continue along this
row until we reach the column α = 0.05. The value 1.812 suggests
that P(T10 ≥ 1.812) = 0.05. Due to the symmetry of the t distribution,
we can infer that P(T10 ≤ −1.812) = 0.05. Figure 5.5 shows these
results graphically. Also, because the area under the entire tdf
distribution sums to one, we can infer that P(T10 < 1.812) = 1 − 0.05
= 0.95, which also equals P(T10 > −1.812).

FIGURE 5.5 Graph of the probability α = 0.05 on both sides of T10

page 188

Sometimes the exact probability cannot be determined from the t

table. For example, given df = 10, the exact probability P(T10 ≥
1.562) is not included in the table. However, this probability is
between 0.05 and 0.10 because the value 1.562 falls between 1.372
and 1.812. Similarly, P(T10 < 1.562) is between 0.90 and 0.95. We
can use Excel, R, and other statistical packages to find exact
probabilities.
As mentioned earlier, we construct the confidence interval for μ
as point estimate ± margin of error, where the margin of error
accounts for the desired confidence level of the interval and the
standard error of the sample mean. It can be shown (derivations not
provided) that the margin of error is calculated as the product of
tα/2,df and where tα/2,df is often referred to as the reliability factor.

While it is common to report the 95% confidence interval, in

theory we can construct an interval of any level of confidence. We
now formally let α (alpha) denote the allowed probability of error; in
the next section this is referred to as the signficance level. This is the
probability that the estimation procedure will generate an interval that
does not contain μ. The confidence coefficient (1 − α) is interpreted
as the probability that the estimation procedure will generate an
interval that contains μ. Thus, the probability of error α is related to
the confidence coefficient and the confidence level as follows:
Confidence coefficient = 1 − α, and
Confidence level = 100(1 − α)%.
For example, the confidence coefficient of 0.95 implies that the
probability of error α equals 1 − 0.95 = 0.05 and the confidence level
equals 100(1 − 0.05)% = 95%. Similarly, for the 90% confidence
interval, the confidence coefficient equals 0.90 and α = 1 − 0.90 =
0.10. The following statement generalizes the construction of a
confidence interval for μ.

CONFIDENCE INTERVAL FOR μ

A 100(1 − α)% confidence interval for the population mean μ is
computed as

where s is the sample standard deviation and df = n − 1. This

formula is valid only if (approximately) follows a normal
 distribution.

Confidence intervals are often misinterpreted; we need to

exercise care in characterizing them. For instance, suppose we
construct a 90% confidence interval for μ. This does not imply that
the probability that μ falls in the confidence interval is 0.90.
Remember that μ is a constant, although its value is not known. It
either falls in the interval (probability equals one) or does not fall in
the interval (probability equals 0). The randomness comes from

not μ, because many possible sample means can be derived from a

population. Therefore, it is incorrect to say that the probability that μ
falls in some interval is 0.90. The 90% confidence interval simply
implies that if numerous samples of some predetermined size are
drawn from a given population, then 90% of the intervals formed by
the estimating procedure (formula) will contain μ. Keep in mind that
we only use a single sample to derive the estimates. Because there
are many possible samples, we will be right 90% of the time, thus
giving us 90% confidence.

EXAMPLE 5.5
In a sample of 25 ultra-green cars, we find that the mean miles
per gallon (mpg) is with a standard deviation of s =

10.70. Use this information to construct the 90% confidence

interval for the population mean. Assume that mpg follows a
normal distribution.

page 189
 SOLUTION: The condition that follows a normal distribution is
satisfied since we assumed that mpg is normally distributed.
We construct the confidence interval as For the 90%

confidence interval, α = 0.10, α/2 = 0.05, and, given n = 25, df

= 25 − 1 = 24. Thus, t0.05,24 = 1.711. The 90% confidence
interval for μ is computed as

The 90% confidence interval for the average mpg of all ultra-
green cars is between 92.86 mpg and 100.18 mpg.

Using Excel and R to Construct a Confidence

Interval for μ
Excel and R are quite useful when constructing confidence intervals.
Consider the following example.

EXAMPLE 5.6
Amazon Prime is a $119-per-year service that gives the
company’s customers free two-day shipping and discounted
rates on overnight delivery. Prime customers also get other
perks, such as free e-books. Table 5.3 shows a portion of the
annual expenditures (in $) for 100 Prime customers. Use Excel
and R to construct the 95% confidence interval for the average
 annual expenditures of all Prime customers. Summarize the
results.
FILE
Prime

TABLE 5.3 Annual Prime Expenditures (in $)

Customer Expenditures

1 1272

2 1089

⋮ ⋮

100 1389

SOLUTION: We compute or, equivalently, we find the

lower and upper limits of the confidence interval:

Using Excel

a. Open the Prime data file. Note that the observations for the
Expenditures variable are in cells B2 through B101.
b. For a 95% confidence interval, with n = 100, we find t0.025,99
using = T.INV(0.975,99). Thus, in order to obtain the lower
limit, we enter =AVERAGE(B2:B101) − T.INV(0.975,99)
*STDEV.S(B2:101)/SQRT(100). For the upper limit, we enter
=AVERAGE(B2:B101) + T.INV(0.975,99)
*STDEV.S(B2:101)/SQRT(100).
 Note: For a one-step approach to constructing a confidence
interval in Excel, we can use the Descriptive Statistics option
in its Analysis Toolpak which we discussed in Chapter 3. In the
Descriptive Statistics dialog box, we select Summary statistics
and Confidence Interval for Mean. (By default, the confidence
level is set at 95%, but you easily enter another level.) In the
table that Excel returns, we find the mean and the margin of
error which is labeled Confidence Level(95.0%).

page 190
Using R

a. Import the Prime data into a data frame (table) and label it
myData.
b. For a 95% confidence interval, with n = 100, we find t0.025,99
using the qt function. Thus, in order to obtain the lower and
upper limits, we enter:

lower <− mean(myData$Expenditures) − qt(0.975, 99,

lower.tail = TRUE)*sd(myData$Expenditures)/sqrt(100)
upper <− mean(myData$Expenditures) + qt(0.975, 99,
lower.tail = TRUE)*sd(myData$Expenditures)/sqrt(100)

Note: For a one-step approach to constructing a confidence

interval in R, refer to Example 5.12.
Summary
With 95% confidence, we conclude that the average annual
expenditures of all Prime customers fall between $1,240.24
and $1,373.64

Confidence Interval for the Population

Proportion p
 LO 5.4
Construct a confidence interval for the population proportion.

Recall that while the population mean μ describes a numerical

variable, the population proportion p is the essential descriptive
measure for a categorical variable. The parameter p represents the
proportion of successes in the population, where success is defined
by a particular outcome.
As in the case of the population mean, we estimate the
population proportion on the basis of its sample counterpart. In
particular, we use the sample proportion as the point estimator of

the population proportion p. Also, although the sampling distribution

of is based on the binomial distribution, we can approximate it by

the normal distribution for large samples, according to the central

limit theorem. This approximation is valid when the sample size n is
such that np ≥ 5 and n(1 − p) ≥ 5.

Using the normal approximation for with and

and analogous to the construction of the confidence

interval for the population mean, a 100(1 − α)% confidence interval

for the population proportion is where the notation zα/2

(the reliability factor) is the z value associated with the probability of

α/2 in the upper tail of the standard normal distribution.
This confidence interval is theoretically sound; however, it cannot
be implemented because it uses p in the derivation, which is
unknown. Because we always use large samples for the normal
distribution approximation, we can also conveniently replace p with
its estimate in the construction of the interval. Therefore, for

we substitute This substitution yields a feasible confidence

interval for the population proportion. The following statement

generalizes the construction of a confidence interval for p.

CONFIDENCE INTERVAL FOR p

A 100(1 − α)% confidence interval for the population proportion
p is computed as

This formula is valid only if (approximately) follows a normal

distribution.

page 191

The normality condition is evaluated at the sample proportion

In other words, for constructing a confidence interval for the

population proportion p, we require that and

EXAMPLE 5.7
In a sample of 25 ultra-green cars, seven of the cars obtained
over 100 miles per gallon (mpg). Construct 90% and 99%
confidence intervals for the population proportion of all ultra-
green cars that obtain over 100 mpg.

SOLUTION: The point estimate for the population proportion is

Note that the normality condition is satisfied

because np ≥ 5 and n(1 − p) ≥ 5, where p is evaluated at

With the 90% confidence level, α/2 = 0.10/2 = 0.05;

thus, we find zα/2 = z0.05 = 1.645. Substituting the appropriate

values into

With 90% confidence, the percentage of cars that obtain over

100 mpg is between 13.2% and 42.8%.
For the 99% confidence interval, we use α/2 = 0.01/2 =
0.005 and zα/2 = z0.005 = 2.576 to obtain

At a higher confidence level of 99%, the interval for the

percentage of cars that obtain over 100 mpg becomes 4.9% to
51.1%. Given the current sample size of 25 cars, we gain
confidence (from 90% to 99%) at the expense of precision, as
 the corresponding margin of error increases from 0.148 to
0.231.

EXERCISES 5.2
Applications
16. A popular weight loss program claims that with its recommended healthy diet
regimen, users lose significant weight within a month. In order to estimate the
mean weight loss of all customers, a nutritionist takes a sample of 18 dieters and
records their weight loss one month after joining the program. He computes the
mean and the standard deviation of weight loss as 12.5 pounds and 9.2 pounds,
respectively. He believes that weight loss is likely to be normally distributed.
a. Calculate the margin of error with 95% confidence.
b. Calculate the 95% confidence interval for the population mean.
c. How can the margin of error reported in part a be reduced?
17. FILE Customers. The manager of The Cheesecake Factory in Memphis reports
that on six randomly selected weekdays, the number of customers served was
120, 130, 100, 205, 185, and 220. She believes that the number of customers
served on weekdays follows a normal distribution.
a. Calculate the margin of error with 90% confidence.
b. Construct the 90% confidence interval for the average number of customers
served on weekdays.
c. How can the margin of error reported in part a be reduced?
18. According to a recent survey, high school girls average 100 text messages daily.
Assume that the survey was based on a random sample of 36 high school girls.
The sample standard deviation is computed as 10 text messages daily.
a. Calculate the margin of error with 99% confidence.
b. What is the 99% confidence interval for the population mean texts that all high
school girls send daily?
19. The Chartered Financial Analyst (CFA) designation is fast becoming a
requirement for serious investment professionals. Although it
requires a successful completion of three levels of grueling
page 192
exams, the designation often results in a promising career with a lucrative salary.
A student of finance is curious about the average salary of a CFA charterholder.
He takes a random sample of 36 recent charterholders and computes a mean
salary of $158,000 with a standard deviation of $36,000. Use this sample
information to determine the 95% confidence interval for the average salary of a
CFA charterholder.
20. The sudoku puzzle has become very popular all over the world. It is based on a 9
× 9 grid and the challenge is to fill in the grid so that every row, every column, and
every 3 × 3 box contains the digits 1 through 9. A researcher is interested in
estimating the average time taken by a college student to solve the puzzle. He
takes a random sample of eight college students and records their solving times
(in minutes) as 14, 7, 17, 20, 18, 15, 19, 28.
a. Construct the 99% confidence interval for the average time taken by a college
student to solve a sudoku puzzle.
b. What assumption is necessary to make this inference?
21. FILE Stock_Price. The monthly closing stock price for a large technology firm
for the first six months of the year are reported in the following table.

Month Price

January 71

February 73

March 76

April 78

May 81

June 75

a. Calculate the sample mean and the sample standard deviation.

b. Calculate the 90% confidence interval for the mean stock price of the firm,
assuming that the stock price is normally distributed.
c. What happens to the margin of error if a higher confidence level is used for the
interval estimate?
22. Suppose the 90% confidence interval for the mean SAT scores of applicants at a
business college is given by [1690, 1810]. This confidence interval uses the
sample mean and the sample standard deviation based on 25 observations. What
are the sample mean and the sample standard deviation used when computing
the interval?
23. A teacher wants to estimate the mean time (in minutes) that students take to go
from one classroom to the next. His research assistant uses the sample time of
36 students to report the confidence interval as [8.20, 9.80].
a. Find the sample mean time used to compute the confidence interval.
b. Determine the confidence level if the sample standard deviation used for the
interval is 2.365.
24. In order to attract more Millennial customers (those born between 1980 and
2000), a new clothing store offers free gourmet coffee and pastry to its customers.
The average daily revenue over the past five-week period has been $1,080 with a
standard deviation of $260. Use this sample information to construct the 95%
confidence interval for the average daily revenue. The store manager believes
that the coffee and pastry strategy would lead to an average daily revenue of
$1,200. Use the 95% interval to determine if the manager is wrong.
25. FILE Debt_Payments. The accompanying table shows a portion of average
monthly debt payments (Debt in $) for residents of 26 metropolitan areas.
Construct 90% and 95% confidence intervals for the population mean. Comment
on the width of the intervals.

City Debt

Washington, D.C. 1285

Seattle 1135

⋮ ⋮

Pittsburgh 763

26. FILE Economics. An associate dean of a university wishes to compare the

means on the standardized final exams in microeconomics and macroeconomics.
He has access to a random sample of 40 scores from each of these two courses.
A portion of the data is shown in the accompanying table.

Micro Macro

85 48

78 79

⋮ ⋮

75 74

a. Construct 95% confidence intervals for the mean score in microeconomics and
the mean score in macroeconomics.
b. Explain why the widths of the two intervals are different.
27. FILE Math_Scores. For decades, people have believed that boys are innately
more capable than girls in math. In other words, due to the intrinsic differences in
brains, boys are believed to be better suited for doing math than girls. Recent
 research challenges this stereotype, arguing that gender differences in math
performance have more to do with culture than innate aptitude. Others argue,
however, that while the average may be the same, there is more variability in
math ability for boys than girls, resulting in some boys with soaring math skills. A
portion of sample data on math scores of boys and girls is shown
in the accompanying table.
page 193

Boys Girls

74 83

89 76

⋮ ⋮

66 74

a. Construct 95% confidence intervals for the mean scores of boys and the mean
scores of girls. Explain your assumptions.
b. Explain why the widths of the two intervals are different.
28. FILE Startups. Many of today’s leading companies, including Google, Microsoft,
and Facebook, are based on technologies developed within universities. Lisa
Fisher is a business school professor who believes that a university’s research
expenditure (Research in $ millions) and the age of its technology transfer office
(Duration in years) are major factors that enhance innovation. She wants to know
what the average values are for the Research and the Duration variables. She
collects data from 143 universities on these variables, a portion of which is shown
in the accompanying table.

Research Duration

145.52 23

237.52 23

⋮ ⋮

154.38 9

a. Construct and interpret the 95% confidence interval for the mean research
expenditure of all universities.
b. Construct and interpret the 95% confidence interval for the mean duration of all
universities.
29. A survey of 1,026 people were asked about what they would do with an
unexpected cash gift. Forty-seven percent responded that they would pay off
debts.
a. With 95% confidence, what is the margin of error?
b. Construct the 95% confidence interval for the population proportion of people
who would pay off debts with an unexpected cash gift.
30. A sample of 5,324 Americans were asked about what matters most to them in a
place to live. Thirty-seven percent of the respondents felt job opportunities matter
most.
a. Construct the 90% confidence interval for the proportion of Americans who feel
that good job opportunities matter most in a place to live.
b. Construct the 99% confidence interval for the proportion of Americans who feel
that good job opportunities matter most in a place to live.
c. Which of the above two intervals has a higher margin of error? Explain why.
31. An economist reports that 560 out of a sample of 1,200 middle-income American
households actively participate in the stock market.
a. Construct the 90% confidence interval for the proportion of middle-income
Americans who actively participate in the stock market.
b. Can we conclude that the percentage of middle-income Americans who actively
participate in the stock market is not 50%?
32. In a survey of 1,116 adults, 47% approved of the job that President Trump was
doing in handling the economy (Opinion Today, July 1, 2019).
a. Compute the 90% confidence interval for the proportion of Americans who
approved of President Trump’s handling of the economy.
b. What is the resulting margin of error?
c. Compute the margin of error associated with the 99% confidence level.
33. In a recent poll of 760 homeowners in the United States, one in five homeowners
reports having a home equity loan that he or she is currently paying off. Using a
confidence coefficient of 0.90, construct the confidence interval for the proportion
of all homeowners in the United States that hold a home equity loan.
34. Obesity is generally defined as 30 or more pounds over a healthy weight. A recent
study of obesity reports 27.5% of a random sample of 400 adults in the United
States to be obese.
a. Use this sample information to compute the 90% confidence interval for the
adult obesity rate in the United States.
b. Is it reasonable to conclude with 90% confidence that the adult obesity rate in
the United States differs from 30%?
35. An accounting professor is notorious for being stingy in giving out good letter
grades. In a large section of 140 students in the fall semester, she gave out only
5% A’s, 23% B’s, 42% C’s, and 30% D’s and F’s. Assuming that this was a
representative class, compute the 95% confidence interval of the probability of
getting at least a B from this professor.
 page 194

5.3 HYPOTHESIS TESTING

Each and every, day people make decisions based on their beliefs
about the true state of the world. They hold certain things to be true
and others to be false, and then act accordingly. For example, an
engineer believes that a certain steel cable has a breaking strength
of 5,000 pounds or more, and then permits its use at a construction
site; a manufacturer believes that a certain process yields capsules
that contain precisely 100 milligrams of a drug, and then ships the
capsules to a pharmacy; a manager believes that an incoming
shipment contains 2%, or fewer, of defects, and then accepts the
shipment.
In these cases, and many more, the formation of these beliefs
may have started as a mere conjecture, an informed guess, or a
proposition tentatively advanced as true. When people formulate a
belief in this way, we refer to it as a hypothesis. Sooner or later,
however, every hypothesis eventually confronts evidence that either
substantiates or refutes it. Determining the validity of an assumption
of this nature is called hypothesis testing.
We use hypothesis testing to resolve conflicts between two
competing hypotheses on a particular population parameter of
interest. We refer to one hypothesis as the null hypothesis,
denoted H0, and the other as the alternative hypothesis, denoted
HA. We think of the null hypothesis as corresponding to a presumed
default state of nature or status quo. The alternative hypothesis, on
the other hand, contradicts the default state or status quo.

NULL HYPOTHESIS VERSUS ALTERNATIVE HYPOTHESIS

 When constructing a hypothesis test, we define a null
hypothesis, denoted H0, and an alternative hypothesis, denoted
HA. We conduct a hypothesis test to determine whether or not
sample evidence contradicts H0.

The hypothesis testing procedure enables us to make one of two

decisions. If sample evidence is inconsistent with the null hypothesis,
we reject the null hypothesis. Conversely, if sample evidence is not
inconsistent with the null hypothesis, then we do not reject the null
hypothesis. It is not correct to conclude that “we accept the null
hypothesis” because while the sample information may not be
inconsistent with the null hypothesis, it does not necessarily prove
that the null hypothesis is true.
Consider an example from the medical field where the null is
defined as “an individual is free of a particular disease.” Suppose a
medical procedure does not detect this disease. On the basis of this
limited information, we can only conclude that we are unable to
detect the disease (do not reject the null hypothesis). It does not
necessarily prove that the person does not have the disease (accept
the null hypothesis).

Defining the Null and the Alternative Hypotheses

A very crucial step in a hypothesis test concerns the formulation of
the two competing hypotheses because the conclusion of the test
depends on how the hypotheses are stated. As a general guideline,
whatever we wish to establish is placed in the alternative hypothesis,
whereas the null hypothesis includes the status quo. If we are unable
to reject the null hypothesis, then we maintain the status quo or
“business as usual.” However, if we reject the null hypothesis, this
establishes that the evidence supports the alternative hypothesis,
which may require that we take some kind of action. For instance, if
we reject the null hypothesis that an individual is free of a particular
disease, then we conclude that the person is sick, for which
treatment is prescribed.
 In most applications, we require some form of the equality sign in
the null hypothesis. (The justification for the equality sign will be
provided later.) In general, any statement including one of the three
signs “=”, “≤”, or “≥” is valid for the null hypothesis. Given that the
alternative hypothesis states the opposite of the null hypothesis, the
alternative hypothesis is then specified with a “≠”, “>”, or “<” sign.

page 195

CONSTRUCTION OF THE COMPETING HYPOTHESES

As a general guideline, we use the alternative hypothesis as a
vehicle to establish something new—that is, contest the status
quo. In most applications, the null hypothesis regarding a
particular population parameter of interest is specified with one
of the following signs: =, ≤, or ≥; the alternative hypothesis is
then specified with the corresponding opposite sign: ≠, >, or <.

A hypothesis test can be one-tailed or two-tailed. A two-tailed test

is defined when the alternative hypothesis includes the sign “≠”. For
example, H0: μ = μ0 versus HA: μ ≠ μ0 and H0: p = p0 versus HA: p ≠
p0 are examples of two-tailed tests, where μ0 and p0 represent
hypothesized values of the population mean and the population
proportion, respectively. If the null hypothesis is rejected, it suggests
that the true parameter does not equal the hypothesized value.
A one-tailed test, on the other hand, involves a null hypothesis
that can only be rejected on one side of the hypothesized value. For
example, consider H0: μ ≤ μ0 versus HA: μ > μ0. Here we can reject
the null hypothesis only when there is substantial evidence that the
population mean is greater than μ0. It is also referred to as a right-
tailed test because rejection of the null hypothesis occurs on the
right side of the hypothesized mean. Another example is a left-tailed
test, H0: μ ≥ μ0 versus HA: μ < μ0, where the null hypothesis can only
 be rejected on the left side of the hypothesized mean. One-tailed
tests for the population proportion are defined similarly.
In general, we follow three steps when formulating the competing
hypotheses:
. Identify the relevant population parameter of interest.
2. Determine whether it is a one- or two-tailed test.
3. Include some form of the equality sign in the null hypothesis and
use the alternative hypothesis to establish a claim.
In the next two examples, our objective is to state the appropriate
competing hypotheses.

EXAMPLE 5.8
A trade group predicts that back-to-school spending will
average $606.40 per family this year. A different economic
model is needed if the prediction is wrong. Specify the null and
the alternative hypotheses to determine if a different economic
model is needed.

SOLUTION: Given that we are examining average back-to-school

spending, the parameter of interest is the population mean μ.
Because we want to be able to determine if the population
mean differs from $606.40 (μ ≠ 606.40), we need a two-tailed
test and formulate the null and alternative hypotheses as

The trade group is advised to use a different economic model if

the null hypothesis is rejected.
 page 196

EXAMPLE 5.9
A television research analyst wishes to test a claim that more
than 50% of the households will tune in for a TV episode.
Specify the null and the alternative hypotheses to test the
claim.

SOLUTION: This is an example of a one-tailed test regarding the

population proportion p. Given that the analyst wants to
determine whether p > 0.50, this claim is placed in the
alternative hypothesis, whereas the null hypothesis is just its
opposite.

The claim that more than 50% of the households will tune in for
a TV episode is valid only if the null hypothesis is rejected.

Type I and Type II Errors

Because the decision of a hypothesis test is based on limited sample
information, we are bound to make errors. Ideally, we would like to
be able to reject the null hypothesis when the null hypothesis is false
and not reject the null hypothesis when the null hypothesis is true.
However, we may end up rejecting or not rejecting the null
hypothesis erroneously. In other words, sometimes we reject the null
hypothesis when we should not, or not reject the null hypothesis
when we should.
 We consider two types of errors in the context of hypothesis
testing: a Type I error and a Type II error. A Type I error is
committed when we reject the null hypothesis when the null
hypothesis is actually true. On the other hand, a Type II error is
made when we do not reject the null hypothesis when the null
hypothesis is actually false.
Table 5.4 summarizes the circumstances surrounding Type I and
Type II errors. Two correct decisions are possible: not rejecting the
null hypothesis when the null hypothesis is true and rejecting the null
hypothesis when the null hypothesis is false. Conversely, two
incorrect decisions (errors) are also possible: rejecting the null
hypothesis when the null hypothesis is true (Type I error) and not
rejecting the null hypothesis when the null hypothesis is false (Type
II error).

TABLE 5.4 Type I and Type II Errors

Null hypothesis Null hypothesis

Decision
is true is false

Reject the null Type I error Correct decision

hypothesis

Do not reject the null Correct decision Type II error

hypothesis

EXAMPLE 5.10
An online retailer is deciding whether or not to build a brick-
and-mortar store in a new marketplace. A market analysis
determines that the venture will be profitable if average
pedestrian traffic exceeds 500 people per day. The competing
hypotheses are specified as follows.
 Discuss the consequences of a Type I error and a Type II error.

page 197

SOLUTION: A Type I error occurs when the conclusion to the

hypothesis test is to reject H0 so the retailer builds the brick-
and-mortar store, but, in reality, average pedestrian traffic does
not exceed 500 people per day and the venture will not be
profitable. A Type II error occurs when the conclusion to the
hypothesis test is to not reject H0 so the retailer does not build
the brick-and-mortar store, but, in reality, average pedestrian
traffic exceeds 500 people per day and the venture would have
been profitable. Arguably, the consequences of a Type I error
in this example are more serious than those of a Type II error.

It is not always easy to determine which of the two errors has

more serious consequences. For given evidence, there is a trade-off
between these errors; by reducing the likelihood of a Type I error, we
implicitly increase the likelihood of a Type II error, and vice versa.
The only way we can reduce both errors is by collecting more
evidence. Let us denote the probability of a Type I error by α, the
probability of a Type II error by β, and the strength of the evidence by
the sample size n. The only way we can lower both α and β is by
increasing n. For a given n, however, we can reduce α only at the
expense of a higher β and reduce β only at the expense of a higher
α. The optimal choice of α and β depends on the relative cost of
these two types of errors, and determining these costs is not always
easy. Typically, the decision regarding the optimal level of Type I and
Type II errors is made by the management of a firm where the job of
a data analyst is to conduct the hypothesis test for a chosen value of
α.

Hypothesis Test for the Population Mean μ

There are two approaches to implementing a hypothesis test—the p-
value approach and the critical value approach. The critical value
approach is attractive when a computer is unavailable and all
calculations must be done by hand. Most researchers and
practitioners, however, favor the p-value approach because virtually
every statistical software package reports a p-value. We too will
focus on the p-value approach. We implement a four-step procedure
that is valid for one- and two-tailed tests regarding the population
mean, the population proportion, or any other population parameter
of interest.
A hypothesis test regarding the population mean μ is based on
the sampling distribution of the sample mean In particular, it uses

the fact that and As mentioned in the last section,

because σ is rarely known, we estimate Also, in order to

implement the test, it is essential that is normally distributed.

Recall that is normally distributed when the underlying population

is normally distributed. If the underlying population is not normally
distributed, then, by the central limit theorem, is approximately
normally distributed if the sample size is sufficiently large—that is, n
≥ 30.

LO 5.5
Conduct a hypothesis test for the population mean.

The basic principle of hypothesis testing is to first assume that

the null hypothesis is true and then determine if sample evidence
contradicts this assumption. Suppose a hiring manager wants to
establish that the mean retirement age is greater than 67 (μ > 67). It
is assumed that retirement age is normally distributed. We can
investigate the manager’s belief by specifying the competing
hypotheses as

Let a random sample of 25 retirees produce an average retirement

age of 71 with a standard deviation of 9—that is, and s = 9.
This sample evidence casts doubt on the validity of the null
hypothesis because the sample mean is greater than the
hypothesized value, μ0 = 67. However, the discrepancy between
and μ0 does not necessarily imply that the null hypothesis is false.
Perhaps the discrepancy can be explained by pure chance. It is
common to evaluate this discrepancy in terms of the appropriate test
statistic.

page 198

TEST STATISTIC FOR μ

The value of the test statistic for the hypothesis test of the
population mean μ is computed as
 where μ0 is the hypothesized value of the population mean and
the degrees of freedom df = n − 1. This formula is valid only if

(approximately) follows a normal distribution.

Note that the value of the test statistic tdf is evaluated at μ = μ0,
which explains why we need some form of the equality sign in the
null hypothesis. Given that the population is normally distributed with
n = 25 (so df = 24), and s = 9, we compute the value of the test

statistic as Therefore, comparing with 67 is

identical to comparing t24 = 2.22 with 0, where 67 and 0 are the

means of and T24, respectively.

We now find the p-value, which is the likelihood of obtaining a

sample mean that is at least as extreme as the one derived from the
given sample, under the assumption that the null hypothesis is true
as an equality—that is, μ0 = 67. Because in this example we

define the extreme value as a sample mean of 71 or higher, so we

find the p-value as Table 5.5 shows a portion

of the t table. Referencing Table 5.5 for df = 24, we find that the
exact probability P(T24 ≥ 2.22) cannot be determined. Because 2.22
lies between 2.064 and 2.492, this implies that the p-value is
between 0.01 and 0.025. Using Excel or R (instructions shown
shortly), we find that the exact p-value is 0.018. Figure 5.6 shows the
computed p-value.

TABLE 5.5 Portion of the t Table

FIGURE 5.6 The p-value for a right-tailed test with t24 = 2.22

Note that when the null hypothesis is true, there is only a 1.8%
chance that the sample mean will be 71 or more. This seems like a
very small chance, but is it small enough to allow us to reject the null
hypothesis in favor of the alternative hypothesis? Let’s see how we
define “small enough.”

page 199

Remember that a Type I error occurs when we reject the null

hypothesis when it is actually true. We define the allowed probability
of making a Type I error as α; we refer to 100α% as the significance
level. The p-value, on the other hand, is referred to as the observed
probability of making a Type I error. When using the p-value
approach, the decision rule is:
Reject the null hypothesis if the p-value < α, or
Do not reject the null hypothes if the p-value ≥ α.
We generally choose a value for α before implementing a
hypothesis test; that is, we set the rules of the game before playing.
Care must be exercised in choosing α because important decisions
are often based on the results of a hypothesis test, which in turn
depend on α. Most hypothesis tests are conducted using a
significance level of 1%, 5%, or 10%, using α = 0.01, 0.05, or 0.10,
respectively. For example, α = 0.05 means that we allow a 5%
chance of rejecting a true null hypothesis.
In the retirement age example, suppose we had chosen α = 0.05
to conduct the hypothesis test. At this significance level, we reject
the null hypothesis because 0.018 < 0.05. This means that the
sample data support the manager’s claim that the average
retirement age is greater than 67 years old. Individuals may be
working past the normal retirement age of 67 because of poor
savings and/or because this generation is expected to outlive any
previous generation and needs jobs to pay the bills. We should note
that if α had been set at 0.01, then the findings would have been
different. At this smaller significance level, the evidence does not
allow us to reject the null hypothesis (0.018 > 0.01). At the 1%
significance level, we cannot conclude that the mean retirement age
is greater than 67.
In the retirement age example of a right-tailed test, we calculated
the p-value as P(Tdf ≥ tdf). Analogously, for a left-tailed test, the p-
value is given by P(Tdf ≤ tdf). For a two-tailed test, the extreme
values exist on both sides of the distribution of the test statistic.
Given the symmetry of the tdf distribution, the p-value for a two-tailed
test is twice that of the p-value for a one-tailed test. It is calculated as
2P(Tdf ≥ tdf) if tdf > 0 or as 2P(Tdf ≤ tdf) if tdf < 0.

THE p-VALUE APPROACH

Under the assumption that μ = μ0, the p-value is the likelihood
of observing a sample mean that is at least as extreme as the
one derived from the given sample. Its calculation depends on
the specification of the alternative hypothesis.

Alternative
p-value
Hypothesis

HA: μ > μ0 Right-tail probability: P(Tdf ≥ tdf)

HA: μ < μ0 Left-tail probability: P(Tdf ≤ tdf)

HA: μ ≠ μ0 Two-tail probability: 2P(Tdf ≥ tdf) if tdf > 0 or

2P(Tdf ≤ tdf) if tdf < 0

The decision rule is:

Reject H0 if the p-value < α, or
Do not reject H0 if the p-value ≥ α.

Figure 5.7 shows the three different scenarios of determining the

p-value depending on the specification of the competing hypotheses.
Figure 5.7a shows the p-value for a left-tailed test. Because the
appropriate test statistic follows the tdf distribution, we compute the
p-value as P(Tdf ≤ tdf). When calculating the p-value for a right-tailed
test (see Figure 5.7b), we find the area to the right of the value of the
test statistic tdf or, equivalently, P(Tdf ≥ tdf). Figure 5.7c shows the p-
value for a two-tailed test, calculated as 2P(Tdf ≤ tdf) when tdf < 0 or
as 2P(Tdf ≥ tdf) when tdf > 0.

page 200

FIGURE 5.7 The p-values for one- and two-tailed tests

It is important to note that we cannot reject H0 for a right-tailed

test if or, equivalently, tdf ≤ 0. Consider, for example, a right-

tailed test with the hypotheses specified as H0: μ ≤ 67 versus HA: μ >

67. Here, if there is no need for formal testing because we

have no discrepancy between the sample mean and the

hypothesized value of the population mean. Similarly, we cannot
reject H0 for a left-tailed test if or, equivalently, tdf ≥ 0. We will

now summarize the four-step procedure using the p-value approach.

THE FOUR-STEP PROCEDURE USING THE p-VALUE

APPROACH
Step 1. Specify the null and the alternative hypotheses. We
identify the relevant population parameter of interest, determine
whether it is a one- or a two-tailed test and, most importantly,
include some form of the equality sign in the null hypothesis
and place whatever we wish to establish in the alternative
hypothesis.

Step 2. Specify the significance level. Before implementing a

hypothesis test, we first specify α, which is the allowed
probability of making a Type I error.

Step 3. Calculate the value of the test statistic and the p-value.
When testing the population mean μ, the value of the test
statistic is where μ0 is the hypothesized value of the

population mean. For a right-tailed test, the p-value is P(Tdf ≥

tdf), and for a left-tailed test, the p-value is P(Tdf ≤ tdf). For a
two-tailed test, the p-value is 2P(Tdf ≥ tdf) if tdf > 0, or 2P(Tdf ≤
tdf) if tdf < 0.

Step 4. State the conclusion and interpret the results. The

decision rule is to reject the null hypothesis when the p-value <
α and not reject the null hypothesis when the p-value ≥ α.
Clearly interpret the results in the context of the application.

page 201
 EXAMPLE 5.11
FILE
Study_Hours

In the introductory case to this chapter, the dean at a large

university in California wonders if students at her university
study less than the 1961 national average of 24 hours per
week. She randomly selects 35 students and asks their
average study time per week (in hours). From their responses,
she calculates a sample mean of 16.3714 hours and a sample
standard deviation of 7.2155 hours.
a. Specify the competing hypotheses to test the dean’s concern.
b. Calculate the value of the test statistic.
c. Find the p-value.
d. At the 5% significance level, what is the conclusion to the
hypothesis test?
SOLUTION:

a. This is an example of a one-tailed test where we would like to

determine if the mean hours studied is less than 24; that is, μ
< 24. We formulate the competing hypotheses as

b. Recall that for any statistical inference regarding the

population mean, it is essential that the sample mean is
normally distributed. This condition is satisfied because the
sample size is greater than 30, specifically n = 35. The
 degrees of freedom, df = n − 1 = 34. Given and s =
7.2155, we compute the value of the test statistic as

c. Because this is a left-tailed test, we compute the p-value as

P(T34 ≤ −6.255). Referencing the t table for df = 34, we can
infer that the probability P(T34 ≤ −6.255), which is equivalent
to P(T34 ≥ 6.255), is less than 0.005. In other words, we
approximate the p-value as P(T34 ≤ −6.255) < 0.005. Using
Excel or R (instructions shown next), we find that the exact
probability is 2.013 × 10−7.
d. We reject the null hypothesis because the p-value is less than
α = 0.05. At the 5% significance level, we conclude that the
average study time at the university is less than the 1961
national average of 24 hours per week.

Using Excel and R to Test μ

Again we find that functions in Excel and R are quite useful when
calculating the value of the test statistic and the exact p-value.
Consider the following example.

EXAMPLE 5.12
FILE
Study_Hours
 As the introductory case to this chapter mentions, research
finds that today’s undergraduates study an average of 14 hours
per week. Using the sample data from Table 5.1, the dean at a
large university in California would like to test if the mean study
time of students at her university differs from 14 hours per
week. At the 5% significance level, use Excel and R to find the
value of the test statistic and the p-value. Summarize the
results.

page 202

SOLUTION:Because the dean would like to test if the mean study

time of students at her university differs from 14 hours per
week, we formulate the competing hypotheses as

Using Excel

a. Open the Study_Hours data file. Note that the observations

for hours studied are in cells A2 through A36.

b. We find the value of the test statistic, using =

(AVERAGE(A2:A36)−14)/(STDEV.S(A2:A36)/SQRT(36)). We
obtain t34 = 1.9444.
c. Even though Excel offers a number of functions that generate
p-values, we use the T.DIST.RT function. Here, for a two-tailed
test and a positive value for the test statistic, we enter
=2*T.DIST.RT(1.9444, 34) and Excel returns 0.0602.
Using R

a. Import the Study_Hours data into a data frame (table) and

label it myData.
b. We use R’s t.test function to obtain both the test statistic and
the p-value. For options within this function, we use alternative
to denote the specification of the alternative hypothesis
(denoted as “two.sided” for a two-tailed test, “less” for a left-
tailed test, and “greater” for a right-tailed test) and mu to
denote the hypothesized value of the mean. Another feature
of this function is that it automatically provides the 95%
confidence interval for the mean by default; other levels can
be found using the option conf.level. We enter

> t.test(myData$Hours, alternative=“two.sided”, mu=14)

FIGURE 5.8 R’s output using t.test function

Summary
Because the p-value of 0.0602 is not less than α = 0.05, we do
not reject the null hypothesis. At the 5% significance level, we
cannot conclude that the mean study time of students at this
large university in California differs from 14 hours per week.

page 203

SYNOPSIS OF INTRODUCTORY CASE

 A report claims that undergraduates are studying far less today as compared to six
decades ago (The Wall Street Journal, April 10, 2019). In 1961, students invested
24 hours per week in their academic pursuits, whereas today’s students study an
average of 14 hours per week. In an attempt to determine whether or not this
national trend is present at a large university in California, 35 students are randomly
selected and asked their average study time per week (in hours). The sample
produces a mean of 16.37 hours with a standard deviation of 7.22 hours.

Asia Images Group/Getty Images

Two hypothesis tests are conducted. The first test examines whether the mean
study time of students at this university is below the 1961 national average of 24
hours per week. At the 5% significance level, the sample data suggest that the mean
is less than 24 hours per week. The second test investigates whether the mean
study time of students at this university differs from today’s national average of 14
hours per week. At the 5% significance level, the results do not suggest that the
mean study time is different from 14 hours per week.
Thus, the sample results support the overall findings of the report:
undergraduates study, on average, 14 hours per week, far below the 1961 average
of 24 hours per week. The present analysis, however, does not explain why that
might be the case. For instance, it cannot be determined whether students have just
become lazier or if, with the advent of the computer, they can access information in
less time.

Hypothesis Test for the Population

Proportion p
 LO 5.6
Conduct a hypothesis test for the population proportion.

As discussed earlier, sometimes the variable of interest is categorical

rather than numerical. Recall that the population proportion p is the
essential descriptive measure for a qualitative variable, and that it is
estimated on the basis of its sample counterpart, the sample
proportion Although is based on a binomial distribution, it can be

approximated by a normal distribution in large samples. This

approximation is considered valid when np ≥ 5 and n(1 − p) ≥ 5.
Because p is not known, we typically test the sample size
requirement under the hypothesized value of the population
proportion p0. In most applications, the sample size is large and the
normal distribution approximation is justified. However, when the
sample size is not deemed large enough, the statistical methods
suggested here for inference regarding the population proportion are
no longer valid. The test statistic for p is defined as follows.

TEST STATISTIC FOR p

The value of the test statistic for the hypothesis test of the
population proportion p is computed as

where p0 is the hypothesized value of the population

proportion. This formula is valid only if (approximately) follows
a normal distribution.
 page 204

The following example elaborates on the four-step procedure for

a hypothesis test for the population proportion.

EXAMPLE 5.13
Driven by growing public support, the legalization of marijuana
in America has been moving at a very rapid rate. Today, 57% of
adults say the use of marijuana should be made legal
(www.pewresearch.org, October 12, 2016). A health
practitioner in Ohio collects data from 200 adults and finds that
102 of them favor marijuana legalization.
a. The health practitioner believes that the proportion of adults
who favor marijuana legalization in Ohio is not representative
of the national proportion. Specify the competing hypotheses
to test her claim.
b. Calculate the value of the test statistic and the p-value.
c. At the 10% significance level, do the sample data support the
health practitioner’s belief?
SOLUTION:

a. The parameter of interest is the population proportion p. The

health practitioner wants to test if the population proportion of
those who favor marijuana legalization in Ohio differs from the
national proportion of 0.57. We construct the competing
hypotheses as
b. When evaluated at p0 = 0.57 with n = 200, the normality
requirement that np ≥ 5 and n(1 − p) ≥ 5 is easily satisfied. We
use the sample proportion to compute the

value of the test statistic as

Given a two-tailed test and z < 0, we compute the p-value as

2P(Z ≤ z) = 2P(Z ≤ − 1.71) = 0.0872.
c. Because the p-value of 0.0872 is less than α = 0.10, we reject
the null hypothesis. Therefore, at the 10% significance level,
the proportion of adults who favor marijuana legalization in
Ohio differs from the national proportion of 0.57.

EXERCISES 5.3
Applications
36. Construct the null and the alternative hypotheses for the following tests:
a. Test if the mean weight of cereal in a cereal box differs from 18 ounces.
b. Test if the stock price increases on more than 60% of the trading days.
c. Test if Americans get an average of less than seven hours of sleep.
d. Define the consequences of Type I and Type II errors for part a.
37. Construct the null and the alternative hypotheses for the following claims:
a. “I am going to get the majority of the votes to win this election.”
b. “I suspect that your 10-inch pizzas are, on average, less than 10 inches in size.”

page 205
c. “I will have to fine the company because its tablets do not
contain an average of 250 mg of ibuprofen as advertised.”
d. Discuss the consequences of Type I and Type II errors for part a.
38. The manager of a large manufacturing firm is considering switching to new and
expensive software that promises to reduce its assembly costs. Before
purchasing the software, the manager wants to conduct a hypothesis test to
determine if the new software does reduce its assembly costs.
a. Would the manager of the manufacturing firm be more concerned about a Type
I error or a Type II error? Explain.
b. Would the software company be more concerned about a Type I error or a Type
II error? Explain.
39. A polygraph (lie detector) is an instrument used to determine if an individual is
telling the truth. These tests are considered to be 95% reliable. In other words, if
an individual lies, there is a 0.95 probability that the test will detect a lie. Let there
also be a 0.005 probability that the test erroneously detects a lie even when the
individual is actually telling the truth. Consider the null hypothesis, “the individual
is telling the truth,” to answer the following questions.
a. What is the probability of a Type I error?
b. What is the probability of a Type II error?
c. What are the consequences of Type I and Type II errors?
d. What is wrong with the statement, “I can prove that the individual is telling the
truth on the basis of the polygraph result”?
40. FILE Wait_Time. The manager of a small convenience store does not want her
customers standing in line for too long prior to a purchase. In particular, she is
willing to hire an employee for another cash register if the average wait time of the
customers is more than five minutes. She randomly observes the wait time (in
minutes) of customers during the day as:

a. Set up the null and the alternative hypotheses to determine if the manager
needs to hire another employee.
b. Calculate the value of the test statistic and the p-value. What assumption
regarding the population is necessary to implement this step?
c. Decide whether the manager needs to hire another employee at α = 0.10.
41. A machine that is programmed to package 1.20 pounds of cereal in each cereal
box is being tested for its accuracy. In a sample of 36 cereal boxes, the mean and
the standard deviation are calculated as 1.22 pounds and 0.06 pound,
respectively.
a. Set up the null and the alternative hypotheses to determine if the machine is
working improperly—that is, it is either underfilling or overfilling the cereal
boxes.
b. Calculate the value of the test statistic and the p-value.
c. At the 5% level of significance, can you conclude that the machine is working
improperly? Explain.
42. This past year, home prices in the Midwest increased by an average of 6.6%. A
Realtor collects data on 36 recent home sales in the West. He finds an average
increase in home prices of 7.5% with a standard deviation of 2%. Can he
conclude that the average increase in home prices in the West is greater than the
increase in the Midwest? Use a 5% significance level for the analysis.
43. Based on the average predictions of 45 economists, the U.S. gross domestic
product (GDP) will expand by 2.8% this year. Suppose the sample standard
deviation of their predictions was 1%. At the 5% significance level, test if the
mean forecast GDP of all economists is less than 1%.
44. FILE MPG. The data accompanying this exercise show miles per gallon (MPG)
for a sample of 50 hybrid cars.
a. State the null and the alternative hypotheses in order to test whether the
average MPG differs from 50.
b. Calculate the value of the test statistic and the p-value.
c. At α = 0.05, can you conclude that the average MPG differs from 50?
45. A car manufacturer is trying to develop a new sports car. Engineers are hoping
that the average amount of time that the car takes to go from 0 to 60 miles per
hour is below six seconds. The manufacturer tested 12 of the cars and clocked
their performance times. Three of the cars clocked in at 5.8 seconds, five cars at
5.9 seconds, three cars at 6.0 seconds, and 1 car at 6.1 seconds. At the 5% level
of significance, test if the new sports car is meeting its goal to go from 0 to 60
miles per hour in less than six seconds. Assume a normal distribution for the
analysis.
46. FILE Highway_Speeds. A police officer is concerned about speeds on a certain
section of Interstate 90. The data accompanying this exercise show the speeds of
40 cars on a Saturday afternoon.
a. The speed limit on this portion of Interstate 90 is 65 mph. Specify the competing
hypotheses in order to determine if the average speed of cars on Interstate 90
is greater than the speed limit.
b. Calculate the value of the test statistic and the p-value.
c. At α = 0.01, are the officer’s concerns warranted? Explain.
47. FILE Debt_Payments. The data accompanying this exercise show the average
monthly debt payments (Debt, in $) for residents of 26 metropolitan areas.
a. State the null and the alternative hypotheses in order to test whether average
monthly debt payments are greater than $900.
b. What assumption regarding the population is necessary to implement this step?

page 206
c. Calculate the value of the test statistic and the p-value.
d. At α = 0.05, are average monthly debt payments greater than $900? Explain.
48. Some first-time home buyers—especially millennials—are raiding their retirement
accounts to cover the down payment on a home. An economist is concerned that
the percentage of millennials who will dip into their retirement accounts to fund a
 home now exceeds 20%. He randomly surveys 190 millenials with retirement
accounts and finds that 50 are borrowing against them.
a. Set up the null and the alternative hypotheses to test the economist’s concern.
b. Calculate the value of the test statistic and the p-value.
c. Determine if the economist’s concern is justifiable at α = 0.05.
49. FILE Lottery. A 2012 survey by Business Week found that Massachusetts
residents spent an average of $860.70 on the lottery, more than three times the
U.S. average. A researcher at a Boston think tank believes that Massachusetts
residents spend less than this amount. He surveys 100 Massachusetts residents
and asks them about their annual expenditures on the lottery.
a. Specify the competing hypotheses to test the researcher’s claim.
b. Calculate the value of the test statistic and the p-value.
c. At the 10% significance level, do the data support the researcher’s claim?
Explain.
50. The margarita is one of the most common tequila-based cocktails, made with
tequila mixed with triple sec and lime or lemon juice, often served with salt on the
glass rim. A common ratio for a margarita is 2:1:1, which includes 50% tequila,
25% triple sec, and 25% fresh lime or lemon juice. A manager at a local bar is
concerned that the bartender uses incorrect proportions in more than 50% of
margaritas. He secretly observes the bartender and finds that he used the correct
proportions in only 10 out of 30 margaritas. Test if the manager’s suspicion is
justified at α = 0.05.
51. A politician claims that he is supported by a clear majority of voters. In a recent
survey, 24 out of 40 randomly selected voters indicated that they would vote for
the politician. Is the politician’s claim justified at the 5% level of significance?
52. With increasing out-of-pocket healthcare costs, it is claimed that more than 60%
of senior citizens are likely to make serious adjustments to their lifestyle. Test this
claim at the 1% level of significance if in a survey of 140 senior citizens, 90
reported that they have made serious adjustments to their lifestyle.
53. A movie production company is releasing a movie with the hopes that many
viewers will return to see the movie in the theater for a second time. Their target is
to have 30 million viewers, and they want more than 30% of the viewers to want
to see the movie again. They show the movie to a test audience of 200 people,
and after the movie they asked them if they would see the movie in theaters
again. Of the test audience, 68 people said they would see the movie again.
a. At the 5% level of significance, test if more than 30% of the viewers will return
to see the movie again.
b. Repeat the analysis at the 10% level of significance.
54. FILE Silicon_Valley. According to a 2018 report by CNBC on workforce
diversity, about 60% of the employees in high-tech firms in Silicon Valley are white
and about 20% are Asian. Women, along with African Americans and Latinxs, are
highly underrepresented. Just about 30% of all employees are women, with
African Americans and Latinx accounting for only about 15% of the workforce.
 Tara Jones is a recent college graduate working for a large high-tech firm in
Silicon Valley. She wants to determine if her firm faces the same diversity as in
the report. She collects sex and ethnicity information on 50 employees in her firm.
A portion of the data is shown in the accompanying table.

Sex Ethnicity

Woman White

Man White

⋮ ⋮

Man Nonwhite

a. At the 5% level of significance, determine if the proportion of women in Tara’s

firm is different from 0.30.
b. At the 5% level of significance, determine if the proportion of whites in Tara’s
firm is more than 0.50.

page 207

5.4 WRITING WITH DATA

When using big data, we typically do not construct confidence intervals or perform
hypothesis tests. Why is this the case? It turns out that if the sample size is
sufficiently large, there is little difference in the estimates of or the estimates of
generated by different random samples.
Ariel Skelley/Blend Images LLC

Recall that we use to gauge the variability in and to

gauge the variability in . In both cases, the variability depends on the size of the

sample on which the value of the estimator is based. If the sample size is sufficiently
large, then the variability virtually disappears, or, equivalently, and

approach zero. Thus, with big data, it is not very meaningful to construct confidence
intervals for the population mean or the population proportion because the margin of
error also approaches zero; under these circumstances, when estimating μ or p, it is
sufficient to use the estimate of the relevant point estimator.
Recall too that when testing the population mean, the value of the test statistic is
calculated as and when testing the population proportion, the

value of the test statistic is calculated as With big data, the

value of the respective test statistic increases, leading to a small p-value, and thus
rejection of the null hypothesis in virtually any scenario.
Thus, if the sample size is sufficiently large, statistical inference is not very useful.
In this Writing with Data section, we focus on a case study where the sample size is
relatively small.

Case Study
According to a 2018 paper released by the Economic Policy Institute, a nonprofit,
nonpartisan think tank in Washington, D.C., income inequality continues to grow in
the United States. Over the years, the rich have become richer while working-class
wages have stagnated. A local Latino politician has been vocal regarding his concern
about the welfare of Latinx. In various speeches, he has stated that the mean salary
of Latinx households in his county has fallen below the 2017 mean of approximately
$50,000. He has also stated that the proportion of Latinx households making less
than $30,000 has risen above the 2017 level of 20%. Both of his statements are
based on income data for 36 Latinx households in the county. A portion of the data is
shown in Table 5.6.

FILE
Latinx_Income

TABLE 5.6 Latinx Household Income (in $1,000s)

Income

  23  

  63  

  ⋮  

  47  

Trevor Jones is a newspaper reporter who is interested in verifying the concerns of

the local politician. Trevor wants to use the sample information to
1. Determine if the mean income of Latinx households has fallen below the 2017 level
of $50,000.
2. Determine if the percentage of Latinx households making less than $30,000 has
risen above 20%.

page 208

Sample Report— Income Inequality in the

United States
One of the hotly debated topics in the United States is that of growing income
inequality. This trend, which has picked up post Great Recession, is a
reversal of what was seen during and after the Great Depression, where the
gap between rich and poor narrowed. Market forces such as increased trade
and technological advances have made highly skilled and well-educated
workers more productive, thus increasing their pay. Institutional forces, such
as deregulation, the decline of unions, and the stagnation of the minimum
wage, have contributed to income inequality. Arguably, this income inequality
has been felt by minorities, especially African Americans and Latinxs,
because a very high proportion of both groups is working class.
A sample of 36 Latinx households resulted in a mean household income
of $47,278 with a standard deviation of $19,524. The sample mean is below
the 2017 level of $50,000. In addition, eight Latinx households, or
approximately 22%, make less than $30,000. Based on these results, a
politician concludes that current market conditions continue to negatively
impact the welfare of Latinxs. However, it is essential to provide statistically
significant evidence to substantiate these claims. Toward this end, formal
tests of hypotheses regarding the population mean and the population
proportion are conducted. The results of the tests are summarized in Table
5.7.

TABLE 5.7 Test Statistic Values and p-Values for Hypothesis Tests

Hypotheses Test Statistic Value p-value

0.204

0.369

When testing whether the mean income of Latinx households has fallen
below the 2017 level of $50,000, a test statistic value of −0.837 is obtained.
Given a p-value of 0.204, the null hypothesis regarding the population mean,
specified in Table 5.7, cannot be rejected at any reasonable level of
significance. Similarly, given a p-value of 0.369, the null hypothesis regarding
the population proportion cannot be rejected. Therefore, sample evidence
does not support the claims that the mean income of Latinx households has
fallen below $50,000 or that the proportion of Latinx households making less
than $30,000 has risen above 20%. Perhaps the politician’s remarks were
 based on a cursory look at the sample statistics and not on a thorough
statistical analysis.

Suggested Case Studies

Report 5.1 FILE Fidelity_Returns. The accompanying data are the annual returns
for two mutual funds offered by the investment giant Fidelity for the years 2001–
2017. The Fidelity Select Automotive mutual fund invests primarily in companies
engaged in the manufacturing, marketing, or sales of automobiles, trucks, specialty
vehicles, parts, tires, and related services. The Fidelity Gold mutual fund invests
primarily in companies engaged in exploration, mining, processing, or dealing in gold
and, to a lesser degree, in other precious metals and minerals. In a report, use the
sample information to
Calculate descriptive statistics to compare the returns of the mutual funds.
Assess reward by constructing and interpreting 95% confidence intervals for the
population mean return. What assumption did you make when constructing the
confidence intervals?

Report 5.2 FILE Field_Choice. A 2018 Pew Research Center survey finds that
more than half of Americans (52%) believe that young people do not pursue a
degree in science, technology, engineering, or math (STEM)
because they feel that these subjects are too hard. Sixty college-
page 209
bound students are asked about the field they would like to pursue in college. The
choices offered in the questionnaire are STEM, Business, and Other; information on
a teenager’s sex (Male or Female) is also collected. The accompanying data file
contains the responses. In a report, use the sample information to
Compare the 95% confidence interval for the proportion of students who would like
to pursue STEM with the proportion who would like to pursue Business. Do the
results appear to support the survey’s finding? Explain.
Construct and interpret the 95% confidence interval for the proportion of female
students who are college bound.

Report 5.3 FILE Wellbeing. The Gallup-Healthways Well-Being Index provides an

assessment measure of health and well-being of U.S. residents. The overall
composite score is calculated on a scale from 0 to 100, where 100 represents fully
realized well-being. In 2017, the overall well-being of American residents was
reported as 61.5—a decline from 62.1 in 2016. The accompanying data file shows
the overall well-being score for a random sample of 35 residents from South Dakota
—the state with the highest level of well-being. In a report, use the sample
information to
Determine whether the well-being score of South Dakotans is more than the
national average of 61.5 at the 5% significance level.
Determine if fewer than 40% of South Dakotans report a score below 50 at the 5%
significance level.
Comment on the well-being of South Dakotans given your results.
 page 210

6 Regression Analysis

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 6.1  Estimate and interpret a linear regression
model.
LO 6.2  Interpret goodness-of-fit measures.
LO 6.3  Conduct tests of significance.
LO 6.4  Address common violations of the OLS
assumptions.

s discussed in Chapter 1, we often group the study of business

A analytics into three broad categories: descriptive analytics,
predictive analytics, and prescriptive analytics. In this chapter,
we introduce regression analysis, which is one of the most widely
used techniques in predictive analytics. We use regression analysis
for two primary purposes: (1) to assess the relationship between
variables and (2) to predict the outcome of a target variable on the
basis of several input variables. For example, if a firm increases
advertising expenditures by $100,000, then we may want to know
the likely impact on sales. Or, we may want to predict the price of a
house based on its size and location. Regression analysis can be
applied in both of these scenarios.
We first explore the ordinary least squares (OLS) method for
estimating a linear regression model, where the relationship between
the target and the input variables is assumed to be linear. We then
examine a number of goodness-of-fit measures and conduct
hypothesis tests in order to assess which input variables matter the
most. Finally, we examine the importance of the assumptions on the
statistical properties of the OLS estimator, as well as the validity of
the testing procedures. We address common violations to the model
assumptions, the consequences when these assumptions are
violated, and offer some remedial measures.

page 211

©Rawpixel.com/Shutterstock
INTRODUCTORY CASE
College Scorecard
With college costs and student debt on the rise, the choices
that families make when searching for and selecting a college
have never been more important. Yet, students and parents
struggle to find clear, reliable data on critical questions of
college affordability and value. For these reasons, the
Department of Education (DOE) published a redesigned
College Scorecard that reports the most reliable national data
on college costs and students’ outcomes at specific colleges.
Fiona Schmidt, a college counselor, believes that the
information from the College Scorecard can help her as she
advises families. Fiona wonders what college factors influence
post-college earnings and wants answers to the following
questions: If a college costs more or has a higher graduation
rate, should a student expect to earn more after graduation? If
a greater percentage of the students are paying down debt
after college, does this somehow influence post-college
earnings? And finally, does the location of a college affect post-
college earnings?
To address these questions, Fiona gathers information from
116 colleges on annual post-college earnings (Earnings in $),
the average annual cost (Cost in $), the graduation rate (Grad
in %), the percentage of students paying down debt (Debt in
%), and whether or not a college is located in a city (City
equals 1 if a city location, 0 otherwise). Table 6.1 shows a
portion of the data.

TABLE 6.1 College Scorecard Data, May 2016

 FILE
College

Fiona would like to use the information in Table 6.1 to:

1. Make predictions for post-college earnings using regression
analysis.
2. Interpret goodness-of-fit measures for the post-college
earnings model.
3. Determine which factors are statistically significant in
explaining post-college earnings.
A synopsis of this case is provided at the end of Section 6.3.

page 212

6.1 THE LINEAR REGRESSION

MODEL
LO 6.1
Estimate and interpret a linear regression model.

Regression analysis is one of the most widely used techniques in

predictive analytics. We use it to capture the relationship between
two or more variables and to predict the outcome of a target variable
based on several input variables. Regression analysis allows us to
make assessments and robust predictions by determining which of
the relationships matter most and which we can ignore.
We formulate a mathematical model that relates the outcome of a
target variable, called the response variable, to one or more other
input variables, called the predictor variables. Consequently, we
use information on the predictor variables to describe and/or predict
changes in the response variable.
In the introductory case, Fiona wonders what college factors may
influence post-college earnings. In another scenario, we may want to
predict a firm’s sales based on its advertising; estimate an
individual’s salary based on education and years of experience;
predict the selling price of a house on the basis of its size and
location; or describe auto sales with respect to consumer income,
interest rates, and price discounts. In all of these examples, we can
use regression analysis to describe the relationship between the
variables of interest.
No matter the response variable that we choose to examine, we
cannot expect to predict its exact (unique) value. If the value of the
response variable is uniquely determined by the values of the
predictor variables, we say that the relationship between the
variables is deterministic. This is often the case in the physical
sciences. For example, momentum p is the product of the mass m
and the velocity v of an object; that is, p = mv. In most fields of
research, however, we tend to find that the relationship between the
response variable and the predictor variables is stochastic, due to
the omission of relevant factors (sometimes not measurable) that
influence the response variable. For instance, when trying to predict
an individual’s salary, the individual’s natural ability is often omitted
since it is extremely difficult, if not impossible, to quantify.

DETERMINISTIC VERSUS STOCHASTIC RELATIONSHIPS

The relationship between the response variable and the
predictor variables is deterministic if the value of the response
 variable is uniquely determined by the predictor variables;
otherwise, the relationship is stochastic.

Our objective is to develop a mathematical model that captures the

relationship between the response variable y and the k predictor
variables x1, x2, . . . , xk. The model must also account for the
randomness that is a part of real life. In order to develop a linear
regression model, we start with a deterministic component that
approximates the relationship we want to model, and then add a
random error term to it, making the relationship stochastic.

The Components of the Linear Regression

Model
We use economic theory, intuition, and statistcal measures to
determine which predictor variables might best explain the response
variable. A regression model treats all predictor variables as
numerical, where observations of a categorical variable are first
converted into numerical values. Recall from Chapters 1 and 2 that
the observations of a numerical variable represent meaningful
numbers, whereas the observations of a categorical variable
represent different categories. For example, income is a numerical
variable, whereas marital status is a categorical variable.

page 213

Recall, too, that a dummy variable, also referred to as an

indicator or a binary variable, takes on numerical values of 1 or 0 to
describe two categories of a categorical variable. For a predictor
variable that is a dummy variable, it is common to refer to the
category that assumes a value of 0 as the reference category or
the benchmark category. All comparisons are made in relation to
the reference category. For example, in the case of a dummy
variable categorizing a person’s sex, we can define 1 for male and 0
for female. In this case, the reference category would be females.
Alternatively, we can define 1 for female and 0 for male, where
males would be the reference category.

A DUMMY VARIABLE USED AS A PREDICTOR VARIABLE

A dummy variable takes on numerical values of 1 or 0 to
describe two categories of a categorical variable. For a
predictor variable that is a dummy variable, we refer to the
category that assumes a value of 0 as the reference or
benchmark category. All comparisons in the regression analysis
are made in relation to the reference category.

If a linear regression model uses only one predictor variable, then

the model is referred to as a simple linear regression model and is
represented as y = β0 + β1x1 + ɛ, where y is the response variable,
β0 and β1 (the Greek letters read as betas) are the unknown
intercept and slope parameters, respectively, x1 is the predictor
variable, and ɛ is the random error term (the Greek letter read as
epsilon). (Technically, we should include the subscript i on all
variables, or yi = β0 + β1x1i + ɛi . For now, we exclude the subscript i.)
A fundamental assumption underlying the simple linear
regression model is that the expected value of y lies on a straight
line, denoted by β0 + β1x1. The expression β0 + β1x1 is the
deterministic component of the simple linear regression model,
which can be thought of as the expected value of y for a given value
of x. In other words, conditional on x, E(y) = β0 + β1x1. The slope
parameter β1 determines whether the linear relationship between x
and E(y) is positive (β1 > 0) or negative (β1 < 0); β1 = 0 indicates that
there is no linear relationship between x and E(y). Figure 6.1 shows
the expected value of y for various values of the intercept β0 and the
slope β1 parameters.
FIGURE 6.1 Examples of a simple linear regression model

We can easily extend the simple linear regression to include

more than one predictor variable. The resulting model is referred to
as a multiple linear regression model and is represented as y = β0
+ β1x1 + β2x2 + . . . + βkxk + ɛ; a common notation is to replace an x
with a d if a given predictor variable is a dummy variable. For
simplicity, we will refer to the simple linear regression model or the
multiple linear regression model as the linear regression model.

page 214

The population parameters β0, β1, β2, . . . , βk used in the linear

regression model are unknown and, therefore, must be estimated.
As always, we use sample data to estimate the population
parameters of interest. Here, sample data consist of n observations
on y, x1, x2, . . . , xk.

THE LINEAR REGRESSION MODEL

The linear regression model is defined as
 where y is the response variable; x1, x2, . . . , xk are the k
predictor variables; and ɛ is the random error term. The
coefficients β0, β1, β2, . . . , βk are the unknown parameters to
be estimated.

Let b0, b1, b2, . . . , bk represent the estimates of β0, β1, β2, . . . ,

βk, respectively. We form the sample regression equation as = b0

+ b1x1 + b2x2 + . . . bkxk, where (read as y-hat) is the predicted

value of the response variable given specified values of the predictor

variables. For given values of the predictor variables, the observed
and the predicted values of the response variable are likely to be
different because many factors besides the included predictor
variables influence y. We refer to the difference between the
observed and the predicted values of y, that is, y – , as the residual

e.
A common approach to obtaining estimates for β0, β1, β2, . . . , βk
is to use the ordinary least squares (OLS) method. OLS
estimators have many desirable properties if certain assumptions
hold. (These assumptions are discussed in Section 6.4.) The OLS
method chooses the sample regression equation whereby the error
sum of squares, SSE, is minimized, where SSE = Σ(y – )2 = Σe2.

SSE is the sum of the squared differences between an observed

value y and its predicted value , or, equivalently, the sum of the

squared distances from the regression equation. Thus, using this

distance measure, we say that the OLS method produces the
equation that is “closest” to the data.
 Using calculus, equations have been developed for b0, b1, b2, . . .
, bk that satisfy the OLS criterion. Virtually every statistical software
package computes the necessary output to construct a sample
regression equation. In addition, values of all relevant statistics for
assessing the model, discussed later in this chapter, are also
included with the output.
In order to graphically depict what we have just described with
just one predictor variable, consider Figure 6.2 which shows a
scatterplot of y against x. Superimposed on the scatterplot is the
sample regression line, yˆ = b0 + b1x, that was obtained using OLS.

We also show the residual, e = yˆ, for one observation.

FIGURE 6.2 Scatterplot of y against x with superimposed sample regression line

It is important to be able to interpret the estimated regression

coefficients. Mathematically, the intercept estimate b0 represents the

predicted value of when each predictor variable xj (j = 1, . . . , k)

assumes a value of zero. However, in many applications it is not

possible to provide a meaningful interpretation for the intercept. For
each predictor variable xj ( j = 1, . . . , k), the page 215
corresponding slope coefficient bj is the estimate of
βj. The slope coefficient bj measures the change in the predicted
value of the response variable given a unit increase in the

associated predictor variable xj, holding all other predictor variables

constant. In other words, it represents the partial influence of xj on .

FILE
College

EXAMPLE 6.1
Using the data from Table 6.1, estimate the linear regression
model, Earnings = β0 + β1Cost + β2Grad + β3Debt + β4City + ɛ,
where Earnings is annual post-college earnings (in $), Cost is
the average annual cost (in $), Grad is the graduation rate (in
%), Debt is the percentage of students paying down debt (in
%), and City assumes a value of 1 if the college is located in a
city, 0 otherwise.
a. What is the sample regression equation?
b. Interpret the slope coefficients.
c. Predict annual post-college earnings if a college’s average
annual cost is $25,000, its graduation rate is 60%, its
percentage of students paying down debt is 80%, and it is
located in a city.

SOLUTION: Table 6.2 shows the Excel output from estimating this
model. We will provide Excel and R instructions for obtaining
regression output at the end of this section.

TABLE 6.2 Regression Results for Example 6.1

a. As Table 6.2 shows, Excel produces quite a bit of information.
In order to answer the questions in Example 6.1, we only need
the estimated coefficients, which we have put in boldface in
the lower portion of the table. We will address the remaining
information in later sections of this chapter. The coefficient
estimates, rounded to four decimal places, are b0 =
10,004.9665, b1 = 0.4349, b2 = 178.0989, b3 = 141.4783, and
b4 = 2,526.7888. Thus, the sample regression equation is

page 216
b. All coefficients are positive, suggesting a positive
influence of each predictor variable on the response variable.
Specifically:
 Holding all other predictor variables constant, if average
annual costs increase by $1, then, on average, predicted
earnings are expected to increase by b1; that is, by about
$0.4349. This result suggests that spending more on college
pays off.
If the graduation rate increases by 1%, then, on average,
predicted earnings are expected to increase by $178.10,
holding the other predictor variables constant. Policymakers
often debate whether graduation rate can be used as a
proxy for a college’s academic quality. Perhaps it is not the
academic quality but the case that colleges with a higher
graduation rate have more motivated students, which
translates into higher earnings for these students.
A one percentage point increase in the percentage of
students paying down debt is expected to increase predicted
earnings by approximately $141.48, holding the other three
predictor variables constant. As it turns out, Debt is not
statistically significant at any reasonable level; we will
discuss such tests of significance in Section 6.3.
All else constant, predicted earnings are $2,526.79 higher
for graduates of colleges located in a city. This difference is
not surprising seeing that students who attend college in a
city likely have more internship opportunities, which then
often translate into higher-paying jobs after graduation.
c. If a college’s average annual cost is $25,000, its graduation
rate is 60%, its percentage of students paying down debt is
80%, and it is located in a city, then average post-college
earnings for its students are
Estimating a Linear Regression Model with
Excel or R
Using Excel
In order to obtain the regression output in Table 6.2 using Excel, we
follow these steps.
FILE
College

A. Open the College data file.

B. Choose Data > Data Analysis > Regression from the menu. (Recall from Chapter 3
that if you do not see the Data Analysis option under Data, you must add in the
Analysis Toolpak option.)
C. See Figure 6.3. In the Regression dialog box, click on the box next to Input Y Range,
and then select the data for Earnings. Click on the box next to Input X Range, and then
simultaneously select the data for Cost, Grad, Debt, and City. Select Labels because we
are using Earnings, Cost, Grad, Debt, and City as headings.

FIGURE 6.3 Excel’s Regression dialog box for Example 6.1

D. Click OK.

Using R
In order to obtain the regression output in Table 6.2 using R, we
follow these steps.
A. Import the College data into a data frame (table) and label it myData.
B. By default, R will report the regression output using scientific notation. We opt to turn this
option off using the following command:
> options(scipen=999)
In order to turn scientific notation back on, we would enter options(scipen=0) at the
prompt.

page 217
C. Use the lm function to create a linear model, which we label Multiple.
Note that we use the ‘+’ sign to add predictor variables, even if we believe that a
negative relationship may exist between the response variable and the predictor
variables. You will not see output after you implement this step.
> Multiple <- lm(Earnings ~ Cost + Grad + Debt + City, data = myData)
D. Use the summary function to view the summary regression output. Figure 6.4 shows the
R regression output. We have put the intercept and the slope coefficients in boldface. As
expected, these values are identical to the ones obtained using Excel.
> summary(Multiple)

FIGURE 6.4 R’s regression output for Example 6.1

page 218

Categorical Variables with Multiple

Categories
So far, we have used a dummy variable to describe a categorical
variable with only two categories. Sometimes, a categorical variable
may be defined by more than two categories. In such cases, we use
multiple dummy variables to capture all categories. For example, the
mode of transportation used to commute to work may be described
by three categories: Public Transportation, Driving Alone, and Car
Pooling. We can define two dummy variables d1 and d2, where d1
equals 1 for Public Transportation, 0 otherwise, and d2 equals 1 for
Driving Alone, 0 otherwise. For this three-category case, we need to
define only two dummy variables; Car Pooling, the reference
category, is indicated when d1 = d2 = 0.
Given the intercept term, we exclude one of the dummy variables
from the regression, where the excluded variable represents the
reference category against which the others are assessed. If we
include as many dummy variables as there are categories, then their
sum will equal one. For instance, if we add a third dummy d3 that
equals 1 to denote Car Pooling, then for all observations, d1 + d2 +
d3 = 1. This creates the problem called perfect multicollinearity, a
topic discussed in Section 6.4. As we will see, such a model cannot
be estimated. This situation is sometimes referred to as the dummy
variable trap.

AVOIDING THE DUMMY VARIABLE TRAP

If the linear regression model includes an intercept, the number
of dummy variables representing a categorical variable should
be one less than the number of categories of the variable.

It may be the case that a categorical variable has many categories.

Not only is it cumbersome to create many dummy variables, it may
not necessarily be informative. In this instance, it typically makes
sense to group the categories based on some similarity. For
example, when analyzing traffic volume, significant differences may
be found between weekdays and weekends instead of between the
different days of the week. Therefore, we can simply create one
dummy variable to indicate whether it is a weekday or a weekend
instead of creating six (7 − 1) dummy variables to indicate the
different days of the week.
FILE
Retail
 EXAMPLE 6.2
A government researcher is analyzing the relationship between
retail sales (Sales in $ millions) and the gross national product
(GNP in $ billions). He also wonders whether there are
significant differences in retail sales related to the quarters of
the year. He collects 10 years of quarterly data and creates
dummy variables for the quarters. A portion is shown in Table
6.3.

TABLE 6.3 Data for Example 6.2

Year Quarter Sales GNP

Year Quarter Sales GNP

2007 1  921266 14301.854

2007 2 1013371 14512.945

⋮ ⋮ ⋮ ⋮

2016 4 1299699 19134.463

a. Estimate the model y = β0 + β1x + β2d1 + β3d2 + β4d3 + ɛ,

where y and x represent retail sales and GNP, respectively; d1
equals 1 if quarter 1, 0 otherwise; d2 equals 1 if quarter 2, 0
otherwise; and d3 equals 1 if quarter 3, 0 otherwise. Note that
the reference category is quarter 4.
b. Interpret the slope coefficient for quarter 1.
c. What are predicted sales in quarter 2 if GNP is $18,000 (in
billions)? For the same value of GNP, what are predicted sales
 in quarter 4?

page 219
SOLUTION:

a. In order to estimate this model, we have to first create three

dummy variables d1, d2, and d3 from the Quarter variable. For
this example, d1 will take on the value 1 if Quarter 1, 0
otherwise; d2 will take on the value 1 if Quarter 2, 0 otherwise;
and d3 will take on the value 1 if Quarter 3, 0 otherwise. We
report a portion of the regression results of this model in Table
6.4.

TABLE 6.4 Portion of Regression Results for Example 6.2

b. All else equal, retail sales in quarter 1 are expected to be

approximately $108,765 million less than sales in quarter 4.

c. Given x = 18,000, we set d1 = 0, d2 = 1, d3 = 0 and calculate

= 47,095.6859 + 65.0548 × 18,000 − 30,486.2947 =

1,187,594.98. Or, equivalently, retail sales are predicted to be
approximately $1,187,595 (in millions). For the same GNP
value, predicted retail sales in quarter 4 (d1 = d2 = d3 = 0) are

= 47,095.6859 + 65.0548 × 18,000 = 1,218,081.28, or

approximately $1,218,081 (in millions).

EXERCISES 6.1
Applications
1. Using data from 50 workers, a researcher estimates Wage = β0 + β1Education +
β2Experience + β3Age + ɛ, where Wage is the hourly wage rate and Education,
Experience, and Age are the years of higher education, the years of experience,
and the age of the worker, respectively. A portion of the regression results is
shown in the following table.

a. Interpret the estimates for β1 and β2.

b. Predict the hourly wage rate for a 30-year-old worker with four years of higher
education and three years of experience.
2. A sociologist believes that the crime rate in an area is significantly influenced by
the area’s poverty rate and median income. Specifically, she hypothesizes crime
will increase with poverty and decrease with income. She collects data on the
crime rate (crimes per 100,000 residents), the poverty rate (in %), and the median
income (in $1,000s) from 41 New England cities. A portion of the regression
results is shown in the following table.

a. Are the signs as expected on the slope coefficients?

b. Predict the crime rate in an area with a poverty rate of 20% and a median
income of $50,000.
3. Osteoporosis is a degenerative disease that primarily affects women over the age
of 60. A research analyst wants to forecast sales of StrongBones, a prescription
drug for treating this debilitating disease. She uses the model Sales = β0 +
β1Population + β2Income + ɛ, where Sales refers to the sales of StrongBones (in
$1,000,000s), Population is the number of women over the age of 60 (in millions),
and Income is the average income of women over the age of 60 (in $1,000s). She
collects data on 38 cities across the United States and obtains the following
regression results:

page 220
a. Interpret the slope coefficients.
b. Predict sales if a city has 1.5 million women over the age of 60 and their
average income is $44,000.
4. An executive researcher wants to better understand the factors that explain
differences in salaries for marketing majors. He decides to estimate two models: y
= β0 + β1d1 + ɛ (Model 1) and y = β0 + β1d1 + β2d2 + ɛ (Model 2). Here y
represents salary, d1 is a dummy variable that equals 1 for male employees, and
d2 is a dummy variable that equals 1 for employees with an MBA.
a. What is the reference group in Model 1?
b. What is the reference group in Model 2?
c. In these two models, would it matter if d1 equaled 1 for female employees?
5. House price y is estimated as a function of the square footage of a house x and a
dummy variable d that equals 1 if the house has ocean views. The estimated

house price, measured in $1,000s, is given by .

a. Compute the predicted price of a house with ocean views and square footage of
2,000 and 3,000, respectively.
b. Compute the predicted price of a house without ocean views and square
footage of 2,000 and 3,000, respectively.
c. Discuss the impact of ocean views on the house price.
6. FILE GPA. The director of graduate admissions at a large university is analyzing
the relationship between scores on the math portion of the Graduate Record
Examination (GRE) and subsequent performance in graduate school, as
measured by a student’s grade point average (GPA). She uses a sample of 24
 students who graduated within the past five years. A portion of the data is as
follows:

GPA GRE

3.0 700

3.5 720

⋮ ⋮

3.5 780

a. Find the sample regression equation for the model: GPA =

β0 + β1GRE + ɛ.
b. What is a student’s predicted GPA if he/she scored 710 on
the math portion of the GRE?
7. FILE Education. A social scientist would like to analyze the relationship between
educational attainment (in years of higher education) and annual salary (in
$1,000s). He collects data on 20 individuals. A portion of the data is as follows:

Salary Education

40 3

53 4

⋮ ⋮

38 0

a. Find the sample regression equation for the model: Salary = β0 + β1Education
+ ɛ.
b. Interpret the coefficient for Education.
c. What is the predicted salary for an individual who completed seven years of
higher education?
8. FILE Consumption. The consumption function, first developed by John Maynard
Keynes, captures one of the key relationships in economics. It expresses
consumption as a function of disposable income, where disposable income is
income after taxes. The accompanying table shows a portion of quarterly data for
 average U.S. annual consumption (Consumption in $) and disposable income
(Income in $) for the years 2000–2016.

Date Consumption Income

Q1, 2000 28634 31192

Q2, 2000 28837 31438

⋮ ⋮ ⋮

Q4, 2016 35987 39254

a. Find the sample regression equation for the model: Consumption = β0 +

β1Income + ɛ.
b. In this model, the slope coefficient is called the marginal propensity to consume.
Interpret its meaning.
c. What is predicted consumption if disposable income is $35,000?
9. FILE Car_Prices. The accompanying table shows a portion of data consisting of
the price, the age, and the mileage for 20 used sedans.

Price Age Mileage

13590 6 61485

13775 6 54344

⋮ ⋮ ⋮

11988 8 42408

a. Estimate the sample regression equation that enables us to predict the price of
a sedan on the basis of its age and mileage.
b. Interpret the slope coefficient of Age.
c. Predict the price of a five-year-old sedan with 65,000 miles.
10. FILE Engine. The maintenance manager at a trucking company wants to build a
regression model to forecast the time until the first engine overhaul (Time in
years) based on four predictor variables: (1) annual miles driven (Miles in 1,000s),
(2) average load weight (Load in tons), (3) average driving speed (Speed in mph),
and (4) oil change interval (Oil in 1,000s of miles). Based on driver logs and
onboard computers, data have been obtained for a sample of 25 trucks. A portion
of the data is shown in the accompanying table.
 page 221
a. For each predictor variable, discuss whether it is likely to have
a positive or negative influence on time until the first engine overhaul.
b. Find the sample regression equation for the regression model (use all four
predictor variables).
c. Based on part a, are the signs of the regression coefficients logical?
d. Predict the time before the first engine overhaul for a particular truck driven
60,000 miles per year with an average load of 22 tons, an average driving
speed of 57 mph, and 18,000 miles between oil changes.
11. FILE MCAS. Education reform is one of the most hotly debated subjects on both
state and national policymakers’ list of socioeconomic topics. Consider a linear
regression model that relates school expenditures and family background to
student performance in Massachusetts using 224 school districts. The response
variable is the mean score on the MCAS (Massachusetts Comprehensive
Assessment System) exam given to 10th graders. Four predictor variables are
used: (1) STR is the student-to-teacher ratio in %, (2) TSAL is the average
teacher’s salary in $1,000s, (3) INC is the median household income in $1,000s,
and (4) SGL is the percentage of single-parent households. A portion of the data
is shown in the accompanying table.

a. For each predictor variable, discuss whether it is likely to have a positive or

negative influence on Score.
b. Find the sample regression equation. Are the signs of the slope coefficients as
expected?
c. What is the predicted score if STR = 18, TSAL = 50,INC = 60, and SGL = 5?
d. What is the predicted score if everything else is the same as in part c except
INC = 80?
12. FILE Electricity_Cost. The facility manager at a pharmaceutical company wants
to build a regression model to forecast monthly electricity cost. Three main
 variables are thought to dictate electricity cost (in $): (1) average outdoor
temperature (Avg Temp in °F), (2) working days per month, and (3) tons of
product produced. A portion of the past year’s monthly data is shown in the
accompanying table.

Cost Avg Temp Work Days Tons Produced

24100 26 24 80

23700 32 21 73

⋮ ⋮ ⋮ ⋮

26000 39 22 69

a. For each predictor variable, discuss whether it is likely to have a positive or

negative influence on monthly electricity cost.
b. Find the sample regression equation for the regression model (use all three
predictor variables).
c. What is the predicted electricity cost in a month during which the average
outdoor temperature is 65°, there are 23 working days, and 76 tons are
produced?
13. FILE SAT_1. The SAT has gone through many revisions over the years. People
argue that female students generally do worse on math tests but better on writing
tests. Consider the following portion of data on 20 students who took the SAT test
last year. Information includes each student’s score on the writing and math
sections of the exam, the student’s GPA, and a Female dummy variable that
equals 1 if the student is female, 0 otherwise.

Writing Math GPA Female

620 600 3.44 0

570 550 3.04 0

⋮ ⋮ ⋮ ⋮

540 520 2.84 0

a. Estimate a linear regression model with Writing as the response variable and
GPA and Female as the predictor variables. Compute the predicted writing
score for a male student with a GPA of 3.5. Repeat the computation for a
female student.
 b. Estimate a linear regression model with Math as the response variable and
GPA and Female as the predictor variables. Compute the predicted math score
for a male student with a GPA of 3.5. Repeat the computation for a female
student.
14. FILE Franchise. A president of a large chain of fast-food restaurants collects
data on 100 franchises.
a. Estimate the model: y = β0 + β1x1 + β2x2 + ɛ where y is net profit, x1 is counter
sales, and x2 is drive-through sales. All variables are measured in millions of
dollars.
b. Interpret the coefficient attached to drive-through sales.
c. Predict the net profit of a franchise that had counter sales of $6 million and
drive-up sales of $4 million.

page 222
15. FILE Arlington. A Realtor in Arlington, Massachusetts, is
analyzing the relationship between the sale price of a home (Price in $), its square
footage (Sqft), the number of bedrooms (Beds), the number of bathrooms (Baths),
and a Colonial dummy variable (Colonial equals 1 if a colonial-style home, 0
otherwise). She collects data on 36 sales in Arlington for the analysis. A portion of
the data is shown in the accompanying table.

a. Estimate the model: Price = β0 + β1Sqft + β2Beds + β3Baths + β4Colonial + ɛ.

b. Interpret the coefficients attached to Beds and Colonial.
c. Predict the price of a 2,500-square-foot, colonial-style home with three
bedrooms and two bathrooms.
16. FILE Startups. Many of today’s leading companies, including Google, Microsoft,
and Facebook, are based on technologies developed within universities. Lisa
Fisher is a business school professor who would like to analyze university factors
that enhance innovation. She collects data on 143 universities for a regression
where the response variable is the number of startups (Startups), which is used
as a measure for innovation. The predictor variables include the university’s
research expenditure (Research in $ millions), the number of patents issued
(Patents), and the age of its technology transfer office (Duration in years). A
portion of the data is shown in the accompanying table.

Startups Research Patents Duration

 1 145.52 8 23

1 237.52 16 23

⋮ ⋮ ⋮ ⋮

1 154.38 3 9

a. Estimate: Startups = β0 + β1Research + β2Patents + β3Duration + ɛ.

b. Predict the number of startups for a university that spent $120 million on
research, was issued eight patents, and has had a technology transfer office for
20 years.
c. How much more research expenditure is needed for the university to have an
additional predicted startup, with everything else being the same?
17. FILE Quarterbacks. American football is the highest-paying sport on a per-game
basis. Given that the quarterback is considered the most important player on the
team, he is typically well-compensated. A sports statistician examines the factors
that influence a quarterback’s salary. He believes that a quarterback’s pass
completion rate is the most important variable affecting salary. He also wonders
how total touchdowns scored and a quarterback’s age might impact salary. The
statistician collects data on salary (Salary in $ millions), pass completion rate (PC
in %), total touchdowns scored (TD), and age for 32 quarterbacks during a recent
season. A portion of the data is shown in the accompanying table.

a. Estimate: Salary = β0 + β1PC + β2TD + β3Age + ɛ.

b. Interpret the slope coefficient attached to TD.
c. Player 8 earned 12.9895 million dollars during the season. According to the
model, what is his predicted salary if PC = 70.6, TD = 34, and Age = 30?
d. Player 16 earned 8.0073 million dollars during the season. According to the
model, what is his predicted salary if PC = 65.5, TD = 28, and Age = 32?
e. Compute and interpret the residual salary for Player 8 and Player 16.
6.2 MODEL SELECTION
LO 6.2
Interpret goodness-of-fit measures.

By simply observing the sample regression equation, we cannot

assess how well the predictor variables explain the variation in the
response variable. However, several objective “goodness-of-fit”
measures do exist that summarize how well the sample regression
equation fits the data.
Recall that in the introductory case study, we are interested in
analyzing factors that may influence post-college annual earnings
(Earnings) for 116 colleges. Here, we will estimate three models,
using a combination of four predictor variables, to determine which
sample regression equation ‘best’ explains Earnings. The four
predictor variables are the average annual cost (Cost), the
graduation rate (Grad), the percentage of students page 223
paying down debt (Debt), and whether or not a
college is located in a city (City equals 1 if a city location and 0
otherwise). Let the models be specified as follows:

(To simplify, we use the same notation to refer to the coefficients in

Models 1, 2, and 3. The coefficients and their estimates may have a
different meaning depending on which model we reference. Also, we
note that when explaining Earnings, other models do exist; for
instance, we could have considered other combinations of predictor
variables.)
 If you had to choose one of these models to predict annual post-
college earnings, which model would you choose? It may be that by
using more predictor variables, you can better describe the response
variable. However, for a given sample, more is not always better. In
order to select the preferred model, we examine several goodness-
of-fit measures: the standard error of the estimate (denoted as se),
the coefficient of determination (denoted as R2), and the adjusted
coefficient of determination (denoted as adjusted R2). We first
discuss these measures in general, and then determine whether
Model 1, Model 2, or Model 3 is the preferred model.

The Standard Error of the Estimate, se

To simplify, we describe goodness-of-fit measures in the context of a
simple linear regression model, and we will refer to Model 1 for this
purpose. Figure 6.5 shows a scatterplot of Earnings (y) against Cost
(x), as well as the superimposed sample regression line. Recall that
the residual ei represents the difference between the observed value
and the predicted value of the response variable for the ith
observation—that is, ei = yi − i. If all the data points had fallen on the

line, then each residual would be zero; in other words, there would
be no dispersion between the observed and the predicted values.
Because in practice we rarely, if ever, obtain this result, we evaluate
models on the basis of the relative magnitude of the residuals. The
sample regression equation provides a good fit when the dispersion
of the residuals is relatively small.

FIGURE 6.5 Scatterplot of Earnings against Cost

A numerical measure that gauges the dispersion from the sample
regression equation is the sample variance of the residual, denoted
. This measure is defined as the average squared difference

between yi and i. The numerator of the formula is the error sum of

squares, Dividing SSE by its respective degrees of

freedom n − k − 1 yields . Recall that k denotes the number of

predictor variables in the linear regression model; thus, page 224

for a simple linear regression model, k equals one.
Instead of , we generally report the standard deviation of the

residual, denoted se, more commonly referred to as the standard

error of the estimate. As usual, se is the positive square root of .

The standard error of the estimate is measured in the same units of

measurement as the response variable. When comparing models
with the same response variable, we prefer the model with a smaller
se. A smaller se implies that there is less dispersion of the observed
values from the predicted values.

THE STANDARD ERROR OF THE ESTIMATE

The standard error of the estimate se is calculated as

where SSE is the error sum of squares. Theoretically, se can

assume any value between zero and infinity, 0 ≤ se < ∞. When
comparing models with the same response variable, the model
with the smaller se is preferred.

For a given sample size n, increasing the number k of the predictor

variables reduces both the numerator (SSE) and the denominator (n
− k − 1) in the formula for se. The net effect, shown by the value of
se, allows us to determine if the added predictor variables improve
the fit of the model. Virtually all statistical software packages report
se. Excel reports se in the Regression Statistics portion of the
regression output and refers to it as ‘Standard Error.’ R reports se in
the bottom portion of the regression output and refers to it as
‘Residual standard error.’

The Coefficient of Determination, R2

Like the standard error of the estimate, the coefficient of
determination R2 evaluates how well the sample regression
equation fits the data. In particular, R2 quantifies the sample variation
in the response variable y that is explained by the sample regression
equation. It is computed as the ratio of the explained variation of the
response variable to its total variation. For example, if R2 = 0.72, we
say that 72% of the sample variation in y is explained by the sample
regression equation. Other factors, which have not been included in
the model, account for the remaining 28% of the sample variation.
We use analysis of variance (ANOVA) in the context of the linear
regression model to derive R2. We denote the total variation in y as
, which is the numerator in the formula for the variance of y.

This value, called the total sum of squares, SST, can be broken
down into two components: explained variation and unexplained
variation. Figure 6.6 illustrates the decomposition of the total
variation of y into its two components for a simple linear regression
model.

FIGURE 6.6 Total, explained, and unexplained differences

For illustration purposes, we show a scatterplot with all the points

removed except one, point A. Point A refers to the observation (xi,
yi). The blue line represents the estimated regression equation
based on the entire sample data; the horizontal and vertical green
lines represent the sample means and , respectively. The vertical

distance between point A and point C is the total difference . For

each observation, we square these differences and then find their

sum—this amounts to , or, equivalently, the total

variation in y.
Now, we focus on the distance between point B and point C
which is referred to as the explained difference, . In other words,

this difference can be explained by changes in x. Squaring all such

differences and summing them yields the regression sum of squares,
SSR, where . SSR is a measure of the explained

variation in y.

page 225

The distance between point A and point B is the unexplained

difference. This is the portion that remains unexplained. Squaring all
such differences and summing them yields the familiar error sum of
squares, .

Thus, the total variation in y can be decomposed into explained

and unexplained variation as follows:

We derive the formula for R2 by dividing both sides by SST and

rearranging:
 THE COEFFICIENT OF DETERMINATION, R2
The coefficient of determination R2 is the proportion of the
sample variation in the response variable that is explained by
the sample regression equation. We compute R2 as

where , and .

The value of R2 falls between zero and one; the closer the
value is to one, the better the fit.

Virtually all statistical software packages report R2. Excel reports R2

in the Regression Statistics section of the regression output and
refers to it as R Square; R reports R2 below the coefficient estimates
and refers to it as Multiple R Squared.
Our objective in adding another predictor variable to a linear
regression model is to increase the model’s usefulness. It turns out
that we cannot use R2 for model comparison when the competing
models do not include the same number of predictor variables. This
occurs because R2 never decreases as we add more predictor
variables to the model. A popular model selection method in such
situations is to choose the model that has the highest adjusted R2
value, a topic that we discuss next.

The Adjusted R2
Because R2 never decreases as we add more predictor variables to
the linear regression model, it is possible to increase its value
unintentionally by including a group of predictor variables that may
have no economic or intuitive foundation in the linear regression
model. This is true especially when the number of predictor variables
k is large relative to the sample size n. In order to avoid page 226
the possibility of R2 creating a false impression, virtually
all software packages include adjusted R2. Unlike R2, adjusted R2
explicitly accounts for the sample size n and the number of predictor
variables k. It is common to use adjusted R2 for model selection
because it imposes a penalty for any additional predictor variable
that is included in the analysis. When comparing models with the
same response variable, we prefer the model with the higher
adjusted R2, implying that the model is able to explain more of the
sample variation in y.

ADJUSTED R2
The adjusted coefficient of determination is calculated as

Adjusted R2 is used to compare competing linear regression

models with different numbers of predictor variables; the higher
the adjusted R2, the better the model.

If SSE is substantially greater than zero and k is large compared to

n, then adjusted R2 will differ substantially from R2. Adjusted R2 may
be negative, if the correlation between the response variable and the
predictor variables is sufficiently low. Both the standard error of the
estimate se and the adjusted R2 are useful for comparing linear
regression models with different numbers of predictor variables.
Adjusted R2, however, is the more commonly used criterion for
model selection.
FILE
College

EXAMPLE 6.3
Table 6.5 provides goodness-of-fit measures from estimating
the following models:

TABLE 6.5 Goodness-of-Fit Measures for Models 1, 2, and 3

Model 1 Model 2 Model 3

Standard error of 6,271.4407 5,751.8065 5,645.8306

the estimate se

Coefficient of 0.2767 0.4023 0.4292

determination R2

Adjusted R2 0.2703 0.3862 0.4087

a. Based on two goodness-of-fit measures, which of the three

models is the preferred model.
b. Interpret the coefficient of determination for the preferred
model.
c. What percentage of the sample variation in annual post-
college earnings is unexplained by the preferred model?
SOLUTION:

a. Model 3 has the lowest standard error of the estimate (se =

5,645.8306) as compared to Model 1 and Model 2. Model 3
also has the highest adjusted R2 (adjusted R2 = 0.4087).
Thus, Model 3 is the preferred model. Note that we cannot use
the coefficient of determination R2 to compare the three
models because the models have different numbers of
predictor variables. page 227
b. The coefficient of determination R2 for Model 3 is
0.4292, which means that 42.92% of the sample variation in
Earnings is explained by the regression model.
c. If 42.92% of the sample variation in Earnings is explained by
changes in the predictor variables in Model 3, then 57.08% (1
− 0.4292) is unexplained by the regression equation. It should
not come as a surprise that the regression equation leaves
quite a bit of the sample variation in annual post-college
earnings unexplained; there are many factors (grade-point
average, field of emphasis, natural ability, etc.) that are likely
to explain earnings that have not been included in the model.

One Last Note on Goodness-of-Fit Measures

We note that goodness-of-fit measures discussed in this section use
the same sample to build the model as well as to assess it.
Unfortunately, this procedure does not help us gauge how well the
estimated model will predict in an unseen sample. In Chapters 7
through 10, we discuss cross-validation techniques that evaluate
predictive models by dividing the original sample into a training set to
build (train) the model and a validation set to evaluate (validate) it.
The validation set is used to provide an independent performance
assessment by exposing the model to unseen data.

EXERCISES 6.2
Applications
18. An analyst estimates the sales of a firm as a function of its advertising
expenditures using the model Sales = β0 + β1Advertising + ɛ. Using 20
observations, he finds that SSR = 199.93 and SST = 240.92.
a. What proportion of the sample variation in sales is explained by advertising
expenditures?
b. What proportion of the sample variation in sales is unexplained by advertising
expenditures?
19. FILE Test_Scores. The accompanying data file shows the midterm and final
scores for 32 students in a statistics course.
a. Estimate a student’s final score as a function of his/her midterm score.
b. Find the standard error of the estimate.
c. Find and interpret the coefficient of determination.
20. The director of college admissions at a local university is trying to determine
whether a student’s high school GPA or SAT score is a better predictor of the
student’s subsequent college GPA. She formulates two models:

She estimates these models and obtains the following goodness-of-fit measures.

Model 1 Model 2

R2 0.5595 0.5322

Adjusted R2 0.5573 0.5298

se 40.3684 41.6007

Which model provides a better fit for y? Justify your response with two goodness-
of-fit measures.
21. FILE Property_Taxes. The accompanying data file shows the square footage
and associated property taxes for 20 homes in an affluent suburb 30 miles outside
ofNew York City.
a. Estimate a home’s property taxes as a linear function of the size of the home
(measured by its square footage).
b. What proportion of the sample variation in property taxes is explained by the
home’s size?
c. What proportion of the sample variation in property taxes is unexplained by the
home’s size?
22. FILE Car_Prices. The accompanying data file shows the selling price of a used
sedan, its age, and its mileage. Estimate two models:

Which model provides a better fit for y? Justify your response with two goodness-
of-fit measures.
23. For a sample of 41 New England cities, a sociologist studies the crime rate in
each city (crimes per 100,000 residents) as a function of its poverty rate (in %)
and its median income (in $1,000s). He finds that SSE = 4,182,663 and SST =
7,732,451.
a. Calculate the standard error of the estimate.
b. What proportion of the sample variation in crime rate is explained by the
variability in the predictor variables? What proportion is unexplained?

page 228
24. For a large restaurant chain, a financial analyst uses the
following model to estimate a franchise’s net profit: y = β0 + β1x1 + β2x2 + ɛi,
where y is net profit, x1 is counter sales, and x2 is drive-through sales. For a
sample of 100 franchises, she finds that SSE = 4.4600 and SST = 18.3784.
a. Calculate the standard error of the estimate.
b. Calculate and interpret the coefficient of determination.
c. Calculate the adjusted R2.
25. FILE Football. Is it defense or offense that wins football games? Consider the
following portion of data, which includes a team’s winning record (Win in %), the
average number of yards gained, and the average number of yards allowed
during a recent NFL season.

Team Win Yards Gained Yards Allowed

Arizona Cardinals 46.9 366.8 305.2

 Team Win Yards Gained Yards Allowed

Atlanta Falcons 68.8 415.8 371.2

⋮ ⋮ ⋮ ⋮

Washington Redskins 53.1 403.4 377.9

a. Compare two simple linear regression models where Model 1 predicts Win as a
function of Yards Gained and Model 2 predicts Win as a function of Yards
Allowed.
b. Estimate a linear regression model, Model 3, that applies both Yards Gained
and Yards Allowed to forecast Win. Is this model an improvement over the other
two models? Explain.
26. FILE Ownership. In order to determine if the homeownership rate in the U.S. is
linked with income, state-level data on the homeownership rate (Ownership in %)
and median household income (Income in $) were collected. A portion of the data
is shown in the accompanying table.

State Ownership Income

Alabama 73.2 40933

Alaska 65.7 57848

⋮ ⋮ ⋮

Wyoming 73.0 52201

a. Estimate and interpret the model Ownership = β0 + β1Income + ɛ. Is the

coefficient attached to Income as expected? Explain.
b. What is the standard error of the estimate?
c. Interpret the coefficient of determination. Does this seem like a promising
model? Explain.
27. FILE Return. A research analyst is trying to determine whether a firm’s price-
earnings (P/E) and price-sales (P/S) ratios can explain the firm’s stock
performance over the past year. A P/E ratio is calculated as a firm’s share price
compared to the income or profit earned by the firm per share. Generally, a high
P/E ratio suggests that investors are expecting higher earnings growth in the
future compared to companies with a lower P/E ratio. The P/S ratio is calculated
by dividing a firm’s share price by the firm’s revenue per share for the trailing 12
months. In short, investors can use the P/S ratio to determine how much they are
paying for a dollar of the firm’s sales rather than a dollar of its earnings (P/E ratio).
 In general, the lower the P/S ratio, the more attractive the investment. Consider
the following portion of data, which includes the year-to-date returns (Return in %)
and the P/E and P/S ratios for 30 firms.

Firm Return P/E P/S

1 4.4 14.37 2.41

2 –4.5 11.01 0.78

⋮ ⋮ ⋮ ⋮

30 16.3 13.94 1.94

a. Estimate Return = β0 + β1P/E + β2P/S + ɛ. Are the signs on the coefficients as

expected? Explain.
b. Interpret the slope coefficient of the P/S ratio.
c. What is the predicted return for a firm with a P/E ratio of 10 and a P/S ratio of
2?
d. What is the standard error of the estimate?
e. Interpret R2.
28. FILE SAT_2. It is generally believed that the wealthier a student’s family, the
higher the student’s Scholastic Aptitude Test (SAT) score. Another commonly
used predictor for SAT scores is the student’s grade point average (GPA).
Consider the following portion of data collected on 24 students.

SAT Income GPA

1651 47000 2.79

1581 34000 2.97

⋮ ⋮ ⋮

1940 113000 3.96

a. Estimate three models:

(i) SAT = β0 + β1Income + ɛ,
(ii) SAT = β0 + β1GPA + ɛ, and
(iii) SAT = β0 + β1Income + β2GPA + ɛ.
 b. Use goodness-of-fit measures to select the best-fitting model.
c. Use the preferred model to predict SAT given the mean value of the predictor
variable(s).

page 229

6.3 TESTS OF SIGNIFICANCE

LO 6.3
Conduct tests of significance.

In this section, we continue our assessment of the linear regression

model by turning our attention to hypothesis tests about the unknown
parameters (coefficients) β0, β1, . . . , βk. In particular, we test for joint
and individual significance to determine whether there is evidence of
a linear relationship between the response variable and the predictor
variables. We note that for the tests to be valid, the OLS estimators
b0, b1, . . . , bk must be normally distributed. This condition is
satisfied if the random error term, ɛ, is normally distributed. If we
cannot assume the normality of ɛ, then the tests are valid only for
large sample sizes.

Test of Joint Significance

Consider the following linear regression model, which links the
response variable y with k predictor variables x1, x2, . . . , xk:
If all of the slope coefficients equal zero, then all of the predictor
variables drop out of the model; this implies that none of the
predictor variables has a linear relationship with the response
variable. Conversely, if at least one of the slope coefficients does not
equal zero, then at least one predictor variable influences the
response variable.
When we assess a linear regression model, a test of joint
significance is often regarded as a test of the overall usefulness of a
regression. This test determines whether the predictor variables x1,
x2, . . . , xk have a joint statistical influence on y. Following the
hypothesis testing methodology introduced in Chapter 5, if we reject
the null hypothesis that all slope coefficients equal zero, then we are
able to conclude that at least one predictor variable influences the
response variable. The competing hypotheses for a test of joint
significance take the following form:

The test statistic for a test of joint significance follows the F

distribution. Like the tdf distribution, the F distribution is characterized
by a family of distributions; however, each distribution depends on
two degrees of freedom: the numerator degrees of freedom df1 and
the denominator degrees of freedom df2. It is common to refer to it
as the F (df1,df2) distribution.
We calculate the test statistic for a test of joint significance as

, where df1 = k, df2 = n = k − 1, MSR = SSR/k and MSE =

SSE/(n − k − 1). In other words, the F (df1,df2) test statistic is a ratio of

the mean square regression MSR to the mean square error MSE.
The test statistic measures how well the regression equation
explains the variability in the response variable. We always employ a
right-tailed F test when conducting a test of joint significance since a
larger value of F (df1,df2) provides us with more evidence to reject the
null hypothesis.

A TEST OF JOINT SIGNIFICANCE

For the linear regression model, y = β0 + β1x1 + β2x2 + . . . +
βkxk + ɛ, the following competing hypotheses are used for a test
of joint significance:

The value of the test statistic is calculated as:

page 230

where df1 = k, df2 = n − k − 1, SSR is the regression sum of

squares, SSE is the error sum of squares, MSR is the mean
square regression, and MSE is the mean square error.
If the null hypothesis is not rejected, then the predictor
variables are not jointly significant in explaining the response
variable; the model is not useful.
If the null hypothesis is rejected, then the predictor variables
are jointly significant in explaining the response variable; the
model is useful.
 Most statistical computer packages, including Excel and R,
produce an ANOVA table that decomposes the total variability of the
response variable y into two components: (1) the variability
explained by the regression and (2) the variability that is
unexplained. In addition, the value for the F (df1,df2) test statistic and
its p-value are also provided. Table 6.6 shows the general format of
an ANOVA table. Excel explicitly provides an ANOVA table with its
regression output, with the p-value reported under the heading
Significance F. R provides the value for the F (df1,df2) test statistic and
its p-value with its regression output, which is sufficient to conduct a
test of joint significance. If we use R’s anova function, R generates
an ANOVA table that is even more detailed than the one provided by
Excel; R’s ANOVA table breaks down the contribution of each
predictor variable in explaining the total variability in the response
variable.

TABLE 6.6 General Format of an ANOVA Table for Regression

FILE
College

EXAMPLE 6.4
Let’s revisit Model 3, from Section 6.2.
Model 3: Earnings = β0 + β1Cost + β2Grad + β3Debt + β4City
+ɛ

Recall that we chose Model 3 to predict annual post-college

earnings because it had the lowest standard error of the
estimate and the highest adjusted R2. We reproduce the
ANOVA portion of the regression results in Table 6.7. Conduct
a test to determine if the predictor variables are jointly
significant in explaining Earnings at α = 0.05.

TABLE 6.7 ANOVA portion of Regression Results for Model 3

SOLUTION: When testing whether the predictor variables are

jointly significant in explaining Earnings, we set up the following
competing hypotheses:

page 231

Given n = 116 and k = 4, we find that df1 = k = 4 and df2 = n − k

− 1 = 111. From Table 6.7, we find that
 The p-value, P(F(4,111) ≥ 20.868), is approximately equal to
zero; its exact value is 7.56 × 10−13. Because the p-value is
less than α = 0.05, we reject H0. At the 5% significance level,
the predictor variables are jointly significant in explaining
Earnings.

Test of Individual Significance

In addition to testing all slope coefficients jointly, we often want to
conduct tests on a single coefficient. Again consider the following
linear regression model, which links the response variable y with k
predictor variables x1, x2, . . . , xk:

If, for example, the slope coefficient β1 equals zero, then the
predictor variable x1 basically drops out of the equation, implying that
x1 does not influence y. In other words, if β1 equals zero, then there
is no linear relationship between x1 and y. Conversely, if β1 does not
equal zero, then x1 influences y.
In general, when we want to test whether the population
coefficient βj is different from, greater than, or less than βj0, where βj0
is the hypothesized value of βj, then the competing hypotheses take
one of the following forms:

Two-Tailed Test Right-Tailed Test Left-Tailed Test

H0: βj = βj 0 H0: βj ≤ βj 0 H0: βj ≥ βj 0

HA: βj ≠ βj 0 HA: βj > βj 0 HA: βj < βj 0
When testing whether xj significantly influences y, we set βj0 = 0 and
specify a two-tailed test as H0: βj = 0 and HA: βj ≠ 0. We could easily
specify one-tailed competing hypotheses for a positive linear
relationship (H0: βj ≤ 0 and HA: βj > 0) or a negative linear
relationship (H0: βj ≥ 0 and HA: βj < 0).
Although tests of significance are commonly based on βj0 = 0, in
some situations we might wish to determine whether the slope
coefficient differs from a nonzero value. For instance, if we are
analyzing the relationship between students’ exam scores on the
basis of hours studied, we may want to determine if an extra hour of
review before the exam will increase a student’s score by more than
five points. Here, we formulate the hypotheses as H0: βj ≤ 5 and HA:
βj > 5. Finally, although in most applications we are interested in
conducting hypothesis tests on the slope coefficient(s), there are
instances where we may also be interested in testing the intercept,
β0. The testing framework for the intercept remains the same; that is,
if we want to test whether the intercept differs from zero, we specify
the competing hypotheses as H0: β0 = 0 and HA: β0 ≠ 0.
As in all hypothesis tests, the next step is to define the
appropriate test statistic.

TEST OF INDIVIDUAL SIGNIFICANCE

For the linear regression model, y = β0 + β1x1 + β2x2 + . . . +
βkxk + ε, the following competing hypotheses are used to
conduct a test of individual significance:

H0: βj = βj0

HA: βj ≠ βj0

page 232
 The value of the test statistic is calculated as

where df = n − k − 1, bj is the estimate for βj, se(bj) is the

standard error of the estimator bj, and βj0 is the hypothesized
value of βj. If βj0 = 0, the value of the test statistic reduces to

Suppose the competing hypotheses are H0: βj = 0 versus

HA: βj ≠ 0.

If the null hypothesis is not rejected, then xj is not significant

in explaining y.
If the null hypothesis is rejected, then xj is significant in
explaining y.

We would like to note that while the test of joint significance is

important for a multiple linear regression model, it is redundant for a
simple linear regression model. In fact, for a simple linear regression
model, the p-value of the F-test is identical to that of the t-test on the
single slope coefficient. We advise you to verify this fact.
FILE
College

EXAMPLE 6.5
Let’s again revisit Model 3.
 Model 3: Earnings = β 0 + β 1Cost + β 2Grad + β 3Debt + β
4City + ε

We produce a portion of the regression results in Table 6.8.

Conduct a hypothesis test to determine whether Cost
influences Earnings at the 5% significance level.

TABLE 6.8 Portion of Regression Results for Model 3

SOLUTION: We set up the following competing hypotheses in

order to determine whether Cost influences Earnings:

H0 : β1 = 0

HA : β1 ≠ 0

From Table 6.8, we find that b1 = 0.4349 and se(b1) = 0.1110.

Given earlier information that n = 116, we find df = n − k − 1 =
116 − 4 − 1 = 111. So, using unrounded calculations, we find

the value of the test statistic as . Note

that this calculation is not necessary because virtually all

statistical computer packages automatically provide the value
of the test statistic and its associated p-value. As usual, the
decision rule is to reject H0 if the p-value < α. Because the
reported p-value is 0.0002, we reject H0. At the 5% significance
level, Cost is significant in explaining Earnings.
 page 233

It is important to note that the computer-generated results are valid

only in a standard case where a two-tailed test is implemented to
determine whether a regression coefficient differs from zero. In
Example 6.5 we could use the computer-generated value of the test
statistic as well as the corresponding p-value because it represented
a standard case. For a one-tailed test with βj0 = 0, the value of the
test statistic is valid, but the p-value is not; in most cases, the
computer-generated p-value must be divided in half. For a one- or
two-tailed test to determine if the regression coefficient differs from a
nonzero value, both the computer-generated value of the test
statistic and the p-value become invalid. These facts are
summarized below.

COMPUTER-GENERATED TEST STATISTIC AND THE p-

VALUE
Virtually all statistical packages report a value of the test
statistic and its associated p-value for a two-tailed test that
assesses whether the regression coefficient differs from zero.
If we specify a one-tailed test, then we need to divide the
computer-generated p-value in half.
If we test whether the coefficient differs from a nonzero value,
then we cannot use the value of the computer-generated test
statistic and its p-value.

We would also like to point out that for a one-tailed test with βj0 = 0,
there are rare instances when the computer-generated p-value is
invalid. This occurs when the sign of bj (and the value of the
accompanying test statistic) is not inconsistent with the null
hypothesis. For example, for a right-tailed test, H0: βj ≤ 0 and HA: βj
> 0, the null hypothesis cannot be rejected if the estimate bj (and the
value of the accompanying test statistic tdf) is negative. Similarly, no
further testing is necessary if bj > 0 (and thus tdf > 0) for a left-tailed
test. In these rare instances, the reported p-value is invalid.

A Test for a Nonzero Slope Coefficient

In Example 6.5, the null hypothesis included a zero value for the
slope coefficient—that is, βj0 = 0. We now motivate a test where the
hypothesized value is not zero by using a renowned financial
application referred to as the capital asset pricing model (CAPM).
Let R represent the return on a stock or portfolio of interest.
Given the market return RM and the risk-free return Rf, the CAPM
expresses the risk-adjusted return of an asset, R − Rf, as a function
of the risk-adjusted market return, RM − Rf. It is common to use the
return of the S&P 500 index for RM and the return on a Treasury bill
for Rf. For empirical estimation, we express the CAPM as

We can rewrite the model as y = α + βx + ɛ, where y = R − Rf and x =

RM − Rf. Note that this is essentially a simple linear regression model
that uses α and β, in place of the usual β0 and β1, to represent the
intercept and the slope coefficients, respectively. The slope
coefficient β, called the stock’s beta, measures how sensitive the
stock’s return is to changes in the level of the overall market. When β
equals 1, any change in the market return leads to an identical
change in the given stock return. A stock for which β > 1 is
considered more “aggressive” or riskier than the market, whereas
one for which β < 1 is considered “conservative” or less risky. We
also give importance to the intercept coefficient α, called the stock’s
alpha. The CAPM theory predicts α to be zero, and thus a nonzero
estimate indicates abnormal returns. Abnormal returns are positive
when α > 0 and negative when α < 0.

page 234
FILE
J&J

EXAMPLE 6.6
Johnson & Johnson (J&J) was founded more than 120 years
ago on the premise that doctors and nurses should use sterile
products to treat people’s wounds. Since that time, J&J
products have become staples in most people’s homes.
Consider the CAPM where the J&J risk-adjusted stock return R
− Rf is used as the response variable and the risk-adjusted
market return RM − Rf is used as the predictor variable. A
portion of 60 months of data is shown in Table 6.9.

TABLE 6.9 Risk-Adjusted Stock Return of J&J and Market Return

Month Year R - Rf RM - Rf

Jan 2012 -0.0129 0.0403

Feb 2012 0.0216 0.0304

⋮ ⋮ ⋮ ⋮

Dec 2016 -0.0221 0.0128

a. Because consumer staples comprise many of the products
sold by J&J, its stock is often considered less risky; that is,
people need these products whether the economy is good or
bad. At the 5% significance level, is the beta coefficient less
than one?
b. At the 5% significance level, are there abnormal returns? In
other words, is the alpha coefficient significantly different from
0?

SOLUTION: Using the CAPM notation, we estimate the model, R

− Rf = α + β(RM − Rf) + ε; the relevant portion of the regression
output is presented in Table 6.10.

TABLE 6.10 Portion of CAPM Regression Results for J&J

a. The estimate for the beta coefficient is 0.7503 and its standard
error is 0.1391. In order to determine whether the beta
coefficient is significantly less than one, we formulate the
hypotheses as
H0 : β ≥ 1
HA:β<1
Given 60 data points, df = n − k − 1 = 60 − 1 − 1 = 58. We
cannot use the test statistic value or the p-value reported in
Table 6.10 because the hypothesized value of β is not zero.
Using unrounded calculations, we find the value of the test

statistic as . We can use the t table to

approximate the p-value, P(t58 ≤ −1.796), as a value that is

between 0.025 and 0.05. Using statistical software, we find
 that the exact p-value is 0.039. Because the p-value < α =
0.05, we reject H0 and conclude that β is significantly less than
one; that is, the return on J&J stock is less risky than the
return on the market.
b. Abnormal returns exist when α is significantly different from
zero. Thus, the competing hypotheses are H0: α = 0 versus
HA: α ≠ 0. Because it is a standard case, where the
hypothesized value of the coefficient is zero, we can use the
reported test statistic value of 1.127 with an associated p-
value of 0.2645. We cannot reject H0 at any reasonable level
of significance. Therefore, we cannot conclude that there are
abnormal returns for J&J stock.

page 235

Reporting Regression Results

Regression results are often reported in a “user-friendly” table. Table
6.11 reports the regression results for the three models that attempt
to explain annual post-college earnings (Earnings). For Model 1, the
predictor variable is the average annual cost (Cost); for Model 2, the
predictor variables are Cost, the graduation rate (Grad), and the
percentage of students paying down debt (Debt); and for Model 3,
the predictor variables are Cost, Grad, Debt, and whether or not a
college is located in a city (City equals 1 if a city location, 0
otherwise). If we were supplied with only this table, we would be able
to compare these models, construct the sample regression equation
of the chosen model, and perform a respectable assessment of the
model with the statistics provided. Many tables contain a Notes
section at the bottom explaining some of the notation. We choose to
put the p-values in parentheses under all the estimated coefficients;
however, some analysts place the standard errors of the coefficients
or the values of the test statistics in parentheses. Whichever format
is chosen, it must be made clear to the reader in the Notes section.

TABLE 6.11 Estimates of Alternative Regression Models to Explain Earnings, n = 116

Model 1 Model 2 Model 3

Intercept 28,375.4051* 11819.4747 10,004.9665

(0.000) (0.129) (0.193)

Cost 0.7169 * 0.5050* 0.4349*

(0.000) (0.000) (0.000)

Grad NA 192.6664* 178.0989*

(0.007) (0.011)

Debt NA 104.6573 141.4783

(0.378) (0.230)

City NA NA 2,526.7888*
(0.024)

Se 6,271.4407 5,751.8065 5,645.8306

R2 0.2767 0.4023 0.4292

Adjusted R2 0.2703 0.3862 0.4087

F-test (p-value) 43.608 (0.000) 25.124 (0.000) 20.868(0.000)

NOTES: Parameter estimates are in the top half of the table with the p-values in
parentheses; * represents significance at the 5% level. NA denotes not applicable. The
lower part of the table contains goodness-of-fit measures.

SYNOPSIS OF INTRODUCTORY CASE

The Department of Education published a redesigned College Scorecard that
reports the most reliable national data on college costs and students’ outcomes at
specific colleges. The availability of clear, reliable data was welcomed by families
who are searching for answers to critical questions concerning college affordability
and value.

Africa Studio/Shutterstock

In an attempt to determine which college factors ‘best’ explain annual post-

college earnings, three regression models were estimated. A combination of four
predictor variables was used in the analysis. The four predictor variables were the
average annual cost (Cost), the graduation rate (Grad), the percentage of students
paying down debt (Debt), and a City dummy variable that equals 1 if the college is
located in a city, 0 otherwise. Goodness-of-fit measures suggested that the model
that included all four predictor variables provided the best overall fit, as measured by
its lowest standard error of the estimate and its highest adjusted
page 236
R2 value. The sample regression equation was

. This regression equation implies that if a college’s average annual cost is $25,000,
its graduation rate is 60%, its percentage of students paying down debt is 80%, and
it is located in a city, then average post-college earnings for its students are
$45,409.
 Further testing of this preferred model revealed that the four predictor variables
were jointly significant. Individual tests of significance showed that Cost, Grad, and
City were significant at the 5% level; Debt was not significant in explaining Earnings.
The coefficient of determination, or R2, revealed that approximately 43% of the
sample variability in annual post-college earnings is explained by the model. Thus,
57% of the sample variability in annual post-college earnings remains unexplained.
This is not entirely surprising because factors not included in the model, such as
field of emphasis, grade point average, and natural ability, influence annual post-
college earnings.

EXERCISES 6.3
Applications
29. In order to examine the relationship between the selling price of a used car and its
age, an analyst uses data from 20 recent transactions and estimates Price = β0 +
β1Age + ɛ. A portion of the regression results is shown in the accompanying table.

a. Specify the competing hypotheses in order to determine whether the selling

price of a used car and its age are linearly related.
b. Calculate the value of the test statistic.
c. At the 5% significance level, is the age of a used car significant in explaining its
selling price?
d. Conduct a hypothesis test at the 5% significance level in order to determine if
β1 differs from −1,000. Show all of the relevant steps.
30. For a sample of 20 New England cities, a sociologist studies the crime rate in
each city (crimes per 100,000 residents) as a function of its poverty rate (in %)
and its median income (in $1,000s). A portion of the regression results is shown in
the accompanying table.
 a. Specify the sample regression equation.
b. At the 5% significance level, are the poverty rate and income jointly significant
in explaining the crime rate?
c. At the 5% significance level, show whether the poverty rate and the crime rate
are linearly related.
d. Determine whether income influences the crime rate at the 5% significance
level.
31. Akiko Hamaguchi is a manager at a small sushi restaurant in Phoenix, Arizona.
Akiko is concerned that the weak economic environment has hampered foot traffic
in her area, thus causing a dramatic decline in sales. In order to offset the decline
in sales, she has pursued a strong advertising campaign. She believes
advertising expenditures have a positive influence on sales. To support her claim,
Akiko estimates the following linear regression model: Sales = β0 +
β1Unemployment + β2Advertising + ɛ. A portion of the regression results is shown
in the accompanying table.

a. At the 5% significance level, test whether the predictor

variables jointly influence sales.
b. At the 1% significance level, test whether the unemployment
rate is negatively related with sales.
c. At the 1% significance level, test whether advertising
expenditures are positively related with sales.

page 237
32. A researcher estimates the following model relating the return on
a firm’s stock as a function of its price-to-earnings ratio and its price-to-sales ratio:
Return = β0 + β1P/E + β2P/S + ɛ. A portion of the regression results is shown in
the accompanying table.

a. Specify the sample regression equation.

b. At the 10% significant level, are P/E and P/S jointly significant? Show the
relevant steps of the test.
c. Are both predictor variables individually significant at the 10% significance
level? Show the relevant steps of the test.
33. FILE Fertilizer. A horticulturist is studying the relationship between tomato plant
height and fertilizer amount. Thirty tomato plants grown in similar conditions were
subjected to various amounts of fertilizer (in ounces) over a four-month period,
and then their heights (in inches) were measured.
a. Estimate the regression model: Height = β0 + β1Fertilizer + ɛ.
b. At the 5% significance level, determine if an ounce of fertilizer increases height
by more than three inches. Show the relevant steps of the test.
34. FILE Dexterity. Finger dexterity, the ability to make precisely coordinated finger
movements to grasp or assemble very small objects, is important in jewelry
making. Thus, the manufacturing manager at Gemco, a manufacturer of high-
quality watches, wants to develop a regression model to predict the productivity,
measured by watches per shift, of new employees based on the time required (in
seconds) to place three pins in each of 100 small holes using tweezers. He has
subjected a sample of 20 current employees to the O’Connor dexterity test in
 which the time required to place the pins and the number of watches produced
per shift are measured.
a. Estimate the regression model:Watches = β0 + β1Time + ɛ.
b. The manager claims that for every extra second taken on placing the pins, the
number of watches produced decreases by more than 0.02. Test this claim at
the 5% significance level. Show the relevant steps of the test.
35. FILE Engine. The maintenance manager at a trucking company wants to build a
regression model to forecast the time until the first engine overhaul (Time in
years) based on four predictor variables: (1) annual miles driven (Miles in 1,000s),
(2) average load weight (Load in tons), (3) average driving speed (Speed in mph),
and (4) oil change interval (Oil in 1,000s miles). Based on driver logs and onboard
computers, data have been obtained for a sample of 25 trucks.
a. Estimate the time until the first engine overhaul as a function of all four predictor
variables.
b. At the 10% significance level, are the predictor variables jointly significant?
Show the relevant steps of the test.
c. Are the predictor variables individually significant at the 10% significance level?
Show the relevant steps of the test.
36. FILE Electricity_Cost. The facility manager at a pharmaceutical company wants
to build a regression model to forecast monthly electricity cost. Three main
variables are thought to dictate electricity cost: (1) average outdoor temperature
(Temp in °F), (2) working days per month (Days), and (3) tons of product
produced (Tons).
a. Estimate the regression model.
b. At the 10% significance level, are the predictor variables jointly significant?
Show the relevant steps of the test.
c. Are the predictor variables individually significant at the 10% significance level?
Show the relevant steps of the test.
37. FILE Caterpillar. Caterpillar, Inc., manufactures and sells heavy construction
equipment worldwide. The performance of Caterpillar’s stock is likely to be
strongly influenced by the economy. For example, during the Great Recession,
the value of Caterpillar’s stock plunged dramatically. Monthly data for Caterpillar’s
risk-adjusted return (R − Rf) and the risk-adjusted market return (RM − Rf) are
collected for a five-year period (n = 60). A portion of the data is shown in the
accompanying table.

Month Year R - Rf RM - Rf

Jan 2012 0.051024 0.040289

Feb 2012 -0.068230 0.030432

 Month Year R - Rf RM - Rf

⋮ ⋮ ⋮ ⋮

Dec 2016 0.026386 0.012784

a. Estimate the CAPM model for Caterpillar, Inc. Show the regression results in a
well-formatted table.
b. At the 5% significance level, determine if investment in Caterpillar is riskier than
the market (beta significantly greater than 1).
c. At the 5% significance level, is there evidence of abnormal returns?
38. FILE Arlington. A Realtor examines the factors that influence the price of a
house in Arlington, Massachusetts. He collects data on 36 house sales (Price in
$) and notes each house’s square footage (Sqft), the number of
bedrooms (Beds), the number of bathrooms (Baths), and a
page 238
Colonial dummy variable (Colonial equals 1 if house is colonial, 0 otherwise). A
portion of the data is shown in the accompanying table.

a. Estimate Price = β0 + β1Sqft + β2Beds + β3Baths + β4Colonial + ɛ. Show the

regression results in a well-formatted table.
b. At the 5% significance level, are the predictor variables jointly significant in
explaining Price?
c. At the 5% significance level, are all predictor variables individually significant in
explaining Price?
39. FILE Final_Test. On the first day of class, an economics professor administers a
test to gauge the math preparedness of her students. She believes that the
performance on this math test and the number of hours studied per week on the
course are the primary factors that predict a student’s score on the final exam.
She collects data from 60 students, a portion of which is shown in the
accompanying table.

Final Math Hours

94 92 5
 74 90 3

⋮ ⋮ ⋮

63 64 2

a. Estimate the sample regression equation that enables us to

predict a student’s final exam score on the basis of his/her
math score and the number of hours studied per week.
b. At the 5% significance level, are a student’s math score and
the number of hours studied per week jointly significant in
explaining a student’s final exam score?
c. At the 5% significance level, is each predictor variable
individually significant in explaining a student’s final exam
score?
40. FILE Union_Pay. An automotive workers union, in conjunction with top
management, is negotiating a new hourly pay policy for union workers based on
three variables: (1) job class, (2) years with the company, and (3) years as a union
member at any company. The goal is to develop an equitable model that can
objectively specify hourly pay, thereby reducing pay disparity grievances. Fifty
union workers have been sampled and will be used as the basis for the pay
model. A portion of the data is shown in the accompanying table.

Hourly Pay Job Class Years with Company Years in Union

15.90 24 12 7

23.70 52 17 14

⋮ ⋮ ⋮ ⋮

26.70 43 2 2

a. Report the sample regression equation of the appropriate

model.
b. At the 5% significance level, are the predictor variables
jointly significant? Are they individually significant?
c. Predict hourly pay for a worker in Job Class 48 with 18 years
of experience at the company and 14 years with the union.
41. FILE Yields. While the Federal Reserve controls short-term interest rates, long-
term interest rates essentially depend on supply/demand dynamics, as well as
longer-term interest rate expectations. In order to examine the relationship
between short-term and long-term interest rates, monthly data were collected on
the three-month Treasury yield (in %) and the 10-year Treasury yield (in %). A
portion of the data is shown in the accompanying table.

Month Year 3-month 10-year

Jan 2016 0.26 2.09

Feb 2016 0.31 1.78

⋮ ⋮ ⋮ ⋮

Dec 2017 0.32 2.40

a. Estimate and interpret a sample regression equation using the 10-year yield as
the response variable and the three-month yield as the predictor variable.
b. Interpret the coefficient of determination.
c. At the 5% significance level, is the three-month-yield significant in explaining
the 10-year yield?
d. Many wonder whether a change in the three-month yield implies the same
change in the 10-year yield. At the 5% significance level, is this belief supported
by the data?
42. FILE BMI. According to the World Health Organization, obesity has reached
epidemic proportions globally. While obesity has generally been linked with
chronic disease and disability, researchers argue that it may also affect salaries.
In other words, the body mass index (BMI) of an employee is a predictor for
salary. (A person is considered overweight if his/ her BMI is at least 25 and obese
if BMI exceeds 30.) The accompanying table shows a portion of
salary data (in $1,000s) for 30 college-educated men with their
page 239
respective BMI and a dummy variable that equals 1 for a white man and 0
otherwise.

Salary BMI White

34 33 1

43 26 1
 Salary BMI White

⋮ ⋮ ⋮

45 21 1

a. Estimate a model for Salary using BMI and White as the predictor variables.
Determine if BMI influences salary at the 5% level of significance.
b. What is the estimated salary of a white college-educated man with a BMI of 30?
Compute the corresponding salary of a nonwhite man.
43. FILE Wage. A researcher wonders whether males get paid more, on average,
than females at a large firm. She interviews 50 employees and collects data on
each employee’s hourly wage (Wage in $), years of higher education (Educ),
years of experience (Exper), age (Age), and a Male dummy variable that equals 1
if male, 0 otherwise. A portion of the data is shown in the accompanying table.

a. Estimate Wage = β0 + β1Educ + β2Exper + β3Age + β4Male + ɛ.

b. Predict the hourly wage of a 40-year-old male employee with 10 years of higher
education and five years of experience. Predict the hourly wage of a 40-year-
old female employee with the same qualifications.
c. Interpret the estimated coefficient for Male. Is the variable Male significant at
the 5% level? Do the data suggest that sex discrimination exists at this firm?
44. FILE Quotations. The labor estimation group at Sturdy Electronics, a contract
electronics manufacturer of printed circuit boards, wants to simplify the process it
uses to quote production costs to potential customers. It has identified the primary
drivers for production time (and thus production cost) as being the number of
electronic parts that can be machine-installed and the number of parts that must
be manually installed. Accordingly, it wishes to develop a multiple regression
model to predict production time, measured as minutes per board, using a
random sample of 25 recent product quotations. A portion of the data is shown in
the accompanying table.

Production Time Machine Parts Manual Parts

 9.1 275 14
 Production Time Machine Parts Manual Parts

10.8 446 12

⋮ ⋮ ⋮

15.5 618 16

a. What is the sample regression equation?

b. Predict production time for a circuit board with 475 machine-installed
components and 16 manually installed components.
c. What proportion of the sample variability in production time is explained by the
two predictor variables?
d. At the 5% significance level, are the predictor variables jointly significant? Are
they individually significant?
45. FILE Ice_Cream. A manager at an ice cream store is trying to determine how
many customers to expect on any given day. Overall business has been relatively
steady over the past several years, but the customer count seems to have ups
and downs. He collects data over 30 days and records the number of customers,
the high temperature (in degrees Fahrenheit), and whether the day fell on a
weekend (Weekend equals 1 if weekend, 0 otherwise). A portion of the data is
shown in the accompanying table.

Customers Temperature Weekend

376 75 0

433 78 0

⋮ ⋮ ⋮

401 68 0

a. Estimate Customers = β0 + β1Temperature + β2Weekend +

ɛ.
b. How many customers should the manager expect on a
Sunday with a forecasted high temperature of 80°?
c. Interpret the estimated coefficient for Weekend. Is it
significant at the 5% level? How might this affect the store’s
staffing needs?
46. FILE QuickFix. The general manager of QuickFix, a chain of quick-service, no-
appointment auto repair shops, wants to develop a model to forecast monthly
vehicles served at any particular shop based on four factors: garage bays,
population within five-mile radius (Population in 1,000s), interstate highway
access (Access equals 1 if convenient, 0 otherwise), and time of
year (Winter equals 1 if winter, 0 otherwise). He believes that, all
page 240
else equal, shops near an interstate will service more vehicles and that more
vehicles will be serviced in the winter due to battery and tire issues. A sample of
19 locations has been obtained. A portion of the data is shown in the
accompanying table.

a. Estimate the regression equation relating vehicles serviced to the four predictor
variables.
b. Interpret each of the slope coefficients.
c. At the 5% significance level, are the predictor variables jointly significant? Are
they individually significant? What about at the 10% significance level?
d. What proportion of the variability in vehicles served is explained by the four
predictor variables?
e. Predict vehicles serviced in a nonwinter month for a particular location with five
garage bays, a population of 40,000, and convenient interstate access.
47. FILE Industry. Consider a regression model that links a CEO’s compensation (in
$ millions) with the total assets of the firm (in $ millions) and the firm’s industry.
Dummy variables are used to represent four industries: Manufacturing Technology
d1, Manufacturing Other d2, Financial Services d3, and Nonfinancial Services d4.
A portion of the data for the 455 CEOs is shown in the accompanying table.
a. Estimate the model: y = β0 + β1x + β2d1 + β3d2 + β4d3 + ɛ,
where y and x denote compensation and assets,
respectively. Here the reference category is the nonfinancial
services industry.
b. Interpret the estimated coefficients.
c. Use a 5% level of significance to determine which industries,
relative to the nonfinancial services industry, have different
executive compensation.
d. Reformulate the model to determine, at the 5% significance
level, if compensation is higher in Manufacturing Other than
in Manufacturing Technology. Your model must account for
total assets and all industry types.
48. FILE Retail. A government researcher is analyzing the relationship between
retail sales (in $ millions) and the gross national product (GNP in $ billions). He
also wonders whether there are significant differences in retail sales related to the
quarters of the year. He collects 10 years of quarterly data. A portion is shown in
the accompanying table.

Year Quarter Sales GNP

2007 1  921266 14301.854

2007 2 1013371 14512.945

⋮ ⋮    ⋮ ⋮

2016 4 1299699 19134.463

a. Estimate y = β0 + β1x + β2d1 + β3d2 + β4d3 + ɛ, where y and x represent retail

sales and GNP, respectively; d1 equals 1 if quarter 1, 0 otherwise; d2 equals 1
if quarter 2, 0 otherwise; and d3 equals 1 if quarter 3, 0 otherwise. Note that the
reference category is quarter 4.
b. Predict retail sales in quarters 2 and 4 if GNP equals $16,000 billion.
c. Which of the quarterly sales are significantly different from those of the 4th
quarter at the 5% level?
d. Reformulate the model to determine, at the 5% significance level, if sales differ
between quarter 2 and quarter 3. Your model must account for all quarters.
49. FILE Longevity. According to the Centers for Disease Control and Prevention,
life expectancy at age 65 in the United States is about 18.7 years. Medical
researchers have argued that while excessive drinking is detrimental to health,
drinking a little alcohol every day, especially wine, may be associated with an
increase in life expectancy. Others have also linked longevity with income and a
person’s sex. The accompanying table shows a portion of data relating to the
length of life after 65; average income (in $1,000s) at a retirement age of 65; a
Female dummy variable that equals 1 if the individual is female, 0 otherwise; and
the average number of alcoholic drinks consumed per day.

Life Income Female Drinks

19.00 64 0 1

19.30 43 1 3

⋮ ⋮ ⋮ ⋮

20.24 36 1 0

a. Use the data to model life expectancy at 65 on the basis of Income, Female,
and Drinks.
b. Conduct a one-tailed test at a = 0.01 to determine if females live longer than
males.
c. Predict the life expectancy at 65 of a male with an income of $40,000 and an
alcohol consumption of two drinks per day; repeat the prediction for a female.

page 241
50. FILE SAT_3. A researcher from the Center for Equal
Opportunity wants to determine if SAT scores of admitted students at a large state
university differed by ethnicity. She collects data on SAT scores and ethnic
background for 200 admitted students. A portion of the data is shown in the
accompanying table.

SAT Ethnicity

1515 White

1530 Latinx

⋮ ⋮

1614 White
 a. Estimate the model y = β0 + β1d1 + β2d2 + β3d3 + ɛ, where y represents a
student’s SAT score; d1 equals 1 if the student is white, 0 otherwise; d2 equals
1 if the student is black, 0 otherwise; and d3 equals 1 if the student is Asian, 0
otherwise. Note that the reference category is Latinx. What is the predicted SAT
score for an Asian student? For a Latinx student?
b. At the 5% significance level, determine if the SAT scores of Asian students
differ from those of Latinx students.
c. Reformulate the model to determine if the SAT scores of white students are
lower than the SAT scores of Asian students at the 5% significance level. Your
model must account for all ethnic categories.

6.4 MODEL ASSUMPTIONS AND

COMMON VIOLATIONS
So far we have focused on the estimation and the assessment of
linear regression models. It is important to understand that the
statistical properties of the ordinary least squares (OLS) estimator,
as well as the validity of the testing procedures, depend on the
assumptions of the classical linear regression model. In this section,
we discuss these assumptions. We also address common violations
to the assumptions, discuss the consequences when the
assumptions are violated, and, where possible, offer some remedies.

Required Assumptions of Regression

Analysis
In this chapter, we have estimated the classical linear regression
model using the ordinary least squares (OLS) methodology. We have
also examined various goodness-of-fit measures and conducted
hypothesis tests in order to assess which predictor variables matter
the most. However, when we apply any of these techniques, we
require that certain assumptions hold. The following are the
assumptions that underlie the classical linear regression model:
1. The regression model given by y = β0 + β1x1 + β2x2 + . . . + βkxk + ɛ
is linear in the parameters β0, β1, . . . , βk.
2. Conditional on x1, x2, . . . , xk, the error term has an expected value
of zero, or E(ɛ) = 0. This implies that E(y) = β0 + β1x1 + β2x2 + . . . +
βkxk.
3. There is no exact linear relationship among the predictor variables;
or, in statistical terminology, there is no perfect multicollinearity.
4. Conditional on x1, x2, . . . , xk, the variance of the error term ɛ is the
same for all observations; or, in statistical terminology, there is no
heteroskedasticity. The assumption is violated if observations have
a changing variability.
5. Conditional on x1, x2, . . . , xk, the error term ɛ is uncorrelated
across observations; or, in statistical terminology, there is no serial
correlation. The assumption is violated if observations are
correlated.
6. The error term ɛ is not correlated with any of the predictor variables
x1, x2, . . . , xk; or, in statistical terminology, there is no endogeneity.
In general, this assumption is violated if important predictor
variables are excluded.
7. Conditional on x1, x2, . . . , xk, the error term ɛ is normally
distributed. This assumption allows us to construct interval
estimates and conduct tests of significance. If ɛ is not normally
distributed, the interval estimates and the hypothesis tests are valid
only for large sample sizes.

page 242

Under the assumptions of the classical linear regression model,

the OLS estimators have desirable properties. In particular, the OLS
estimators of the regression coefficients βj are unbiased; that is, E(bj)
= βj. Moreover, among all linear unbiased estimators, they have
minimum variations between samples. These desirable properties of
the OLS estimators become compromised as one or more model
assumptions are violated. Aside from coefficient estimates, the
validity of the significance tests is also impacted by the assumptions.
For certain violations, the estimated standard errors of the OLS
estimators are inappropriate; in these cases, it is not possible to
make meaningful inferences from the t and the F test results.
The assumptions of the classical linear regression model are, for
the most part, based on the error term ɛ. Because the residuals, or
the observed error term, e = y – , contain useful information

regarding ɛ, it is common to use the residuals to investigate the

assumptions. In this section, we will rely on residual plots to detect
some of the common violations to the assumptions. These graphical
plots are easy to use and provide informal analysis of the estimated
regression models. Formal tests are beyond the scope of this text.

RESIDUAL PLOTS
For the regression model, y = β0 + β1x1 + β2x2 + . . . + βkxk + ɛ,

the residuals are computed as e = y , where = b0 + b1x1 +

b2x2 + . . . + bkxk. These residuals can be plotted sequentially

or against a predictor variable xj to look for model
inadequacies.

It is common to plot the residuals e on the vertical axis and the

predictor variable xj or the predicted values on the horizontal axis.

Such plots are useful for detecting departures from linearity as well
as constant variability. If the regression is based on time series data,
we can plot the residuals sequentially to detect if the observations
are correlated.
Residual plots can also be used to detect outliers. Recall from
Chapter 3 that outliers are observations that stand out from the rest
of the data. For an outlier observation, the resulting residual will
appear distinct in a plot; it will stand out from the rest. While outliers
can greatly impact the estimates, it is not always clear what to do
with them. Outliers may indicate bad data due to incorrectly recorded
(or included) observations in the data set. In such cases, the relevant
observation should be corrected or simply deleted. Alternatively,
outliers may just be due to random variations, in which case the
relevant observations should remain. In any event, residual plots
help us identify potential outliers so that we can take corrective
actions, if needed.
In Figure 6.7, we present a hypothetical residual plot when none
of the assumptions has been violated. Note that all the points are
randomly dispersed around the 0 value of the residuals. Also, there
is no evidence of outliers because no residual stands out from the
rest. Any discernible pattern of the residuals indicates that one or
more assumptions have been violated.

FIGURE 6.7 Residual plot of a correctly specified model

page 243

We discuss how to obtain residual plots in Excel and R at the end

of this section, but first we describe common violations of the
assumptions and offer remedies.
Common Violation 1: Nonlinear Patterns
LO 6.4
Address common violations of the OLS assumptions.

Linear regression models are often justified on the basis of their

computational simplicity. The simple linear regression model y = β0 +
β1x + ɛ implies that if x goes up by one unit, we expect y to change
by β1, irrespective of the value of x. However, in many applications,
the relationship cannot be represented by a straight line and,
therefore, must be captured by an appropriate curve. It is always
good to rely on economic theory and intuition to determine if the
linearity assumption is appropriate. We confirm our intuition by
analyzing scatterplots or residual plots. The OLS estimates can be
quite misleading if there are obvious nonlinear patterns in the data.

Detection
We can use residual plots to identify nonlinear patterns. Linearity is
justified if the residuals are randomly dispersed across the values of
a predictor variable. A discernible trend in the residuals is indicative
of nonlinear patterns.
FILE
Happiness_Age

EXAMPLE 6.7
A sociologist wishes to study the relationship between age and
happiness. He interviews 24 individuals and collects data on
age and happiness, measured on a scale from 0 to 100. A
portion of the data is shown in Table 6.12. Examine the linearity
assumption in the regression model, Happiness = β0 + β1Age +
ɛ.
TABLE 6.12 Happiness and Age

Happiness Age

62 49

66 51

⋮ ⋮

72 69

SOLUTION:We start the analysis with a scatterplot of Happiness

against Age. Figure 6.8 shows the scatterplot and the
superimposed trend line, which is based on the sample

regression equation . It is fairly clear from

Figure 6.8 that the linear regression model does not

appropriately capture the relationship between Happiness and
Age. In other words, it is misleading to conclude that a person’s
happiness increases by 0.28 unit every year.

FIGURE 6.8 Scatterplot and the superimposed trendline (Example 6.7)

 page 244

Figure 6.9 is a residual plot that further explores the linearity

assumption of the regression model.

FIGURE 6.9 Residual plot against Age (Example 6.7)

It shows that there is an obvious trend with the residuals

decreasing until the age of 50 and steadily increasing
thereafter. The linear regression model is inappropriate as it
underestimates at lower and higher age levels and
overestimates in the middle. This result is consistent with a
2018 study in The Journal of Happiness that shows that
 happiness initially decreases with age and then increases with
age.

Remedy
Linear regression models are often used as a first pass for most
empirical work. In many instances, they provide a very good
approximation for the actual relationship. However, if residual plots
exhibit strong nonlinear patterns, the inferences made by a linear
regression model can be quite misleading. In such instances, we
should employ nonlinear regression methods based on simple
transformations of the response and the predictor variables; these
methods are discussed in the next chapter.

Common Violation 2: Multicollinearity

Perfect multicollinearity exists when two or more predictor variables
have an exact linear relationship. Consider the model y = β0 + β1x1 +
β2x2 + ɛ, where y is bonus, x1 is the number of cars sold, and x2 is
the number of cars remaining in the lot. If all car salesmen started
with the same inventory, we have a case of perfect multicollinearity
(x2 = Constant − x1). Perfect multicollinearity is easy to detect
because the model cannot be estimated. However, if x2 represents
the proportion of positive reviews from customers, we have some
multicollinearity because the number of cars sold and the proportion
of positive reviews are likely to be correlated. In most applications,
some degree of correlation exists between the predictor variables.
The problem with (nonperfect) multicollinearity is similar to that of
small samples. Multicollinearity does not violate any of the
assumptions; however, its presence results in imprecise estimates of
the slope coefficients. In other words, multicollinearity makes it
difficult to disentangle the separate influences of the predictor
variables on the response variable. If multicollinearity is severe, we
may find insignificance of important predictor variables. Seemingly
wrong signs of the estimated regression coefficients may also
indicate multicollinearity.

Detection
The detection methods for multicollinearity are mostly informal. If we
find a high R2 and a significant F statistic coupled with individually
insignificant predictor variables, then multicollinearity may be an
issue. We can also examine the correlations between the predictor
variables to detect severe multicollinearity. One page 245
guideline suggests that multicollinearity is severe if the
sample correlation coefficient between any two predictor variables is
more than 0.80 or less than −0.80. The variance inflation factor (VIF)
is another measure that can detect a high correlation between three
or more predictor variables even if no pair of predictor variables has
a particularly high correlation. The smallest possible value of VIF is
one (absence of multicollinearity). Multicollinearity may be a problem
if the VIF exceeds five or 10. A more detailed discussion of the VIF is
beyond the scope of this text.
FILE
Home_Values

EXAMPLE 6.8
Examine the multicollinearity issue in a linear regression model
that uses median home values (in $) as the response variable
and median household incomes (in $), per capita incomes (in
$), and the percentage of owner-occupied homes as the
predictor variables. A portion of 2010 data for all states in the
United States is shown in Table 6.13.

TABLE 6.13 Home Values and Other Factors

SOLUTION: We estimate three models to examine the
multicollinearity issue; Table 6.14 presents the regression
results.

TABLE 6.14 Summary of Model Estimates (Example 6.8)

Variable Model 1 Model 2 Model 3

Intercept 417,892.04* 348,187.14* 285,604.08

(0.001) (0.002) (0.083)

HH Income 9.04* 7.74* NA

(0.000) (0.000)

Per Cap Inc –3.27 NA 13.21*

(0.309) (0.000)

Pct Owner Occ –8,744.30* –8,027.90* –6,454.08*

(0.000) (0.000) (0.001)

Adjusted R2 0.8071 0.8069 0.6621

Notes: The table contains parameter estimates with p-values in parentheses; *

represents significance at the 5% level. NA denotes not applicable. Adjusted R2,
reported in the last row, is used for model selection.
 Model 1 uses all three predictor variables to explain home
values. Surprisingly, the per capita income variable has a
negative estimated coefficient of −3.27 and, with a p-value of
0.31, is not even statistically significant at the 5% level.
Multicollinearity might be the reason for this surprising result
because household income and per capita income are likely to
be correlated. We compute the sample correlation coefficient
between these two variables as 0.8582, which suggests that
multicollinearity is severe. We estimate two more models
where one of these collinear variables is removed; Model 2
removes per capita income and Model 3 removes household
income. Note that per capita income in Model 3 now exerts a
positive and significant influence on home values. Between
these two models, Model 2 is preferred to Model 3 because of
its higher adjusted R2 (0.8069 > 0.6621). The page 246
choice between Model 1 and Model 2 is unclear.
In general, Model 1, with the highest adjusted R2 value of
0.8071, is preferred if the sole purpose of the analysis is to
make predictions. However, if the coefficient estimates need to
be evaluated, then Model 2 may be the preferred choice.

Remedy
Inexperienced researchers tend to include too many predictor
variables in their quest not to omit anything important and, in doing
so, may include redundant variables that essentially measure the
same thing. When confronted with multicollinearity, a good remedy is
to drop one of the collinear variables. The difficult part is to decide
which of the collinear variables is redundant and, therefore, can
safely be removed. Another option is to obtain more data because
the sample correlation may get weaker as we include more
observations. Sometimes it helps to express the predictor variables
differently so that they are not collinear. At times, the best approach
may be to do nothing when there is a justification to include all
predictor variables. This is especially so if the estimated model yields
a high R2, which implies that the estimated model is good for making
predictions.

Common Violation 3: Changing Variability

The assumption of constant variability of observations often breaks
down in studies with cross-sectional data. Consider the model y = β0
+ β1x + ɛ, where y is a household’s consumption expenditure and x
is its disposable income. It may be unreasonable to assume that the
variability of consumption is the same across a cross-section of
household incomes. For example, we would expect higher-income
households to have a higher variability in consumption as compared
to lower-income households. Similarly, home prices tend to vary
more as homes get larger, and sales tend to vary more as firm size
increases.
In the presence of changing variability, the OLS estimators are
still unbiased. However, the estimated standard errors of the OLS
estimators are inappropriate. Consequently, we cannot put much
faith in the standard t or F tests because they are based on these
estimated standard errors.

Detection
We can use residual plots to gauge changing variability. The
residuals are generally plotted against each predictor variable xj; for
a multiple regression model, we can also plot them against the
predicted value . There is no violation if the residuals are randomly

dispersed across the values of xj. On the other hand, there is a

violation if the variability increases or decreases over the values of
xj.
FILE
Convenience_Stores
EXAMPLE 6.9
Consider a simple regression model that relates monthly sales
(Sales in $1,000s) from a chain of convenience stores with the
square footage (Sqft) of the store. A portion of the data used
for the analysis is shown in Table 6.15. Estimate the model and
use a residual plot to determine if the observations have a
changing variability.

TABLE 6.15 Sales and Square Footage of Convenience Stores

Sales Sqft

140 1810

160 2500

⋮ ⋮

110 1470

page 247

SOLUTION: The sample regression is given by

where we have put the standard errors of the coefficients in

parentheses. We will reference these values when we
incorporate R instructions at the end of this section.
 A residual plot of the estimated model is shown in Figure
6.10. Note that the residuals seem to fan out across the
horizontal axis. Therefore, we conclude that changing
variability is a likely problem in this application relating sales to
square footage. This result is not surprising because you would
expect sales to vary more as square footage increases. For
instance, a small convenience store is likely to include only
bare essentials for which there is a fairly stable demand. A
larger store, on the other hand, may include specialty items,
resulting in more fluctuation in sales.

FIGURE 6.10 Residual plot against square footage (Example 6.9)

Remedy
As mentioned earlier, in the presence of changing variability, the
OLS estimators are unbiased, but their estimated standard errors are
inappropriate. Therefore, OLS still provides reasonable coefficient
estimates, but the t and the F tests are no longer valid. This has
prompted some researchers to use the OLS estimates along with a
correction for the standard errors, often referred to as robust
standard errors. Unfortunately, the current version of Excel does not
include a correction for the standard errors. However, it is available
on many statistical computer packages, including R. At the end of
this section, we use R to make the necessary correction in Example
6.9. With robust standard errors, we can then perform legitimate t-
tests.

Common Violation 4: Correlated

Observations
When obtaining the OLS estimators, we assume that the
observations are uncorrelated. This assumption often breaks down
in studies with time series data. Variables such as sales,
employment, and asset returns exhibit business cycles. As a
consequence, successive observations are likely to be correlated.
In the presence of correlated observations, the OLS estimators
are unbiased, but their estimated standard errors are inappropriate.
Generally, these standard errors are distorted downward, making the
model look better than it really is with a spuriously high R2.
Furthermore, the t and F tests may suggest that the predictor
variables are individually and jointly significant when this is not true.

Detection
We can plot the residuals sequentially over time to look for
correlated observations. If there is no violation, then the residuals
should show no pattern around the horizontal axis. A violation is
indicated when a positive residual in one period is followed by
positive residuals in the next few periods, followed by page 248
negative residuals for a few periods, then positive
residuals, and so on. Although not as common, a violation is also
indicated when a positive residual is followed by a negative residual,
then a positive residual, and so on.
FILE
Sushi_Restaurant

EXAMPLE 6.10
Consider y = β0 + β1x1 + β2x2 + ɛ, where y represents sales (in
$1,000s) at a sushi restaurant and x1 and x2 represent
advertising costs (in $) and the unemployment rate (in %),
respectively. A portion of monthly data from January 2018 to
May 2019 is given in Table 6.16. Inspect the behavior of the
residuals to comment on serial correlation.

TABLE 6.16 Sales, Advertising Costs, and Unemployment Data

SOLUTION: The model is estimated as

where we have put the standard errors of the coefficients in

parentheses. We will reference these values when we
incorporate R instructions at the end of this section.
In order to detect serial correlation, we plot the residuals
sequentially against time t, where t is given by 1, 2, . . . , 17 for
the 17 months of time series data. Figure 6.11 shows a
wavelike movement in the residuals over time, first clustering
below the horizontal axis, then above the horizontal axis, and
so on. Given this pattern around the horizontal axis, we
conclude that the observations are correlated.

FIGURE 6.11 Scatterplot of residuals against time t

Remedy
As mentioned earlier, in the presence of correlated observations, the
OLS estimators are unbiased, but their standard errors are
inappropriate and generally distorted downward, making the model
look better than it really is. Therefore, OLS still provides reasonable
coefficient estimates, but the t and the F tests are no longer valid.
This has prompted some researchers to use the OLS estimates but
correct the standard errors, often referred to as robust page 249
standard errors. As in the case of changing variability,
the current version of Excel does not include this correction.
However, it is available on many statistical computer packages,
including R. At the end of this section, we use R to make the
necessary correction in Example 6.10. We can perform legitimate t-
tests once the standard errors have been corrected.

Common Violation 5: Excluded Variables

Another crucial assumption in a linear regression model is that the
error term is not correlated with the predictor variables. In general,
this assumption breaks down when important predictor variables are
excluded. If one or more of the relevant predictor variables are
excluded, then the resulting OLS estimators are biased. The extent
of the bias depends on the degree of the correlation between the
included and the excluded predictor variables—the higher the
correlation, the greater the bias.
Suppose we want to estimate y = β0 + β1x + ɛ, where y is salary
and x is years of education. This model excludes innate ability, which
is an important ingredient for salary. Note that the omitted innate
ability is correlated with years of education. Now consider someone
who is highly educated and also commands a high salary. The model
will associate high salary with education, when, in fact, it may be the
person’s unobserved high level of innate ability that has raised both
education and salary. In sum, this violation leads to unreliable
coefficient estimates; some estimates may even have the wrong
signs.

Remedy
It is important that we include all relevant predictor variables in the
regression model. An important first step before running a regression
model is to compile a comprehensive list of potential predictor
variables. We can then build down to perhaps a smaller list of
predictor variables using the adjusted R2 criterion. Sometimes, due
to data limitations, we are unable to include all relevant variables.
For example, innate ability may be an important predictor variable for
a model that explains salary, but we are unable to include it because
innate ability is not observable. In such instances, we use a
technique called the instrumental variable technique, which is
outside the scope of this text.

Summary
It takes practice to become an effective user of the regression
methodology. We should think of regression modeling as an iterative
process. We start with a clear understanding of what the regression
model is supposed to do. We define the relevant response variable
and compile a comprehensive list of potential predictor variables.
The emphasis should be to pick a model that makes economic and
intuitive sense and avoid predictor variables that more or less
measure the same thing, thus causing multicollinearity. We then
apply this model to data and refine and improve its fit. Specifically,
from the comprehensive list, we build down to perhaps a smaller list
of predictor variables using significance tests and goodness-of-fit
measures such as the standard error of the estimate and the
adjusted R2. It is important that we explore residual plots to look for
signs of changing variability and correlated observations in cross-
sectional and time series studies, respectively. If we identify any of
these two violations, we can still trust the point estimates of the
regression coefficients. However, we cannot place much faith in the
standard t or F tests of significance unless we employ the necessary
correction.

Using Excel and R to Construct Residual

Plots
We first replicate Figure 6.10 from Example 6.9 in Excel and R.
Because we found that changing variability was an issue in Example
6.9, we show how to calculate robust standard errors in R.
We then replicate Figure 6.11 from Example 6.10. Because we
found that correlated observations were an issue in Example 6.10,
we show how to calculate robust standard errors in R.

page 250
FILE
Convenience_Stores

Using Excel to Replicate Figure 6.10

A. Open the Convenience_Stores data file.
B. Choose Data > Data Analysis > Regression from the menu.
C. For Input Y Range, select the Sales data, and for Input X Range, select the Sqft data.
D. Select Residual Plots.
E. Click OK. You should see a graph that resembles Figure 6.10. Formatting (regarding
colors, axes, etc.) can be done by selecting Format from the menu.
Using R to Replicate Figure 6.10 and Obtain Robust
Standard Errors
A. Import the Convenience_Stores data into a data frame (table) and label it myData.
B. Install and load the sandwich package. We use this package to calculate robust standard
errors. Enter:
> install.packages(“sandwich”)
> library(sandwich)
C. Use the lm function to create a linear regression model (object), labeled Simple. Enter:
> Simple <- lm(Sales~Sqft, data=myData)
D. Use the resid function to obtain the residuals from the simple linear regression model,
labeled Simple_Residuals. Enter:
> Simple_Residuals <- resid(Simple)
E. Use the plot function to create a scatterplot of the residuals against the predictor
variable, Sqft. We also use the xlab and ylab options to add labels to the x- and y-axes.
Enter:
> plot(Simple_Residuals ~ myData$Sqft, xlab = “Sqft”, ylab = “e”)
The scatterplot that R returns should resemble Figure 6.10.
F. Use the vcovHC function in the sandwich package to calculate robust standard errors
for the OLS estimators. By using the option type = “HC1”, we are asking R to apply a
widely used formula for calculating robust standard errors (other designations within type
are available in R). Enter:
> vcovHC(Simple, type=“HC1”)
R returns:

(Intercept) Sqft
(Intercept) 91.45127988 -6.095285e-02
Sqft -0.06095285 4.149344e-05

The above output represents the variance-covariance matrix, where the diagonal
elements contain the variances and the off-diagonal elements contain the covariances of
the OLS estimators. Because we are interested in the standard errors, we simply take
the square roots of the diagonal values of the matrix (see values in boldface). In order to
find the standard errors, labeled as Simple_SE, enter:
> Simple_SE <- diag(vcovHC(Simple,type=“HC1”))^0.5 > Simple_SE

(Intercept) Sqft
9.56301625 0.00644154
 The corrected standard errors for the intercept and Sqft are 9.5630 and 0.0064,
respectively. Recall from Example 6.9 that the OLS-generated standard errors for the
intercept and Sqft were 10.4764 and 0.0057, respectively. So now we can easily
compute the t-test of significance using the OLS estimates with the corrected standard
errors.

page 251

Using Excel to Replicate Figure 6.11

FILE
Sushi_Restaurant

A. Open the Sushi_Restaurant data file.

B. Choose Data > Data Analysis > Regression from the menu.
C. For Input Y Range, select the Sales data, and for Input X Range, simultaneously select
the AdsCost and Unemp data.
D. Select Residuals. Click OK.
E. Given the regression output, select the residual data and choose Insert > Scatter;
choose the option on the top left. (If you are having trouble finding this option after
selecting Insert, look for the graph with data points above Charts.) You should see a
graph that resembles Figure 6.11. Formatting (regarding colors, axes, etc.) can be done
by selecting Format from the menu.

Using R to Replicate Figure 6.11 and Obtain Robust

Standard Errors
A. Follow steps A through D from the instructions for Using R to Replicate Figure 6.10,
except import the Sushi_Restaurant data file and create a linear regression model
(object), labeled Multiple, which expresses Sales as a function of AdsCost and Unemp.
Retrieve the residuals and label them as Multiple_Residuals.
B. First use the seq function to create a time variable, labeled as T, that has the same
number of observations as Multiple_Residuals; then use the plot function. We also use
the abline function to insert a line at the x-axis. Enter:
> T <- seq(from = 1, to = length(Multiple_Residuals))
> plot(Multiple_Residuals ~ T, xlab = “time”, ylab = “e”)
> abline(h = 0)
The scatterplot that R returns should resemble Figure 6.11.
C. Use the NeweyWest function in the sandwich package to calculate robust standard
errors for the OLS estimators. By using the option prewhite=FALSE, we are asking for
the most basic correction (many other options are available in R). Enter:
> NeweyWest(Multiple, prewhite=FALSE)
R returns:
 (Intercept) AdsCost Unemp
(Intercept) 18.10230474 -2.342228e-02 -5.592667e-01
AdsCost 18.10230474 -2.342228e-02 -5.592667e-01
Unemp -0.55926668 -9.819913e-05 9.036499e-02

The output represents the variance-covariance matrix, where the diagonal elements
contain the variances and the off-diagonal elements contain the covariances of the OLS
estimators. Because we are interested in the standard errors, we simply take the square
roots of the diagonal values of the matrix (see values in boldface). In order to find the
standard errors, labeled as Multiple_SE, enter:
> Multiple_SE <- diag(NeweyWest(Multiple))^0.5
> Multiple_SE
(Intercept) AdsCost Unemp
4.773375748 0.006961572 0.354646071
The corrected standard errors for the intercept, AdsCost, and Unemp are 4.7734,
0.0070, and 0.3546, respectively. Recall from Example 6.10 that the OLS-generated
standard errors for the intercept, AdsCost, and Unemp were 3.9817, 0.0068, and 0.2997,
respectively. The corrected standard errors are all higher than the OLS estimates, which
is typically what we expect when correlated observations are an issue in a regression
model. We can now easily compute the t-test of significance using the OLS estimates
with the corrected standard errors.

page 252

EXERCISES 6.4
Mechanics
51. Using 20 observations, the multiple regression model y = β0 + β1x1 + β2x2 + ɛ
was estimated. A portion of the regression results is as follows:
 a. At the 5% significance level, are the predictor variables jointly significant?
b. At the 5% significance level, is each predictor variable individually significant?
c. What is the likely problem with this model?
52. A simple linear regression, y = β0 + β1x + ɛ, is estimated with cross-sectional
data. The resulting residuals e along with the values of the predictor variable x are
shown in the accompanying table.

a. Graph the residuals e against the values of the predictor

variable x and look for any discernible pattern.
b. Which assumption is being violated? Discuss its
consequences and suggest a possible remedy.
53. A simple linear regression, y = β0 + β1x + ɛ, is estimated with time series data.
The resulting residuals e and the time variable t are shown in the accompanying
table.

a. Graph the residuals against time and look for any discernible pattern.
b. Which assumption is being violated? Discuss its consequences and suggest a
possible remedy.
 Applications
54. FILE Television. Like books and stories, television not only entertains, it also
exposes a child to new information about the world. While watching too much
television is harmful, a little bit may actually help. Researcher Matt Castle gathers
information on the grade point average (GPA) of 28 middle-school children and
the number of hours of television they watched per week. Examine the linearity
assumption in the regression model GPA = β0 + β1Hours + ɛ.
55. FILE Delivery. Quick2U, a delivery company, would like to standardize its
delivery charge model for shipments (Charge in $) such that customers will better
understand their delivery costs. Three predictor variables are used: (1) distance
(in miles), (2) shipment weight (in lbs), and (3) number of boxes. A sample of 30
recent deliveries is collected; a portion of the data is shown in the accompanying
table.

Charge Distance Weight Boxes

92.50 29 183 1

157.60 96 135 3

⋮ ⋮ ⋮ ⋮

143.00 47 117 7

a. Estimate the model Charge = β0 + β1Distance + β2Weight +

β3Boxes + ɛ and examine the joint and individual
significance of the predictor variables at the 1% level.
b. Is there any evidence of multicollinearity?
c. Graph the residuals against the predicted values and
determine if there is any evidence of changing variability.
56. Consider the results of a survey where students were asked about their GPA and
also to break down their typical 24-hour day into study, leisure (including work),
and sleep. Consider the model GPA = β0 + β1Study + β2Leisure + β3Sleep + ɛ.

a. What is wrong with this model?

b. Suggest a simple way to reformulate the model.
57. FILE AnnArbor_Rental. Consider the monthly rent (Rent in $) of a home in Ann
Arbor, Michigan, as a function of the number of bedrooms (Beds), the number of
bathrooms (Baths), and square footage (Sqft).
a. Estimate Rent = β0 + β1Beds + β2Baths + β3Sqft + ɛ.
b. Which of the predictor variables might cause changing
variability? Explain.
c. Use residual plots to verify your economic intuition.
58. FILE Work_Experience. Consider the data on salary (in $) and work experience
(in years) of 100 employees in a marketing firm. Estimate the model Salary = β0 +
β1Experience + ɛ.

a. Explain why you would be concerned about changing

variability in this application.
b. Use a residual plot to confirm your economic intuition.

59. FILE Healthy_Living. Healthy living has always been an important goal for any
society. Consider a regression model that conjectures that fruits and vegetables
(FV) and regular exercising have a positive effect on health and smoking has a
negative effect on health. The sample consists of the percentage
of these variables observed in various states in the United
page 253
States. A portion of the data is shown in the accompanying table.

a. Estimate the model Healthy = β0 + β1FV + β2Exercise +

β3Smoke + ɛ.
b. Analyze the data to determine if multicollinearity and
changing variability are present.
60. FILE J&J. A capital asset pricing model (CAPM) for Johnson & Johnson (J&J)
was discussed in Example 6.6. The model uses the risk-adjusted stock return R −
Rf for J&J as the response variable and the risk-adjusted market return RM − Rf
as the predictor variable. Because serial correlation may occur with time series
data, it is prudent to inspect the behavior of the residuals. Construct a scatterplot
of the residuals against time to comment on correlated observations.
61. FILE Consumption. The consumption function is one of the key relationships in
economics, where consumption is a function of disposable income. The
 accompanying table shows a portion of quarterly data for average U.S. annual
consumption (Consumption in $) and disposable income (Income in $) for the
years 2000–2016.

Date Consumption Income

Q1, 2000 28634 31192

Q2, 2000 28837 31438

⋮ ⋮ ⋮

Q4, 2016 35987 39254

a. Estimate Consumption = β0 + β1Income + ɛ. Plot the

residuals against time to determine if there is a possibility of
correlated observations.
b. Discuss the consequences of correlated observations and
suggest a possible remedy.
62. FILE Mowers. The marketing manager at Turfco, a lawn mower company,
believes that monthly sales across all outlets (stores, online, etc.) are influenced
by three key variables: (1) outdoor temperature (in °F), (2) advertising
expenditures (in $1,000s), and (3) promotional discounts (in %). A portion of the
monthly sales data for the past two years is shown in the accompanying table.

Sales Temperature Advertising Discount

17235 33 15 5.0

19854 42 25 5.0

⋮ ⋮ ⋮ ⋮

22571 44 21 5.0

a. Estimate Sales = β0 + β1Temperature + β2Advertising +

β3Discount + ɛ, and test for the joint and individual
significance of the predictor variables at the 5% level.
b. Examine the data for evidence of multicollinearity. Provide
two reasons why it might be best to do nothing about
multicollinearity in this application.
c. Examine the residual plots for evidence of changing
variability.

6.5 Writing With Big Data

When using regression analysis with big data, we often do not place much value in
tests of significance. Why is this the case? It turns out that if the sample size is
sufficiently large, then the relationship between each predictor variable and the
response variable will be statistically significant even when the relationship is not
economically meaningful.
What occurs in these instances is that there is little difference in the estimates of
βj generated by different large random samples. Recall that we use se(bj) to gauge
the variability in bj, and this variability depends on two factors: (1) how closely the
members of the population mirror the relationship between xj and y that is implied by
βj, and (2) the size of the sample on which the value of the estimator bj is based. If
the sample size is sufficiently large, then the variability virtually disappears, or,
equivalently, se(bj) approaches zero. As se(bj) approaches zero, the value of the tdf
test statistic (calculated as tdf = bj /se(bj)) increases, leading to a small p-value, and
thus rejection of the null hypothesis of insignificance.

page 254
Therefore, if the sample size is sufficiently large, statistical significance does not
necessarily imply that a relationship is economically meaningful. It is for this reason
that when confronted with big data and assessing various models, we tend to rely on
economic intuition and model validation rather than tests of significance. In Chapters
7 through 10, we will rely on cross-validation techniques in order to assess various
models.

FILE
House_Price
Case Study
Develop a predictive model for the price of a house in the college town of Ames,
Iowa. Before evaluating various models, you first have to filter out the House_Price
data to get the appropriate subset of observations for selected variables. After you
have obtained the preferred model, summarize your findings as well as predict the
price of a house in Ames, Iowa, given typical values of the predictor variables.

Sample Report—Investing In College Town

Real Estate
Investing in college town real estate can be a smart move. First, students
offer a steady stream of rental demand as many cash-strapped public
universities are unable to house their students beyond freshman year.
Second, this demand is projected to grow. The National Center for Education
Statistics predicts that college enrollment in the U.S. will reach 19.8 million
students by 2025, an increase of 14% from its 2014 enrollment of 17.3
million.
A regression analysis is conducted to determine the factors that influence
the sale price of a single-family house in Ames, Iowa—home to Iowa State
University. For a sample of 209 single-family houses in 2016, the following
data are collected: the house’s sale price (Price in $), the number of
bedrooms (Beds), the number of bathrooms (Baths), the square footage
(Sqft), and the lot size (LSize in square feet). Table 6.17 shows the mean
values of the relevant variables for newer houses (those built in 2000 or
after), old houses (those built prior to 2000), and for all houses in the sample.
Median values are shown in parentheses.

Shutterstock/Dmytro Zinkevych
TABLE 6.17 The Mean (Median) of Variables for New, Old, and All
Houses

New Old All

Variables
Houses Houses Houses

Price 326,134 209,552 230,191

(292,000) (190,500) (215,000)

Beds 3.68 3.22 3.30

(4.00) (3.00) (3.00)

Baths 3.06 2.20 2.36

(3.00) (2.00) (2.00)

Sqft 1,867 1,596 1,644

(1,691) (1,444) (1,515)

LSize 18,137 19,464 19,229

(10,361) (10,123) (10,171)

Number of 37 172 209

Observations

The average sale price for the newer houses is substantially more than
that for the older houses. For all houses, given that the mean is higher than
the median, the house price distribution is positively skewed, indicating that a
few expensive houses have pulled up the mean above the median. The
square footage and the lot size are also positively skewed. Finally, relatively
newer houses have more bedrooms, bathrooms, and square footage but a
smaller lot size. This is consistent with a 2017 article in Building magazine
that found that newer houses have become 24% bigger over the past 15
years, while lot sizes have shrunk 16%.
 page 255
In order to analyze the factors that may influence the price of a house, the
following linear regression model with all the referenced predictor variables is
considered:

where New is a dummy variable that equals 1 if the house was built in 2000
or after, 0 otherwise. It is expected that Beds, Bath, Sqft, and Lsize will have
a positive relationship with Price; that is, a house with more bedrooms and
bathrooms is expected to obtain a higher price than one with fewer bedrooms
and bathrooms. Similarly, a bigger house, or one on a bigger lot, is expected
to obtain a higher price. A newer house, one with all the latest updates, is
expected to obtain a higher price as compared to an older house in need of
work. Column 2 of Table 6.18 shows the regression results from estimating
this complete model.

TABLE 6.18 Estimates of Alternative Regression Models to Predict

House Price, n = 209

Variables Complete Model Restricted Model

Intercept 95.82 5,815.38

(0.996) (0.672)

Beds 3,124.63 NA
(0.623)

Baths 30,971.41* 31,985.61*

(0.000) (0.000)

Sqft 76.98* 78.36*

(0.000) (0.000)
 Variables Complete Model Restricted Model

Lsize 0.43* 0.43*

(0.000) (0.000)

New 68,248.89* 68,415.54*

(0.000) (0.000)

se 64,612.08 64,492.0062

R2 0.6689 0.6685

Adjusted R2 0.6608 0.6620

F-stat (p- 82.03* (0.000) 102.86* (0.000)

value)

Notes: Parameter estimates are in the top half of the table with the p-values
in parentheses; * represents significance at the 5% level. NA denotes not
applicable. The lower part of the table contains goodness-of-fit measures.

All predictor variables with the exception of Beds are correctly signed and
statistically significant. Perhaps the lack of significance of Beds is due to
multicollinearity because the number of bedrooms is likely to be correlated
with the number of bathrooms as well as square footage. An alternative
explanation might be that additional bedrooms add value only in houses with
large square footage. For comparison, a restricted model is estimated that
omits Beds from the list of predictor variables; see Column 3 of Table 6.18 for
the results.
The following observations are made:

The restricted model is preferred because it has the

lower standard error of the estimate se and the higher
adjusted R2.
 Holding other factors constant, an additional bathroom
adds about $31,986 in value. Similarly, a 100-square-
foot increase in a house adds $7,836 in value and a
1,000-square-foot increase in the lot size adds $430 in
value. Finally, there is a premium of $68,416 for
relatively newer houses.
The coefficient of determination reveals that 66.85% of
the variability in sale price is explained by the predictor
variables, implying that approximately 33.15% is
unexplained. This is not surprising because other
factors, such as the condition of the house or its
proximity to nearby amenities, are likely to influence the
sales price.
Suppose a 1,600-square-foot house with two bathrooms
sits on a 15,000-square-foot lot. Given the preferred
model, its predicted sale price is $269,969 for a
relatively newer house and $201,553 for an older house.

page 256

Suggested Case Studies

Many different regression models can be estimated and assessed with the big data
that accompany this text. Here are some suggestions.

Report 6.1 FILE House_Price. Choose two comparable college towns. Find the
model that best predicts the sale price of a house. Make sure to include a dummy
variable that accounts for the possibility of differences in the sale price due to the
location.

Report 6.2 FILE College_Admissions. Choose a college of interest and use the
sample of enrolled students to best predict a student’s college grade point average.
Use goodness-of-fit measures to find the best predictive model. In order to estimate
these models, you have to first filter the data to include only the enrolled students.

Report 6.3 FILE NBA. Develop and compare two models for predicting a player’s
salary based on offensive performance measures and defensive performance
measures, respectively. In order to estimate these models, you have to first filter the
data to include only career statistics based on regular seasons. Exclude players with
no information on salary.

Report 6.4 FILE Longitudinal_Surveys. Develop a linear regression model for

predicting the weight of a respondent in 1981. Use goodness-of-fit measures to find
the appropriate predictor variables. In order to estimate these models, you have to
first filter the data to exclude missing observations.

Report 6.5 FILE TechSales_Reps. Develop a linear regression model for predicting
the salary of a sales representative for both software and hardware industries.
Create dummy variables to capture personality type and discuss their role, along with
the other relevant predictor variables, in predicting salary.

page 257
 page 258

7 Advanced Regression
Analysis

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 7.1  Estimate and interpret regression models
with interaction variables.
LO 7.2  Estimate and interpret nonlinear
regression models.
LO 7.3  Estimate and interpret linear probability
and logistic regression models.
LO 7.4  Use cross-validation techniques to
evaluate regression models.

n Chapter 6, we discussed linear regression models where the

Ipartial effect of a predictor variable on the response variable was
constant and did not depend on other predictor variables in the
model. For example, when evaluating the price of a house, we
assumed that an additional bedroom resulted in the same increase
in the price of a house regardless of its square footage.
In this chapter, we estimate and interpret regression models with
interaction variables. These variables allow the partial effect of a
predictor variable to depend on the value of the other predictor
variables. Continuing with the house price example, an additional
bedroom may have a higher impact on a house price with a larger
square footage. We then turn our attention to regression models for
nonlinear relationships. These models are useful when the
relationship between the predictor variable and the response
variable cannot be represented by a straight line. In the house price
example, an additional bedroom may have a higher impact on the
price of a house when going from one to two bedrooms than from
four to five bedrooms.
So far, we have used regression models to predict the value of a
numerical response variable. We extend the analysis to estimate and
interpret the linear and the logistic regression models for predicting
the value of a binary or dummy response variable. For example, we
may want to predict the probability that mortgage applicants will
default on their payments, based on predictor variables such as the
loan-to-value ratio, income level of the borrower, etc. Finally, we
employ cross-validation techniques for regression models where we
partition the original sample into a training set to train the model and
a validation set to evaluate the model.

page 259
©JrCasas/Shutterstock

INTRODUCTORY CASE
Gender Gap in Manager Salaries
The salary difference between men and women has shrunk
over the years, particularly among younger workers, but it still
persists. According to a Pew Research Center analysis,
women earned 82% of what men earned in 2017
(http://www.pewresearch.org, April 9, 2018). Ara Lily is
completing her MBA degree from Bentley University, located
just outside Boston. She is upset that the gender gap in
salaries continues to exist in the American workplace. For a
class project, she decides to analyze the gender gap in salaries
of project managers. Ara gains access to the salary (in
$1,000s) for 200 project managers in small- to middle-sized
 firms in the Boston area. In addition, she has data on the firm
size, the manager’s experience (in years), whether the
manager is a female (Female equals 1 if female, 0 otherwise),
and whether the manager has a graduate degree (Grad equals
1 if graduate degree, 0 otherwise). Table 7.1 shows a portion of
the data.
FILE
Gender_Gap

TABLE 7.1 Salary and Other Information on Project Managers (n = 200)

Ara would like to use the information in Table 7.1 to:

1. Analyze the determinants of a project manager’s salary.
2. Estimate and interpret a regression model with relevant
interaction variables.
3. Determine whether there is evidence of a gender gap in
salaries.
A synopsis of this case is provided at the end of Section 7.1.

page 260

7.1 REGRESSION MODELS WITH

INTERACTION VARIABLES
 LO 7.1
Estimate and interpret regression models with interaction variables.

As discussed in Chapter 6, regression analysis is one of the most

widely used techniques in predictive analytics for predicting the value
of the response (target) variable on the basis of several predictor
variables; other techniques are discussed in Chapters 8, 9, and 10.
Recall that the linear regression model is specified as y = β0 + β1x1 +
β2x2 + . . . + βkxk + ε, where y is the response variable, x1, x2, . . . , xk
are the k predictor variables, and ε is the random error term. The
coefficients β0, β1, β2, . . . , βk are the unknown parameters to be
estimated. We use the method of ordinary least squares to derive
the sample regression equation as where

is the predicted value of the response variable given specified values

of the predictor variables. For each predictor variable xj (j = 1, . . . ,
k), the corresponding slope coefficient bj measures the change in the

predicted value of the response variable given a unit increase in

the associated predictor variable xj, holding all other predictor

variables constant. In other words, bj represents the partial (or

marginal) effect of xj on (Note that the partial effect is simply the

partial derivative of with respect to xj.)

It is important to note that the partial effect of a predictor variable

in the above regression model does not depend on the values of any
other predictor variable. Sometimes, it is natural for the partial effect
to depend on another predictor variable. For example, when
evaluating the price of a house, it is commonly assumed that an
additional bedroom results in the same increase in the house price
regardless of the square footage of the house. However, this
assumption may be unrealistic because, for larger houses, an
additional bedroom often results in a higher increase in house prices.
In other words, there is an interaction effect between the number of
bedrooms and the square footage of the house. We capture this
effect by incorporating interaction variables in the regression model.

INTERACTION EFFECT IN A REGRESSION MODEL

The interaction effect in a regression model occurs when the
partial effect of a predictor variable on the response variable
depends on the value of another predictor variable.

Consider a regression model with two predictor variables: y = β0

+ β1x1 + β2x2 + ε. (The analysis can easily be extended to include
more predictor variables.) The interaction variable x1x2 is defined as
a product of the two predictor variables and is included in the
regression as: y = β0 + β1x1 + β2x2 + β3x1x2 + ε. This regression

model is estimated as Note that the partial

effect of x1 on is now given by b1 + b3x2, which depends on the

value of x2. Similarly, the partial effect of x2 on given by b2 + b3x1,

depends on the value of x1.

As discussed in Chapter 6, the predictor variables can be both
numerical and dummy variables, denoted by x and d, respectively.
Recall that we use a dummy variable to describe a categorical
variable with two categories. We will consider three types of
interaction variables: (a) the interaction of two dummy variables, (b)
the interaction of a dummy variable with a numerical variable, and (c)
the interaction of two numerical variables.

The Interaction of Two Dummy Variables

Consider a regression model with two dummy variables, d1 and d2,
along with an interaction variable d1d2: y = β0 + β1d1 + β2d2 + β3d1d2

+ ε. This model is estimated as The partial

page 261
effect of d1 on given by b1 + b3d2, equals b1 if d2 = 0

and b1 + b3 if d2 = 1. Similarly the partial effect of d2 on is b2 +

b3d1, which depends on the value of d1. Models with the interaction
variables are easy to estimate. In addition, tests of significance can
be conducted on all variables, including the interaction variable.
Example 7.1 illustrates the interaction between two dummy
variables.

EXAMPLE 7.1
According to the National Association of Colleges and
Employers, in 2016 the top-paid business graduates at the
bachelor’s degree level were expected to be those with a
concentration in management information systems (MIS). This
was due, in large part, to the concentration’s linkage with the
exploding field of data analytics. At a University of California
campus, data were collected on the starting salary of business
graduates (Salary in $1,000s) along with their cumulative GPA,
whether they have an MIS concentration (MIS = 1 if yes, 0
otherwise), and whether they have a statistics minor (Statistics
= 1 if yes, 0 otherwise). A portion of the data is shown in Table
7.2.
FILE
Salary_MIS

TABLE 7.2 Starting Salary of Business Majors (n =120)

 Salary GPA MIS Statistics

72 3.53 1 0

66 2.86 1 0

⋮ ⋮ ⋮ ⋮

66 3.65 0 0

a. Estimate and interpret the effect of GPA, MIS, and Statistics

on Salary. Predict the salary of a business graduate with and
without the MIS concentration and the statistics minor. Use a
GPA of 3.5 for making predictions.
b. Extend the model from part a to include the interaction
between MIS and Statistics. Predict the salary of a business
graduate with and without the MIS concentration and the
statistics minor. Use a GPA of 3.5 for making predictions.
SOLUTION:
Table 7.3 shows the regression results for part a (see Model 1)
and for part b (see Model 2).

TABLE 7.3 Regression Results of the Starting Salary of Business Graduates

Variable Model 1 Model 2

Constant 44.0073* 44.0993*

(0.000) (0.000)

GPA 6.6227* 6.7109*

(0.000) (0.000)
 Variable Model 1 Model 2

MIS 6.6071* 5.3250*

(0.000) (0.000)

Statistics 6.7309* 5.5350*

(0.000) (0.000)

MIS × Statistics NA 3.4915*

(0.004)

Adjusted R2 0.7901 0.8029

Notes: Parameter estimates are followed with the p-values in parentheses; NA

denotes not applicable; * represents significance at the 5% level.

page 262

a. We refer to the second column of Table 7.3 to derive the

estimated equation as

Note that with

the positive coefficients and the p-values of approximately 0,

all predictor variables exert a positive and significant influence
on the starting salary of business graduates. The MIS
concentration and a statistics minor are predicted to fetch an
additional starting salary of $6,607 and $6,731, respectively.
For a GPA of 3.5, we compute the predicted salary for a
business graduate with neither an MIS concentration nor a
Statistics minor as 44.0073 + 6.6227 × 3.5 + 6.6071 × 0 +
6.7309 × 0 = $67,187. For a GPA of 3.5, the predicted salaries
for business graduates are $73,794 with an MIS concentration
 only, $73,918 with a Statistics minor only, and $80,525 with
both an MIS concentration and a Statistics minor.
b. To estimate Model 2, we need to create a data column for the
interaction variable MIS × Statistics. The entries for the new
column are computed by multiplying the column entries of MIS
and Statistics. For example, the first entry is computed as 1 ×
0 = 0; other entries are computed similarly. We refer to the
third column of Table 7.3 to derive the estimated equation as

Like the earlier predictor variables, the

interaction variable is also positive and statistically significant

at the 5% level. All else constant, the slope coefficient
attached to the interaction variable implies that the value of an
MIS concentration is enhanced by about $3,492 when it is
accompanied with a Statistics minor. Likewise, the value of a
Statistics minor is enhanced by $3,492 when accompanied
with an MIS concentration. Model 2 provides a better fit
because it has a higher value of the adjusted R2 than Model 1
(0.8029 > 0.7901). For a GPA of 3.5, we compute the
predicted salary for a business graduate who has neither an
MIS concentration nor a Statistics minor as 44.0993 + 6.7109
× 3.5 + 5.3250 × 0 + 5.5350 × 0 + 3.4915(0 × 0) = $67,588.
For a GPA of 3.5, the predicted salaries for business
graduates are $72,913 with an MIS concentration only,
$73,123 with a Statistics minor only, and $81,939 with both an
MIS concentration and a Statistics minor.

The Interaction of a Dummy Variable and a

Numerical Variable
Consider a regression model with a numerical variable x and a
dummy variable d along with an interaction variable xd: y = β0 + β1x

+ β2d + β3xd + ε. This model is estimated as

The partial effect of x on given by b1 + b3d, equals b1 if d = 0 and

b1 + b3 if d = 1. The partial effect of d on given by b2 + b3x, is

difficult to interpret because it depends on the value of the numerical

variable x. It is common to interpret this partial effect at the variable’s
sample mean (or some other value of interest). Therefore, at the

average value of x, the partial effect of d on is equal to The

next example illustrates the interaction between a dummy variable

and a numerical variable.

EXAMPLE 7.2
Important risk factors for high blood pressure reported by the
National Institute of Health include weight and ethnicity. High
blood pressure is common in adults who are overweight and
are African American. According to the American Heart
Association, the systolic pressure (top number) should be
below 120. In a recent study, a public policy page 263
researcher in Atlanta surveyed 110 adult men
about 5′10″ in height and in the 55–60 age group. Data were
collected on their systolic pressure, weight (in pounds), and
race (Black = 1 for African American, 0 otherwise); a portion of
the data is shown in Table 7.4.
FILE
BP_Race
 TABLE 7.4 Systolic Pressure of Adult Men (n = 110)

Systolic Weight Black

196 254 1

151 148 0

⋮ ⋮ ⋮

170 228 0

a. Estimate and interpret the effect of Weight and Black on

systolic pressure. Predict the systolic pressure of black and
non-black adult men with a weight of 180 pounds.
b. Extend the model in part a to include the interaction between
Weight and Black. Predict the systolic pressure of black and
nonblack adult men with a weight of 180 pounds.
SOLUTION:
Table 7.5 shows the regression results for part a (see Model 1)
and for part b (see Model 2).

TABLE 7.5 Regression Results of the Systolic Pressure of Adult Men

Variable Model 1 Model 2

Constant 80.2085* 70.8312*

(0.000) (0.000)

Weight 0.3901* 0.4362*

(0.000) (0.000)
 Variable Model 1 Model 2

Black 6.9082* 30.2482*

(0.001) (0.006)

Weight × Black NA −0.1118*

(0.029)

Adjusted R2 0.7072 0.7175

Notes: Parameter estimates are followed with the p-values in parentheses; NA

denotes not applicable; * represents significance at the 5% level.

a. We refer to the second column of Table 7.5 to derive the

estimated equation as

Both predictor variables are statistically significant at the 5%

level (p-values < 0.05). For any given weight, black men are
predicted to have about seven units of higher systolic pressure
than their nonblack counterparts. This result is consistent with
the National Institute of Health report. For a weight of 180
pounds, the predicted systolic pressure of a black man is
80.2085 + 0.3901 × 180 + 6.9082 × 1 = 157. The
corresponding systolic pressure for a nonblack man is 150.
b. To estimate Model 2, we need to create a data column for the
interaction variable Weight × Black. The data entries for the
new column are computed by multiplying the column entries of
Weight and Black. For example, the first entry is computed as
254 × 1 = 254; other entries are computed similarly. We refer
to the third column of Table 7.5 to derive the estimated
 equation as

page 264
The interaction variable is negative and
statistically significant at the 5% level. The negative coefficient
is interesting as it implies that black men carry their weight
better in terms of the systolic pressure than their nonblack
counterparts. Model 2 is more suitable for prediction because
it has a higher value of the adjusted R2 than Model 1 (0.7175
> 0.7072). For a weight of 180 pounds, the predicted systolic
pressure of a black man is 70.8312 + 0.4362 × 180 + 30.2482
× 1 − 0.1118(180 × 1) = 159. The corresponding systolic
pressure for a nonblack man is 149.

The Interaction of Two Numerical Variables

Consider a regression model with two numerical variables, x1 and x2,
along with an interaction variable x1x2: y = β0 + β1x1 + β2x2 + β3x1x2

+ ε. This model is estimated as Here the partial

effects of both predictor variables are difficult to interpret. For

example, the partial effect of x1 on given by b1 + b3x2, depends on

the given value of x2. Similarly, the partial effect of x2 on given by

b2 + b3x1, depends on the given value of x1. As in the case of the

interaction of a dummy variable and a numerical variable, it is
common to evaluate the partial effects at the sample means and
(or some other values of interest). Therefore, we can use the sample
means to evaluate the partial effect of x1 on as and the

partial effect of x2 on as The next example illustrates the

interaction between two numerical variables.

EXAMPLE 7.3
The Master’s in Business Administration (MBA), once a
flagship program in business schools, has lost its appeal in
recent years (The Wall Street Journal, June 5, 2019). While
elite schools like Harvard, Wharton, and Chicago are still
attracting applicants, other schools are finding it much harder
to entice students. As a result, business schools are focusing
on specialized master’s programs to give graduates the extra
skills necessary to be career ready and successful in more
technically challenging fields.
An educational researcher is trying to analyze the
determinants of the applicant pool for the specialized Master of
Science in Accounting (MSA) program at medium-sized
universities in the United States. Two important determinants
are the marketing expense of the business school and the
percentage of the MSA alumni who were employed within three
months after graduation. Consider the data collected on the
number of applications received (Applicants), marketing
expense (Marketing, in $1,000s), and the percentage employed
within three months (Employed); a portion of the data is shown
in Table 7.6.
FILE
Marketing_MSA

TABLE 7.6 Applications Received for MSA (n = 80)

 Applicants Marketing Employed

60 173 61

71 116 83

⋮ ⋮ ⋮

69 70 92

a. Estimate and interpret the effect of Marketing and Employed

on the number of applications received. For a given marketing
expense of $80,000, predict the number of applications
received if 50% of the graduates were employed within three
months. Repeat the analysis with 80% employed within three
months.
page 265
b. Extend the model in part a to include the
interaction between Marketing and Employed. For a given
marketing expense of $80,000, predict the number of
applications received if 50% of the graduates were employed
within three months. Repeat the analysis with 80% employed
within three months.
SOLUTION:
Table 7.7 shows the regression results for part a (see Model 1)
and for part b (see Model 2).

TABLE 7.7 Regression Results of the Applications Received for MSA

Variable Model 1 Model 2

 Variable Model 1 Model 2

Constant −49.5490* −16.6356

(0.000) (0.316)

Marketing  0.3550*  0.0865

(0.000) (0.460)

Employed  1.0149*  0.5405*

(0.000) (0.019)

Marketing × Employed NA  0.0039*

(0.020)

Adjusted R2  0.7254  0.7410

Notes: Parameter estimates are followed with the p-values in parentheses; NA

denotes not applicable; * represents significance at the 5% level.

a. We refer to the second column of Table 7.7 to derive the

estimated equation as

Both predictor

variables are statistically significant at the 5% level (p-values <

0.05). For any given value of Employment, a $10,000 increase
in the marketing expense is predicted to bring in 3.55 more
applicants. Similarly, for any value of Marketing, a 10%
increase in Employed brings in 10.149 more applicants. If
Marketing equals 80 and Employed equals 50, then the
predicted number of applicants is −49.5490 + 0.3550 × 80 +
1.0149 × 50 = 30. For the same marketing expenditure, if
there is a 30 percentage point increase in Employed (= 80 −
 50), then the predicted number of applicants increases by 30
(= 1.0149 × 30) to 60 applicants.
b. To estimate Model 2, we need to create a data column for the
interaction variable Marketing × Employed. The data entries
for the new column are computed by multiplying the column
entries of Marketing and Employed. For example, the first
entry is computed as 173 × 61 = 10,533; other entries are
computed similarly. We refer to the third column of Table 7.7 to
derive the estimated equation as

Interestingly, the

Marketing variable is no longer significant at the 5% level;

however, its interaction with Employed is significant. In other
words, the marketing effort by the business schools is
effective only if the school has a good record in placing
students after graduation. Model 2 is the preferred model for
predictions because it has a higher value of the adjusted R2
than Model 1 (0.7410 > 0.7254). If Marketing equals 80 and
Employed equals 50, then the predicted number of applicants
is −16.6356 + 0.0865 × 80 + 0.5405 × 50 + 0.0039(80 × 50) =
33. For the same marketing expenditure, if the percentage of
MSA alumni finding employment within three months jumps to
80%, then the predicted number of applicants is 59.

page 266

EXAMPLE 7.4
 The objective outlined in the introductory case is to analyze a
possible gender gap in the salaries of project managers. Use
the data in Table 7.1 for the following analysis.
a. Evaluate the determinants of a project manager’s salary.
b. Estimate and interpret a regression model with relevant
interaction variables.
c. Determine whether there is evidence of a gender gap in
salaries.
SOLUTION:

a. In addition to the predictor variables Size, Experience,

Female, and Grad denoting, respectively, the firm size, the
manager’s experience, whether the manager is a female, and
whether the manager has a graduate degree, we also
consider important interactions. By interacting Female with
Experience and Female with Grad, the influence of experience
on salary or a graduate degree on salary now depends on
whether the manager is female or male. Similarly, by
interacting Experience with Size, we consider the partial effect
of Experience to depend on Size and the partial effect of Size
to depend on Experience.
b. In order to estimate the regression, we first create three new
interaction variables. Table 7.8 shows the regression results.
FILE
Gender_Gap
TABLE 7.8 Regression Results for the Project Manager’s
Salary; n = 200

Variable Coefficient p-value

Constant 41.9298* 0.002

Size −0.0432  0.388

 Variable Coefficient p-value

Experience 1.9286* 0.035

Female −4.3688  0.654

Grad 15.4357* 0.000

Female × Experience −1.3915* 0.026

Female × Grad 4.2128  0.269

Size × Experience 0.0105* 0.002

R2 0.7437

F-test (p-value) 79.5934* (0.000)

Notes: * represents significance at the 5% level.

Overall, 74.37% of the sample variations in salaries is
explained by the sample regression equation. The predictor
variables are jointly significant at any significance level, given
the F-statistic value of 79.5934 with an approximate p-value of
0. Because Size and Female are not individually significant at
the 5% level, one may be tempted to infer that there are no
salary differences between males and females and between
firms of different sizes. This inference, however, is erroneous
as it ignores the significance of the interaction variables
Female × Experience and Size × Experience. The negative
coefficient of Female × Experience implies that for every year
of experience, the gender gap increases by about $1,392.
Similarly, the positive coefficient of Size × Experience implies
that managers get a higher salary increase for every year of
experience in relatively larger firms. Project managers with a
graduate degree make about $15,436 more in salaries; this
additional salary is not statistically different between male and
female managers.
page 267
To better understand the gender gap, Figure
7.1 shows the predicted salary for male and female managers
with respect to experience using the sample average firm size
of 244.72 employees and a manager with a graduate degree.
Qualitatively, the results will not change if we reference a
different sized firm and/or an employee without a graduate
degree. We find that initially the salary gap between male and
female managers is minimal; however, the gap increases with
experience. For example, with 20 years of experience, the
predicted salaries of male and female managers are about
$137,000 and $109,000, respectively.

FIGURE 7.1 Salary of male and female managers

Similarly, to better understand the effect of firm size on salary,

Figure 7.2 shows the predicted salary of a male employee with
a graduate degree in firms with 200 and 400 employees. We
find that, initially, the predicted salary is relatively higher in
 smaller firms with 200 employees as compared to firms with
400 employees. This, however, changes with experience. For
example, with 20 years of experience, the predicted salaries of
male managers in firms with 200 and 400 employees are
about $129,000 and $163,000, respectively.

FIGURE 7.2 Salary of male managers in firms with 200 and

400 employees

c. We find evidence of a gender gap in the salaries of project

managers despite the fact that the Female variable is
insignificant at the 5% level. The significance of the interaction
of Female and Experience implies that for every year of
experience, the gender gap increases by about $1,392,
becoming $28,000 after 20 years of experience.

page 268

SYNOPSIS OF INTRODUCTORY CASE

Millennial women are closing the salary gap with men; however, the gender gap in
the United States still persists (Washington Post, May 14, 2019). It is well
documented that women are less likely than men to hold management roles and
those who do make about 20% less than men. In this report, the gender gap in
salaries of project managers is analyzed for small- to middle-sized firms in the
Boston area. The predictor variables used for the analysis include firm size, work
experience, female and graduate degree dummy variables, and interactions
between female with experience, female with graduate degree, and experience with
firm size. Overall, 74.37% of the sample variations in salaries is explained by the
sample regression equation, and the predictor variables are jointly significant at any
level.
Several interesting results emerge from the regression analysis. First, over the
course of their career, salaries of both male and female managers increase but not
at the same rate. For every year of experience, male managers get about $1,392
more than their female counterparts. Therefore, while the estimated difference in
salaries between male and female managers is only $1,548 with one year of
experience, it increases to $27,987 with 20 years of experience. Second, project
managers with a graduate degree make about $15,436 more in salaries, and this
premium for a graduate degree is not statistically different between male and female
managers. Finally, experience pays off, especially so in relatively larger firms. For
example, the predicted salary of a male manager with a graduate degree and 20
years of experience is about $129,000 in firms with 200 employees, whereas it is
about $163,000 in firms with 400 employees.

EXERCISES 7.1
Mechanics
1. Consider a linear regression model where y represents the response variable and
x, d1, and d2 are the predictor variables. Both d1 and d2 are dummy variables,
each assuming values 1 or 0. A regression model with x, d1, d2, and d1d2, where

d1d2 is the interaction variable, is estimated as

a. Compute for x = 5, d1 = 1, and d2 equal to 0 and 1.

b. Compute for x = 5, d1 = 0, and d2 equal to 0 and 1.

2. Consider a linear regression model where y represents the response variable and
x and d are the predictor variables; d is a dummy variable assuming values 1 or 0.
 A model with x, d, and the interaction variable xd is estimated as

a. Compute for x = 10 and d = 1.

b. Compute for x = 10 and d = 0.

3. Consider a linear regression model where y represents the response variable and
x1 and x2 are the predictor variables. A regression model with x1, x2, and x1x2,

where x1x2 is the interaction variable, is estimated as

a. Compute for x1 = 12 and x2 equal to 20 and 30.

b. Compute for x1 = 24 and x2 equal to 20 and 30.

4. FILE Exercise_7.4. The accompanying data file contains 20 observations on the

response variable y along with the predictor variables x, d1, and d2.
a. Estimate a regression model with the predictor variables x, d1, and d2, and then
extend it to also include the interaction variable d1d2.

b. Use the preferred model to compute given x = 20, d1 = 1, and d2 equal to 0

and 1.
5. FILE Exercise_7.5. The accompanying data file contains 20 observations on the
response variable y along with the predictor variables x and d.
a. Estimate a regression model with the predictor variables x and d, and then
extend it to also include the interaction variable xd.

b. Use the preferred model to compute given x = 15 and d equal to 0 and 1.

6. FILE Exercise_7.6. The accompanying data file contains 20 observations on the

response variable y along with the predictor variables x1 and x2.
a. Estimate a regression model with the predictor variables x1 and x2, and then
extend it to also include the interaction variable x1x2.
 b. Use the preferred model to compute given x1 = 30 with x2 evaluated at 10

and 20.

Applications
7. FILE Overweight. According to the U.S. Department of Health and Human
Services, African American women have the highest rates of being overweight
compared to other groups in the United States. Individuals are
considered overweight if their body mass index (BMI) is 25 or
page 269
greater. Data are collected from 120 individuals. The following table shows a
portion of data on each individual’s BMI; a Female dummy variable that equals 1 if
the individual is female, 0 otherwise; and a Black dummy variable that equals 1 if
the individual is African American, 0 otherwise.

BMI Female Black

28.70 0 1

28.31 0 0

⋮ ⋮ ⋮

24.90 0 1

a. Estimate the model, BMI = β0 + β1Female + β2Black + β3(Female × Black) + ɛ,

to predict the BMI for white males, white females, black males, and black
females.
b. Is the difference between white females and white males statistically significant
at the 5% level?
c. Is the difference between white males and black males statistically significant at
the 5% level?
8. FILE Diversity_Silicon. Silicon Valley tech companies are under pressure to
increase diversity as most of their employees are white and Asian men
(Bloomberg, August 16, 2017). Black and Hispanic people, especially women, are
grossly underrepresented. In order to analyze if the lack of diversity is influenced
by a lack of diversity in the founding leadership team, data are collected on the
percentage of white and Asian men (Majority Male) in Silicon Valley tech
companies and whether the founding members included a woman (Female)
and/or a Black or Hispanic man (Minority Male). A portion of the data is shown in
the accompanying table.

Majority Male Female Minority Male

 Majority Male Female Minority Male

100 0 1

80  1 0

⋮  ⋮ ⋮

80  0 0

a. Estimate and interpret the effect of Female, Minority Male, and the interaction
between Female and Minority Male.
b. Predict the percentage of White and Asian men if the founding members
included a woman but no Black or Hispanic men. Repeat the analysis if the
founding members included a woman and a Hispanic man.
9. FILE IceCream. In a recent survey, ice cream truck drivers in Cincinnati, Ohio,
reported that they make about $280 in income on a typical summer day. The
income was generally higher on days with longer work hours, particularly hot
days, and on holidays. Irma follows an ice cream truck driver for five weeks. She
collects the information on the driver’s daily income (Income), number of hours on
the road (Hours), whether it was a particularly hot day (Hot = 1 if the high
temperature was above 85°F, 0 otherwise), and whether it was a Holiday (Holiday
= 1, 0 otherwise). A portion of the data is shown in the accompanying table.

Income Hours Hot Holiday

196 5 1 0

282 8 0 0

⋮ ⋮ ⋮ ⋮

374 6 1 1

a. Estimate and interpret the effect of Hours, Hot, and Holidays on Income. Predict
the income of a driver working 6 hours on a hot holiday. What if it was not a
holiday?
b. Extend the above model to include the interaction between Hot and Holiday.
Predict the income of a driver working 6 hours on a hot holiday. What if it was
not a holiday?
10. FILE Mobile_Devices. Americans are spending an average of two hours and 37
minutes daily on smartphones and other mobile devices to connect to the world of
digital information (Business Insider, May 25, 2017). The usage of mobile devices
 is especially high for affluent, college-educated urban/suburban dwellers. A
survey was conducted on customers in the 50-mile radius of Chicago. Participants
were asked the average daily time they spent on mobile devices (Usage, in
minutes), their household income (Income, in $1,000s), if they lived in a rural area
(Rural =1 if rural, 0 otherwise), and if they had a college degree (College = 1 if
college graduate, 0 otherwise). A portion of the data is shown in the
accompanying table.

Usage Income Rural College

172 146 0 0

165 198 0 1

⋮ ⋮ ⋮ ⋮

110 31 0 1

a. Estimate and interpret a regression model for the mobile device usage based
on Income, Rural, College, and the interaction between Rural and College.
Explain the rationale for using the interaction variable.
b. Predict the mobile device usage for a college-educated person with a
household income of $120,000 and living in a rural area. What would be the
corresponding usage for someone living in a urban/suburban area?
c. Discuss the impact of a college degree on salary.
11. FILE Urban. A sociologist is looking at the relationship between consumption
expenditures of families in the United States (Consumption in $), family income
(Income in $), and whether or not the family lives in an urban or rural community
(Urban = 1 if urban, 0 otherwise). She collects data on 50
families across the United States, a portion of which is shown in
page 270
the accompanying table.

Consumption Income Urban

62336 87534  0

60076 94796  1

⋮ ⋮ ⋮

59055 100908 1
 a. Estimate Consumption = β0 + β1Income + ɛ. Compute the predicted
consumption expenditures of a family with income of $75,000.
b. Include a dummy variable Urban to predict consumption for a family with
income of $75,000 in urban and rural communities.
c. Include a dummy variable Urban and an interaction variable (Income × Urban)
to predict consumption for a family with income of $75,000 in urban and rural
communities.
d. Which of the preceding models is most suitable for the data? Explain.
12. FILE Pick_Errors. The distribution center for an online retailer has been
experiencing quite a few “pick errors” (i.e., retrieving the wrong item). Although
the warehouse manager thinks most errors are due to inexperienced workers, she
believes that a training program also may help to reduce them. Before sending all
employees to training, she examines data from a pilot study of 30 employees.
Information is collected on the employee’s annual pick errors (Errors), experience
(Exper in years), and whether or not the employee attended training (Train equals
1 if the employee attended training, 0 otherwise). A portion of the data is shown in
the accompanying table.

Errors Exper Train

13 9 0

3 27 0

⋮ ⋮ ⋮

4 24 1

a. Estimate two models:

Errors = β0 + β1Exper + β2Train + ɛ, and
Errors = β0 + β1Exper + β2Train + β3 (Exper × Train) + ɛ.

b. Which model provides a better fit in terms of adjusted R2 and the significance of
the predictor variables at the 5% level?
c. Use the chosen model to predict the number of pick errors for an employee with
10 years of experience who attended the training program, and for an employee
with 20 years of experience who did not attend the training program.
d. Give a practical interpretation for the positive interaction coefficient.
13. FILE BMI. According to the World Health Organization, obesity has reached
epidemic proportions globally. While obesity has generally been linked with
chronic disease and disability, researchers argue that it may also affect wages. In
other words, the body mass index (BMI) of an employee is a predictor for salary.
(A person is considered overweight if his/her BMI is at least 25 and obese if BMI
 exceeds 30.) The accompanying table shows a portion of salary data (in $1,000s)
for 30 college-educated men with their respective BMI and a dummy variable that
represents 1 for a white man and 0 otherwise.

Salary BMI White

34 33 1

43 26 1

⋮ ⋮ ⋮

45 21 1

a. Estimate a model for Salary with BMI and White as the predictor variables.
b. Reestimate the model with BMI, White, and a product of BMI and White as the
predictor variables.
c. Which of the models is more suitable? Explain. Use this model to estimate the
salary for a white college-educated man with a BMI of 30. Compute the
corresponding salary for a nonwhite man.
14. FILE Compensation. To encourage performance, loyalty, and continuing
education, the human resources department at a large company wants to develop
a regression-based compensation model (Comp in $ per year) for midlevel
managers based on three variables: (1) business-unit profitability (Profit in $1000s
per year), (2) years with the company (Years), and (3) whether or not the manager
has a graduate degree (Grad equals 1 if graduate degree, 0 otherwise). The
accompanying table shows a portion of data collected for 36 managers.

Comp Profit Years Grad

118100 4500 37 1

90800  5400  5 1

⋮ ⋮ ⋮ ⋮

85000  4200 29 0

a. Estimate the following model for compensation:

Comp = β0 + β1Profit + β2Years + β3Grad + β4(Profit × Grad) + β5(Years ×
Grad) + ɛ.
b. At the 5% significance level, is the overall regression model significant?
c. Which predictor variables and interaction terms are significant at α = 0.05?
 d. Use the (full) model to determine compensation for a manager having 15 years
with the company, a graduate degree, and a business-unit profit of $4,800(000)
last year.
15. FILE IPO. One of the theories regarding initial public offering (IPO) pricing is that
the initial return (the percentage change from offer to open price)
on an IPO depends on the price revision (the percentage change
page 271
from pre-offer to offer price). Another factor that may influence the initial return is
a dummy variable that equals 1 for high-tech firms and 0 otherwise. The following
table shows a portion of data on 264 IPO firms.

Initial Return Price Revision High Tech

33.93   7.14 0

18.68 −26.39 0

⋮ ⋮ ⋮

 0.08 −29.41 1

a. Estimate a model with the initial return as the response variable and the price
revision and the high-tech dummy variable as the predictor variables.
b. Reestimate the model with price revision along with the dummy variable and the
product of the dummy variable and the price revision.
c. Which of these models is the preferred model? Explain. Use this model to
estimate the initial return for a high-tech firm with a 15% price revision.
Compute the corresponding initial return for a firm that is not high tech.
16. FILE GPA_College. It is often claimed that the SAT is an important indicator of
the grades students will earn in the first year of college. The admission officer at a
state university wants to analyze the relationship between the first-year GPA in
college (College GPA) and the SAT score of the student (SAT), the unweighted
high school GPA (HS GPA), and the student’s race (White equals 1 if the student
is white, 0 otherwise). A portion of the data is shown in the accompanying table.

College GPA SAT HS GPA White

4.00 1560 4.00 1

1.78 1060 2.18 0

⋮ ⋮ ⋮ ⋮

3.38 1460 4.00 0

 a. Estimate and interpret a regression model for college GPA using SAT, HS GPA,
White, and the interactions of SAT with HS GPA and SAT with White as
predictor variables.
b. Predict the first-year college GPA of a white student with a high school GPA of
3.2 and SAT scores of 1200, 1300, and 1400. Make a corresponding prediction
for a non-white student.
17. FILE Health_Factors. According to the World Health Organization, the health of
an individual, to a large extent, is determined by factors such as income,
education, genetics, and social connections. In a survey, 120 American adults are
asked how they rated their health (Health) and social connections (Social) on a
scale of 1 to 100. Information is also gathered on their household income
(Income, in $1,000s), and college education (College equals 1 if they have
completed bachelor’s degree, 0 otherwise). A portion of the data is shown in the
accompanying table.

Health Social Income College

52 58 80 0

55 68 43 0

⋮ ⋮ ⋮ ⋮

58 90 35 0

a. Estimate and interpret a regression model for Health using Social, Income, and
College as the predictor variables. Predict the health rating of a college-
educated person given Social = 80 and Income = 100. What if the person is not
college-educated?
b. Estimate and interpret an extended model that includes the interactions of
Social with Income and Social with College. Predict the health rating of a
college-educated person given Social = 80 and Income = 100. What if the
person is not college-educated?
c. Explain which of the above two models is preferred for making predictions.
18. FILE College. With college costs and student debt on the rise, the choices that
families make when searching for and selecting a college have never been more
important. Consider the information from 116 colleges on annual post-college
earnings of graduates (Earnings in $), the average annual cost (Cost in $), the
graduation rate (Grad in %), the percentage of students paying down debt (Debt
in %), and whether or not a college is located in a city (City equals 1 if a city
location, 0 otherwise). A portion of the data is shown in the accompanying table.
 a. Estimate a regression model for the annual post-college earnings on the basis
of Cost, Grad, Debt, City, and the interaction between Cost and Grad. Interpret
the estimated coefficient of the interaction variable.
b. Predict the annual post-college earnings of graduates of a college in a city with
a graduation rate of 60%, Debt of 80% and the average annual cost of $20,000,
$30,000, and $40,000.
c. Repeat the above analysis with a graduation rate of 80%.
19. FILE Rental. Jonah has worked as an agent for rental properties in Cincinnati,
Ohio, for almost 10 years. He understands that the premium for additional
bedrooms is more in larger as compared to smaller homes. Consider the monthly
data on rent (Rent, in $) along with square footage (Sqft), number of bedrooms
(Bed), and number of bathrooms (Bath). A portion of the data is shown in the
accompanying table.

page 272
Rent Sqft Bed Bath

2300 2050 3 3

1240 680 2 1

⋮ ⋮ ⋮ ⋮

2800 2370 4 3

a. Estimate a regression model for monthly rent on the basis of Bed, Bath, Sqft,
and the interaction between Bed and Sqft. Interpret the estimated coefficient of
the interaction variable.
b. Predict the monthly rent of homes with a square footage of 1,600, 2 baths, and
the number of bedrooms equal to 2, 3, and 4. Compute the incremental
predicted rent as the number of bedrooms increases from 2 to 3 and 3 to 4.
c. Repeat the above analysis with a square footage of 2,400.
7.2 REGRESSION MODELS FOR
NONLINEAR RELATIONSHIPS
LO 7.2
Estimate and interpret nonlinear regression models.

Regression analysis not only empirically validates whether a

relationship exists between variables, but also quantifies the strength
of the relationship. So far, we have considered only linear regression
models. There are numerous applications where the relationship
between the predictor variable and the response variable cannot be
represented by a straight line and, therefore, must be captured by an
appropriate curve. In fact, the choice of a functional form is a crucial
part of specifying a regression model.
In this section, we discuss some common nonlinear regression
models by making simple transformations of the variables. These
transformations include squares and natural logarithms, which
capture interesting nonlinear relationships while still allowing easy
estimation within the framework of the linear regression model. We
use goodness-of-fit measures to choose between alternative model
specifications.

The Quadratic Regression Model

Linear regression models are often justified on the basis of their
computational simplicity. An implication of a simple linear regression
model, y = β0 + β1x + ε, is that if x goes up by one unit, the expected
value of y changes by β1, irrespective of the value of x. However, in
many applications, the relationship between the variables cannot be
represented by a straight line. We note that the linearity assumption
discussed in Chapter 6 places the restriction of linearity on the
parameters and not on the variables. Consequently, we can capture
many interesting nonlinear relationships, within the framework of the
linear regression model, by simple transformations of the response
and/or the predictor variables.
If you ever studied microeconomics, you may have learned that a
firm’s (or industry’s) average cost curve tends to be “U-shaped.” Due
to economies of scale, the average cost y of a firm initially decreases
as output x increases. However, as x increases beyond a certain
point, its impact on y turns positive. Other applications show the
influence of the predictor variable initially positive but then turning
negative, leading to an “inverted U-shape.” The quadratic
regression model is appropriate when the slope, capturing the
influence of x on y, changes in magnitude as well as sign.
A quadratic regression model with one predictor variable is
specified as y = β0 + β1x + β2x2 + ε; we can easily extend it to
include multiple predictor variables. The expression β0 + β1x + β2x2
is the deterministic component of a quadratic regression model. In
other words, conditional on x, E(y) = β0 + β1x + β2x2.
Figure 7.3 shows two scatterplots of sample data with
superimposed trendlines that were generated by estimating the
quadratic regression model. Although the estimated simple linear
regression model is not included in either panel, it is clear that the
quadratic regression model provides a better fit for both scatterplots.

page 273

FIGURE 7.3 Scatterplot of y against x with trendline generated from estimating the
quadratic regression model
It is important to be able to determine whether a quadratic regression
model provides a better fit than the linear regression model. Recall
from Chapter 6 that we cannot compare these models on the basis
of their respective R2 values because the quadratic regression
model uses one more parameter than the linear regression model.
For comparison purposes, we use adjusted R2, which imposes a
penalty for the additional parameter.
In order to estimate the quadratic regression model y = β0 + β1x +
β2x2 + ε, we have to first create a variable x2 that contains the
squared values of x. The quadratic model is estimated in the usual
way as where b1 and b2 are the estimates of β1 and

β2, respectively.

Interpretation of coefficients in the quadratic regression model:

It does not make sense to think of b1 in the estimated quadratic
regression equation as being the effect of changing x by one unit,
holding the square of x constant. In nonlinear models, the sample
regression equation is best interpreted by calculating, and even
graphing, the predicted effect on the response variable over a range
of values for the predictor variable. We will elaborate on this point in
Examples 7.5 and 7.6.

Evaluating the marginal effect of x on y in the quadratic

regression model: It is important to evaluate the estimated marginal
(partial) effect of the predictor variable x on the predicted value of the
response variable; that is, we want to evaluate the change in due

to a one-unit increase in x. In the estimated linear regression

equation the partial (marginal) effect is constant, estimated

by the slope coefficient b1. In a quadratic regression model, it can be

shown that the partial effect of x on can be approximated by b1 +

2b2x. This partial effect, unlike in the case of a linear regression

model, depends on the value at which x is evaluated. In addition,

reaches a maximum (b2 < 0) or minimum (b2 > 0) when the partial
effect equals zero. The value of x when this happens is obtained

from solving the equation b1 + 2b2x = 0, as (Note that the

optimum is obtained where the first derivative, b1 + 2b2x, equals zero

and the second derivative, 2b2, is negative for a maximum and
positive for a minimum.)

THE QUADRATIC REGRESSION MODEL

In a quadratic regression model y = β0 + β1x + β2x2 + ε, the
coefficient β2 determines whether the relationship between x
and y is U-shaped (β2 > 0) or inverted U-shaped (β2 < 0).

Predictions with a quadratic model are made by

It is advisable to use unrounded coefficients for making

predictions.

EXAMPLE 7.5
Consider the quadratic regression of the average cost (AC, in
$) on annual output (Output, in millions of units). The estimated
regression equation using the predictor variables Output and
Output2 is given by It is
 found that both predictor variables are statistically page 274
significant at the 5% level, thus confirming the
quadratic effect. Use the estimates to answer the following
questions.
a. What is the change in average cost going from an output level
of four million units to five million units?
b. What is the change in average cost going from an output level
of eight million units to nine million units? Compare this result
to the result found in part a.
c. What is the output level that minimizes average cost?
SOLUTION:

a. The predicted average cost for a firm that produces four

million units is:

The predicted average cost for a firm that produces five million
units is:

An increase in output from four to five million units results in a

$0.12 decrease in predicted average cost.
b. The predicted average cost for a firm that produces eight
million units is:

The predicted average cost for a firm that produces nine

million units is:
 An increase in output from eight to nine million units results in
a $0.05 increase in predicted average cost. Comparing this
result to the one found in part a, we note that a one-unit
change in x depends on the value at which x is evaluated.
c. Given b1 = −0.3073 and b2 = 0.0210, the output level that

minimizes average cost is

Let’s now turn to an example with an inverted U-shaped relationship.

EXAMPLE 7.6
In the United States, age discrimination is illegal, but its
occurrence is hard to prove (The New York Times, August 7,
2017). Even without discrimination, it is widely believed that
wages of workers decline as they get older. A young worker
can expect wages to rise with age only up to a certain point,
beyond which wages begin to fall. Ioannes Papadopoulos
works in the human resources department of a large
manufacturing firm and is examining the relationship between
wages (in $), years of education, and age. Specifically, he
wants to verify the quadratic effect of age on wages. He
gathers data on 80 workers in his firm with information on their
hourly wage, education, and age. A portion of the data is
shown in Table 7.9.
FILE
Wages
 TABLE 7.9 Hourly Wage of Americans (n = 80)

Wage Education Age

17.54 12 76

20.93 10 61

⋮ ⋮ ⋮

23.66 12 49

page 275

a. Plot Wage against Age and evaluate whether the linear or the
quadratic regression model better captures the relationship.
Verify your choice by using the appropriate goodness-of-fit
measure.
b. Use the appropriate model to predict hourly wages for
someone with 16 years of education and age equal to 30, 50,
or 70.
c. According to the model, at what age will someone with 16
years of education attain the highest wages?
SOLUTION:

a. Figure 7.4 shows a scatterplot of Wage against Age. We

superimpose linear and quadratic trendlines on the scatterplot.
It seems that the quadratic regression model provides a better
fit for the data as compared to the linear regression model.

FIGURE 7.4 Scatterplot of Wage versus Age

We estimate two models.
Linear Regression Model: Wage = β0 + β1Education + β2Age
+ɛ
Quadratic Regression Model: Wage = β0 + β1Education +
β2Age + β3Age2 + ɛ
For ease of exposition, we use the same notation for the
coefficients in the linear and the quadratic models even
though they have a different meaning depending on the model
we reference. Table 7.10 shows the relevant regression
results for the linear and the quadratic regression models.
TABLE 7.10 Estimates of the Linear and the Quadratic
Regression Models for Example 7.6

Linear Regression Quadratic Regression

Variable
Model Model

Intercept 2.6381 (0.268) −22.7219* (0.000)

Education 1.4410* (0.000) 1.2540* (0.000)

Age 0.0472 (0.127) 1.3500* (0.000)

 Linear Regression Quadratic Regression
Variable
Model Model

Age2 NA −0.0133* (0.000)

Adjusted 0.6088 0.8257

R2

Notes: Parameter estimates are in the main body of the table

with the p-values in parentheses; NA denotes not applicable; *
represents significance at the 5% level. The last row presents
adjusted R2 for model comparison.
Note that in the linear regression model, Age has an estimated
coefficient of only 0.0472, which is not statistically significant
(p-value = 0.127) even at the 10% significance level. However,
results change dramatically when Age2 is included along with
Age. In the quadratic regression model, both of these
variables, with p-values of approximately 0, are statistically
significant at any reasonable level. Also, the adjusted R2 is
higher for the quadratic regression model (0.8257 page 276
> 0.6088), making it a better choice for prediction.
This conclusion is consistent with our visual impression from
the scatterplot in Figure 7.4, which suggested a weak linear
but strong quadratic relationship between age and hourly
wage.
b. The predicted hourly wage for a 30-year-old person with 16
years of education is
 Similarly, the predicted hourly wage for a 50- and a 70-year-
old person is $31.59 and $26.67, respectively.
c. In part b, we predicted the hourly wage for 30-, 50-, and 70-
year-old persons with 16 years of education. Therefore, of the
three ages considered, a 50-year-old person earns the highest
wage. In Figure 7.5, we plot the predicted wage with 16 years
of education and vary age from 20 to 80 with increments of 1.

FIGURE 7.5 Predicted wages with 16 years of education and

varying age

In order to determine the optimal age at which the wage is

maximized, we also solve The optimal

age at which the wage is maximized is about 51 years, with a

wage of about $31.60. It is worth noting that at a different
education level, predicted wages will not be the same, yet the
highest wage will still be achieved at the same 51 years of
age. We advise you to plot a similar graph with 12 years of
education and varying age levels.

Regression Models with Logarithms

Consider an estimated linear regression of annual food expenditure
y on annual income x: An estimated slope coefficient

value of b1 = 0.20 implies that a $1,000 increase in annual income

would lead to a $200 increase in annual food expenditure,
irrespective of whether the income increase is from $20,000 to
$21,000 or from $520,000 to $521,000. This is yet another example
where the linearity assumption is not justified. Because we would
expect the impact to be smaller at high income levels, it may be
more meaningful to analyze what happens to food expenditure as
income increases by a certain percentage rather than by a certain
dollar amount.
Another commonly used transformation is based on the natural
logarithm that converts changes in a variable into percentage
changes.

THE LOGARITHMIC TRANSFORMATION

The natural logarithm converts changes in a variable into
percentage changes.

page 277

The logarithmic transformation is useful because many

relationships are naturally expressed in terms of percentages. For
instance, it is common to log-transform variables such as incomes,
house prices, and sales. On the other hand, variables such as age,
experience, and scores are generally expressed in their original
form. We rely on both economic intuition as well as statistical
measures to find the appropriate form for the variables.
We first illustrate log models with only one predictor variable,
which we later extend to a multiple regression model.
The Log-Log Regression Model
In the log-log regression model, both the response variable and
the predictor variable are transformed into natural logs. We can write
this model as

where ln(y) is the log-transformed response variable and ln(x) is the

log-transformed predictor variable. With these transformations, the
relationship between y and x is captured by a curve whose shape
depends on the sign and magnitude of the slope coefficient β1.
Figure 7.6 shows two scatterplots of sample data with superimposed
trendlines. Each trendline was generated by estimating the log-log
regression model. Although the estimated simple linear regression
model is not included in either panel, it is clear that the log-log
regression model provides a better fit for both scatterplots.

FIGURE 7.6 Scatterplot of y against x with trendline generated from estimating the
log-log regression model

For 0 < β1 < 1, the log-log regression model implies a positive

relationship between x and E(y); as x increases, E(y) increases at a
slower rate. This may be appropriate in the food expenditure
example where we expect food expenditure to react positively to
changes in income, with the impact diminishing at higher income
levels. If β1 < 0, it suggests a negative relationship between x and
E(y); as x increases, E(y) decreases at a slower rate. Finally, β1 > 1
implies a positive and increasing relationship between x and y; this
case is not shown in Figure 7.6. For any application, the estimated
value of β1 is determined by the data.
Note that while the log-log regression model is nonlinear in the
variables, it is still linear in the coefficients, thus satisfying the
requirement of the linear regression model. The only requirement is
that we have to first transform both variables into logs before running
the regression. We should also point out that in the log-log
regression model, the slope coefficient β1 measures the approximate
percentage change in y for a small percentage change in x. In other
words, β1 is a measure of elasticity. In the log-log regression model,
if y represents the quantity demanded of a particular good and x is
its unit price, then β1 measures the price elasticity of demand, a
parameter of considerable economic interest. Suppose β1 = −1.2;
then a 1% increase in the price of this good is expected to lead to
about a 1.2% decrease in its quantity demanded.
Finally, even though the response variable is transformed into
logs, we still make predictions in regular units. Given

you may be tempted to use the anti-log function to make predictions

page 278
in regular units as where

b0 and b1 are the coefficient estimates. However, this transformation

is known to systematically underestimate the expected value of y.
One relatively simple correction is to make predictions as
where se is the standard error of the estimate

from the log-log regression model. This correction is easy to

implement because virtually all statistical packages report se.
 THE LOG-LOG REGRESSION MODEL
The log-log regression model is specified as ln(y) = β0 + β1ln(x)
+ ε, and β1 measures the approximate percentage change in
E(y) when x increases by 1%.

Predictions with the log-log regression model are made by

where b0 and b1 are the coefficient

estimates and se is the standard error of the estimate. It is

advisable to use unrounded coefficients for making predictions.

The Logarithmic Regression Model

The log-log specification transforms all variables into logs. It is also
common to employ the semi-log regression model, in which not all
variables are transformed into logs. We will discuss two types of
semi-log regression models in the context of the simple linear
regression model. A semi-log model that transforms only the
predictor variable is often called the logarithmic regression model,
and the semi-log model that transforms only the response variable is
often called an exponential regression model. We can have many
variants of semi-log regression models when we extend the analysis
to include multiple predictor variables.
The logarithmic regression model is defined as

Like the log-log regression model, the logarithmic regression model

implies that an increase in x will lead to an increase (β1 > 0) or
decrease (β1 < 0) in E(y) at a decreasing rate. The logarithmic
regression model is especially attractive when only the predictor
variable is better captured in percentages. Figure 7.7 shows two
scatterplots of sample data with superimposed trendlines. Each
trendline was generated by estimating the logarithmic regression
model. Although the estimated simple linear regression model is not
included in either panel, it is clear that the logarithmic regression
model provides a better fit for both scatterplots. Because the log-log
and the logarithmic regression models can allow similar shapes, the
choice between the two models can be tricky. We will compare
models later in this section.

FIGURE 7.7 Scatterplot of y against x with trendline generated from estimating the
logarithmic regression model

In the logarithmic regression model, the response variable is

specified in regular units, but the predictor variable is transformed
into logs. Therefore, β1 × 0.01 measures the approximate unit
change in E(y) when x increases by 1%. For example, if β1 = 5,000,
then a 1% increase in x leads to a 50 unit (= 5,000 × 0.01) increase
in E(y). Because the response variable is already specified in regular
units, no further transformation is necessary when making
predictions.

page 279

THE LOGARITHMIC REGRESSION MODEL

 The logarithmic regression model is specified as y = β0 +
β1ln(x) + ε, and β1 × 0.01 measures the approximate change in
E(y) when x increases by 1%.

Predictions with the logarithmic model are made by

where b0 and b1 are the coefficient estimates. It is

advisable to use unrounded coefficients for making predictions.

The Exponential Regression Model

Unlike the logarithmic regression model, in which we were interested
in finding the unit change in E(y) for a 1% increase in x, the
exponential regression model allows us to estimate the percent
change in E(y) when x increases by one unit. The exponential
regression model is defined as

Figure 7.8 shows shows two scatterplots of sample data with

superimposed trendlines. Each trendline was generated by
estimating the exponential regression model. Although the estimated
simple linear regression model is not included in either panel, it is
clear that the exponential regression model provides a better fit for
both scatterplots.

FIGURE 7.8 Scatterplot of y against x with trendline generated from estimating the
exponential regression model
 For the exponential model, β1 × 100 measures the approximate
percentage change in E(y) when x increases by one unit; the exact
percentage change can be calculated as (exp(β1) − 1) × 100. For
example, a value of β1 = 0.05 implies that a one-unit increase in x
leads to an approximate 5% (= 0.05 × 100) increase in E(y), or more
precisely, a 5.1271% (= (exp(0.05) − 1) × 100) increase in E(y). In
applied work, we often see this model used to describe the rate of
growth of certain economic variables, such as population,
employment, salaries, and sales. As in the case of the log-log
regression model, we make a correction for making predictions
because the response variable is measured in logs.

THE EXPONENTIAL REGRESSION MODEL

The exponential regression model is specified as ln(y) = β0 +
β1x + ε, and β1 × 100 measures the approximate percentage
change in E(y) when x increases by one unit.

Predictions with the exponential regression model are made by

where b0 and b1 are the coefficient

estimates and se is the standard error of the estimate. It is

advisable to use unrounded coefficients for making predictions.
 While these log models are easily estimated within the framework
of a linear regression model, care must be exercised in making
predictions and interpreting the estimated slope coefficient. When
interpreting the slope coefficient, keep in mind that logs essentially
convert changes in variables into percentage changes. Table 7.11
summarizes the results.

page 280

TABLE 7.11 Summary of the Linear, Log-Log, Logarithmic, and Exponential

Regression Models

Predicted
Model Estimated Slope Coefficient
Value

y = β0 +
b1 measures the change in when x
β1x + ε
increases by one unit.

ln(y) = β0 b1 measures the approximate percentage

+ β1ln(x)
change in when x increases by 1%.
+ε

y = β0 + b1 × 0.01 measures the approximate

β1ln(x) +
change in when x increases by 1%.
ε

ln(y) = β0 b1 × 100 measures the approximate

+ β1x + ε
percentage change in when x increases

by one unit.
It is advisable to use unrounded coefficients for making predictions.
FILE
AnnArbor

EXAMPLE 7.7
Real estate investment in college towns promises good returns
(The College Investor, August 22, 2017). First, students offer a
steady stream of rental demand as cash-strapped public
universities are unable to house their students beyond
freshman year. Second, this demand is projected to grow as
more children of baby boomers head to college. Table 7.12
shows a portion of rental data for Ann Arbor, Michigan, which is
home to the main campus of the University of Michigan. The
data include the monthly rent (Rent, in $), the number of
bedrooms (Beds), the number of bathrooms (Baths), and the
square footage (Sqft) for 40 rentals.

TABLE 7.12 Rental Data for Ann Arbor, Michigan (n = 40)

Rent Beds Baths Sqft

645 1 1 500

675 1 1 648

⋮ ⋮ ⋮ ⋮

2400 3 2.5 2700

a. Plot rent against each of the three predictor variables and

evaluate whether the relationship is best captured by a line or
 a curve. Identify variables that may require a log-
transformation.
b. Estimate the linear and the relevant log regression models to
predict rent for a 1,600-square-foot rental with 3 bedrooms
and 2 bathrooms.
SOLUTION:
Given the nature of Beds and Baths, we will specify these
variables only in regular units. We will, however, consider log-
transformations for Rent and Sqft because their changes are
often expressed in percentages.
a. In Figure 7.9, we plot Rent against (a) Beds and (b) Baths and
superimpose linear and exponential trendlines (recall that the
exponential regression model log-transforms only the
response variable).

FIGURE 7.9 Comparing Rent against (a) Beds and (b) Baths

It is hard to tell from Figure 7.9 whether the relationship

between Rent and Beds or Rent and Baths is better captured
by a line or a curve. We will use goodness-of-fit measures for
the selection.
We now plot Rent against Sqft in Figure 7.10. Here, it appears
that the relationship between Rent and Sqft is better captured
by a curve than a line. The logarithmic regression page 281
 model that log-transforms Sqft fits the data better than the
linear model, suggesting that as square footage increases,
rent increases at a decreasing rate. In other words, the
increase in Rent is higher when Sqft increases from 1,000 to
2,000 than from 2,000 to 3,000. Two other models worth
considering are the exponential regression model, where only
Rent is log-transformed, and the log-log regression model,
where both Rent and Sqft are log-transformed. In order to
avoid a “cluttered” figure, these trendlines are not
superimposed on the scatterplot; however, we will formally
evaluate all models.

FIGURE 7.10 Comparing Rent against Sqft

b. While the preceding visual tools are instructive, we evaluate

four models and use goodness-of-fit measures to select the
most appropriate model for prediction.
In order to estimate these models, we first log-transform Rent
and Sqft; see the last two columns of Table 7.13.

TABLE 7.13 Transforming Rent and Sqft into Logs

page 282
In Models 1 and 2, we use Rent as the response
variable with Beds and Baths, along with Sqft in Model 1 and
ln(Sqft) in Model 2, as the predictor variables. Similarly, in
Models 3 and 4, we use ln(Rent) as the response variable with
Beds and Baths, along with Sqft in Model 3 and ln(Sqft) in
Model 4, as the predictor variables. Model estimates are
summarized in Table 7.14.

TABLE 7.14 Regression Results for Example 7.7

Notes: Parameter estimates are followed with the p-values in

parentheses; NA denotes not applicable; * represents
significance at the 5% level.
 For the most part, the number of bedrooms and the square
footage are statistically significant at the 5% level, while the
number of bathrooms is insignificant. We use the model
results to predict the rent for a 1,600-square-foot rental with 3
bedrooms and 2 bathrooms. In order to make a prediction with
Models 3 and 4, which are both based on ln(Rent), we will add
the correction term

The predicted rent ranges from $1,463 in Model 3 to $1,540 in

Model 2. We would like to know which model provides the
best prediction, as we discuss next.

Comparing Linear and Log-Transformed

Regression Models
As seen in Example 7.7, it is often not clear which regression model
is best suited for an application. While we can use economic intuition
and scatterplots for direction, we also justify our selection on the
basis of goodness-of-fit measures. In Chapter 6, we introduced R2 to
compare models based on the same number of predictor variables;
we compared adjusted R2 if the number of predictor variables was
different. Such comparisons are valid only when the response
variable of the competing models is the same. Because R2
measures the percentage of sample variations of the response
variable explained by the model, we cannot compare the percentage
of explained variations of y with that of ln(y). Comparing models
based on the computer-generated R2, which does not differentiate
between y and ln(y), is like comparing apples with oranges. For a
valid comparison, we need to compute the percentage page 283
of explained variations of y, the appropriate R2, even
though the estimated model uses ln(y) as the response variable. In
Chapter 6, we calculated R2 = SSR/SST = 1 − SSE/SST. Here, we
will use an alternative way by squaring the sample correlation
coefficient of y and

AN ALTERNATIVE WAY TO CALCULATE THE COEFFICIENT

OF DETERMINATION, R2
The coefficient of determination, R2, can be computed as
where is the sample correlation coefficient between

y and

Example 7.8 elaborates on the method with the use of Excel and
R.

Using Excel and R to Compare Linear and

Log-Transformed Regression Models
FILE
AnnArbor
EXAMPLE 7.8
Revisit the four regression models in Example 7.7 and
determine which model is best suited for making predictions.
SOLUTION:
From Table 7.14, Model 4 has the highest computer-generated
R2 value of 0.8613. However, this does not mean that Model 4
is necessarily the best, since R2 is based on Rent for Models 1
and 2 and on ln(Rent) for Models 3 and 4. Therefore, while we
can infer that Model 2 is superior to Model 1 (0.8482 > 0.8092)
and Model 4 is superior to Model 3 (0.8613 > 0.8095), we
cannot directly compare Models 2 and 4 based on the
computer-generated R2. For a valid comparison, we compute
R2 for Model 4 from scratch; that is, R2 is based on Rent, even
though it uses ln(Rent) for estimation.
For Model 4, we first compute
for the given sample

values of the predictor variables. For example, for the first

sample observation, with Beds = 1, Baths = 1, and Sqft = 500,
the predicted rent is computed as

The predicted rent for other sample observations is calculated

similarly. Excel and R are quite useful in performing the
calculations for the predicted values for Rent, as well as
generating a value of R2 for Model 4 based on Rent rather than
ln(Rent).
Using Excel
a. Open the AnnArbor data file. Convert the Rent and Sqft
variables into their respective logarithms by using Excel’s ln
function.
b. Estimate Model 4 by choosing Data > Data Analysis >
Regression from the menu. For Input Y Range, select the
ln(Rent) data, and for Input X Range, simultaneously select
the Beds, Baths, and ln(Sqft) data. If you check the Residuals
box, then Excel provides the predicted values for the ln(Rent),
or Select this box, and then click OK.

c. For convenience, paste the values for next to the

observed values for Rent on the spreadsheet; the first two

columns of Table 7.15 show a portion of the results. Next, we
want to calculate

Given the Excel-produced predicted values for ln(Rent), this

page 284
equation simplifies to

Substituting the standard error of the estimate se from Model 4

into this equation yields (For a more

precise estimate, you can use the unrounded value for se from
the regression output.) The third column of Table 7.15 shows a
portion of the results.
TABLE 7.15 Excel-Produced Predicted Values for Model 4

Rent

645 6.4778 655.7334

 Rent

675 6.6007 742.5210

⋮ ⋮ ⋮

2400 7.5648 1944.5760

d. Finally, use Excel’s CORREL function to calculate the

correlation between Rent and (columns 1 and 3 in Table

7.15) as Square the sample correlation coefficient to

find the coefficient of determination, R2 = (0.8691)2 = 0.7554.

We can now compare this value with the computer-generated
value for Model 2 that is correctly based on Rent and conclude
that Model 2 is better suited for making predictions because
0.8482 > 0.7554.
Using R

a. Import the AnnArbor data into a data frame (table) and label it
myData.
b. As shown in Chapter 6, we use the lm function to create a
regression model or, in R terminology, an object. We label this
object as Model4. We also use the log function for the natural
log transformation of the variables Rent and Sqft. Enter:

> Model4 <- lm(log(Rent) ~ Bed+Bath+log(Sqft),

data=myData)
> summary(Model4)
c. We then generate predicted values for Model4, or

using the predict function. Within the function, we indicate the

name of the model or object. We label as Pred_lnRent.

Enter:

> Pred_lnRent<-predict(Model4)
d. Next, we want to calculate the predicted values for Rent:

We label as Pred_Rent. We could

use the standard error of the estimate of 0.1262 that is

provided in the model summary output. However, we first
obtain the unrounded value, which provides a more precise
estimate; we label this value as SE. Enter:

> SE<-summary(Model4)$sigma
> Pred_Rent<-exp(Pred_lnRent+SE^2/2)
e. Finally, we use R’s cor function to calculate the correlation
between Rent and We square this value to find the

coefficient of determination, R2. Enter:

> cor(myData$Rent, Pred_Rent)^2
And R returns 0.7554188. This result is the same as the one
that we obtained with Excel; it would be identical to the last
digit if we had used the unrounded value for se from the Excel
output. We again conclude that Model 2, with an R2 equal to
0.8482, is better suited than Model 4 for making predictions.
 page 285

EXERCISES 7.2
Mechanics
20. Consider the estimated quadratic model

a. Predict y when x equals 10, 20, and 30.

b. Find the value of x at which the predicted y is optimized. At this x value, is the
predicted y maximized or minimized.
21. Consider the following sample regressions for the linear and quadratic models
along with their respective R2 and adjusted R2.

Linear Quadratic

Intercept 13.3087  1.7656

x 0.3392  4.0966

x2 NA −0.2528

R2 0.1317  0.5844

Adjusted R2 0.0232  0.4657

a. Use the appropriate goodness-of-fit measure to justify which model fits the data
better.
b. Given the best-fitting model, predict y for x = 4, 8, and 12.
22. Consider the sample regressions for the linear, the logarithmic, the exponential,
and the log-log models. For each of the estimated models, predict y when x
equals 50.
23. Consider the following sample regressions for the linear and the logarithmic
models.

Linear Logarithmic

Intercept 6.7904 −5.6712

x 1.0607 NA

ln(x) NA 10.5447*

se 2.4935 1.5231

R2 0.8233 0.9341

Adjusted R2 0.8013 0.9259

a. Justify which model fits the data better.

b. Use the selected model to predict y for x = 10.
24. Consider the following sample regressions for the log-log and the exponential
models.

Log-Log Exponential

Intercept 1.8826 2.0219

x NA 0.0513

ln(x) 0.3663 NA
 Log-Log Exponential

se 0.3508 0.2922

R2 0.5187 0.6660

Adjusted R2 0.4585 0.6242

a. Justify which model fits the data better.

b. Use the selected model to predict y for x = 20.

Applications
25. FILE Television. Numerous studies have shown that watching too much
television hurts school grades. Others have argued that television is not
necessarily a bad thing for children (Psychology Today, October 22, 2012). Like
books and stories, television not only entertains, it also exposes a child to new
information about the world. While watching too much television is harmful, a little
bit may actually help. Researcher Matt Castle gathers information on the grade
point average (GPA) of 28 middle school children and the number of hours of
television they watched per week. A portion of the data is shown in the
accompanying table.

GPA Hours

3.24 19

3.10 21

⋮ ⋮

3.31  4

a. Estimate a quadratic regression model where the GPA of middle school children
is regressed on hours and hours-squared.
b. Is the quadratic term in this model justified? Explain.
c. Find the optimal number of weekly hours of TV for middle school children.
26. FILE Sales_Reps. Brendan Connolly manages the human resource division of a
high-tech company. He has access to the salary information of 300 sales reps
along with their age, sex (Female), and the net promoter score (NPS) that
indicates customer satisfaction. A portion of the data is shown in the
accompanying table.
 page 286
Salary Age Female NPS

97000 44 0 9

50000 34 0 4

⋮ ⋮ ⋮ ⋮

88000 36 0 10

a. Estimate and interpret a quadratic model using the natural log of salary as the
response variable and Age, Age2, Female, and NPS as the predictor variables.
b. Determine the optimal level of age at which the natural log of salary is
maximized.
c. At the optimal age, predict the salary of male and female sales reps with NPS =
8.
27. FILE Fertilizer2. A horticulturist is studying the relationship between tomato
plant height and fertilizer amount. Thirty tomato plants grown in similar conditions
were subjected to various amounts of fertilizer (in ounces) over a four-month
period, and then their heights (in inches) were measured. A portion of the data is
shown in the accompanying table.

Height Fertilizer

20.4 1.9

29.1 5.0

⋮ ⋮

36.4 3.1

a. Estimate the linear regression model Height = β0 + β1Fertilizer + ɛ.

b. Estimate the quadratic regression model Height = β0 + β1Fertilizer +
β2Fertilizer2 + ɛ. Find the fertilizer amount at which the height reaches a
minimum or maximum.
c. Use the best-fitting model to predict, after a four-month period, the height of a
tomato plant that received 3.0 ounces of fertilizer.
28. FILE Circuit_Boards. The operators manager at an electronics company
believes that the time required for workers to build a circuit board is not
necessarily proportional to the number of parts on the board. He wants to develop
a regression model to predict time (in minutes) based on part quantity. He has
collected data for the last 25 boards. A portion of this data is shown in the
accompanying table.

Time Parts

30.8 62

9.8 32

⋮ ⋮

29.8 60

a. Estimate the linear regression model to predict time as a function of the number
of parts (Parts). Then estimate the quadratic regression model to predict time
as a function of Parts and Parts squared.
b. Evaluate the two models in terms of variable significance (α = 0.05) and
adjusted R2.
c. Use the best-fitting model to predict how long it would take to build a circuit
board consisting of 48 parts.
29. FILE Inventory_Cost. The inventory manager at a warehouse distributor wants
to predict inventory cost (Cost in $) based on order quantity (Quantity in units).
She thinks it may be a nonlinear relationship because its two primary components
move in opposite directions: (1) order processing cost (costs of procurement
personnel, shipping, transportation), which decreases as order quantity increases
(due to fewer orders needed), and (2) holding cost (costs of capital, facility,
warehouse personnel, equipment), which increases as order quantity increases
(due to more inventory held). She has collected monthly inventory costs and order
quantities for the past 36 months. A portion of the data is shown in the
accompanying table.

Cost Quantity

844 54.4

503 52.1

⋮ ⋮
 Cost Quantity

870 55.5

a. Create a scatterplot of inventory cost as a function of quantity. Superimpose the

linear trendline and the quadratic trendline.
b. Estimate the linear regression model to predict inventory cost as a function of
order quantity. Then estimate the quadratic regression model to predict
inventory cost as a function of order quantity and order quantity squared.
c. Evaluate the two models in terms of significance tests (α = 0.05) and adjusted
R2.
d. Use the best-fitting model to predict monthly inventory cost for an order quantity
of 800 units.
30. FILE Dexterity. A manufacturing manager uses a dexterity test on 120 current
employees in order to predict watch production based on time to completion (in
seconds) and a Male dummy variable. A portion of the data is shown in the
accompanying table.

Watches Time Male

28 503 1

33 534 1

⋮ ⋮ ⋮

24 518 1

a. Estimate the linear model Watches = β0 + β1Time + β2Male + ɛ. Interpret the

slope coefficient for Time. If the time required to complete the dexterity test is
550 seconds, what is the predicted watch production for male and female
employees?
b. Estimate the logarithmic model Watches = β0 + β1ln(Time) + β2Male + ɛ.
Interpret the slope coefficient for ln(Time). If the time required
to complete the dexterity test is 550 seconds, what is the
page 287
predicted watch production for male and female employees?
c. Which model provides a better fit? Explain.
31. FILE Wine_Pricing. Professor Orley Ashenfelter of Princeton University is a
pioneer in the field of wine economics. He claims that, contrary to old orthodoxy,
the quality of wine can be explained mostly in terms of weather conditions. Wine
romantics accuse him of undermining the whole wine-tasting culture. In an
interesting co-authored paper that appeared in Chance magazine in 1995, he ran
 a multiple regression model where quality, measured by the average vintage price
relative to 1961, is used as the response variable y. The predictor variables were
the average temperature x1 (in degrees Celsius), the amount of winter rain x2 (in
millimeters), the amount of harvest rain x3 (in millimeters), and the years since
vintage x4. A portion of the data is shown in the accompanying table.

a. Estimate the linear model y = β0 + β1x1 + β2x2 + β3x3 + β4x4 + ɛ. What is the
predicted price if x1 = 16, x2 = 600, x3 = 120, and x4 = 20?
b. Estimate the exponential model ln(y) = β0 + β1x1 + β2x2 + β3x3 + β4x4 + ɛ.
What is the predicted price if x1 = 16, x2 = 600, x3 = 120, and x4 = 20?

c. Use R2 to select the appropriate model for prediction.

32. FILE Electricity. The facility manager at a pharmaceutical company wants to
build a regression model to forecast monthly electricity cost (Cost in $). Three
main variables are thought to influence electricity cost: (1) average outdoor
temperature (Temp in °F), (2) working days per month (Days), and (3) tons of
product produced (Tons). A portion of the past monthly data on 80 observations is
shown in the accompanying table.

Cost Temp Days Tons

16747 46 22 75

7901 31 24 98

⋮ ⋮ ⋮ ⋮

11380 56 28 84

a. Estimate the linear model Cost = β0 + β1Temp + β2Days + β3Tons + ɛ. What is

the predicted electricity cost in a month during which the average outdoor
temperature is 65°, there are 23 working days, and 76 tons are produced?
b. Estimate the exponential model ln(Cost) = β0 + β1Temp + β2Days + β3Tons +
ɛ. What is the predicted electricity cost in a month during which the average
 outdoor temperature is 65°, there are 23 working days, and 76 tons are
produced?
c. Based on R2, which model provides the better fit?
33. FILE Davis_Rental. Chad Dobson has heard about the positive outlook for real
estate investment in college towns. He is interested in investing in Davis,
California, which houses one of the University of California campuses. He has
access to monthly rents (in $) for 27 houses, along with three characteristics of
the home: number of bedrooms (Beds), number of bathrooms (Baths), and square
footage (Sqft). A portion of the data is shown in the accompanying table.

Rent Beds Baths Sqft

2950 4 4 1453

2400 4 2 1476

⋮ ⋮ ⋮ ⋮

744 2 1  930

a. Estimate the linear model that uses Rent as the response variable. Estimate the
exponential model that uses log of Rent as the response variable.
b. Compute the predicted rent for a 1,500-square-foot house with 3 bedrooms and
2 bathrooms for the linear and the exponential models (ignore the significance
tests).
c. Use R2 to select the appropriate model for prediction.
34. FILE Savings_Rate. The accompanying table shows a portion of the monthly
data on the personal savings rate (Savings in %) and personal disposable income
(Income in $ billions) in the U.S. from January 2007 to November 2010.

Date Savings Income

Jan 2007 2.2 10198.2

Feb 2007 2.3 10252.9

⋮ ⋮ ⋮

Nov 2010 5.5 11511.9

a. Estimate the linear model Savings = β0 + β1Income + ɛ and the log-log model,
ln(Savings) = β0 + β1ln(Income) + ɛ. For each model, predict Savings if Income
 = $10,500.
b. Which is the preferred model? Explain.
35. FILE Learning_Curve. Learning curves are used in production operations to
estimate the time required to complete a repetitive task as an operator gains
experience. Suppose a production manager has compiled 30 time values (in
minutes) for a particular operator as she progressed down the learning curve
during the first 100 units. A portion of this data is shown in the accompanying
table.

page 288
Time per Unit Unit Number

18.30 3

17.50 5

⋮ ⋮

5.60 100

a. Create a scatterplot of time per unit against units built. Superimpose the linear
trendline and the logarithmic trendline to determine visually the best-fitting
model.
b. Estimate the simple linear regression model and the logarithmic regression
model for predicting time per unit using unit number as the predictor variable.
c. Based on R2, use the best-fitting model to predict the time that was required for
the operator to build Unit 50.
36. FILE Happiness. Numerous attempts have been made to understand
happiness. Because there is no unique way to quantify it, researchers often rely
on surveys to capture a subjective assessment of well-being. One study finds that
holding everything else constant, people seem to be least happy when they are in
their 40s (Psychology Today, April 27, 2018). Another study suggests that money
does buy happiness, but its effect diminishes as incomes rise above $75,000 a
year (Money Magazine, February 14, 2018). Consider survey data of 100 working
adults’ self-assessed happiness on a scale of 0 to 100, along with their age and
annual income. A portion of the data is shown in the accompanying table.

Happiness Age Income

69 49  52000

83 47 123000
 Happiness Age Income

⋮ ⋮ ⋮

79 31 105000

a. Estimate and interpret a regression model for Happiness based on Age, Age2,
and ln(Income).
b. Predict happiness with Income equal to $80,000 and Age equal to 30, 45, and
60 years.
c. Predict happiness with Age equal to 60 and Income equal to $25,000, $75,000,
$125,000.
37. FILE Production_Function. Economists often examine the relationship between
the inputs of a production function and the resulting output. A common way of
modeling this relationship is referred to as the Cobb-Douglas production function.
This function can be expressed as ln(Q) = β0 + β1ln(L) + β2ln(K) + ɛ, where Q
stands for output, L for labor, and K for capital. The accompanying table lists a
portion of data relating to the U.S. agricultural industry in the year 2004.

State Output Labor Capital

AL 3.1973 2.7682 3.1315

AR 7.7006 4.9278 4.7961

⋮ ⋮ ⋮ ⋮

WY 1.2993 1.6525 1.5206

a. Estimate ln(Q) = β0 + β1ln(L) + β2ln(K) + ɛ. What is the predicted change in

output if labor increases by 1%, holding capital constant?
b. Holding capital constant, can we conclude at the 5% level that a 1% increase in
labor will increase the output by more than 0.5%?
38. FILE Smoking. A nutritionist wants to understand the influence of income and
healthy food on the incidence of smoking. He collects data on the percentage of
smokers in each state in the U.S., the percentage of the state’s population that
regularly eats fruits and vegetables, and the state’s median income (in $). A
portion of the data is shown in the accompanying table.

State Smoke Fruits/Vegetables Income

 State Smoke Fruits/Vegetables Income

AK 14.6 23.3 61604

AL 16.4 20.3 39980

⋮ ⋮ ⋮ ⋮

WY 15.2 23.3 52470

a. Estimate Smoke = β0 + β1Fruits/Vegetables + β2Income + ɛ.

b. Compare this model with a model that log-transforms the income variable.

7.3 LINEAR PROBABILITY AND

LOGISTIC REGRESSION MODELS
LO 7.3
Estimate and interpret linear probability and logistic regression models.

So far we have considered regression models where dummy (binary)

variables are used as predictor variables. In this section, we analyze
binary choice (classification) models where the response variable
is a binary variable.. The consumer choice literature is replete with
applications such as whether or not to buy a house, join a health
club, or go to graduate school. At the firm level, managers make
decisions such as whether or not to run a marketing campaign,
restructure debt, or approve a loan. In all such applications, the
response variable is binary, where one of the choices page 289
can be designated as 1 and the other as 0. Usually, this
choice can be related to a host of factors—the predictor variables.
For instance, whether or not a family buys a house depends on
predictor variables such as household income, mortgage rates, and
so on.
The Linear Probability Model
Consider a simple linear regression model y = β0 + β1x + ε, where y
is a binary variable; we can easily extend the model to include
multiple predictor variables. The linear regression model applied to a
binary response variable is called the linear probability model
(LPM).

THE LINEAR PROBABILITY MODEL

The linear probability model is specified as y = β0 + β1x + ε,
where y assumes a 1 or 0 value. Predictions with this model
are made by where b0 and b1 are the estimates of

the population parameters β0 and β1, and is the predicted

probability of success.

EXAMPLE 7.9
FILE
Mortgage

The subprime mortgage crisis of 2006 that eventually lead to

the Great Recession forced financial institutions to be extra
stringent in granting mortgage loans. Thirty recent mortgage
applications are obtained to analyze the mortgage approval
rate. The response variable y equals 1 if the mortgage loan is
approved, 0 otherwise. It is believed that approval depends on
the percentage of the down payment x1 and the percentage of
the income-to-loan ratio x2. Table 7.16 shows a portion of the
data.
 TABLE 7.16 Mortgage Application Data

y x1 x2

1 16.35 49.94

1 34.43 56.16

⋮ ⋮ ⋮

0 17.85 26.86

a. Estimate and interpret the linear probability model y = β0 +

β1x1 + β2x2 + ε.
b. Predict the loan approval probability for an applicant with a
20% down payment and a 30% income-to-loan ratio. What if
the down payment was 30%?
SOLUTION:

a. Table 7.17 shows a portion of the regression results. The

estimated regression equation is
With p-values of 0.012 and

0.000, respectively, both predictor variables exert a positive

and statistically significant influence on loan approval, at a 5%
level. Also, holding the income-to-loan ratio constant, b1 =
0.0188 implies that a 1-percentage-point increase in down
payment increases the approval probability by 0.0188, or by
1.88%. Similarly, holding down payment constant, a 1-
percentage-point increase in the income-to-loan ratio
increases the approval probability by 0.0258, or by 2.58%.
page 290
 TABLE 7.17 The Linear Probability Model Results for
Example 8.9

b. The predicted loan approval probability for an applicant with a

20% down payment and a 30% income-to-loan ratio is
Similarly, the predicted

loan approval probability with a down payment of 30% is

In other words, as

down payment increases by 10 percentage points, the

predicted probability of loan approval increases by 0.1880 (=
0.4698 − 0.2818), which is essentially the estimated slope,
0.0188, multiplied by 10. The estimated slope coefficient for
the percentage of income-to-loan variable can be interpreted
similarly.

Although it is easy to estimate and interpret the linear probability

model, it has some shortcomings. The major shortcoming is that it
can produce predicted probabilities that are greater than 1 or less
than 0. For instance, for a down payment of 60%, with the same
income-to-loan ratio of 30%, we get a predicted loan approval
probability of a

probability greater than 1! Similarly, for a down payment of 5%, the

model predicts a negative probability,
 Furthermore, the linearity of

the relationship may also be questionable. For instance, we would

expect a big increase in the probability of loan approval if the
applicant makes a down payment of 30% instead of 20%. This
increase in probability is likely to be much smaller if the same 10-
percentage-point increase in down payment is from 60% to 70%.
The linear probability model cannot differentiate between these two
scenarios. For these reasons, we introduce the logistic regression
model, which is a more appropriate probability model for binary
response variables.

The Logistic Regression Model

Recall that in an estimated linear probability model, the

influence of x on captured by the estimated slope b1, is constant.

Furthermore, for any given slope, we can find some value of x for
which the predicted probability is outside the [0,1] interval. For
meaningful analysis, we would like a nonlinear specification that
constrains the predicted probability between 0 and 1.
Consider the following logistic specification

where and e ≈ 2.718. The logistic specification

ensures that the estimated probability is between 0 and 1 for all

values of x.
 The logistic regression model cannot be estimated with standard
ordinary least squares (OLS) procedures. Instead, we rely on the
method of maximum likelihood estimation (MLE). While the MLE
of the logistic regression model is not supported by Excel, it can
easily be estimated with most statistical packages, including Analytic
Solver and R. Given the relevance of the logistic regression model in
business applications, it is important to be able to interpret and make
predictions with the estimated model.

page 291

THE LOGISTIC REGRESSION MODEL

The logistic regression model is a nonlinear model that can be
estimated with most statistical packages. Predictions with this
model are made by

where b0 and b1 are the estimates of the population parameters

β0 and β1 and is the predicted probability of success.

Figure 7.11 highlights the relationship between the predicted

probability and the predictor variable x for the linear probability

model and the logistic regression model, given b1 > 0. Note that in
the linear probability model, the probability falls below 0 for small
values of x and exceeds 1 for large values of x. The probabilities
implied by the logistic regression model, however, are always
constrained in the [0,1] interval. (For ease of exposition, we use the
same notation to refer to the coefficients in the linear probability
model and the logistic regression model. We note, however, that
these coefficients and their estimates have a different meaning
depending on which model we are referencing.)

FIGURE 7.11 Predicted probabilities for linear probability and logistic regression
models, with b1 > 0

It is important to be able to interpret the regression coefficients of

the logistic regression model. In a linear probability model, the
interpretation of a regression coefficient is straightforward. For
instance, if the estimated linear probability model is it

implies that for every 1-unit increase in x, the predicted probability

increases by 0.03. We note that increases by 0.03, whether x

increases from 10 to 11 or from 20 to 21.

Now consider the estimated logistic regression model,

Because the regression coefficient b1 = 0.18 is

positive, we can infer that x exerts a positive influence on

However, the exact impact based on the estimated regression

coefficient is not obvious. A useful method to interpret the estimated
coefficient is to highlight the changing impact of x on For instance,

given x = 10, we compute the predicted probability as

Similarly, for x = 11, the predicted probability

is Therefore, as x increases by one unit from 10 to 11, the

predicted probability increases by 0.0444. However, the increase in

will not be the same if x increases from 20 to 21. We can show that

increases from 0.8176 when x = 20 to 0.8429 when x = 21, for a

smaller increase of 0.0253. Note that when x is relatively large, its
reduced influence on the predicted probability is consistent with the
depiction of logistic probabilities in Figure 7.11.

page 292

EXAMPLE 7.10
Let’s revisit Example 7.9.
a. Estimate and interpret the logistic regression model for the
loan approval outcome y based on the applicant’s percentage
of down payment x1 and the income-to-loan ratio x2.
b. For an applicant with a 30% income-to-loan ratio, predict loan
approval probabilities with down payments of 20% and 30%.
c. Compare the predicted probabilities based on the estimated
logistic regression model with those from the estimated linear
probability model in Example 7.9.
SOLUTION:

a. We use R to produce the logistic regression results shown in

Table 7.18. (Instructions for estimating the logistic regression
model with Analytic Solver and R are provided at the end of
this section.)

TABLE 7.18 Logistic Regression Results for Example 7.10

As in the case of the linear probability model, both predictor

variables exert a positive and statistically significant influence
on loan approval at a 5% level, given positive estimated
coefficients and p-values of 0.035 and 0.006, respectively. (In
maximum likelihood estimation, the significance tests are valid
only with large samples. Consequently, we conduct the z test,
in place of the usual t test, to evaluate the statistical
significance of a coefficient.)
b. The estimated probability equation is computed as

The predicted loan approval probability with x1 = 20 and x2 =

30 is
 Similarly, the predicted loan approval probability with x1 = 30
and x2 = 30 is 0.5065. Note that, given an income-to-loan ratio
of 30%, the predicted probability increases by 0.2962 (=
0.5065 − 0.2103) when the down payment increases from
20% to 30%. For the same income-to-loan ratio of 30%, it can
be shown that the increase in the predicted probability is only
0.0449 (= 0.9833 − 0.9384) when the down payment
increases from 50% to 60%.
c. Table 7.19 provides predicted probabilities for the linear
probability and the logistic regression models for selected
values of x1 given x2 = 30.
page 293
TABLE 7.19 Predicted Probabilities for the
Linear Probability and the Logistic Regression Models

x1 x2 Linear Probability Logistic

 4 30 −0.02   0.03

20 30 0.28  0.21

30 30 0.47  0.51

50 30 0.85  0.94

60 30 1.04  0.98
 As discussed earlier, with the linear probability model, the
predicted probabilities can be negative or greater than one.
The probabilities based on the logistic regression model stay
between zero and one for all possible values of the predictor
variables. Therefore, whenever possible, it is preferable to use
the logistic regression model over the linear probability model
in binary choice models.

Sometimes, analysts prefer to interpret the estimated logistic

model in terms of the odds ratio. The odds ratio is defined as the
ratio of the probability of success P(y = 1) to the probability of failure
P(y = 0), or simply The odds ratio metric is especially popular

in sports and gambling.

Accuracy of Binary Choice Models

There is no universal goodness-of-fit measure for binary choice
models that can be used to assess how well the model fits the data.
Unlike in the case of the linear regression models, we cannot assess
the performance of binary choice models based on the standard
error of the estimate se, the coefficient of determination R2, or
adjusted R2. The residual analysis is meaningless because the
response variable can only take values 0 and 1, which then
corresponds to two values for the residuals: for y = 0 and for y

= 1.
It is common to assess the performance of linear probability and
logistic regression models on the basis of the accuracy rates defined
as the percentage of correctly classified observations. Using a
default cutoff of 0.5, we compare the binary values of the response
variable y with the binary predicted values that equal one if

and zero if where is the estimated probability. The accuracy

rate is calculated as the number of correct predictions divided by the

total number of observations.

THE ACCURACY RATE OF A BINARY CHOICE MODEL

Using the default cutoff of 0.5, the binary predicted values are
calculated as 1 for These values are then

compared with the y values to compute the accuracy rate of the

model as:

Although the accuracy measure is useful, it can sometimes be

misleading. In applications with many ones and a few zeros, or many
zeros and a few ones, the percentage of correctly classified
observations can be high even when the model predicts the less
likely outcome poorly. For example, in the case of fraud detection,
we are interested primarily in identifying the fraudulent cases where
the number of fraudulent cases often account for a very small
percentage (less than 1%) of the total number of cases. A fraud
detection model that classifies all the nonfraudulent cases correctly
but misses most of the fraudulent cases is not very page 294
useful despite having an extremely high accuracy
rate (> 99%). For this reason, it is better to report the percentage of
correctly classified observations for both outcomes of the binary
response variable, referred to as the sensitivity and the specificity
measures.
Note that the relevance of performance measures, including
accuracy, sensitivity, and specificity, depends on their related
misspecification costs. Although the default for the cutoff value is
0.50, the analyst may choose a higher or lower cutoff value to
improve a particular performance measure. Further details on
performance measures and cutoff values are provided in Chapter 8.

EXAMPLE 7.11
For the mortgage approval example, compare the accuracy
rates of the estimated linear probability model (LPM) with the
estimated logistic regression model.
SOLUTION:
In order to compute the accuracy rates, we first find the
predicted approval probabilities given the sample values of the
predictor variables. For the first sample observation, we find
the predicted approval as:

Because the predicted values for both models are greater than
0.5, their corresponding binary predicted values are one.
Predictions for other sample observations are computed
similarly; see Table 7.20 for a portion of these predictions.

TABLE 7.20 Computing the Accuracy Rates of Binary Choice Models

 We find that out of 30 observations, the binary predicted values
match the values of y in 25 and 26 of the cases for the linear
probability model and the logistic regression model,
respectively. Therefore, the accuracy rate is 83.33% for the
linear probability model and 86.67% for the logistic regression
model. We infer that the logistic model provides better
predictions than the linear probability model.

Using Analytic Solver and R to Estimate the

Logistic Regression Model
FILE
Mortgage

Analytic Solver and R are useful when estimating the logistic

regression model and finding the corresponding accuracy rates; we
use the Mortgage data to replicate the results for Examples 7.10 and
7.11.

Using Analytic Solver

Before following the Analytic Solver instructions, make sure that you
have read Appendix B (“Getting Started with Excel and Excel Add-
Ins”).
. Open the Mortgage data file.
. Choose Data Mining > Classify > Logistic page 295
Regression from the menu.

. See Figure 7.12. Click on the ellipsis next to the Data range and

highlight cells A1:C31. Make sure that the box preceding First Row
Contains Headers is checked. The Variables in Input Data box will
populate. Select and move variables x1 and x2 to Selected
Variables box and y to Output Variable box. Accept other defaults
and click Next.

FIGURE 7.12 Logistic regression dialog box

Source: Microsoft Excel

. Check Variance-Covariance Matrix under Regression: Display; this

is used to compute the standard errors for the significance tests.
Click Next and accept the defaults. Click Finish.
. Analytic Solver will produce a lot of output in separate worksheets.
The top part of Figure 7.13 shows the relevant portion from the
LogReg_Output worksheet.

FIGURE 7.13 Analytic Solver’s relevant output for Examples 7.10

and 7.11

Source: Microsoft Excel

page 296
Note that the coefficient estimates are the same as
those reported in Table 7.18. In order to find the standard errors, we
take the positive square root of the diagonal elements of the
variance-covariance matrix. For example, the standard error of
which is the same as in Table 7.18. The z-

value is calculated as

. The bottom part of Figure 7.13 shows output from the

LogReg_TrainingScore worksheet. We find that 26 observations, or
86.66667%, are correctly classified. Again, these results are the
same as we derived in Example 7.11.
Using R
. Import the Mortgage data into a data frame (table) and label it
myData.
. We use the glm function to construct a logistic regression model
object, which is a generalized version of the lm function; we label
this object as Logistic_Model. Within the function, we first specify y
as a function of x1 and x2 and then use the family option to denote
the type of model. Finally, we specify the data frame. Like the lm
function, we will not see any immediate output after we enter the
command. We use the summary function to obtain the output.
Enter:
> Logistic_Model <- glm(y ∼ x1 + x2, family = binomial(link = logit),
data = myData)
> summary(Logistic_Model)
Figure 7.14 shows the R regression output. Note that the estimated
logistic regression model results are the same as those reported in
Table 7.18.

FIGURE 7.14 R’s relevant output for Example 7.11

. We use the predict function to compute the predicted probabilities

for the given sample values, labeled as Pred. Enter:

> Pred <- predict(Logistic_Model, type = “response”)

. We use the round function to construct the binary predicted values,
labeled as Binary. If a predicted probability is 0.50 or more, then its
 corresponding binary value is 1; if a predicted probability is less
than 0.5, then its corresponding binary value is 0. Enter:

> Binary <- round(Pred)

. Both y and Binary now contain 0 and 1 values. We want to find the
proportion of the values in the sample for which y equals Binary.
Recall from Chapter 2 that the double equal sign (==) is used to
compare the values of y and Binary. If the two values are the same,
the operator returns 1, and 0 otherwise. We use the mean function
to compute the proportion of correctly classified observations; we
multiply by 100 to get a percentage. Enter:

> 100 ∗ mean(myData$y = = Binary)

R returns: [1] 86.66667.

page 297

EXERCISES 7.3
Mechanics
39. Consider a binary response variable y and a predictor variable x that varies
between 0 and 4. The linear probability model is estimated as

a. Compute the estimated probability for x = 2 and x = 3.

b. For what values of x is the estimated probability negative or greater than 1?
40. Consider a binary response variable y and a predictor variable x. The following
table contains the parameter estimates of the linear probability model (LPM) and
the logistic regression model, with the associated p-values shown in parentheses.

Variable LPM Logistic

Intercept −0.72  −6.20 

(0.04) (0.04)

x 0.05 0.26
(0.06) (0.02)
 a. Test for the significance of the intercept and the slope coefficients at the 5%
level in both models.
b. What is the predicted probability implied by the linear probability model for x =
20 and x = 30?
c. What is the predicted probability implied by the logistic regression model for x =
20 and x = 30?
41. Consider a binary response variable y and two predictor variables x1 and x2. The
following table contains the parameter estimates of the linear probability model
(LPM) and the logistic regression model, with the associated p-values shown in
parentheses.

Variable LPM Logistic

Intercept −0.40  −2.20 

(0.03) (0.01)

x1 0.32 0.98
(0.04) (0.06)

x2 −0.04  −0.20 
(0.01) (0.01)

a. Comment on the significance of the variables.

b. What is the predicted probability implied by the linear probability model for x1 =
4 with x2 equal to 10 and 20?
c. What is the predicted probability implied by the logistic regression model for x1
= 4 with x2 equal to 10 and 20?
42. Using 30 observations, the following output was obtained when estimating the
logistic regression model.

a. What is the predicted probability when x = 0.40?

b. Is x significant at the 5% level?
43. FILE Exercise_7.43. The accompanying data file contains 20 observations on
the binary response variable y along with the predictor variables x1 and x2.
 a. Estimate the linear probability model to compute for x1 = 12 and x2 = 8.

b. Estimate the logistic regression model to compute for x1 = 12 and x2 = 8.

44. FILE Exercise_7.44. The accompanying data file contains 20 observations on

the binary response variable y along with the predictor variables x1 and x2.
a. Estimate and interpret the linear probability model and the logistic regression
model.
b. Compute the accuracy rates of both models.

c. Use the preferred model to compute for x1 = 60 and x2 = 18.

Applications
45. FILE Purchase. Annabel, a retail analyst, has been following Under Armour,
Inc., the pioneer in the compression-gear market. Compression garments are
meant to keep moisture away from a wearer’s body during athletic activities in
warm and cool weather. Annabel believes that the Under Armour brand attracts a
younger customer, whereas the more established companies, Nike and Adidas,
draw an older clientele. In order to test her belief, she collects data on the age of
the customers and whether or not they purchased Under Armour (Purchase; 1 for
purchase, 0 otherwise). A portion of the data is shown in the accompanying table.

Purchase Age

1 30

0 19

⋮ ⋮

1 24

a. Estimate the linear probability model using Under Armour as the response
variable and Age as the predictor variable.
b. Compute the predicted probability of an Under Armour purchase for a 20-year-
old customer and a 30-year-old customer.
c. Test Annabel’s belief that the Under Armour brand attracts a younger customer,
at the 5% level.
46. FILE Purchase. Refer to the previous exercise for a description of the data set.
Estimate the logistic regression model where the Under Armour purchase
depends on age.
 a. Compute the predicted probability of an Under Armour purchase for a 20-year-
old customer and a 30-year-old customer.

page 298
b. Test Annabel’s belief that the Under Armour brand attracts a
younger customer, at the 5% level.
47. FILE Parole. More and more parole boards are using risk assessment tools
when trying to determine an individual’s likelihood of returning to crime (Prison
Legal News, February 2, 2016). Most of these models are based on a range of
character traits and biographical facts about an individual. Many studies have
found that older people are less likely to re-offend than younger ones. In addition,
once released on parole, women are not likely to re-offend. A sociologist collects
data on 20 individuals who were released on parole two years ago. She notes if
the parolee committed another crime over the last two years (Crime equals 1 if
crime committed, 0 otherwise), the parolee’s age at the time of release, and the
parolee’s sex (Male equals 1 if male, 0 otherwise). The accompanying table
shows a portion of the data.

Crime Age Male

1 25 1

0 42 1

⋮ ⋮ ⋮

0 30 1

a. Estimate the linear probability model where crime depends on age and the
parolee’s sex.
b. Are the results consistent with the claims of other studies with respect to age
and the parolee’s sex?
c. Predict the probability of a 25-year-old male parolee committing another crime;
repeat the prediction for a 25-year-old female parolee.
48. FILE Parole. Refer to the previous exercise for a description of the data set.
a. Estimate the logistic regression model where crime depends on age and the
parolee’s sex.
b. Are the results consistent with the claims of other studies with respect to age
and the parolee’s sex?
c. Predict the probability of a 25-year-old male parolee committing another crime;
repeat the prediction for a 25-year-old female parolee.
49. FILE Health_Insurance. According to the 2017 census, just over 90% of
Americans have health insurance (CNBC, May 22, 2018). However, a higher
percentage of Americans on the lower end of the economic spectrum are still
 without coverage. Consider a portion of data in the following table relating to
insurance coverage (1 for coverage, 0 for no coverage) for 30 working individuals
in Atlanta, Georgia. Also included in the table is the percentage of the premium
paid by the employer and the individual’s income (in $1,000s).

Premium
Insurance Income
Percentage

1 0 88

0 0 60

⋮ ⋮ ⋮

0 60 60

a. Analyze the linear probability model for insurance coverage with premium
percentage and income used as the predictor variables.
b. Consider an individual with an income of $60,000. What is the probability that
she has insurance coverage if her employer contributes 50% of the premium?
What if her employer contributes 75% of the premium?
50. FILE Health_Insurance. Refer to the previous exercise for a description of the
data set. Estimate the logistic regression model where insurance coverage
depends on premium percentage and income. Consider an individual with an
income of $60,000. What is the probability that she has insurance coverage if her
employer contributes 50% of the premium? What if her employer contributes 75%
of the premium?
51. FILE Assembly. Because assembly line work can be tedious and repetitive, it is
not suited for everybody. Consequently, a production manager is developing a
binary choice regression model to predict whether a newly hired worker will stay
in the job for at least one year (Stay equals 1 if a new hire stays for at least one
year, 0 otherwise). Three predictor variables will be used: (1) Age; (2) a Female
dummy variable that equals 1 if the new hire is female, 0 otherwise; and (3) an
Assembly dummy variable that equals 1 if the new hire has worked on an
assembly line before, 0 otherwise. The accompanying table shows a portion of
data for 32 assembly line workers.

Stay Age Female Assembly

0 35 1 0

0 26 1 0
 Stay Age Female Assembly

⋮ ⋮ ⋮ ⋮

1 38 0 1

a. Estimate and interpret the linear probability model and the logistic regression
model where being on the job one year later depends on Age, Female, and
Assembly.
b. Compute the accuracy rates of both models.
c. Use the preferred model to predict the probability that a 45-year-old female who
has not worked on an assembly line before will still be on the job one year later.
What if she has worked on an assembly line before?
52. FILE CFA. The Chartered Financial Analyst (CFA) designation is the de facto
professional certification for the financial industry. Employers
encourage their prospective employees to complete the CFA
page 299
exam. Daniella Campos, an HR manager at SolidRock Investment, is reviewing
10 job applications. Given the low pass rate for the CFA Level 1 exam, Daniella
wants to know whether or not the 10 prospective employees will be able to pass
it. Historically, the pass rate is higher for those with work experience and a good
college GPA. With this insight, she compiles the information on 263 current
employees who took the CFA Level I exam last year, including the employee’s
success on the exam (1 for pass, 0 for fail), the employee’s college GPA, and
years of work experience. A portion of the data is shown in the accompanying
table.

Pass GPA Experience

1 3.75 18

0 2.62 17

⋮ ⋮ ⋮

0 2.54 4

a. Estimate the linear probability model to predict the probability of passing the
CFA Level I exam for a candidate with a college GPA of 3.80 and five years of
experience.
b. Estimate the logistic regression model to predict the probability of passing the
CFA Level I exam for a candidate with a college GPA of 3.80 and five years of
experience.
53. FILE Admit. Unlike small selective colleges that pay close attention to personal
statements, teacher recommendations, etc., large, public state university systems
primarily rely on a student’s grade point average (GPA) and scores on the SAT or
ACT for the college admission decisions. Data were collected for 120 applicants
on college admission (Admit equals 1 if admitted, 0 otherwise) along with the
student’s GPA and SAT scores. A portion of the data is shown in the
accompanying table.

Admit GPA SAT

1 3.10 1550

0 2.70 1360

⋮ ⋮ ⋮

1 4.40 1320

a. Estimate and interpret the appropriate linear probability and the logistic
regression models.
b. Compute the accuracy rates of both models.
c. Use the preferred model to predict the probability of admission for a college
student with GPA = 3.0 and SAT = 1400. What if GPA = 4.0?
54. FILE Divorce. Divorce has become an increasingly prevalent part of American
society. According to a 2019 Gallup poll, 77% of U.S. adults say divorce is morally
acceptable, which is a 17-point increase since 2001 (Gallup, May 29, 2019). In
general, the acceptability is higher for younger adults who are not very religious. A
sociologist conducts a survey in a small Midwestern town where 200 American
adults are asked about their opinion on divorce (Acceptable equals 1 if morally
acceptable, 0 otherwise), religiosity (Religious equals 1 if very religious, 0
otherwise), and their age. A portion of the data is shown in the accompanying
table.

Acceptable Age Religious

1 78 0

1 20 0

⋮ ⋮ ⋮

1 22 0
 a. Estimate and interpret the appropriate linear probability and the logistic
regression models.
b. Compute the accuracy rates of both models.
c. Use the preferred model to predict the probability that a 40-year old, very
religious adult will find divorce morally acceptable. What if the adult is not very
religious?
55. FILE STEM. Several studies have reported lower participation in the science,
technology, engineering, and mathematics (STEM) careers by female and
minority students. A high school counselor surveys 240 college-bound students,
collecting information on whether the student has applied to a STEM field (1 if
STEM, 0 otherwise), whether or not the student is female (1 if female, 0
otherwise), white (1 if white, 0 otherwise), and Asian (1 if Asian, 0 otherwise). Also
included in the survey is the information on the student’s high school GPA and the
SAT scores. A portion of the data is shown in the accompanying table.

a. Estimate and interpret the logistic regression model using STEM as the
response variable, and GPA, SAT, White, Female, and Asian as the predictor
variables.
b. Find the predicted probability that a white male student will apply to a STEM
field with GPA = 3.4 and SAT = 1400. Find the corresponding probabilities for
an Asian male and a male who is neither white nor Asian.
c. Find the predicted probability that a white female student will apply to a STEM
field with GPA = 3.4 and SAT = 1400. Find the corresponding probabilities for
an Asian female and a female who is neither white nor Asian.

page 300

7.4 CROSS-VALIDATION METHODS

LO 7.4
Use cross-validation techniques to evaluate regression models.

In Section 6.2, we discussed goodness-of-fit measures—the

standard error of the estimate se, the coefficient of determination R2,
and the adjusted R2—that summarize how well the sample
regression equation fits the data in linear regression models.
Similarly, in Section 7.3, we assessed the performance of binary
choice models on the basis of their accuracy rates. All these
measures help us assess predictability in the sample data that was
also used to build the model. Unfortunately, these measures do not
help us gauge how well an estimated model will predict in an unseen
sample.
It is possible for a model to perform really well with the data set
used for estimation, but then perform miserably once a new data set
is used. Recall that in linear regression models, we can easily
improve the fit, as measured by R2, by including a whole bunch of
predictor variables, some of which may have no economic
justification. We often call modelling of this sort overfitting.
Overfitting occurs when an estimated model begins to describe the
quirks of the data rather than the real relationships between
variables. By making the model conform too closely to the given
data, the model’s predictive power is compromised. The overfit
model will fail to describe the behavior in a new sample that has its
own quirks. Although the adjusted R2 imposes a penalty for
additional predictor variables, the measure still suffers from the fact
that the same data are used for both estimation and assessment.

OVERFITTING
Overfitting occurs when a regression model is made overly
complex to fit the quirks of given sample data. By making the
model conform too closely to the sample data, its predictive
power is compromised.
 A useful method to assess the predictive power of a model is to
test it on a data set not used in estimation. Cross-validation is a
technique that evaluates predictive models by partitioning the
original sample into a training set to build (train) the model and a
validation set to evaluate (validate) it. Although training set
performance can be assessed, it may result in overly optimistic
estimates because it is based on the same data as were used to
build the model. Therefore, a validation set is used to provide an
independent performance assessment by exposing the model to
unseen data. Sometimes, the data are partitioned into an optional
third set called a test data set; further detail is provided in Chapters
8, 9, and 10.

CROSS-VALIDATION
Cross-validation is a technique in which the sample is
partitioned into a training set to estimate the model and a
validation set to assess how well the estimated model predicts
with unseen data.

In this section, we will examine two cross-validation methods: the

holdout method and the k-fold cross-validation method. For both
methods, it is common to use the root mean squared error (RMSE)
on the validation set to assess the linear regression models, where

and n* is the number of observations in the validation

set. Recall that RMSE is the square root of mean squared error
(MSE). In addition to RMSE, other important performance measures
include the mean absolute deviation (MAD) and the mean absolute
percentage error (MAPE); these measures are discussed in Chapter
8. To assess binary choice models, we use the accuracy rate
computed as the percentage of correctly classified observations in
the validation set; other performance measures are discussed in
Chapter 8.
 page 301

Note that the predictability is likely to be higher in the training set

as compared to the validation set because the model is built from the
training set. In other words, the RMSE will be lower, and the
accuracy rate will be higher, in the training set as compared to the
validation set. In general, a large discrepancy between the RMSE (or
the accuracy rates) of the training and validation sets is indicative of
the overfitting problem discussed earlier; the model fits the training
data set very well but fails to generalize to an unseen data set.

The Holdout Method

The simplest of the cross-validation methods is the holdout method.
The sample data set is partitioned into two independent and mutually
exclusive data sets—the training set and the validation set. Figure
7.15 shows the partitioning of the sample data into two distinct sets.
There is no rule as to how the sample data should be partitioned.
Ideally, we should make random draws when partitioning the data.
For illustration purposes, we will use the first 75% of the
observations for the training set and the remaining 25% of the
observations for the validation set. It is easy to make random draws
in Analytic Solver and R for the partitioning of the data.

FIGURE 7.15 Partitioning the sample for the holdout method

The holdout method is implemented as follows:

. We partition the sample data into two parts, labeled training set and
validation set.
. We use the training set to estimate competing models.
. We use the estimates from the training set to predict the response
variable in the validation set.
. We calculate the RMSE (or the accuracy rate) for each competing
model. The preferred model will have the smallest RMSE (or the
largest accuracy rate).
Ideally, we would like the model with the best performance in the
training set to also have the best performance in the validation set.
For example, we would like that the linear regression model with the
highest adjusted R2 also have the lowest RMSE in the validation set.
As discussed earlier, conflicting results between the two is a sign of
overfitting. Data analytics professionals generally rely on cross-
validation for model selection if they have a sufficient number of
observations in the training and the validation sets for effective
estimation and assessment, respectively. After selecting the best
model, we generally reestimate this model with the entire data set for
making predictions.
FILE
Gender_Gap

EXAMPLE 7.12
The objective outlined in the introductory case is to analyze the
gender gap in salaries of project managers. Consider two
models to analyze a project manager’s salary. For predictor
variables, Model 1 uses Size, Experience, Female, and Grad,
whereas Model 2 also includes the interactions between
Female with Experience, Female with Grad, and Size with
Experience. Use the holdout method to compare the
predictability of both models using the first 150 observations for
training and the remaining 50 observations for validation.
SOLUTION:
We use the training set with 150 observations to estimate
Model 1 and Model 2. The estimates are presented in Table
7.21.

page 302

TABLE 7.21 Estimates of the Competing Models, Example 7.12

Variable Model 1 Model 2

Constant   7.4453 56.1077

Size   0.1136 −0.1185

Experience   4.2250  1.0891

Female −23.2909  7.3771

Grad  15.6116 15.2640

Female × Experience NA −2.0428

Female × Grad NA  2.2042

Size × Experience NA  0.0150

Adjusted R2   0.6896  0.7284

Notes: Parameter estimates; NA denotes not applicable.

Based on the training set, Model 2 is more suitable for

prediction because it has a higher adjusted R2 than Model 1
(0.7284 > 0.6896). There is no guarantee, however, that Model
2 will also have a better predictability in the validation set. We
 use the estimated models to predict the response variable in

the validation set and compute the where n* =

50 is the number of observations in the validation set. A portion

of the calculations is shown in Table 7.22.

TABLE 7.22 Analysis of the Holdout Method, Example 7.12

Observation y

151 117 113.2913 105.8106

152 125 121.3257 126.1153

⋮ ⋮ ⋮ ⋮

200 117 110.6685 110.4045

RMSE 11.31  12.54 

Interestingly, Model 2, with a better performance in the training

set, does poorer in the validation set (12.54 > 11.31). Based on
the results, Model 1 would be preferred for prediction.

FILE
AnnArbor

EXAMPLE 7.13
In Example 7.8, we used in-sample measures to compare two
models, where Model 1 used Rent and Model 2 used ln(Rent)
as the response variable. Recall that these models were
referred to as Models 2 and 4, respectively, and the predictor
variables for both models included Beds, Baths, and ln(Sqft).
Use the holdout method to compare the predictability of the two
competing models, using the first 30 observations for training
and the remaining 10 observations for validation.
SOLUTION:
We use the training set with 30 observations to estimate Model
1 and Model 2. The estimates are presented in Table 7.23.

page 303

TABLE 7.23 Estimates of the Competing Models, Example 7.13

Variable Model 1 Model 2

Constant −2674.2552 2.5742

Beds   123.6672 0.0975

Baths    18.0267 −0.0025 

Ln(Sqft)   496.9671 0.6004

se   104.2082 0.0854

Notes: Standard error of the estimate se is used for making predictions with Model
2.
 We use the estimated models to predict the response variable

in the validation set and compute the resulting

where n* = 10 is the number of observations in the validation

set. A portion of the calculations is shown in Table 7.24.

TABLE 7.24 Analysis of the Holdout Method, Example 7.13

Observation y

31 1518 1447.4587 1523.5103

32 1600 1328.9161 1385.7901

⋮ ⋮ ⋮ ⋮

40 2400 1668.3540 2013.7291

RMSE 362.52 422.25

Thus, Model 1 is considered a better predictor than Model 2

(362.52 < 422.25) based on this cross-validation method.
These results are consistent with the analysis with in-sample
measures in Example 7.8.

In Example 7.14, we use the holdout method to assess binary

choice models.
FILE
Spam
 EXAMPLE 7.14
Peter Derby works as a cyber security analyst at a private
equity firm. His colleagues at the firm have been inundated by
a large number of spam e-mails. Peter has been asked to
implement a spam detection system on the company’s e-mail
server. He analyzes a sample of 500 spam and legitimate e-
mails with the following relevant variables: spam (1 if spam, 0
otherwise), the number of recipients, the number of hyperlinks,
and the number of characters in the message. A portion of the
data is shown in the accompanying table.

TABLE 7.25 Spam Data (n = 500)

a. Use the holdout method to compare the accuracy rates of two

competing logistic models for spam, using the first 375
observations for training and the remaining 125 page 304
observations for validation. Model 1 uses
Recipients, Hyperlinks, and Characters as predictor variables,
whereas Model 2 drops the predictor variable(s) found to be
statistically insignificant in Model 1.
b. Re-estimate the preferred model to predict the probability of
spam if the number of recipients, hyperlinks, and characters
are 20, 5, and 60, respectively.
SOLUTION:

a. We use the training set with 375 observations to estimate two

logistic models; see Table 7.26 for the estimates. Note that in
Model 1, the variable, Characters, is not statistically significant
at the 5% level and is, therefore, dropped in Model 2.
TABLE 7.26 Estimates of the pam Data (n = 500)

Variable Model 1 Model 2

Constant −5.1778* −5.8045*

(0.000) (0.000)

Recipients 0.1765* 0.1806*

(0.000) (0.000)

Hyperlinks 0.5473* 0.5402*

(0.000) (0.000)

Characters −0.0104 NA
(0.074)

Notes: Parameter estimates are followed with the p-values in

parentheses; NA denotes not applicable; * represents
significance at the 5% level.
To compute the accuracy rates of the estimated logistic binary
choice models, we first compute the predicted probabilities of
the models in the validation set; a portion of the calculations is
shown in Table 7.27.

TABLE 7.27 Analysis of the Holdout Method, Example 7.14

 By converting the predicted probabilities into binary
predictions and comparing them with the actual spam data, we
find that the accuracy rates for Model 1 and Model 2 are
68.00% and 65.60%, respectively. Based on these rates, we
infer that Model 1 is superior to Model 2 (68.00% > 65.60%).
b. We re-estimate Model 1 with all 500 observations to predict
the spam probability with the number of recipients, hyperlinks,
and characters as 20, 5, and 60, respectively, as:

page 305

Using Analytic Solver and R for the Holdout

Method for the Logistic Regression Model
FILE
Spam

Analytic Solver and R are easy to use for partitioning the data,
estimating models with the training set, and deriving the necessary
cross-validation measures. As mentioned earlier, we generally use
random draws for partitioning the sample data into the training and
validation sets. Here, for replicating purposes, we will continue to
use the latter part of the sample data for the validation set. We use
the Spam data to replicate the results in Example 7.14 for Model 1;
results for Model 2 can be derived similarly.

Using Analytic Solver

. Open the Spam data file. In column F of the data file, create a
variable called Flag with the letter T for the first 375 observations
and the letter V for the remaining 125 observations; here T and V
denote training and validation, respectively.
. From the menu, choose Data Mining > Partition > Standard
Partition.
Specify the Data range by highlighting cells A1:F501. Select and
move variables Spam, Recipients, Hyperlinks, and Characters to
the Selected Variables box. Although there is an option to pick up
rows randomly, for Partitioning Options, we select Use partition
variable. Select and move Flag to the Use partition variable box;
see the relevant portion of the screen shot in Figure 7.16. Click OK.
The STDPartition worksheet now contains partitioned data with 375
observations in the training set and 125 observations in the
validation set.

FIGURE 7.16 Partitioning the sample with Analytic Solver

 Source: Microsoft Excel

. Make sure the STDPartition worksheet is active, then choose Data

Mining > Classify > Logistic Regression. Select and move
variables Recipients, Hyperlinks, and Characters to the Selected
Variables box and Spam to the Output Variable box. Click Finish. In
the LogReg_ValidationScore worksheet, you will find Accuracy
(%correct) equal to 68. This is the same as derived for Model 1 in
Example 7.14.

page 306

Using R
. Import the Spam data into a data frame (table) and label it myData.
. We partition the sample into training and validation sets, labeled
TData and VData, respectively. Enter:
 > TData <- myData[1:375,]
> VData <- myData[376:500,]
. We use the training set, TData, to estimate Model 1. Enter:
> Model1 <- glm(Spam ∼ Recipients+Hyperlinks+Characters,
family=binomial(link = logit), data = TData)
. We use the estimates to make predictions for VData and then
convert them into a binary prediction. Finally, we compute the
accuracy rate in the validation set. Enter:
> Pred1 <- predict(Model1, VData, type=“response”)
> Binary1 <- round(Pred1)
> 100*mean(VData$Spam == Binary1)
R returns: [1] 68
This is the same as derived for Model 1 in Example 7.14.

The k-Fold Cross-Validation Method

The main limitation of the holdout method is that the experiment is
performed only once where the estimates obtained from the training
set are used to make predictions for the response variable with the
validation set. Therefore, the choice of the model will be sensitive to
how the data are partitioned. A popular method that is less sensitive
to data partition is called the k-fold cross-validation method. Here,
the original sample is partitioned into k subsets, and the one that is
left out in each iteration is the validation set. In other words, we
perform the holdout method k times. In Figure 7.17, we show the
experiments with k = 4.

FIGURE 7.17 Partitioning the sample for the k-fold method

As mentioned above, in each experiment, one of the k subsets is
used as the validation set and the remaining k − 1 subsets are put
together to form a training set. The average of the assessment
measures is used to select the best-performing model. Note that the
greater the k, the greater will be the reliability of the k-fold method
and the greater will be its computational cost. When k equals the
sample size, the resulting method is also called the leave-one-out
cross-validation method, where you leave out just one observation
for the validation set.

THE HOLDOUT AND THE k-FOLD CROSS-VALIDATION

METHODS
In the holdout method, the sample data set is partitioned into
two independent and mutually exclusive data sets—the training
set and the validation set. In the k-fold cross-validation method,
the sample data are partitioned into k subsets, where one of the
k subsets is used as the validation set and the page 307
remaining k − 1 subsets are put together to form a
training set. By averaging the assessment measures, the k-fold
cross-validation method is less sensitive to data partitioning
than the holdout method.

FILE
Gender_Gap
EXAMPLE 7.15
In Example 7.12, we used the holdout method to assess two
models for analyzing a project manager’s salary. For
predictors, Model 1 used Size, Experience, Female, and Grad.
Model 2 extends the model by also including the interactions
between Female with Experience, Female with Grad, and Size
with Experience. Use the k-fold cross-validation method to
compare the predictability of the models using k = 4.
SOLUTION:
We assess both models four times with the validation set
formed by the observations 151–200, 101–150, 51–100, and
1–50, respectively. Each time the training set includes the
remaining observations for estimating the regression models.

In Table 7.28, we also present for both models,

where n* = 50 is the number of observations in the validation

set.

TABLE 7.28 RMSE for the k-Fold Cross-Validation Method with k = 4, Example
7.15

Observations in Validation Set Model 1 Model 2

151–200 11.31 12.54

101–150 12.26 11.88

51–100 13.30 12.09

1–50 13.76 13.46

 Observations in Validation Set Model 1 Model 2

Average 12.66 12.49

In Example 7.12 we had noted that Model 1 performed better in

the validation set even though it did poorer in the testing set.
With the k-fold cross-validation method, we observe that the
performance of Model 1 is better only when observations 151–
200 are used in the validation set. For every other partitioning,
Model 2 is better, giving it a lower average RMSE (12.49 <
12.66). We conclude that Model 2 is a superior model for
prediction. Whenever possible, it is better to use the more
reliable k-fold cross-validation method.

We can similarly assess the predictability of binary choice models

with the k-fold cross-validation method, using the average accuracy
rate for model comparison.
As you may have noticed, the k-fold cross-validation method is
quite involved. Software packages such as R provide several ways
to automatize the process. In this chapter, we focus primarily on the
partitioning based on fixed draws of the observations. For instance,
in Example 7.15, we partitioned the data into four equal parts, based
on how the data appeared in the sample. Software packages often
use random draws of the observations for the partitioning.
Consequently, the results from these packages are not identical to
the ones we obtain with fixed partitioning; for the most part, however,
the decision is not impacted by the method used. In Appendix 7.1,
we introduce the popular caret package in R for implementing the k-
fold cross-validation method, which we will revisit in Chapters 9 and
10.

page 308

EXERCISES 4
 EXERCISES 7.4
Mechanics
56. FILE Exercise_7.56. The accompanying data file contains 40 observations on
the response variable y along with the predictor variables x and d. Consider two
linear regression models where Model 1 uses the variables x and d and Model 2
extends the model by including the interaction variable xd. Use the holdout
method to compare the predictability of the models using the first 30 observations
for training and the remaining 10 observations for validation.
57. FILE Exercise_7.57. The accompanying data file contains 40 observations on
the response variable y along with the predictor variables x1 and x2. Use the
holdout method to compare the predictability of the linear model with the
exponential model using the first 30 observations for training and the remaining
10 observations for validation.
58. FILE Exercise_7.58. The accompanying data file contains 40 observations on
the binary response variable y along with the predictor variables x1 and x2. Use
the holdout method to compare the accuracy rates of the linear probability model
with the logistic regression model using the first 30 observations for training and
the remaining 10 observations for validation.

Applications
59. FILE IceCream. The accompanying data file contains 35 observations for an ice
cream truck driver’s daily income (Income in $), number of hours on the road
(Hours), whether it was a particularly hot day (Hot = 1 if the high temperature was
above 85°F, 0 otherwise), and whether it was a Holiday (Holiday = 1, 0 otherwise).
Consider two models where Model 1 predicts Income on the basis of Hours, Hot,
and Holiday and Model 2 also includes the interaction between Hot and Holiday.
Use the holdout method to compare the predictability of the models using the first
24 observations for training and the remaining 11 observations for validation.
60. FILE Mobile_Devices. The accompanying data file contains survey data for 80
participants with information on average daily time spent on mobile devices
(Usage, in minutes), household income (Income, in $1,000s), lived in a rural area
(Rural =1 if rural, 0 otherwise), and had a college degree (College = 1 if college
graduate, 0 otherwise). Consider two predictive models for mobile device usage
where Model 1 is based on Income, Rural, and College and Model 2 also includes
the interaction between Rural and College.
a. Use the holdout method to compare the predictability of the models using the
first 60 observations for training and the remaining 20 observations for
validation.
b. Use the k-fold cross-validation method to compare the predictability of the
models using k = 4.
61. FILE BMI. The accompanying data file contains salary data (in $1,000s) for 30
college-educated men with their respective BMI and a dummy variable that
represents 1 for a white man and 0 otherwise. Model 1 predicts Salary using BMI
and White as predictor variables. Model 2 includes BMI and White along with an
interaction between the two.
a. Use the holdout method to compare the predictability of the models using the
first 20 observations for training and the remaining 10 observations for
validation.
b. Use the k-fold cross-validation method to compare the predictability of the
models using k = 3.
62. FILE Pick_Errors. The accompanying data file contains information on 30
employee’s annual pick errors (Errors), experience (Exper in years), and whether
or not the employee attended training (Train equals 1 if the employee attended
training, 0 otherwise). Model 1 predicts Errors using Exper and Train as predictor
variables. Model 2 includes Exper and Train along with an interaction between the
two.
a. Use the holdout method to compare the predictability of the models using the
first 20 observations for training and the remaining 10 observations for
validation.
b. Use the k-fold cross-validation method to compare the predictability of the
models using k = 3.
63. FILE Health_Factors. The accompanying data file contains survey information
on 120 American adults who rated their health (Health) and social connections
(Social) on a scale of 1 to 100. The file also contains information on their
household income (Income, in $1,000s) and college education (College equals 1
if they have completed a bachelor’s degree, 0 otherwise). Consider two linear
regression models for Health. Model 1 uses Social, Income, and College as the
predictor variables, whereas Model 2 also includes the interactions of Social with
Income and Social with College.
a. Use the holdout method to compare the predictability of the models using the
first 90 observations for training and the remaining 30 observations for
validation.
b. Use the k-fold cross-validation method to compare the predictability of the
models using k = 4.
64. FILE Rental. The accompanying data file contains monthly data on rent (Rent, in
$) along with square footage (Sqft), number of bedrooms (Bed), and number of
bathrooms (Bath) of 80 rental units. Consider two linear regression models for
Rent. Model 1 uses Bed, Bath, and Sqft, whereas Model 2 also allows the
interaction between Bed and Sqft.
a. Use the holdout method to compare the predictability of the models using the
first 60 observations for training and the remaining 20 observations for
validation.
b. Use the k-fold cross-validation method to compare the predictability of the
models using k = 4.
 page 309
65. FILE Crew_Size. The accompanying data file contains weekly
data on crew size (the number of workers) and productivity (jobs/week) over the
past 27 weeks. A linear regression model and a quadratic regression model are
considered for predicting productivity on the basis of crew size.
a. Use the holdout method to compare the predictability of the models using the
first 18 observations for training and the remaining 9 observations for validation.
b. Use the k-fold cross-validation method to compare the predictability of the
models using k = 3.
66. FILE Happiness. The accompanying data file contains information on 100
working adults’ self-assessed happiness on a scale of 0 to 100, along with their
age and annual income. Consider two quadratic models for happiness. Model 1 is
based on age, age2, and income. Model 2 is based on age, age2, and ln(income).
a. Use the holdout method to compare the predictability of the models using the
first 75 observations for training and the remaining 25 observations for
validation.
b. Use the k-fold cross-validation method to compare the predictability of the
models using k = 4.
67. FILE Electricity. Consider a regression model for predicting monthly electricity
cost (Cost in $) on the basis of average outdoor temperature (Temp in °F),
working days per month (Days), and tons of product produced (Tons). The
accompanying data file contains data on 80 observations. Use the holdout
method to compare the predictability of the linear and the exponential regression
models using the first 60 observations for training and the remaining 20
observations for validation.
68. FILE Arlington_Homes. The accompanying data file contains information on the
sale price (in $) for 36 single-family homes in Arlington, Massachusetts. In order
to analyze house price, the predictor variables include the house’s square footage
(Sqft), the number of bedrooms (Beds), the number of bathrooms (Baths), and
whether or not it is a colonial (Col = 1 if colonial, 0 otherwise).
a. Use the holdout method to compare the predictability of the linear and the
exponential models using the first 24 observations for training and the
remaining 12 observations for validation.
b. Use the k-fold cross-validation method to compare the predictability of the
models using k = 3.
69. FILE Purchase. Consider the accompanying data to predict an Under Armour
purchase (Purchase; 1 for purchase, 0 otherwise) on the basis of customer age.
a. Use the holdout method to compare the accuracy rates of the linear probability
model (Model 1) and the logistic regression model (Model 2) using the first 20
observations for training and the remaining 10 observations for validation.
b. Use the k-fold cross-validation method to compare the accuracy of the models
using k = 3.
70. FILE Divorce. Consider the accompanying data to analyze how people view
divorce (Acceptable equals 1 if morally acceptable, 0 otherwise) on the basis of
age and religiosity (Religious equals 1 if very religious, 0 otherwise).
a. Use the holdout method to compare the accuracy rates of two competing
logistic models for divorce, using the first 150 observations for training and the
remaining 50 observations for validation. Model 1 uses Age and Religious as
predictor variables, whereas Model 2 also incudes the interaction between Age
and Religious.
b. Use the k-fold cross-validation method to compare the accuracy of the models
using k = 4.

7.5 WRITING WITH BIG DATA

FILE
College_Admission

As mentioned in Chapter 6, when using regression analysis with big data, the
emphasis is placed on the model’s predictability and not necessarily on tests of
significance. In this section, we will use logistic regression models for predicting
college admission and enrollment decisions. Before running the models, we have to
first filter out the College_Admission data to get the appropriate subset of
observations for selected variables. We encourage you to replicate the results in the
report.

Case Study
Create a sample report to analyze admission and enrollment decisions at the school
of arts & letters in a selective four-year college in North America. For predictor
variables, include the applicant’s sex, ethnicity, grade point average, and SAT scores.
Make predictions for the admission probability and the enrollment probability using
typical values of the predictor variables.

page 310
Sample Report—College Admission and
Enrollment
College admission can be stressful for both students and parents as there is
no magic formula when it comes to admission decisions. Two important
factors considered for admission are the student’s high school record and
performance on standardized tests. According to the National Association for
College Admission Counseling (NACAC), a student’s high school record
carries more weight than standardized test scores.

Rawpixel.com/Shutterstock

Just as prospective students are anxious about receiving an acceptance

letter, most colleges are concerned about meeting their enrollment targets.
The number of acceptances a college sends out depends on its enrollment
target and admissions yield, defined as the percentage of students who enroll
at the school after being admitted. It is difficult to predict admissions yield as it
depends on the college’s acceptance rate as well as the number of colleges
to which students apply. As the number of applications for admission and the
number of acceptances increase, the yield decreases.
In this report, we analyze factors that affect the probability of college
admission and enrollment at a school of arts & letters in a selective four-year
college in North America. Predictors include the applicant’s high school GPA,
SAT score,1 and the Male, White, and Asian dummy variables capturing the
applicant’s sex and ethnicity. In Table 7.29, we present the representative
applicant profile.
TABLE 7.29 Applicant Profile for the School of Arts & Letters

Variable Applied Admitted Enrolled

Male Applicant (%) 30.76 27.37 26.68

White Applicant (%) 55.59 61.13 69.83

Asian Applicant (%) 12.42 11.73 8.73

Other Applicant (%) 31.99 27.14 21.45

High School GPA 3.50 3.86 3.74

(Average)

SAT Score (Average) 1,146 1,269 1,229

Number of 6,964 1,739 401

Applicants

Of the 6,964 students who applied to the school of arts & letters, 30.76%
were males; in addition, the percentages of white and Asian applicants were
55.59% and 12.42%, respectively, with about 32.00% from other ethnicities.
The average applicant had a GPA of 3.50 and an SAT score of 1146. Table
7.29 also shows that 1,739 (or 24.97%) applicants were granted admission,
of which 401 (23.06%) decided to enroll. As expected, the average GPA and
SAT scores of admitted applicants are higher than those who applied and
those who enrolled, but to a lesser extent.
Two logistic regression models are estimated using the same predictor
variables, one for predicting the admission probability and the other for
predicting the enrollment probability. The entire pool of 6,964 applicants is
used for the first regression, whereas 1,739 admitted applicants are used for
the second regression. The results are presented in Table 7.30.

TABLE 7.30 Logistic Regressions for College Admission and

Enrollment
 Variable Admission Enrollment

Constant −17.5732* 7.2965*

(−37.41) (8.48)

Male Dummy Variable 0.0459 −0.1433

(0.61) (−1.05)

White Dummy Variable −0.3498* 0.7653*

(−4.43) (5.15)

Asian Dummy Variable −0.4140* −0.0074

(−3.57) (−0.03)

High School GPA 2.7629* −1.4265*

(25.74) (−7.17)

SAT Score 0.0056* −0.0028*

(20.93) (−5.99)

Accuracy (%) 81 77

Number of Observations 6,964 1,739

Notes: Parameter estimates are in the top half of the table with the z-statistics
given in parentheses; * represents significance at the 5% level. Accuracy (%)
measures the percentage of correctly classified observations.

With accuracy rates of 81% and 77%, respectively, both models do a good
job with predicting probabilities. It seems that the sex of the applicant plays
no role in the admission or enrollment decisions.
Interestingly, both white and Asian applicants have a
page 311
lower probability of admission than those from other ethnicities. Perhaps this
is due to affirmative action, whereby colleges admit a proportionally higher
percentage of underrepresented applicants. As expected, quality applicants,
in terms of both GPA and SAT, are pursued for admission.
On the enrollment side, admitted applicants who are white are more likely
to enroll than all other admitted applicants. Finally, admitted applicants with
high GPA and high SAT scores are less likely to enroll at this college. This is
not surprising because academically strong applicants will have many offers,
which lowers the probability that an applicant will accept the admission offer
of a particular college.
In order to further interpret the influence of SAT scores on college
admission and enrollment, we compute predicted admission and enrollment
probabilities for representative males from all ethnicities with a GPA of 3.8
and SAT scores varying between 1000 and 1600. The results are shown in
Figures 7.18 and 7.19.

FIGURE 7.18 Predicted Admission Probability

FIGURE 7.19 Predicted Enrollment Probability

 Consider the case of a representative male applicant with an SAT score of
1300. For a white male, the predicted probabilities of admission and
enrollment are 47% and 24%, respectively. The corresponding probabilities
are 45% and 13%, respectively, for Asians and 55% and 13%, respectively,
for all other ethnicities. The probabilities get closer as SAT scores increase.
Higher admission rate for underrepresented applicants is
consistent with the admission practices at other colleges
page 312
that believe that diversity enriches the educational experience for all.
Unfortunately, despite a higher admission rate, the lower admission yield
at this college for underrepresented applicants makes the percentage of
those who actually enroll even lower than the percentage of those who apply.
As we can observe in Table 7.18, about 23% of underrepresented applicants
applied for admission, but the percentage of those who enrolled dropped to
21.45%. This is consistent with an article in The New York Times (August 24,
2017) that found that black and Hispanic students are more underrepresented
at the nation’s top colleges and universities than they were 35 years ago. The
yield for Asian applicants is the lowest at 8.73%. It is advised that the college
explore the reasons for the especially low admission yield of nonwhite
applicants and finds ways to raise it.

Suggested Case Studies

Many predictive models can be estimated and assessed with the big data that
accompany this text. Here are some suggestions.

Report 7.1 FILE College_Admissions. Choose a college of interest and use the
sample of enrolled students to best predict a student’s college grade point average.
Explore interactions of the relevant predictor variables and use cross validation to
 select the best predictive model. In order to estimate these models, you have to first
filter the data to include only the enrolled students.

Report 7.2 FILE House_Price. Choose two comparable college towns. Develop a
predictive model for the sale price of a house for each college town. Explore log-
linear transformations, dummy variables, and interactions of the relevant predictor
variables. Use cross validation to select the best predictive model.

Report 7.3 FILE NBA. Develop a model for predicting a player’s salary. For
predictor variables, consider age (quadratic effect), height, weight, and relevant
performance measures. Use cross validation to assess the choice of predictor
variables as well as the functional form (linear or exponential). In order to estimate
these models, you will have to first filter the data to include only career statistics
based on regular seasons. Exclude players with no information on salary.

Report 7.4 FILE Longitudinal_Survey. Develop a logistic regression model for

predicting if the respondent is outgoing in adulthood. Use cross validation to select
the appropriate predictor variables. In order to estimate this model, you have to first
handle missing observations using the missing or the imputation strategy.

Report 7.5 FILE TechSales_Reps. The net promoter score (NPS) is a key indicator
of customer satisfaction and loyalty. Use data on employees in the software product
group with a college degree to develop the logistic regression model for predicting if
a sales rep will score an NPS of 9 or more. Use cross validation to select the
appropriate predictor variables. In order to estimate this model, you have to first
construct the (dummy) target variable, representing NPS ≥ 9 and subset the data to
include only the employees who work in the software product group with a college
degree.

page 313

APPENDIX 7.1 The Caret Package in R for

the k-fold Cross-Validation Method
The caret package (short for Classification And Regression
Training) in R is a set of functions that attempt to streamline the
process for creating predictive models. It is extremely useful for
 implementing the k-fold cross-validation method. We introduce it
here, and will revisit it in Chapters 9 and 10. (As discussed in
Appendix C, we note that the following instructions are based on
R version 3.5.3. They may not work for different versions of R.)

Using the Caret Package to Assess the Linear Regression

Model
FILE
Gender_Gap

For illustration, we will use the Gender_Gap data to replicate the

results of Example 7.15.
A. Import the Gender_Gap data into a data frame (table) and label
it myData.
B. Install and load the caret package. Enter:
> install.packages(“caret”)
> library(caret)
On some computers, you might also need to install other
packages that support the caret package using the command
> install.packages(“caret”, dependencies = c(“Depends”,
“Suggests”)).
C. We use the trainControl and the train functions. For options
within the trainControl function, we use method to specify the
sampling method (here denoted as “cv” for cross-validation) and
number to indicate the number of folds. Enter:
> myControl <- trainControl(method = “cv”, number = 4)
D. Within the train function, we specify the model, and then set the
following options: data (to indicate the data frame), method (here
denoted as “lm” for linear model) and trControl (to indicate the
variable defined when using the trainControl function). We show
the steps for Model 1 and Model 2. Enter:
> Model1 <- train(Salary∼ Size + Experience + Female + Grad,
data = myData, method = “lm”, trControl = myControl)
> Model1
> Model2 <- train(Salary∼ Size + Experience + Female + Grad +
Female*Experience + Female*Grad + Size*Experience, data
= myData, method = “lm”, trControl = myControl)
> Model2
R reports the average RMSE as well as other measures. The
average RMSE is 12.60 for Model 1 and 12.38 for Model 2 (results
will vary). These values are fairly close but not identical to those
reported in Table 7.28. The reason for the difference is that R uses
random draws for determining the training and validation sets,
whereas we used fixed draws in Example 7.15; re-running the R
command will give you a slightly different result. With both
methods, however, we find that Model 2 is better for prediction
purposes.
We can easily modify the trainControl function to implement a
repeated k-fold cross-validation method. For instance, suppose
we wanted to implement a four-fold cross-validation that is
repeated five times. We would modify the earlier function as:

> myControl <- trainControl(method = “repeatedcv”, number

= 4, repeats = 5).

Using the Caret Package to Assess the Logistic

Regression Model
FILE
Spam

For illustration, we use the Spam data that was used to assess
Model 1 and Model 2 in Example 7.14. The first three steps are
similar to those used when assessing the linear regression model.

page 314
A. Import the Spam data into a data frame (table) and label it
myData.
B. Install and load the caret package. Enter:
> install.packages(“caret”)
> library(caret)
C. > myControl <- trainControl(method = “cv”, number = 4)
D. Before estimating the logistic regression model, we must convert
y from numeric type into factor type so that R treats it as a
categorical variable with two classes; in other words, a dummy
variable. We use the as.factor function to accomplish this task.
Enter:
> myData$Spam <- as.factor(myData$Spam)
> Model1 <- train(Spam ~ Recipients + Hyperlinks + Characters,
data = myData, trControl = myControl, method = “glm”, family
= binomial(link=logit), metric = “Accuracy”)
> Model1
> Model2 <- train(Spam ~ Recipients + Hyperlinks, data =
myData, trControl = myControl, method = “glm”, family =
binomial(link=logit), metric = “Accuracy”)
> Model2

R reports the accuracy of 78.59% for Model 1 and 76.60% for

Model 2 (results will vary). As in Example 7.14, we find that Model
1 is better for making predictions.

Again, we can implement a four-fold cross-validation that is

repeated five times by modifying the earlier function as:

> myControl <- trainControl(method = “repeatedcv”, number =

4, repeats = 5).
1The higher of SAT and ACT scores is included in the data where for comparison, ACT
scores on reading and math are first converted into SAT scores.

page 315
 page 316

8 Introduction to Data
Mining

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 8.1  Describe the data mining process.
LO 8.2  Implement similarity measures.
LO 8.3  Assess the predictive performance of data
mining models.
LO 8.4  Conduct principal component analysis.

ata mining is the process of applying analytical techniques to a

D data set to find hidden structures, patterns, and relationships
among variables. These techniques go beyond data
visualization and summary measures discussed in Chapter 3 and
linear and logistic regression models discussed in Chapters 6 and 7.
In this chapter, we first provide an overview of data mining, the data
mining process, and the two primary data mining techniques:
supervised (or predictive or directed) and unsupervised (or
descriptive or undirected) data mining.
We then explore three key issues that are relevant to data
mining: similarity measures, predictive performance assessment,
and dimension reduction. Similarity measures gauge whether a
group of observations are similar or dissimilar to one another and are
essential ingredients for data mining techniques. Assessment of
predictive performance is a critical step for model selection in
supervised data mining where we partition the original sample into a
training set to build (train) the model, a validation set to assess
(evaluate) the model, and an optional test data set. Several
performance measures are developed to assess how well an
estimated model performs in an unseen sample.
Finally, principal component analysis, a dimension reduction
technique, is discussed. This technique reduces the number of
variables in the data set to a smaller set of principal components that
retain most of the crucial information in the original data. These
concepts are the prerequisite knowledge for understanding the
supervised and unsupervised data mining techniques discussed later
in Chapters 9, 10, and 11.

page 317
©easy camera/Shutterstock

INTRODUCTORY CASE
Social Media Marketing
Alissa Bridges is the marketing director of FashionTech, an
online apparel retailer that specializes in activewear for both
men and women. The target market for FashionTech includes
individuals between the ages of 18 and 35 with an active
and/or outdoors lifestyle who look for both fashion and value in
their apparel purchase. The company markets its products via
a variety of media channels including TV ads, quarterly
catalogs, product placements, search engines, and social
media. Alissa has hired a social media marketing firm,
 MarketWiz, to develop predictive models that would help
FashionTech acquire new customers as well as increase sales
from existing customers. Using FashionTech’s historical social
media marketing and sales data, MarketWiz develops two
types of predictive models.
A classification model that predicts potential customers’
purchase probability from FashionTech within 30 days of
receiving a promotional message in their social media
account.
Two prediction models that predict the one-year purchase
amounts of customers acquired through social media
channels.
In order to assess the performance of the predictive models,
Alissa’s team would like to use the validation data set to:
1. Evaluate how accurately the classification model classifies
potential customers into the purchase and no-purchase
classes.
2. Compare the performance of prediction models that estimate
the one-year purchase amounts of customers acquired
through social media channels.
A synopsis of this introductory case is presented at the end of
section 8.3.

page 318

8.1 DATA MINING OVERVIEW

The terms artificial intelligence, machine learning, and data mining
describe applications of computer software used to obtain insightful
solutions that traditional data analysis techniques may not be able to
achieve. In a very broad sense, artificial intelligence is used to
describe computer systems that demonstrate human-like intelligence
and cognitive functions, such as deduction, pattern recognition, and
the interpretation of complex data.
In 1959, IBM computer scientist Arthur Samuel coined the term
machine learning. It describes an application of artificial intelligence
that allows the computer to learn automatically without human
intervention or assistance. Machine learning techniques can uncover
hidden patterns and relationships in data and use self-learning
algorithms to evaluate results and improve performance over time.
Uber, a popular ride-hailing service, uses machine learning
algorithms to predict rider demand to strategically dispatch drivers to
various locations.
Data mining describes the process of applying a set of analytical
techniques necessary for the development of machine learning and
artificial intelligence. The goal of data mining is to uncover hidden
patterns and relationships in data, which allows us to gain insights
and derive relevant information to help make decisions. As we will
discuss in the next three chapters, data mining techniques are used
for data segmentation, pattern recognition, classification, and
prediction. For example, a retail company might be interested in
studying consumer behavior and group customers into different
market segments so that it can design and customize a promotion
campaign unique to each segment. The company can also use
information from previous promotion campaigns to help predict future
behaviors of existing and prospective customers.
In many ways, the definitions of these terms overlap, and there
are no definitive boundaries among them. For example, a public
transportation department might use data mining tools to analyze
traffic data from the previous months to set general guidelines for the
timing of traffic lights throughout a city, while employing machine
learning algorithms to make small adjustments on individual traffic
lights based on real-time traffic information during rush hours, and
use artificial intelligence techniques to make decisions when dealing
with occasional problems with traffic congestion and accidents that
involve complex circumstances. As such, data mining is often
recognized as a building block of machine learning and artificial
intelligence.

ARTIFICIAL INTELLIGENCE, MACHINE LEARNING, AND

DATA MINING
The terms artificial intelligence, machine learning, and data
mining are often grouped together or used interchangeably
because their definitions tend to overlap with no clear
boundaries. The following general definitions for these terms
simply reinforce this fact:
Artificial intelligence is a computer system that can
demonstrate human-like intelligence and cognitive functions,
such as deduction, pattern recognition, and the interpretation
of complex data.
Machine learning describes techniques that integrate self-
learning algorithms designed to evaluate results and to
improve performance over time.
Data mining is a process of applying a set of analytical
techniques designed to uncover hidden patterns and
relationships in data.

The Data Mining Process

LO 8.1
Describe the data mining process.

Data mining is a complex process of examining large sets of data for

identifying patterns and then using them for valuable business
insights. Due to the recent explosion in the field of data page 319
mining, there is a growing need for the establishment of
 standards in the area. When conducting data mining analysis,
practitioners generally adopt either the Cross-Industry Standard
Process for Data Mining (CRISP-DM) methodology or the Sample,
Explore, Modify, Model, and Assess (SEMMA) methodology. With
either methodology, it is important to fully understand the
surrounding socioeconomic climate, business goals, and underlying
issues at hand prior to preparing the data and choosing analysis
techniques. We now elaborate on each methodology.

CRISP-DM
CRISP-DM was developed in the 1990s by a group of five
companies: SPSS, TeraData, Daimler AG, NCR, and OHRA. CRISP-
DM consists of six major phases: business understanding, data
understanding, data preparation, modeling, evaluation, and
deployment. The six phases can be summarized as follows:

. Business understanding: The first phase focuses on understanding

the data mining project and its objectives. Information about the
situational context around the data mining project (e.g., resources
and data availability) is gathered during this phase in order to define
specific objectives (e.g., predict the number of units sold to a
customer segment or increase the sales in a particular region) as
well as the project schedule and deliverables.
2. Data understanding: This phase involves collecting relevant data
and conducting a preliminary analysis to understand the data.
Results from these initial analyses may lead to ideas and potential
hypotheses for subsequent data mining phases.
3. Data preparation: Specific tasks in this phase include record and
variable selection, data wrangling, and cleansing for subsequent
analyses. For example, certain data mining techniques may require
subsetting and/or transformation of numerical and categorical data.
These, along with other topics, related to data wrangling were
discussed in Chapter 2.
4. Modeling: This phase involves the selection and execution of data
mining techniques, including linear and logistic regression models
 discussed earlier. Certain analysis techniques require specific
formats and types of variables in the data set. For example, a
traditional naïve Bayes technique (discussed in Chapter 9) can only
work with categorical variables. As such, we may need to step back
to the data preparation phase to convert or transform data. We also
need to document the assumptions that we make (e.g., normal
distribution, treatment of missing values, etc.) and set parameter
values for the model. For data mining, we often need to set aside
portions of a data set for training and validating the model(s).
5. Evaluation: After developing data mining models, we evaluate the
performance of competing models based on specific criteria
(discussed later in this chapter) in order to select the best model
that meets the business objectives of the project. We then review
and interpret the results in the context of the business objectives
described in the first phase. At the end of each regression and data
mining chapter in this text, we include a section called “Writing with
Big Data” where a written report presents the results, interpretation,
and recommendations based on the knowledge gained from the
analysis.
6. Deployment: During this final phase, we develop a set of actionable
recommendations based on the analysis results. Similar to other
business projects, we need a strategy for deployment, monitoring,
and feedback. The CRISP-DM model was conceived as a life cycle,
implying the cyclical nature of data mining projects. Once
recommendations are implemented, we may gain additional insights
that trigger subsequent data mining projects.
Figure 8.1 shows the steps of the CRISP-DM methodology.

page 320

FIGURE 8.1 CRISP-DM data mining methodology

 SEMMA
Some practitioners of data mining prefer the SEMMA methodology.
Developed by the SAS Institute, this methodology focuses on a core
set of tasks and provides a step-by-step process for analyzing data.
The five steps of the SEMMA methodology can be summarized as
follows:

. Sample: The first step focuses on identifying appropriate variables,

merging or dividing data sets, and drawing a sample of data for
subsequent analyses.
2. Explore: Various exploratory data analysis techniques are
implemented, including data visualization and summary measures,
to gain a thorough understanding of the data.
3. Modify: Similar to the data preparation phase in the CRISP-DM
methodology, relevant variables are selected, created, and/or
transformed in order to prepare the data set for subsequent
analyses.
4. Model: Similar to the modeling phase in the CRISP-DM
methodology, analysis techniques and models are chosen and
applied to the data, and results are produced.
5. Assess: Results from different models are presented to the end
users. The users then compare the outcomes and performance of
the competing models. In some cases, new observations are
scored using the insights gained from the model(s).

Figure 8.2 shows the steps of the SEMMA methodology.

FIGURE 8.2 SEMMA data mining methodology

THE DATA MINING PROCESS

Two important methodologies when conducting data mining
analysis are Cross-Industry Standard Process for Data Mining
(CRISP-DM) and Sample, Explore, Modify, Model, and Assess
(SEMMA).
CRISP-DM includes six major phases in the data mining
process: business understanding, data understanding, data
preparation, modeling, evaluation, and deployment.
SEMMA includes five major steps in the data mining process:
sample, explore, modify, model, and assess.

Both CRISP-DM and SEMMA provide good coverage of the

sequential steps, starting with problem identification, data
exploration, data collection, data processing, analysis, performance
evaluation, and implementation. It is important to page 321
note that not every step of these processes is
needed for all data mining applications. Also, in practice, the data
preparation phase in CRISP-DM and the modify step in SEMMA play
a significant role in the data mining process. An analyst or analytics
team tends to spend a sizable portion of the project time (often 80%)
on understanding, cleansing, transforming and preparing data
leading up to the modeling activities. Several surveys conducted by
KD Nuggets, a leading website on data analytics, suggest that the
CRISP-DM methodology is more popular than the SEMMA
methodology. Many find CRISP-DM a true methodology as it offers a
holistic approach to data mining with detailed phases, tasks, and
activities.

Supervised and Unsupervised Data Mining

Data mining uses many kinds of computational algorithms to identify
hidden patterns and relationships in data. These algorithms are
classified into supervised and unsupervised techniques, depending
on the way they “learn” about data to make predictions or identify
patterns. For developing predictive models, one tends to employ
supervised data mining techniques. Unsupervised data mining
techniques are especially effective for data exploration, dimension
reduction, and pattern recognition. The key distinction between
supervised and unsupervised techniques is that, in supervised data
mining, the target variable is identified. In Chapters 6 and 7, we refer
to the target variable as a response variable, which is the name often
used in statistics. The historical values of the target variable exist in
the data set so that data mining algorithms can examine the impact
of the predictor variables on the target variable. On the contrary, in
unsupervised data mining, no target variable is identified.

Supervised Data Mining

Some of the most commonly used supervised data mining
algorithms are based on classic statistical techniques. The linear
regression model and the logistic regression model, discussed in
Chapters 6 and 7, are examples of supervised techniques. These
models are used to predict the outcomes of a target variable based
on several predictor variables. The mathematical model relates the
outcome of the target variable (commonly referred to as y) to one or
more predictor variables (commonly referred to as x1, x2, . . . , xk).
Consequently, we use information on the predictor variables to
predict and/or describe changes in the target variable. A regression
model is therefore “trained” or “supervised” because the known
values of the target variable are used to build the model. In addition,
performance of the model can be evaluated based on the extent to
which the predicted values deviate from the actual values of the
target variable. There are several other machine learning algorithms
that are widely used in supervised data mining. We will discuss three
of these techniques including k-Nearest Neighbors, naïve Bayes,
and Decision Trees in Chapters 9 and 10.
Common applications of supervised data mining include
classification and prediction models. In a classification model, the
target variable is categorical. The objective of a classification model
is to predict the class memberships of new cases. For example, a
marketing manager may classify a list of prospective customers as
buyers and nonbuyers. A financial analyst may classify a stock into
one of the following three classes: buy, hold, and sell. In a prediction
model, the target variable is numerical. Examples of prediction
models include predicting the spending of a customer, the selling
price of a house, or the length of hospital stay after a medical
procedure.

SUPERVISED DATA MINING

In supervised data mining, the target (response) variable is
known. Common applications of supervised learning include
classification and prediction models.
In a classification model, the target variable is categorical.
The objective of the classification model is to predict the class
membership of a new case.
In a prediction model, the target variable is numerical. The
objective of the prediction model is to predict the numeric
value of a new case.
 page 322

Unsupervised Data Mining

Unsupervised data mining requires no knowledge of the target
variable. It is called unsupervised because, unlike supervised
learning, the algorithms allow the computer to identify complex
processes and patterns without any specific guidance from the
analyst. In the field of business analytics, unsupervised learning is
considered to be an important part of exploratory data analysis and
descriptive analytics. It is also used prior to conducting supervised
learning in order to understand the data set, formulate questions, or
summarize data.
Common applications of unsupervised learning include
dimension reduction and pattern recognition. Dimension
reduction (discussed later in this chapter) refers to the process of
converting a set of high-dimensional data (data with a large number
of variables) into data with a smaller number of variables without
losing much of the information in the original data. It is an important
step before deploying other data mining methods to reduce
information redundancy and improve model stability. Dimension
reduction is especially relevant to today’s big data environment. For
example, data collected by sensors or other Internet of things (IoT)
devices tend to have a large number of variables (high-dimensional)
and a lot of redundancy in the information they carry. Reducing the
dimensionality of these data sets will help bring out the important
patterns in data and build more stable predictive models.
Pattern recognition is the process of recognizing patterns using
machine learning techniques (discussed in Chapter 11). A pattern
may be recurring sequences of behaviors of customers, frequent
combinations of words in documents, recognizable features of
objects, or common characteristics of things that belong to the same
group. For example, retail companies use pattern recognition
techniques to group customers into different market segments in
order to customize product and service offerings that are of interest
to each segment.
 UNSUPERVISED DATA MINING
In unsupervised data mining, there is no target variable.
Common applications of unsupervised learning include
dimension reduction and pattern recognition.
Dimension reduction is the process of converting a set of
high-dimensional data into data with lesser dimensions.
Pattern recognition is the process of recognizing patterns in
the data using machine learning techniques.

The next few sections of this chapter explore three key concepts
relevant to data mining: similarity measures, performance evaluation,
and dimension reduction techniques. They are the prerequisite
knowledge for understanding the data mining techniques discussed
in Chapters 9, 10, and 11.

8.2 SIMILARITY MEASURES

We first explore similarity measures, which gauge whether a group
of observations are similar or dissimilar to one another. These
measures are important for understanding both supervised and
unsupervised data mining techniques discussed later in the text. In
Chapter 9, for example, similarity measures are used to identify
observations that are similar to each other or close neighbors in the
k-Nearest Neighbors (KNN) technique. In Chapter 11, cluster
analysis finds similarities among observations in the data and groups
them into meaningful clusters based on similar characteristics.
Clusters are formed in such a way that observations are similar
within a group but dissimilar across groups.

page 323

LO 8.2
Implement similarity measures.

Similarity measures are based on the distance between pairwise

observations (records) of the variables. A small distance between
the observations implies a high degree of similarity, whereas a large
distance between the observations implies a low degree of similarity.
In this chapter, we represent all k variables as x1, x2, . . . , xk.
Consider three observations for the variables, x1 and x2, shown in
Table 8.1.

TABLE 8.1 Three Pairwise Observations for x1 and x2

Observation x1 x2

1 3 4

2 4 5

3 10  1

Figure 8.3 plots the observations. We can clearly see that

observations 1 and 2 are closer to each other, implying that they are
“similar” to each other. Also, we can see that observation 3 is
dissimilar from observations 1 and 2. In this section, we present a
more formal way to measure similarity among observations for
numerical and categorical variables.

FIGURE 8.3 Plot of three pairwise observations

Similarity Measures for Numerical Data
The Euclidean distance is one of the most widely used measures
for evaluating similarity with numerical variables. It is defined as the
length of a straight line between two observations. For a data set
with k variables, the Euclidean distance between the ith and jth
observations is calculated as:

where xki and xkj represent the ith and jth observations for the kth
variable.
The Manhattan distance is the shortest distance between two
observations if you are only allowed to move horizontally or
vertically. Referring to the city of Manhattan, which is laid out in
square blocks, the Manhattan distance between two points is the
shortest path for a vehicle to travel from one point to another in the
city. For a data set with k variables, the Manhattan distance between
the ith and jth observations is calculated as:
where xki and xkj represent the ith and jth observations for the kth
variable.

page 324

EXAMPLE 8.1
Refer back to Table 8.1, where three observations are reported
for variables x1 and x2.
a. Calculate and interpret the Euclidean distance between all
pairwise observations.
b. Calculate and interpret the Manhattan distance between all
pairwise observations.
SOLUTION:

a. The Euclidean distances between all pairwise observations

are calculated as:
Observations 1 and 2:

Observations 1 and 3:

Observations 2 and 3:
 Using the Euclidean measure to gauge similarity, observations
1 and 2 are the most similar with the shortest Euclidean
distance of 1.41. Observations 1 and 3 are the most dissimilar
given the longest Euclidean distance of 7.62. This conclusion
is reinforced by the plot in Figure 8.3.
b. The Manhattan distances between all pairwise observations
are calculated as:

Like the results found using the Euclidean distance,

observations 1 and 2 are the most similar based on the
Manhattan distance. Observations 1 and 2 are now equally
dissimilar to observation 3.

EUCLIDIAN AND MANHATTAN DISTANCES

For numerical data, similarity measures are based on the
distance between pairwise observations (records) of the
variables. For a data set with k variables, the Euclidean and the
Manhattan distances between the ith and jth observations are
calculated as:

and
 where xki and xkj represent the ith and jth observations for the
kth variable.

As shown in Example 8.1, different distance measures do not

always lead to the same inference. The Euclidian distance is known
to be more influenced by outliers than the Manhattan distance. In
Example 8.1, the third observation is clearly an outlier, and,
therefore, the Manhattan distance may be more appropriate. Most
data mining algorithms, however, employ the Euclidian distance.

page 325

Standardization and Normalization

When a data set contains multiple variables, the scale of each
variable can influence the distance measure and how we determine
the similarity between two observations. For example, when
comparing consumer information, data such as annual incomes tend
to have a greater scale than, say, the number of hours spent online
per week simply by the fact that annual incomes are larger numbers.
As a result, annual income will have greater influence on the
similarity measure than the number of hours spent online per week.
This difference in scale distorts the true distance between
observations and can lead to inaccurate results. It is common,
therefore, to make the observations unit-free by standardizing or
normalizing numerical data before calculating the similarity
measure for the data mining techniques. These transformations of
numerical data ensure that each variable receives an equal weight
when the similarity measure is calculated.
One common way to standardize the data is to compute z-
scores, which were discussed in Chapter 3. Recall that the z-score
measures the distance of a given observation from the sample mean
in terms of standard deviations. In order to find the z-score for the ith
observation of the kth variable, denoted zki, we calculate:
where xki is the ith observation of the kth variable, and and sk are

the mean and the standard deviation of the kth variable, respectively.
Another widely used transformation method is min-max
normalization, which subtracts the minimum value from the
observation and then divides the difference by the range (maximum
value − minimum value). This approach rescales each value to be
between 0 and 1. In order to find the min-max normalized value for
the ith observation of the kth variable, denoted qki, we calculate:

where xki is the ith observation of the kth variable, and mink and
rangek are the minimum value and the range of the kth variable,
respectively.

STANDARDIZING AND NORMALIZING NUMERICAL DATA

The z-score standardization and min-max normalization are two
conversion techniques that make the numerical data
independent of scale. The z-score standardization for the ith
observation of the kth variable, zki, is calculated as:
 where xki is the ith observation of the kth variable, and and sk
are the mean and the standard deviation of the kth variable,
respectively. The min-max normalized value for the ith
observation of the kth variable, qki, is calculated as:

where xki is the ith observation of the kth variable, and mink and
rangek are the minimum value and the range of the kth
variable, respectively.

A Note on the Term Normalization

It is important to note that the term ‘normalization’ has several
meanings in statistics. Generally, it refers to a process that makes
the numerical data independent of scale. Sometimes, the term is
even used for the z-score standardization; Analytic Solver uses the
term ‘normalization’ to convert the numerical data into z- page 326
scores. In this text, we prefer to use the term min-max
normalization to avoid confusion.
Example 8.2 demonstrates how to standardize and normalize
data.

EXAMPLE 8.2
Consider a sample of five consumers with their annual income
(Income, in $) and hours spent online per week (Hours Spent)
shown in columns 1–3 of Table 8.2. The distance measures
calculated from raw data can be distorted because the annual
 income values are much larger than the values of hours spent
online per week.

TABLE 8.2 Annual Income and Hours Spent

a. Use z-scores to standardize the observations for Income and

Hours Spent.
b. Use the min-max transformation to normalize the observations
for Income and Hours Spent.
SOLUTION:

a. In order to standardize the observations for Income, we first

calculate its average and the standard deviation as $95,077
and $24,534.45, respectively. We then divide the difference
between each observation and the mean by the standard
deviation. For Jane, we find the standardized Income value
as:

Therefore, Jane’s income is 1.2473 standard deviations above

the average income of all consumers in the sample.
Standardized values for all Income observations are shown in
column 4 of Table 8.2. Kevin’s annual income, for example, is
1.1892 standard deviations below the average income.
 Standardized values for Hours Spent are found similarly
using an average and standard deviation of 7.0 hours and
2.99 hours, respectively. Column 5 of Table 8.2 shows these
standardized values. The standardized values have a similar
scale, and, therefore, they will receive an equal weight when
calculating the distance measure.
b. In order to normalize the observations for Income, we first find
the minimum and maximum values for Income as $65,901 and
$125,678, respectively, and then calculate the range as
$59,777. We then divide the difference between each
observation and the minimum value by the range. For Jane,
we find the min-max normalized value for Income as:

Because the min-max normalization rescales each

observation between 0 and 1, Jane, with the highest annual
income, has a min-max normalized income of 1, and Kevin,
with the lowest annual income, has a min-max normalized
income of 0. The min-max normalized values for Income are
shown in column 6 of Table 8.2.
page 327
The min-max normalization values for Hours
Spent are found similarly using minimum and maximum
values of 2.5 hours and 10.1 hours, respectively. Column 7 of
Table 8.2 shows the min-max normalization values for Hours
Spent. As in the case of standardized values, the min-max
normalized values have a similar scale, and, therefore, they
will receive an equal weight when calculating the distance
measure.
Similarity Measures for Categorical Data
Euclidean and Manhattan distance measures are suitable for
numerical variables. When dealing with categorical variables, we rely
on other measures of similarity. For example, you may be interested
in analyzing types of cars, where the values of the categorical
variable include sedan, sport utility, truck, minivan, etc. Recall that a
categorical variable with only two categories is called a binary
variable. For example, when analyzing loans, the values of the
binary variable include approved or not approved. Two commonly
used measures for categorical and binary data are the matching
coefficient and Jaccard’s coefficient.

Matching Coefficient
The matching coefficient for a categorical variable is based on
matching values to determine similarity among observations
(records). The matching coefficient between two observations is:

The higher the value of the matching coefficient, the more similar
the two observations are. A matching coefficient value of one implies
a perfect match.

EXAMPLE 8.3
Consider the list of college students in Table 8.3. Each record
shows the student’s major, field, sex, and whether or not the
student is on the Dean’s List. Compute the matching coefficient
between all pairs of students and determine how similar or
dissimilar they are from each other.
 TABLE 8.3 Characteristics of College Students

SOLUTION:
Comparing students 1 and 2, only one of the four variables
(i.e., Dean’s List) has a matching value. Therefore, the
matching coefficient equals 1/4 = 0.25. Comparing students 1
and 3, two of the four variables (i.e., Major and Sex) have a
matching value. Therefore, the matching coefficient equals 2/4
= 0.50. Comparing students 2 and 3, none of the four variables
has a matching value. Therefore, the matching coefficient
equals 0/4 = 0.
Based on the matching coefficients for these three students,
students 1 and 3 are the most similar to each other, whereas
students 2 and 3 are the most dissimilar.

Jaccard’s Coefficient
The matching coefficient makes no distinction between positive
outcomes (e.g., made a purchase) and negative outcomes (e.g., did
not make a purchase) and, therefore may provide a page 328
misleading measure of similarity. Consider, for example,
consumer purchasing behavior at a grocery store where thousands
of products are available. Intuitively, two consumers purchasing the
same products exhibit similar behavior. However, if one consumer
purchases only a Hostess cupcake, and the other purchases only
organic apples, most people would conclude that these two
consumers are not similar. However, a simple matching coefficient
will account for the fact that neither of the customers purchased
thousands of other products that are available and, therefore,
erroneously produce a very high value of the matching coefficient.
Jaccard’s coefficient, named after French scientist Paul
Jaccard, is a more appropriate measure of similarity in situations
where negative outcomes are not as important as positive outcomes.
It is computed as:

EXAMPLE 8.4
A retail store collects point of sales information that shows
whether or not five products were included in each sales
transaction. This information is captured in Table 8.4. A ‘yes’
indicates a positive outcome, while a ‘no’ indicates a negative
outcome. For example, Transaction 1 indicates that a
consumer purchases a computer keyboard, a mouse, and a
headphone. Compute and compare matching coefficients and
Jaccard’s coefficients for all sets of pairwise transactions.

TABLE 8.4 Retail Transactions

SOLUTION:
We demonstrate how to calculate the matching coefficient and
the Jaccard coefficient between transactions 1 and 2. For the
matching coefficient, transactions 1 and 2 have three matching
values: keyboard, mouse, and USB drive. Given that there are
five products (so five variables), the matching coefficient is
calculated as 3/5 = 0.6. For the Jaccard’s coefficient, the two
transactions have two matching positive outcomes (keyboard
and mouse) and one matching negative outcome (USB drive).
Therefore, the Jaccard’s coefficient for the pair is calculated as
2/(5 − 1) = 0.50. The coefficients for the remaining pairs are
calculated similarly and are shown in Table 8.5.

TABLE 8.5 Comparing Matching Coefficients with Jaccard’s Coefficients

Matching Jaccard’s
Transactions
coefficients coefficients

1 and 2 3/5 = 0.6 2/4 = 0.50

1 and 3 3/5 = 0.6 1/3 = 0.33

1 and 4 3/5 = 0.6 1/3 = 0.33

2 and 3 1/5 = 0.2 0/4 = 0.00

2 and 4 3/5 = 0.6 1/3 = 0.33

3 and 4 3/5 = 0.6 0/2 = 0.00

Table 8.5 shows that both measures produce the highest

coefficient for transactions 1 and 2, implying that these
transactions are most similar. However, some page 329
notable differences arise. For example, the
matching coefficients yield a higher value when comparing
transactions with matching negative outcomes (for example,
transactions 1 and 3; transactions 1 and 4; etc.). The absence
of particular products in a sales transaction (i.e., negative
outcomes) is usually not as important as the positive outcomes,
and including the negative outcomes in the similarity measure
is misleading. Consider transactions 3 and 4. The fact that the
two customers did not purchase the same three products
(memory card, mouse, and USB drive) may not imply any
similarity between the two transactions or customers at all.
Jaccard’s coefficient makes the appropriate adjustment and
with a Jaccard’s coefficient of 0 suggests that the two
transactions are dissimilar. The matching coefficient of 0.6, on
the other hand, erroneously suggests a relatively high similarity
because of the number of matching negative outcomes.

MATCHING AND JACCARD’S COEFFICIENTS

Matching and Jaccard’s coefficients are similarity measures for
categorical data. For a data set with k variables, the matching
coefficient between the ith and the jth observations (records) is
calculated as:

For a data set with k variables, the Jaccard’s coefficient

between the ith and the jth observations is calculated as:
 The Jaccard’s coefficient is appropriate when it is more
informative to match only positive outcomes between two
observations.

EXERCISES 8.2
Note: For exercises in this section, it is advisable to use unrounded numbers in the
calculations.

Mechanics
1. FILE Exercise_8.1. The accompanying data file contains 10 observations with
two variables, x1 and x2.
a. Using the original values, compute the Euclidean distance between the first two
observations.
b. Using the original values, compute the Manhattan distance between the first
two observations.
c. Based on the entire data set, calculate the sample mean and standard deviation
for x1 and x2. Use z-scores to standardize the values, and then compute the
Euclidean distance between the first two observations.
d. Based on the entire data set, find the minimum and maximum values for x1 and
x2. Use the min-max transformation to normalize the values, and then compute
the Euclidean distance between the first two observations.
2. FILE Exercise_8.2. The accompanying data file contains 10 observations with
three variables, x1, x2, and x3.
a. Using the original values, compute the Euclidean distance between the first two
observations.
b. Using the original values, compute the Manhattan distance between the first
two observations.
c. Based on the entire data set, calculate the sample mean and standard deviation
for the three variables. Use z-scores to standardize the values, and then
compute the Euclidean distance between the first two observations.
 d. Based on the entire data set, find the minimum and maximum values for the
three variables. Use the min-max transformation to normalize the values, and
then compute the Euclidean distance between the first two
observations.
page 330
3. FILE Exercise_8.3. The accompanying data file contains 28 observations with
three variables, x1, x2, and x3.
a. Using the original values, compute the Euclidean distance for all possible pairs
of the first three observations.
b. Based on the entire data set, calculate the sample mean and standard deviation
for the three variables. Use z-scores to standardize the values, and then
compute the Euclidean distance for all possible pairs of the first three
observations. Compare the results with part a.
c. Using the original values, and then the z-score standardized values, compute
the Manhattan distance for all possible pairs of the first three observations.
4. FILE Exercise_8.4. The accompanying data file contains 19 observations with
two variables, x1 and x2.
a. Using the original values, compute the Euclidean distance for all possible pairs
of the first three observations.
b. Based on the entire data set, find the minimum and maximum values for the
three variables. Use the min-max transformation to normalize the values, and
then compute the Euclidean distance for all possible pairs of the first three
observations.
c. Using the original values, and then the min-max normalized values, compute
the Manhattan distance for all possible pairs of the first three observations.
5. FILE Exercise_8.5. The accompanying data file contains 10 observations with
two variables, x1 and x2.
a. Based on the entire data set, calculate the sample mean and standard deviation
of the two variables. Using the original values, and then the z-score
standardized values, compute the Euclidean distance for all possible pairs of
the first three observations.
b. Using the original values, and then the z-score standardized values, compute
the Manhattan distance for all possible pairs of the first three observations.
6. FILE Exercise_8.6. The accompanying data file contains 12 observations with
three variables, x1, x2, and x3.
a. Based on the entire data set, calculate the sample mean and standard deviation
for the three variables. Using the original values, and then the z-score
standardized values, compute the Euclidean distance for all possible pairs of
the first three observations.
b. Based on the entire data set, find the minimum and maximum values of the
three variables. Using the original values, and then the min-max normalized
values, compute the Manhattan distance for all possible pairs of the first three
observations.
 7. FILE Exercise_8.7. The accompanying data file contains five observations with
three categorical variables, x1, x2, and x3.
a. Compute the matching coefficient for all pairwise observations.
b. Identify the observations that are most and least similar to each other.
8. FILE Exercise_8.8. The accompanying file contains four observations with four
categorical variables, x1, x2, x3, and x4.
a. Compute the matching coefficient for all pairwise observations.
b. Identify the pair of observations that are most and least similar to each other.
9. FILE Exercise_8.9. The accompanying file contains six observations with four
binary variables, x1, x2, x3, and x4.
a. Compute the matching coefficient for all pairwise observations. Identify the pair
of observations that are most and least similar to each other.
b. Compute the Jaccard’s coefficient for all pairwise observations. Identify the pair
of observations that are most and least similar to each other.

Applications
10. FILE Employees. Consider the following portion of data that lists the starting
salaries (in $1,000) of newly hired employees and their college GPAs:

Employee Salary GPA

1 72 3.53

2 66 2.86

3 72 3.69

⋮ ⋮ ⋮

11 59 3.49

a. Without transforming the values, compute the Euclidean distance for all
possible pairs of the first three employees. Make sure to exclude the employee
numbers in the calculations.
b. Compute the z-score standardized values for salary and GPA, and then
compute the Euclidean distance for all possible pairs of the first three
employees. Discuss the impact that standardization has on the similarity
distance.
c. Using the z-score standardized values, compute the Manhattan distance for all
possible pairs of the first three employees. Discuss the differences between the
Euclidean and Manhattan distances.
11. FILE Online_Retailers. Consider the following portion of data that lists
information about an online retailer’s annual revenues (Revenue in $), the number
of products available on the retailer’s website (SKUs), and the number of visits to
the website per day (Visits):

Website ID Revenue SKUs Visits

001 2,984,567 34,567 546,799

002 1,230,956 22,398 342,455

⋮ ⋮ ⋮ ⋮

010 2,278,890 28,773 333,219

page 331
a. Without transforming the values, compute the Euclidean
distance for all pairwise observations of websites 001, 003, and 005 based on
annual revenues, SKUs, and visits per day. Make sure to exclude the Website
IDs from your calculation.
b. Compute the min-max normalized values for annual revenues, SKUs, and visits
per day, and then compute the Euclidean distance for all pairwise observations
of websites 001, 003, and 005. Discuss the differences in parts a and b. Make
sure to exclude the Website IDs from your calculation.
c. Using the min-max normalized values, compute the Manhattan distance for all
pairwise observations of websites 001, 003, and 005. Discuss the difference
between the Euclidean and Manhattan distances. Make sure to exclude the
Website IDs from your calculation.
12. FILE Vehicles. Consider the following portion of data that shows vehicle
information based on three variables: Type of vehicle, whether or not the vehicle
has all-wheel drive (AWD), and whether the vehicle’s transmission is automatic or
manual.

Vehicle Type AWD Transmission

1 SUV Yes Automatic

2 Sedan No Manual

⋮ ⋮ ⋮ ⋮

56 SUV No Automatic
 a. Using the first five vehicles, compute the matching coefficient for all pairwise
observations based on Type, AWD, and Transmission.
b. Identify and describe the vehicles that are most and least similar to each other.
Explain the characteristics of these vehicles.
13. FILE Home_Loan. Consider the following portion of data that includes
information about home loan applications. Variables on each application include
(1) whether the application is conventional or subsidized by the federal housing
administration (LoanType), (2) whether the property is a single-family or multi-
family home (PropertyType), and (3) whether the application is for a first-time
purchase or refinancing (Purpose).

Application LoanType PropertyType Purpose

1 Conventional Single-family Purchase

2 Conventional Multi-family Refinancing

⋮ ⋮ ⋮ ⋮

103 Federal Housing Single-family Purchase

a. Using the first five applications, compute the matching coefficients for all
pairwise observations based on LoanType, PropertyType, and Purpose.
b. Identify and describe the loan applications that are most and least similar to
each other. Explain the characteristics of these loan applications.
14. FILE University_Students. Information is collected on university students. The
variables of interest include (1) whether a student is an undergraduate or
graduate student (Level of education), (2) whether or not the student is a math
major, (3) whether or not the student is a statistics minor, and (4) whether a
student is male or female. A portion of the data is shown in the accompanying
table.

a. Using the first five students, compute the matching coefficients for all pairwise
observations based on the four variables.
b. Identify and describe students who are most and least similar to each other.
15. FILE Bookstore. Consider the following data that shows a portion of point of
sales transactions from a local bookstore. Each binary variable indicates whether
or not a particular book genre is purchased within a transaction.

a. Using the first five sales transactions, compute the matching coefficients for all
pairwise observations for the five binary variables.
b. Using the first five sales transactions, compute Jaccard’s coefficients for all
pairwise observations for the five binary variables.
c. Based on the results from parts a and b, identify and describe the transactions
that are most and least similar to each other. Discuss the difference between
the results from parts a and b.
16. FILE Grocery_Store. Consider the following data that show a portion of the
point of sales transactions from a grocery store. Each binary variable indicates
whether or not a product is purchased as part of the transaction.

a. Using the first five sales transactions, compute the matching coefficients for all
pairwise observations for the four binary variables.

page 332
b. Using the first five sales transactions, compute Jaccard’s
coefficients for all pairwise observations for the four binary variables.
c. Based on the results from parts a and b, identify and describe the transactions
that are most and least similar to each other. Discuss the difference between
the results from parts a and b.
17. FILE Fast_Food. Consider the following data that show a portion of the sales
transactions from a local fast-food restaurant. Each binary variable indicates
whether or not a product is purchased as part of the sales transaction.
 Transaction Hamburgers Fries Soda

1 0 1 0

2 1 1 0

⋮ ⋮ ⋮ ⋮

238 0 1 1

a. Using the first five sales transactions, compute the matching coefficients for all
pairwise observations for the three binary variables.
b. Using the first five sales transactions, compute Jaccard’s coefficients for all
pairwise observations for the three binary variables.
c. Based on the results from parts a and b, identify and describe the transactions
that are most and least similar to each other. Discuss the difference between
the results from parts a and b.

8.3 PERFORMANCE EVALUATION

LO 8.3
Assess the predictive performance of data mining models.

In this section we assess the predictive performance of data mining

models, including linear and the logistic regression models
discussed in Chapters 6 and 7. Recall that it is important to develop
performance measures that evaluate how well an estimated model
will perform in an unseen sample, rather than making the evaluation
solely on the basis of the sample data used to build the model.
In Chapter 7, we discussed cross-validation techniques where we
partition the original sample into a training set to build (train) the
model and a validation set to assess (evaluate) the model. We used
the root mean squared error (RMSE) to evaluate linear regression
models and the accuracy rate to evaluate logistic regression models.
In this chapter, we extend the analysis to include several other
performance measures that can be applied to all supervised data
mining models.

Data Partitioning
Data partitioning is the process of dividing a data set into a training,
a validation, and, in some situations, an optional test data set. A
common data partitioning practice is a two-way random partitioning
of the data to generate a training data set and a validation data set.
We use random partitioning, as opposed to fixed draws discussed in
Chapter 7, to avoid any bias in the selection of the training and
validation data sets. The training data set, which often contains the
larger portion of the data, “trains” data mining algorithms to identify
the relationship between the predictor variables and the target
(response) variable. The validation data set, which is not involved in
model building, is used to provide an unbiased assessment of the
predictive performance of data mining models.
The model built from the training data set is used to predict the
values of the target variable in the validation data set. These
predicted values are then compared to the actual values of the target
variable of the validation data set to evaluate the performance of the
model. The performance measures obtained from this process can
be used to help assess model performance, fine-tune the model, or
compare the performance of competing models. A page 333
common practice is to partition 60% of the data into
the training data set and 40% of the data into the validation data set.
However, sometimes, the 70%/30% training/validation partitioning is
used if the model building process can benefit from a larger training
data set. As explained in Chapter 7, we can implement cross-
validation with the holdout as well as k-fold methods.
In supervised data mining, we sometimes use three-way random
partitioning of the data to generate training, validation, and test data
sets. The third data partition, called a test data set, which is not
involved in either model building or model selection, is created to
evaluate how well the final model would perform on a new data set
that it has never seen before. As in the case of two-way partitioning,
data are usually partitioned randomly to create unbiased training,
validation, and test data sets. In the three-way random partitioning
method, 50% training/30% validation/20% test partitioning is often
recommended. The three-way random partitioning is usually
implemented with the holdout method, as opposed to k-fold cross-
validation.

DATA PARTITIONING
Data partitioning is the process of dividing a data set into a
training, a validation, and an optional test data set. The training
data set is used to generate one or more models. The
validation data set is used to fine-tune or compare the
performance of competing models, The optional test data set is
used to assess the performance of the final model on a new
data set.

Oversampling
Sometimes the target class that we wish to classify is very rare,
which reduces the usefulness of classification models. For example,
in order to build a classification model that informs an online retailer
about which prospective customers are likely to respond to a
promotional e-mail, the retailer conducts an experiment by sending
promotional e-mail messages to 10,000 potential customers of which
only 200 customers respond, resulting in a response rate of tow %.
The resulting data from this experiment would consist mostly of non-
responders with little information to distinguish them from the target
class of responders. If this retailer attempts to build classification
models, the best-performing model might be the one that suggests
no prospective customers will respond to the promotional e-mail
message as the model would have an extremely high accuracy rate
of 98%. Nevertheless, the best-performing model in this application
might be useless to the retailer.
 A common solution to this problem is called oversampling. The
oversampling technique involves intentionally selecting more
samples from one class than from the other class or classes in order
to adjust the class distribution of a data set. Rare target class cases
will be more represented in the data set if they are oversampled.
This would lead to predictive models that are more useful in
predicting the target class cases.
In the promotional e-mail example, the retailer may choose to
oversample the rare target class cases by including most, if not all,
responders and reduce the number of non-responders so that the
resulting data set constitutes a healthy combination of responders
and non-responders.

OVERSAMPLING
The oversampling procedure overweights the rare class relative
to the other class or classes in order to adjust the class
distribution of a data set.

page 334

As a common practice, only the training data set is oversampled.

Observations in the validation and test data sets maintain the original
class proportions so that the resulting performance estimates will be
more representative of the reality.

Performance Evaluation in Supervised Data

Mining
In Section 6.2, we used goodness-of-fit measures such as R2 and
adjusted R2 to assess how well the regression model fits or explains
the data. Nevertheless, a model that fits a set of in-sample data well
may not predict new observations accurately due to the possibility of
overfitting. Recall from Section 7.4 that overfitting occurs when a
model corresponds too closely to a set of data but fails to predict
future observations reliably. In other words, an overfit model
accounts for the noise in a particular set of data rather than the
underlying patterns that can be generalized to future observations.
Overfitting may become more severe as the model complexity
increases (for example, as more predictor variables are included in
the model).

OVERFITTING
Overfitting occurs when a predictive model is made overly
complex to fit the quirks of given sample data. By making the
model conform too closely to the sample data, its predictive
power is compromised.

Data partitioning and cross-validation can be used to detect

overfitting and provide objective assessment of the predictive
performance of models. Suppose a model is developed using a
training data set and is then assessed by inspecting the performance
measures obtained from the validation data set. As shown in Figure
8.4, while the performance of the model on the training data set
tends to improve (i.e., predictive error rate decreases) as the model
gets more and more complex, the model performance on the
validation data set will improve initially but deteriorate after a certain
point. In addition to providing an unbiased performance evaluation of
the model, the validation data set is often used to identify the optimal
model complexity in some supervised data mining techniques (e.g.,
KNN and decision trees discussed in Chapters 9 and 10).

FIGURE 8.4 The relationship between error rate and model complexity
 We will now discuss the assessment of classification models
where the target variable is categorical, and prediction models where
the target variable is numerical. Performance measures used in
classification and prediction models are different; therefore, they are
discussed separately in the following subsections.

page 335

Performance Evaluation for Classification

Models
Performance measures for classification models can be computed
from a confusion matrix, a table that summarizes classification
outcomes obtained from the validation data set. Table 8.6 shows a
confusion matrix for a binary classification problem. Assume that the
target class, also called the success class, is Class 1 and that the
nontarget class is Class 0. From the confusion matrix we see that
two correct classifications are possible:
True positive (TP) is a Class 1 observation that is correctly
classified by the model, and
 True negative (TN) is a Class 0 observation that is correctly
classified by the model.
Conversely, two incorrect classifications are possible:
False positive (FP) is a Class 0 observation that is incorrectly
classified as a Class 1 observation by the model, and
False negative (FN) is a Class 1 observation that is incorrectly
classified as a Class 0 observation by the model.

TABLE 8.6 Confusion Matrix

Actual
Predicted Class 1 Predicted Class 0
Class

Class 1 No. of true positives No. of false negatives

(TP) (FN)

Class 0 No. of false positives No. of true negatives

(FP) (TN)

The followings are common performance measures for

classification models:

The misclassification rate, also called error rate, is the overall

proportion of observations that are misclassified. In other
words, the misclassification rate measures how often the
predictive model makes an incorrect prediction. It is computed
using the following equation:
The accuracy rate is the overall proportion of observations that
are classified correctly. As it measures how often the predictive
model makes a correct prediction, we can think of the accuracy
rate as the opposite measure of the misclassification rate. It is
computed using the following equation:

Sensitivity, also called recall, is the proportion of target class

cases that are classified correctly. It is computed using the
following equation:

To better understand the importance of the sensitivity rate,

consider an example of a model, say, for predicting breast
cancer among a group of patients. A predictive model with
100% sensitivity would correctly identify all patients who have
cancer. A model with 60% sensitivity would correctly identify
only 60% of the patients with cancer (true positives) but leave
the remaining 40% with the cancer undetected and untreated
(false negatives). Developing a predictive model with a high
sensitivity is clearly important to help identify patients with a
life-threatening but treatable disease.
Precision, also called positive predictive value, is the
proportion of the predicted target class cases that actually
belong to the target class. It differs from sensitivity in page 336
that the denominator is based on predicted Class 1
cases instead of actual Class 1 cases. Precision is computed
using the following equation:
Consider again the predictive model for breast cancer. While the
sensitivity rate measures the proportion of the patients with the
cancer who test positive (true positives as a proportion of those
who have the cancer), the precision rate measures the
proportion of the patients with a positive test who in fact have
the cancer (true positives as a proportion of those with a
positive test). A predictive model with a low precision rate
would unnecessarily subject many patients who are cancer-
free to undergo further medical tests and treatments (false
positives).
Specificity is the proportion of nontarget class cases that are
classified correctly. It is computed using the following equation:

Again, consider the predictive model for breast cancer. A

predictive model with 100% specificity would correctly identify
all patients without the cancer. A model with 75% specificity
would correctly identify 75% of the patients as cancer-free (true
negatives), but the remaining 25% of patients who are cancer-
free are classified by the model as having the cancer (false
positives). These 25% of patients may be unnecessarily
subjected to further medical tests or treatments.

CLASSIFICATION PERFORMANCE MEASURES

Common performance measures for classification models
include the following:
 The misclassification rate, also called error rate, is the overall
proportion of observations that are misclassified.
The accuracy rate is the overall proportion of observations
that are classified correctly.
The sensitivity, also called recall, is the proportion of target
class cases that are classified correctly.
The precision, also called positive predictive value, is the
proportion of the predicted target class cases that belong to
the target class.
The specificity is the proportion of nontarget class cases that
are classified correctly.

These performance measures collectively represent the

performance of a classification model. When the misclassification
costs are symmetric for both target and nontarget class cases, the
overall misclassification rate and accuracy rate are good indicators
of model performance. However, emphasis may be placed on
sensitivity or specificity measures when the misclassification costs
are asymmetric, and/or the proportion of the target class
observations is very small.
Consider the social media marketing example from the
introductory case. The cost of misclassifying a buying customer as a
nonbuying customer (e.g., loss in profit) is often much higher than
the cost of misclassifying a non-buying customer as a buying
customer (e.g., cost of sending marketing materials). In such cases,
marketers tend to pay more attention to the sensitivity measure than
to the overall accuracy as the sensitivity measure indicates the
proportion of buyers the classification model is able to identify
correctly. In addition, actual buyers tend to represent only a small
portion of all the prospective buyers. Here, the accuracy rate may be
misleading as a high overall accuracy rate may be a result of
classifying the large number of non-buyers correctly, which is not
indicative of a successful marketing campaign.
 page 337

In practice, in order to assess the performance of a classification

model, we often compare the model with a baseline or benchmark. A
common baseline approach is to compare the accuracy rate of the
model with that of a naïve rule where all cases are classified into the
most predominant class. Additional information and examples for the
naïve rule baseline are available in Chapters 9 and 10.

EXAMPLE 8.5
Recall the FashionTech example from the introductory case
where Class 1 (target) is the group of social media users who
purchase at least one product within 30 days of receiving a
promotional message and Class 0 (nontarget) is the group of
social media users who do not make a purchase after receiving
a promotional message. Table 8.7 shows the confusion matrix
that is obtained after applying the classification model to the
validation data set of 200 observations. [Shortly, we show how
we create the confusion matrix in Excel.]

TABLE 8.7 Confusion Matrix for the FashionTech Example

Actual Class Predicted Class 1 Predicted Class 0

Class 1 29  19

Class 0 19 133

a. Interpret the four classifications in the confusion matrix.

b. Calculate and interpret the accuracy rate.
c. Calculate and interpret sensitivity and specificity.
 SOLUTION:

a. The values in the confusion matrix are interpreted as follows:

True positives (TP): There are 29 FashionTech customers
who made a purchase and were correctly classified as
buyers by the classification model, so TP = 29.
True negatives (TN): There are 133 FashionTech customers
who did not make a purchase and were correctly classified
as nonbuyers by the classification model, so TN = 133.
False positives (FP): There are 19 nonpurchasing
FashionTech customers who were incorrectly classified as
buyers by the classification model (the lower left cell of Table
8.7), so FP = 19, and
False negatives (FN): There are 19 FashionTech customers
who made a purchase but were classified by the
classification model as nonbuyers (the upper right cell of
Table 8.7), so FN = 19.

b. The accuracy rate is calculated as

this measure implies that 81% of the observations are

correctly classified.

c. The sensitivity is calculated as this is the

proportion of buyers that the classification model is able to

identify correctly. The specificity is calculated as
this is the proportion of non-buyers that the

classification model is able to identify correctly. As mentioned

earlier, in this application the sensitivity measure is the more
useful performance measure.
Using Excel to Obtain the Confusion Matrix
and Performance Measures
The confusion matrix in Table 8.7 is based on the Class_Prob data
set that was generated by the FashionTech firm from the introductory
case. Each row in the data set includes the actual class page 338
to which the observation belongs (ActualClass) and the
probability of the target class (TargetProb) predicted by an estimated
classification model; FashionTech did not provide the model details.
A portion of the data is shown in Table 8.8.
FILE
Class_Prob

TABLE 8.8 A Portion of the FashionTech Firm Data Set

ActualClass TargetProb

1 0.77585787

0 0.04709009

⋮ ⋮

0 0.38111670

Here, we show step-by-step Excel instructions to build a

confusion matrix for an estimated classification model given the
target probabilities; specifically, we replicate Table 8.7. (In Chapters
9 and 10, we discuss how Analytic Solver and R estimate
classification models, generate a confusion matrix as well as
performance measures)
. Open the Class_Prob data file.
. Derive the predicted class based on a default cutoff of 0.5; in other
words, the predicted class equals one if the target probability is
greater than or equal to 0.5, zero otherwise. Enter the column
heading “PredClass” in cell C1. Enter the formula = IF(B2 >=0.5, 1,
0) in cell C2. Fill the range C3:C201 using the formula in C2.
. Compute the values for the four classifications of the confusion
matrix as follows:
Enter the column heading “True Positive” in cell D1. Enter the
formula = IF (AND(A2 = 1, C2 = 1), 1, 0) in cell D2. Fill the range
D3:D201 using the formula in D2.
Enter the column heading “True Negative” in cell E1. Enter the
formula = IF(AND (A2 = 0, C2 = 0), 1, 0) in cell E2. Fill the range
E3:E201 using the formula in E2.
Enter the column heading “False Positive” in cell F1. Enter the
formula = IF(AND (A2 = 0, C2 = 1), 1, 0) in cell F2. Fill the range
F3:F201 using the formula in F2.
Enter the column heading “False Negative” in cell G1. Enter the
formula = IF (AND(A2 = 1, C2 = 0), 1, 0) in cell G2. Fill the range
G3:G201 using the formula in G2.
. Build the confusion matrix by entering the formulas as shown in
Table 8.9. The resulting table should be identical to Table 8.7.

TABLE 8.9 Excel Formulas for the Confusion Matrix

Actual
Predicted Class 1 Predicted Class 0
Class

Class 1 =COUNTIF(D2:D201, =COUNTIF(G2:G201,

1) 1)

Class 0 =COUNTIF(F2:F201, 1) =COUNTIF(E2:E201, 1)

Selecting Cut-off Values
In the context of the binary choice models presented in section 7.3,
we computed the probability that a new mortgage loan application
will be approved (Class 1) or denied (Class 0). If the predicted
probability value was greater than or equal to 0.5, we classified the
new loan application as belonging to Class 1 (approved), and Class
0 (denied) otherwise. In other words, we used a default cutoff value
of 0.5. Classification techniques in all supervised data mining follow
a similar process where the class membership is determined by
comparing the predicted probability of the target class to a
predetermined cutoff value, the minimum required page 339
probability of belonging to the target class. In some
applications, however, the analyst may choose to increase or
decrease the cutoff value to classify fewer or more observations into
the target class due to asymmetric misclassification costs or uneven
class distributions. As a result, the choice of the cutoff value can
influence the confusion matrix and the resulting performance
measures. Example 8.6 highlights the impact of the cutoff value on
the performance measures of a classification model.

EXAMPLE 8.6
Table 8.10 shows the actual class and the predicted target
class probability for 10 observations of a categorical variable.
The table also includes the predicted target class memberships
based on five different cutoff values.

TABLE 8.10 Actual and Predicted Class Memberships Given Various Cutoffs
a. Calculate the misclassification rate with a cutoff value of 0.25
and a cutoff value of 0.75.
b. Compute the sensitivity and precision with a cutoff value of
0.25 and a cutoff value of 0.75.
c. Report and interpret performance measures with all five cutoff
values.
SOLUTION:

a. Recall that the misclassification rate is the overall proportion of

observations that are misclassified. With a cutoff value of 0.25,
we note that there is only one observation, the 5th, that is not
correctly classified; therefore, the misclassification rate is
computed as 1/10 = 0.10. With a cutoff value of 0.75,
observations 1st, 2nd, and 4th are not correctly classified,
resulting in the misclassification rate of 3/10 = 0.30.
b. Recall that the sensitivity is the proportion of the target class
cases that are classified correctly. With a cutoff value of 0.25,
we find that all five actual Class 1 observations are correctly
classified as the target classes, so the sensitivity is 5/5 = 1.
With a cutoff value of 0.75, the sensitivity decreases to 2/5 =
0.40. Recall that the precision is the proportion of the
 predicted target class cases that actually belong to the target
class. With a cutoff value of 0.25, we note that of the six
predicted target class cases, five of the cases belong to the
target class, so the precision is 5/6 = 0.833. With a cutoff
value of 0.75, the precision increases to 2/2 = 1.
c. Table 8.11 shows performance measures for the different
cutoff values in Table 8.10.

page 340

TABLE 8.11 Classification Performance Measures

As mentioned earlier, and as shown in Table 8.11, the

choice of the cutoff value can greatly influence the confusion
matrix and performance measures. Although the default for the
cutoff value for binary choice models is 0.50, the analyst may
choose a higher or lower cutoff value to classify fewer or more
observations into the target class to adjust for asymmetric
misclassification costs or uneven class distribution.

Performance Charts for Classification

It is sometimes more informative to have graphic representations to
assess the predictive performance of data mining models. The most
popular performance charts are the cumulative lift chart, the decile-
wise lift chart, and the receiver operating characteristic (ROC) curve.

Cumulative Lift Chart

A cumulative lift chart (also called a cumulative gains chart or a lift
chart) shows the improvement that a predictive model provides over
a random selection in capturing the target class cases. It also allows
us to determine the point at which the model’s predictions become
less useful. Figure 8.5 displays a cumulative lift chart for a
classification model. The chart shows the number (percentage) of
the target class cases gained by targeting a number (percentage) of
all the cases. The x axis indicates the number of cases selected. The
y axis indicates the cumulative number of target class (Class 1)
cases identified by the model. The orange diagonal line represents
the baseline model where observations are randomly selected,
without using predictive models. The slope of the diagonal line is the
number of Class 1 cases divided by the total number of cases, or the
probability of a randomly selected case belonging to Class 1.
Suppose 475 of the 5,000 total cases belong to Class 1. The
probability of a randomly selected case belonging to Class 1, or the
slope of the diagonal line, is 475/5,000 = 0.095; note that in Figure
8.5, the slope of the diagonal line is 190/2000 = 0.095.

FIGURE 8.5 The cumulative lift chart

 page 341

The blue upper curve in Figure 8.5 is called the lift curve, which
represents the ability of the classification model to identify Class 1
cases (for example, buyers). To construct a lift curve, the
observations in the validation data set are sorted in descending
order based on their predicted probability of belonging to the target
class (i.e., Class 1). The lift curve in Figure 8.5 shows that if we
select the top 2,000 observations with the highest predicted
probability of belonging to Class 1, we will capture 465 Class 1
observations. If we randomly select 2,000 observations, we will likely
capture only 190 Class 1 observations. The ratio of the number of
Class 1 observations captured by the model to the number captured
by random selection is called the lift.
The cumulative lift chart in Figure 8.5 shows that when we select
2,000 observations, the model provides a lift of 465/190 or 2.447.
This means that among the top 2,000 observations selected by the
model, we are able to capture 2.447 times as many Class 1
observations as compared to the 2,000 observations selected
randomly. Therefore, the cumulative lift chart shows how well a
predictive model can capture the target class cases, compared to
random selection. A high lift number is desirable in most applications
ranging from marketing to fraud detection as it means that we are
able to capture a large proportion of the target class cases by only
focusing on a small portion of cases with a high predicted probability.
A lift curve that is above the baseline is indicative of good predictive
performance of the model. The further above the baseline the lift
curve lies, the better the model’s ability to identify target class cases.

Decile-wise Lift Chart

The decile-wise lift chart conveys similar information as the
cumulative lift chart but presents the information in 10 equal-sized
intervals (e.g., every 10% of the observations). It is usually
presented as a bar chart where the y axis represents the ratio of the
target class cases identified by the model to that identified through
random selection. A decile-wise lift chart is shown in Figure 8.6.

FIGURE 8.6 The decile-wise lift chart

Figure 8.6 shows that the lift for the first 10% of the observations
(first bar) is about 7.1, which means that the top 10% of the
observations selected by the model contain 7.1 times as many Class
1 cases as the 10% of the observations that are randomly selected.
It also allows us to determine the point at which the model’s
predictions become less useful. For example, after the second
decile, the lift values dip below one, suggesting that the model will
capture fewer target class cases than random selection in the
subsequent deciles. Therefore, in this case, we want to focus on the
first 20% of the observations that have the highest predicted
probabilities of belonging to the target class if our goal is to capture
the target class cases (e.g., in a marketing campaign or when
detecting fraudulent credit card transactions).

page 342

Receiver Operating Characteristic (ROC) Curve

The receiver operating characteristic (ROC) curve shows the
sensitivity and specificity measures across all cutoff values and how
accurately the model is able to classify both target and nontarget
class cases overall. An ROC curve is shown in Figure 8.7. The x-
axis of the ROC curve is equal to 1 − Specificity, which ranges from
0 (classifying 100% of the nontarget class cases correctly) to 1
(classifying 0% of the nontarget class cases correctly), and the y-
axis is equal to sensitivity, which also ranges from 0 (classifying 0%
of the target class cases correctly) to 1 (classifying 100% of the
target class cases correctly); therefore, the perfect point on the
diagram is point (0,1) as it indicates that the model is able to
correctly classify 100% of both the target and nontarget class cases
simultaneously.

FIGURE 8.7 The ROC curve

 Figure 8.7 shows the ROC curve along with the orange diagonal
line that represents a baseline model that randomly assigns the
observations to the two classes based on the prior probability of
target class cases. Therefore, a good predictive model would have a
ROC curve that lies above the diagonal line. The greater the area
between the ROC curve and the baseline, the better the model is.
The overall performance of the model can also be assessed using
the area under the curve (AUC) measure, which ranges from 0
(worst possible model) to 1 (perfect model). The diagonal line has an
AUC value of 0.5. The software product you use will provide the
AUC value. The ROC curve in Figure 8.7 has an AUC value of
0.9457, which suggests that the model has very high predictive
performance.

SUMMARY OF CLASSIFICATION PERFORMANCE

DIAGRAMS
The cumulative lift chart represents the ability of the
classification model to identify target class cases. It shows the
improvement that a predictive model provides in capturing the
target class cases when compared with random selection.
 The decile-wise lift chart conveys the same information as the
cumulative lift chart but presents the information in equal-
sized intervals called deciles.
The receiver operating characteristic (ROC) curve shows the
sensitivity and the specificity measures across all cutoff
values and how accurately the model is able to classify both
target and nontarget class cases.

EXAMPLE 8.7
Let’s go back to the introductory case where Alissa and her
team at FashionTech want to assess the performance of
predictive models developed by MarketWiz. Alissa and her
team constructed a cumulative lift chart, a decile-wise lift chart,
and a ROC curve; see Figure 8.8. Interpret these performance
charts. [Shortly, we show how to create these graphs in Excel.]

page 343

FIGURE 8.8 Performance charts for Example 8.7

SOLUTION:
The cumulative lift curve in Figure 8.8a shows that the
classification model is superior compared to the baseline model
because it lies above the baseline model. The chart shows that
if 50 observations (25% of the total observations) with the
highest predicted probability of belonging to Class 1 are
selected, we would be able to capture 30 of the 48 actual Class
1 cases, or 30/48 = 0.625, or 62.5%. In the baseline model
where the cases are randomly selected, we would only be able
to capture 12/48 = 0.25, or 25% of the Class 1 cases.
Therefore, at 50 cases, the lift of the classification model is
0.625/0.25 = 2.5.
The decile-wise lift chart in Figure 8.8b shows the
improvement (lift) that the classification model provides in
capturing the target class cases for each decile compared to
the baseline model. For example, the lift for the first decile is
2.29, which implies that the top 10% of the page 344
observations with the highest predicted
probability of belonging to the target class contains 2.29 times
as many actual target class cases as the 10% of the
observations that are randomly selected. Because the lift
values of the first four deciles are above one, we will be able to
capture a very large portion of the target class cases by
selecting the top 40% of the observations with the highest
predicted probability of belonging to the target class. The
decile-wise lift chart in Figure 8.8b also shows that after eight
deciles, no target class cases can be found. This suggests that
we can ignore the bottom 20% of the observations that have
the lowest predictive probability if our goal is to capture the
target class cases.
The ROC curve in Figure 8.8c shows that the classification
model is superior compared to the baseline model because it
lies above the baseline model and is closer to the point (0,1)
than the baseline model is. The ROC curve suggests that the
classification model has higher sensitivity (ability to classify
target class cases correctly) and specificity (ability to classify
 nontarget class cases correctly) than the baseline model has
across all cutoff values. Please note that while it is difficult to
calculate the AUC measure when you plot the ROC curve
manually, software products, such as Analytic Solver and R,
will provide the AUC measure for the ROC curve automatically.

We now use Excel to replicate the graphs shown in Figure 8.8. The
graphs are based on the Class-Prob data file that we had earlier
used to create the confusion matrix in Excel.

Using Excel to Obtain Performance Charts

for Classification
The Cumulative Lift Chart
FILE
Class_Prob

. Open the original Class_Prob data file.

. Select the data range B1:B201. Choose Home > Sort & Filter >
Sort Largest to Smallest. Select Expand the selection in the Sort
Warning dialog box. Click Sort. This sorts all the observations
based on their predicted probability of belonging to the target class
in descending order. The closer the observation is to the top, the
higher the probability.
. Create a new column with the column heading “# of Cases” in
column D and another new column with the column heading
“Cumulative” in column E. Fill the cells D2:D201with the data
sequence 1 through 200. This column records the number of
observations. Enter the formulas =A2 in cell E2 and =E2 + A3 in cell
E3. Fill the range E4:E201 with the formula in E3. This creates a
column that records the cumulative number of Class 1 cases.
. Select the range D1:E201. Choose Insert > Scatter with Smooth
Lines to create a scatter chart that connects all the data points in a
line.
. To compare the cumulative lift chart of the classification model to
the baseline model, add a diagonal line that connects points (0,0)
and (200,48). As there are 48 actual Class 1 cases among the 200
cases, each case has the probability of 48/200 = 0.24 of belonging
to Class 1. Right-click the chart area and choose Select Data. Click
the Add button in the Select Data Source dialog box. Enter
“Baseline Model” in the Series name: box. Enter “0, 200” in the
Series X values: box. Enter “0, 48” in the Series Y values: box. Click
OK twice. [Formatting regarding the chart title, axis titles, etc. can
be done by selecting Format > Add Chart Element from the
menu.]

The Decile-Wise Lift Chart

FILE
Class_Prob

. Open the original Class_Prob data file and sort the data as
described in part B for the cumulative lift chart.
. We need to compute the ratio of the number of Class 1 cases
identified in each decile (200 × 0.10 = 20 cases in each decile) to
the number of Class 1 cases that we would have identified if 10% of
the cases were selected randomly. First type the page 345
column heading “# of Class 1 Cases in Each Decile” in
D1 and “Decile-Wise Lift” in E1. In order to find the number of Class
1 cases identified in the first decile—the numerator of the ratio for
the first decile—type =SUM(A2:A21) in D2. Similarly, to find the
number of Class 1 cases identified in the second decile, type
=SUM(A22:A41) in D3. The other eight deciles are found in a like
manner. Because there are a total of 48 Class 1 cases, the number
of Class 1 cases among the randomly selected 10% of the cases
would be 48 × 0.10 = 4.8 cases; this is the denominator of the ratio
for all deciles. Enter the formula =D2/4.8 in cell E2 to compute the
 lift for the first decile. Fill the range E3:E11 with the formula in E2.
This computes the lift values for the rest of the 9 deciles.
. Select the range E2:E11. Choose Insert > Insert Column or Bar
Chart > Clustered Columns to create a column chart. [Formatting
(regarding the chart title, axis titles, etc.) can be done by selecting
Format > Add Chart Element from the menu.]

The ROC Curve

Earlier, we used the Class_Prob data to derive the confusion matrix
and the resulting performance measures with a default cutoff value
of 0.5. In order to construct the ROC curve, we follow the same
procedure to find performance measures for the various cutoff
values. We created the Spec_Sens data set, which contains the
specificity and sensitivity measures for cutoff values ranging from 0
to 1 in 0.05 increments. The data set has the following columns: the
cutoff value (Cutoff), the specificity value (Specificity), and the
sensitivity value (Sensitivity). A portion of the data is shown in Table
8.12.
FILE
Spec_Sens

TABLE 8.12 A portion of the Spec_Sens data set

Cutoff Specificity Sensitivity

0 0 1

0.05 0.296053 0.958333

⋮ ⋮ ⋮

1 1 0
. Open the Spec_Sens data file.
. Select column C. Right-click and choose Insert to insert a column.
Type the column heading “1 − Specificity” in the new column. Insert
the formula = 1−B2 in cell C2. Fill the range C3:C22 with the
formula in C2. The new column computes the 1 − Specificity values
for all the cutoff values in column A.
. Select the range C1:D22. Choose Insert > Scatter with Smooth
Lines to create a scatter chart that connects all the data points in a
line. Change the chart title to “ROC Curve”. Change the x-axis title
to “1 − Specificity”. Change the y-axis title to “Sensitivity”.
. To compare the ROC curve of the classification model to that of the
baseline model, you will add a diagonal line that connects points
(0,0) and (1,1). Right-click the chart area and choose Select Data.
Click the Add button in the Select Data Source dialog box. Enter
“Baseline Model” in the Series name: box. Enter “0, 1” in the Series
X values: box. Enter “0, 1” in the Series Y values: box. Click OK
twice.

Performance Evaluation for Prediction

We now consider the performance measures for prediction models
where the target variables have numerical values. Similar to the
performance measures for regression models that were discussed in
Chapters 6 and 7, prediction performance of supervised data mining
models is assessed by computing the prediction error e, which is
the difference between the actual value y and predicted value of

the target variable. It is calculated as

page 346
 Note that prediction errors are calculated in the same way as the
residuals, but they are calculated using the validation data set rather
than the training data set. As with regression models, low values for
the prediction error e are preferred. The following definition box
highlights five performance measures: (1) the root mean square
error (RMSE), (2) the mean error (ME) or the average error (AE), (3)
the mean absolute deviation (MAD) or mean absolute error (MAE),
(4) the mean percentage error (MPE), and (5) the mean absolute
percentage error (MAPE).

PREDICTION PERFORMANCE MEASURES

All performance measures are based on the prediction error,
given n observations in the validation data set.

The root mean square error (RMSE) is similar to the

standard error of the estimate in the linear regression model
except that it is calculated for the validation data set, rather

than the training data set. It is calculated as

Because RMSE uses squared prediction errors, it penalizes

models with large prediction errors when used for comparing
models.
The mean error (ME) or the average error (AE) is the

average of the prediction errors and is calculated as

Because positive and negative errors offset each other,

average error is used as a measure of prediction bias. On
average, a positive ME implies the model underpredicts, a
negative ME implies it overpredicts, and an ME close to zero
implies that the predictions are unbiased.
 The mean absolute deviation (MAD) or mean absolute
error (MAE) is the average absolute error and is calculated

as MAD measures the average magnitude of the

errors, without considering their direction. Because errors are

not squared, models with large errors are not penalized as
heavily as in RMSE.
The mean percentage error (MPE) is the average

percentage error and is calculated as

Because the positive and negative percentage errors can

offset each other, MPE is often used as a measure of
prediction bias; however, it does not reflect the magnitude of
the errors.
The mean absolute percentage error (MAPE) is the
average absolute percentage error and is calculated as

It shows the error as a percentage of the

actual value, giving a sense of the magnitude of the errors.

The two most commonly used performance measures that

capture the magnitude of the errors are RMSE and MAD. They both
range from 0 to ∞, with lower values indicating a better prediction
model. Because RMSE gives a relatively high weight to large errors,
it is more useful when large errors are particularly undesirable. Note
that RMSE will always be larger or equal to MAD; RMSE = MAD only
if all of the errors have the same magnitude.
There is no way to define a “good” value for RMSE and MAD
because they both assume the same units as the original target
variable. For example, the RMSE and MAD values will change
depending on whether the purchase amounts are expressed in U.S.
dollars or in euros. This makes these measures difficult to interpret in
isolation, although they are useful for model selection. The model
with the lowest RMSE or MAD will always be preferred over
competing models, irrespective of the currency used in the
calculations. The main attraction of MAPE is that it is in percentage
terms and, therefore, it is somewhat easy to interpret.

page 347

EXAMPLE 8.8
FILE
Prediction

Recall from the introductory case that Alissa and her team at
FashionTech want to assess the performance of predictive
models developed by MarketWiz. Two prediction models were
designed to predict the one-year purchase amounts for
customers acquired through social media channels. After
applying the prediction models to the validation data set of 200
observations, Alissa and her team generated the Prediction
data set, which contains the predicted one-year purchase
amounts for the 200 observations. Each row of the data set
includes the actual target value (ActVal) and the corresponding
predicted value for two models (PredVal1 and PredVal2). A
portion of the data is shown in Table 8.13.

TABLE 8.13 Actual and Predicted Values

ActVal PredVal1 PredVal2

1326.50 1189.60 1169.60

 ActVal PredVal1 PredVal2

 546.00  493.30  506.30

⋮ ⋮ ⋮

 689.50  481.50  493.50

Using the data from Table 8.13, the five performance

measures, RMSE, ME, MAD, MPE, and MAPE, for both
prediction models were calculated. These measures are shown
in Table 8.14. Use these measures to compare the
performance of the models. [Shortly, we show how we compute
these measures in Excel.]

TABLE 8.14 Prediction Performance Measures

Performance Measure Model 1 Model 2

RMSE 171.3489 174.1758

ME  11.2530  12.0480

MAD 115.1650 117.9920

MPE −2.05% −2.08%

MAPE 15.51% 15.95%

SOLUTION:
For both models, ME and MPE, as measures of prediction bias,
offer conflicting results. The positive values of ME suggest that
 predictions for both models tend to underestimate the purchase
amounts, while negative values of MPE suggest that
predictions tend to slightly overestimate the purchase amounts
percentwise. Because Model 1 exhibits lower values of all
performance measures, we conclude that Model 1 has the
lower prediction bias as well as lower magnitude of the errors.
Therefore, Model 1 is the preferred model to predict the one-
year purchase amounts for customers acquired through social
media channels.

Using Excel to Obtain Performance

Measures for Prediction
FILE
Prediction

We replicate the performance measures for Model 1 shown in Table

8.14; performance measures for Model 2 can be computed similarly.
. Open the Prediction data file.
. Create a new column in column D to compute the prediction error
for each observation. Type the column heading “Error” in cell D1.
Enter the formula = A2−B2 in cell D2. Fill the range D3:D201 with
the formula.
page 348
. Create a new column in column E to compute the
squared prediction error for each observation. Type the column
heading “Squared Error” in cell E1. Enter the formula =D2^2 in cell
E2. Fill the range E3:E201 with the formula.
. Create a new column in column F to compute the absolute
prediction error for each observation. Type the column heading
“Absolute Error” in cell F1. Enter the formula =ABS(D2) in cell F2.
Fill the range F3:F201 with the formula.
. Create a new column in column G to compute the prediction error
as a percentage of the actual value for each observation. Type the
column heading “Percent Error” in cell G1. Enter the formula
=D2/A2 in cell G2. Fill the range G3:G201 with the formula.
. Create a new column in column H to compute the absolute percent
error for each observation. Type the column heading “Absolute
Percent Error” in cell H1. Enter the formula =ABS(G2) in cell H2. Fill
the range H3:H201 with the formula.
. Compute the performance measures anywhere in the worksheet by
entering the formulas as shown in Table 8.15. The resulting table
should be identical to the second column of Table 8.14.

TABLE 8.15 Excel Formulas for Prediction Performance Measures

Performance Measures Formula

RMSE =SQRT(SUM(E2:E201)/200)

ME =SUM(D2:D201)/200

MAD =SUM(F2:F201)/200

MPE =SUM(G2:G201)/200

MAPE =SUM(H2:H201)/200

SYNOPSIS OF INTRODUCTORY CASE

With an objective to help FashionTech acquire new customers as well as increase
sales from existing customers, MarketWiz develops predictive models for the online
apparel retailer. A classification model is estimated that predicts purchase probability
from FashionTech within 30 days of receiving a promotional message in their social
media account. A confusion matrix using 0.5 as the cutoff value is constructed to
assess the performance of the model, using 200 observations in the validation data
set. The confusion matrix shows that 29 buyers and 133 nonbuyers are correctly
classified by the model, suggesting an accuracy rate of 0.81.

Wayhome studio/Shutterstock

A sensitivity measure of 0.6042 and a specificity measure of 0.8750 suggest that

the model is better at classifying nonbuyers correctly than classifying buyers
correctly. Because the confusion matrix and the performance measures are sensitive
to the chosen cutoff value to classify the observations, MarketWiz explains that if the
cutoff value is lowered, then the model will classify more buyers correctly but at the
same time classify more nonbuyers incorrectly. Because the cost of sending social
media messages is low, FashionTech agrees to lower the cutoff to 0.25 to increase
sensitivity to 0.8125.
The performance of the model is also assessed graphically using the cumulative
lift chart, decile-wise lift chart, and ROC curve. The cumulative lift and decile-wise lift
charts show that the model can capture almost all the actual buyers by targeting
only a portion of social media users with the highest predicted probability of
purchasing. The ROC curve shows that the model performs better than the baseline
model (random guess) across all cutoff values in terms of
sensitivity and specificity. Overall, the estimated classification
page 349
model does a good job classifying potential customers into the purchase and no-
purchase classes.
Two models are estimated to predict the one-year purchase amounts for
customers acquired through social media channels. In order to compare the
performance of the models, several performance measures are calculated using 200
observations in the validation data set. Because the first estimated model exhibits
lower values for all performance measures, we conclude that the first model has
lower prediction bias as well as lower magnitude of the errors when compared with
the second model. Therefore, the first model is the preferred model to predict the
one-year purchase amounts for customers acquired through social media channels.
 EXERCISES 8.3
Mechanics
18. FILE Exercise_8.18. Construct a confusion matrix using the accompanying data
set that lists actual and predicted class memberships for 10 observations.
19. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity for the following confusion matrix.

Actual Class Predicted Class 1 Predicted Class 0

Class 1 120  67

Class 0  35 278

20. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity for the following confusion matrix.

Actual Class Predicted Class 1 Predicted Class 0

Class 1 254  10

Class 0  87 649

21. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity for the following confusion matrix.

Actual Class Predicted Class 1 Predicted Class 0

Class 1 295 378

Class 0 379 3948

22. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity for the following confusion matrix.

Actual Class Predicted Class 1 Predicted Class 0

Class 1 1,367    25

 Actual Class Predicted Class 1 Predicted Class 0

Class 0    35 1,573

23. FILE Exercise_8.23. Answer the following questions using the accompanying
data set that lists the actual class memberships and predicted Class 1 (target
class) probabilities for 10 observations.
a. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity using the cutoff value of 0.5.
b. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity using the cutoff value of 0.25.
c. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity using the cutoff value of 0.75.
24. Use the data in the preceding exercise.
a. What is the lift that the classification model provides if 5 observations are
selected by the model compared to randomly selecting 5 observations?
b. What is the lift that the classification model provides if 8 observations are
selected by the model compared to randomly selecting 8 observations?
25. FILE Exercise_8.25. Answer the following questions using the accompanying
data set that lists the actual class memberships and predicted Class 0 (nontarget
class) probabilities for 10 observations.
a. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity using the cutoff value of 0.5.
b. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity using the cutoff value of 0.25.
c. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity using the cutoff value of 0.75.
26. Use the data in the preceding exercise.
a. What is the lift that the classification model provides if 5 observations are
selected by the model compared to randomly selecting 5 observations?
b. What is the lift that the classification model provides if 8 observations are
selected by the model compared to randomly selecting 8 observations?
27. FILE Exercise_8.27. Compute the RMSE, ME, MAD, MPE, and MAPE using the
accompanying data set that lists the actual and predicted values for 10
observations.
28. FILE Exercise_8.28. Compute the RMSE, ME, MAD, MPE, and MAPE using the
accompanying data set that lists the actual and predicted values for 10
observations.

page 350
 Applications
29. A mail-order catalog company has developed a classification model to predict
whether a prospective customer will place an order if he or she receives a catalog
in the mail. If the prospective customer is predicted to place an order, then he or
she is classified in Class 1; otherwise, he or she is classified in Class 0. The
validation data set results in the following confusion matrix.

Actual Class Predicted Class 1 Predicted Class 0

Class 1 75  25

Class 0 84 816

a. Compute the misclassification rate, accuracy rate, sensitivity, precision, and

specificity of the classification model.
b. The confusion matrix was generated based on a cutoff rate of 0.5. If the cutoff
rate is decreased to 0.25, will the sensitivity value increase or decrease? What
about the specificity value?
c. If the profit from an order is much higher than the cost of sending a catalog,
would the company benefit from a higher or lower cutoff rate for classifying
prospective customers in Class 1 and Class 0?
30. Randy Johnson is an insurance adjustor for a national auto insurance company.
Using historical insurance claim data, Randy has built an insurance fraud
detection model with the help of a data analyst. Applying the model on the
validation data set generated the following confusion matrix. A fraudulent
insurance claim is a Class 1 case while a nonfraudulent claim is a Class 0 case.

Actual Class Predicted Class 1 Predicted Class 0

Class 1   130    170

Class 0 2,402 27,298

a. Compute the misclassification rate, accuracy rate, sensitivity, precision, and

specificity of the classification model.
b. Should the data be oversampled to produce a more reliable classification
model?
c. The confusion matrix was generated based on a cutoff rate of 0.5. If the cost of
missing a fraudulent insurance claim is much higher than the cost of
investigating a potentially fraudulent insurance claim, should the cutoff rate be
increased or decreased when classifying claims?
31. FILE Credit_Cards. A national bank has developed a predictive model for
identifying customers who are more likely to accept a credit card offer. If a
customer is predicted to accept the credit card offer, he or she is classified into
Class 1; otherwise, he or she is classified into Class 0. Applying the model to the
validation data set generated a table that lists the actual class membership and
predicted Class 1 probability of the 100 observations in the validation data set. A
portion of the table is shown below.

Customer Actual Class Predicted Class 1 Probability

1 1 0.22

2 0 0.54

⋮ ⋮ ⋮

100 1 0.5 

a. Specify the predicted class membership for the validation data set using the
cutoff values of 0.25, 0.5, and 0.75. Produce a confusion matrix in Excel based
on the classification results from each cutoff value.
b. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity of the classification model for each of the three cutoff values
specified in part a.
c. Create a cumulative lift chart and a decile-wise lift chart for the classification
model.
d. What is the lift that the classification model provides if 20% of the observations
are selected by the model compared to randomly selecting 20% of the
observations?
e. What is the lift that the classification model provides if 50% of the observations
are selected by the model compared to randomly selecting 50% of the
observations?
32. FILE IT_Professionals. Kenzi Williams is the Director of Human Resources of a
high-tech company. In order to manage the high turnover rate of the IT
professionals in her company, she has developed a predictive model for
identifying software engineers who are more likely to leave the company within
the first year. If the software engineer is predicted to leave the company within a
year, he or she is classified into Class 1; otherwise, he or she is classified into
Class 0. Applying the model to the validation data set generated a table that lists
the actual class membership and predicted Class 1 probability of the 100
observations in the validation data set. A portion of the table is shown below.

Employee Actual Class Predicted Class 1 Probability

 Employee Actual Class Predicted Class 1 Probability

1 0 0.09

2 0 0.22

⋮ ⋮ ⋮

100 1 0.41

a. Specify the predicted class membership for the validation data set using the
cutoff values of 0.25, 0.5, and 0.75. Produce a confusion
matrix based on the classification results from each cutoff
page 351
value.
b. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity of the classification model from each of the three cutoff values
specified in part a.
c. Create a cumulative lift chart and a decile-wise lift chart for the classification
model.
d. What is the lift that the classification model provides if 20% of the observations
are selected by the model compared to randomly selecting 20% of the
observations?
e. What is the lift that the classification model provides if 50% of the observations
are selected by the model compared to randomly selecting 50% of the
observations?
33. FILE Gamers. Monstermash, an online game app development company, has
built a predictive model to identify gamers who are likely to make in-app
purchases. The model classifies gamers who are likely to make in-app purchases
in Class 1 and gamers who are unlikely to make in-app purchases in Class 0.
Applying the model on the validation data set generated a table that lists the
actual class and Class 1 probability of the gamers in the validation data set. A
portion of the table is shown below.

Gamer Actual Class Predicted Class 1 Probability

1 0 0.022666

2 0 0.032561

⋮ ⋮ ⋮

920 0 0.040652
 a. Specify the predicted class membership for the validation data set using the
cutoff values of 0.25, 0.5, and 0.75. Produce a confusion matrix based on the
classification results from each cutoff value.
b. Compute the misclassification rate, accuracy rate, sensitivity, precision, and
specificity of the classification model from each of the three cutoff values
specified in part a.
c. Create the cumulative lift chart, decile-wise lift chart and ROC curve for the
classification model.
d. What is the lift that the classification model provides if 20% of the observations
are selected by the model compared to randomly selecting 20% of the
observations?
e. What is the lift that the classification model provides if 50% of the observations
are selected by the model compared to randomly selecting 50% of the
observations?
34. FILE House_Prices. A real estate company has built two predictive models for
estimating the selling price of a house. Using a small test data set of 10
observations, it tries to assess how the prediction models would perform on a new
data set. The following table lists a portion of the actual prices and predicted
prices generated by the two predictive models.

House Actual Price Predicted Price1 Predicted Price 2

1 $230,500 $254,000 $256,000

2 $209,900 $215,500 $223,400

⋮ ⋮ ⋮ ⋮

10 $328,900 $340,000 $324,500

a. Compute the ME, RMSE, MAD, MPE, and MAPE for the two predictive models.
b. Are the predictive models over- or underestimating the actual selling price on
average?
c. Compare the predictive models to a base model where every house is predicted
to be sold at the average price of all the houses in the training data set, which is
$260,500. Do the predictive models built by the real estate company outperform
the base model in terms of RMSE?
d. Which predictive model is the better-performing model?
35. FILE Customer_Spending. Mary Grant, a marketing manager of an online
retailer, has built two predictive models for estimating the annual spending of new
customers. Applying the models to the 100 observations in the validation data set
generates a table that lists the customers’ actual spending and predicted
spending. A portion of the data set is shown below.
 Actual Predicted Spending Predicted Spending
Customer
Spending 1 2

1 555 587 487

2 190 198 182

⋮ ⋮ ⋮ ⋮

100 464 380 312

a. Compute the ME, RMSE, MAD, MPE, and MAPE for the two predictive models.
b. Are the predictive models over- or underestimating the actual selling price on
average?
c. Compare the RMSE and MAD of the two predictive models. Which predictive
model performs better?
d. Compare the better-predictive model to a base model where every customer is
predicted to have the average spending of the cases in the training data set,
which is $290. Does the better-predictive model built by Mary outperform the
base model in terms of RMSE?

page 352

8.4 PRINCIPAL COMPONENT

ANALYSIS
LO 8.4
Conduct principal component analysis.

In many real-world situations, we may find a data set with a large

number of variables. This is especially true in scientific, medical, and
consumer research. Data sets that contain many variables
(dimensions) are referred to as high-dimensional data. Sometimes
the number of variables even exceeds the number of observations in
the data set. In some cases, we may be able to eliminate some of
the redundant variables, but in many other cases simply dropping
the variables might result in losing valuable information.
Principal component analysis (PCA) is a useful dimension
reduction technique that is especially effective when we want to
reduce the number of variables but cannot completely eliminate any
variables from the predictive model without losing important
information. PCA transforms a large number of possibly correlated
variables into a smaller number of uncorrelated variables called
principal components. These principal components allow us to
summarize the data with a smaller number of representative
variables that collectively explain most of the variability in the data.
PCA is considered an unsupervised data mining technique that
allows us to select a relatively small number of principal components
that drive most of the patterns found in the data. It can, for example,
be used to visualize and pre-process data before applying other
unsupervised learning methods such as clustering. We can also use
PCA with supervised learning methods, such as linear regression
and logistic regression, where a large number of predictor variables
can be substituted with a smaller number of principal components.
Given n observations, the main task is to find a sequence of
principal components PC1, PC2, . . . , PCm, which are the weighted
linear combination of the k original variables x1, x2, . . . , xk, where m
≤ k. The derivation of the weights involves a mathematical technique
in linear algebra called eigen decomposition, which is outside of the
scope of this text. Here, we simply use Analytic Solver and R to
derive the weights, the corresponding principal components, and the
variability that they account for in the data. These mutually
uncorrelated principal components are fewer in number than the
original variables, yet they retain most of the information in the data.

PRINCIPAL COMPONENT ANALYSIS

 When faced with a large set of correlated variables, principal
component analysis summarizes this set with a smaller number
of representative variables, called principal components.
Principal components are uncorrelated variables whose values
are the weighted linear combinations of the original variables.
The first principal component accounts for most of the
variability in the data, followed by the second principal
component, and so on.

To illustrate how PCA works, we use a simple example. Table

8.16 shows a portion of data that lists the following nutritional content
for 37 candy bars: the calories per serving page 353
(Calories), total fat (Fat in grams), protein (Protein in
grams), and carbohydrates (Carb in grams). For ease of exposition,
we will show how a single principal component can capture most of
the variability between two of the variables: Fat (denoted as x1) and
Protein (denoted as x2).
FILE
CandyBars

TABLE 8.16 Nutritional Content for Candy Bars (n = 37)

Figure 8.9 shows a scatterplot of x1 (Fat) and x2 (Protein).

FIGURE 8.9 Scatterplot of variables x1 and x2

 When we analyze data, we are usually interested in the
information associated with the variation in the data. If there is no
variation in the data (all data points have the same values), there
would be nothing for us to study! Therefore, the direction where
there is the most variation contains the most information. The first
principal component, represented by the purple line in Figure 8.9, is
designated along the direction where the data vary the most.
Visually, this line captures the largest amount of variability in the
data. Alternatively, we can interpret the first principal component as
the line that is closest to the data.
We can also think of the values of the first principal component
as a weighted linear combination of the x1 and x2 values. In Figure
8.9, there appears to be a linear relationship between x1 and x2, and
so it is reasonable to expect that a linear combination of the x1 and
x2 values will adequately capture that relationship. In other words,
the first principal component appears to capture most of the
information contained in the two variables. As mentioned earlier,
PCA is effective especially when the variables are highly correlated.
The second principal component PC2 is a linear combination of
the variables that is uncorrelated with PC1 and captures the second
largest variability in the data. In Figure 8.9, the blue line represents
the second principal component of the data, where PC2 is
perpendicular to (or, in a more formal term, orthogonal to) or
uncorrelated with PC1.
By applying PCA to a data set, we ensure that all of the resulting
principal components are uncorrelated with each other, while
collectively capturing almost all of the information (variation)
contained in the data. In our example, the first principal component
(purple line) represents a very large portion of the total variation in
the data, and the second principal component (blue line) represents
much of the remaining variation.
In Figure 8.10, the data points have been rotated so that the first
and second principal component directions coincide with the axes.
Note that the data have not changed; we are simply looking at the
data from a different angle. The positions of the data points relative
to each other remain the same. The only things that have changed
are the axes used to derive the coordinates of the data points.
Another difference here is that the old axes were well defined and
easily understood; they represented the fat and protein content of
candy bars. The new axes, however, are more abstract and simply
represent weighted linear combinations of the original variables. The
values of the first and second principal components page 354
for each data point are the projected values of x1
and x2 on the new axes.

FIGURE 8.10 Scatterplot of the rotated data

 In this example with two original variables, or, equivalently, two-
dimensional data, we can construct at most two principal
components. Even so, the first component contains the most
information on x1 and x2. As the large amount of variation in the data
is now captured by the first principal component PC1, substituting the
original variables x1 and x2 with PC1 will result in little loss of
information. In other words, we can replace two original variables x1
and x2 with one principal component PC1 without losing a lot of
information. PCA is sensitive to the scale of the data. In the candy
bar example, because fat has more variability than protein, it will
dominate the PCA results, if not standardized. Therefore, it is
recommended that data be standardized (e.g., converted to z-
scores) prior to performing PCA.
When using Analytic Solver or R to perform PCA, quite a bit of
output is generated. We focus on the weights for computing the PC
scores and the variance distribution. Figure 8.11 summarizes this
portion of output for the candy bar example. (Instructions for
conducting PCA in Analytic Solver and R will be provided shortly.)

FIGURE 8.11 PCA results for the candy bar example

a. Weights for Fat and Protein

 Feature\Component PC1 PC2

Fat (g) 0.707107 −0.70711

Protein (g) 0.707107  0.707107

b. Variance Distribution

PC1 PC2

Variance  1.784387  0.215613

Variance Percentage 89.21933 10.78067

Cumulative Variance % 89.21933 100

c. Original Values, z-Scores, and PC Scores

The PC scores shown in the last two columns of Figure 8.11c are
based on the weights shown in Figure 8.11a. The following equation
is used to compute the mth principal component score for the ith
observation, denoted as PCm,i:
 page 355

where wmk is the weight for the kth variable of the mth principal
component and zki is the standardized value for the ith observation
of the kth variable.
For example, let’s calculate the PC1 score for the first candy bar,
three Musketeers, denoted PC1,3M. Figure 8.11a shows that the
weights for fat and protein are both 0.707107, so w11 = 0.707107
and w12 = 0.707107. Figure 8.11c shows that the standardized
values for three Musketeer’s fat and protein content are −0.7860 and
−0.8982, respectively, so z11 = −0.7860 and z21 = −0.8982. We
calculate PC1,3M as:

The PC1 scores for each candy bar appear in the second-to-last
column in Figure 8.11c. Similarly, the PC2 score for three
Musketeers, denoted PC2,3M, is calculated as:

The PC2 scores for each candy bar appear in the last column in
Figure 8.11c.
Figure 8.11b shows the distribution of variance across all the
principal components. PC1 accounts for 89.22% of the total variance
while PC2 only accounts for 10.78% of the variance. PCA has
redistributed the variance in the data so that using PC1 alone will
result in very little loss of information.
So far, we have worked with two-dimensional data with which we
can construct at most two principal components. If we had k original
variables, x1, x2, . . . , xk, then up to k principal components could be
constructed. For example, with the CandyBars data on calories, fat,
protein and carb, we could have constructed up to four principal
components for the analysis. The amount of variance explained
decreases from the first principal component to the last principal
component. Therefore, the first few principal components often
represent a very large portion of the total variance (or information) in
the data. This allows us to significantly reduce the number of
dimensions by only including the first few principal components
instead of the entire list of original variables without losing a lot of
information. In practice, a predetermined percentage of variance
explained (e.g., 80%) is used as a threshold to determine the
number of principal components to retain. In addition, the principal
components are uncorrelated; therefore, multicollinearity would not
be an issue in subsequent data mining analysis.

Using Analytic Solver and R to Perform

Principal Component Analysis
In Example 8.9, we will demonstrate how to use Analytic Solver and
R to perform PCA with 12 variables.

EXAMPLE 8.9
Consider country-level health and population measures for 38
countries from World Bank’s 2000 Health Nutrition and
Population Statistics database. For each country, the measures
include death rate per 1,000 people (Death Rate, in %), health
expenditure per capita (Health Expend, in US$), life
expectancy at birth (Life Exp, in years), male adult mortality
rate per 1,000 male adults (Male Mortality), female adult
mortality rate per 1,000 female adults (Female Mortality),
annual population growth (Population Growth, in %), female
population (Female Pop, in %), male population (Male Pop, in
%), total population (Total Pop), size of labor force (Labor
Force), births per woman (Fertility Rate), and birth rate per
1,000 people (Birth Rate). A portion of the data is shown in
Table 8.17.

page 356
FILE
Health

TABLE 8.17 Country-Level Health and Population Measures

SOLUTION:
As mentioned earlier, PCA is especially effective when the
variables are highly correlated. The correlation analysis on the
Health data reveals strong correlation between many of the
variables. For example, the correlation coefficient between
male and female mortality rates is 0.95, indicating a strong
positive correlation between the two variables. The correlation
measure tells us that some of the information in male mortality
of a country in the form of variance is duplicated in female
mortality. However, if we remove one of the variables from the
data set, we may lose a sizable portion of the variance
(information) in the data. Applied to a data set with a large
number of correlated variables, PCA can significantly reduce
data dimensions while retaining almost all of the information in
the data.
Next, we demonstrate how to use Analytic Solver and R to
perform PCA on the Health and Population data set. We also
 summarize the findings.
Using Analytic Solver

a. Open the Health data file.

b. From the menu, choose Data Mining > Transform >
Principal Components.
c. See Figure 8.12. Click on the ellipsis button next to the Data
range and highlight cells $A$1:$M$39. Make sure that the box
First Row Contains Headers is checked. The Variables In
Input Data box will populate. Select and move all the variables
except Country Name to Selected Variables box. Accept other
defaults and click Next.

FIGURE 8.12 Principal component analysis dialogue box—

step 1 of 3

Source: Microsoft Excel

d. See Figure 8.13. Analytic Solver offers you two options for the
number of principal components to display in the results. The
first option is to display a fixed number of page 357
principal components. The total number of
principal components that can be calculated is the same
 number as the number of variables. Therefore, the maximum
number of principal components to display in this example is
12. The other option allows you to display the smallest number
of components that collectively explain at least a certain
percentage of the total variance in the data, where the default
threshold is 95%. Choose the option to display all 12
components.
Analytic Solver also offers you the option of choosing one
of the two methods for performing PCA: using the covariance
matrix or the correlation matrix. Recall that the correlation
coefficients are unit-free and, therefore, choosing the Use
Correlation Matrix option is equivalent to using standardized
data for PCA. As mentioned earlier, it is recommended that
data be standardized (converted to z-scores) prior to
performing PCA when the scales are different. The scales of
the data in the Health data set vary significantly; therefore, we
will choose the Use Correlation Matrix option to remove the
impact of data scale on the results. Click Next.

FIGURE 8.13 Principal component analysis dialogue box—

step 2 of 3

Source: Microsoft Excel

e. In the final dialog box, make sure the box next to Show
principal components score is checked. Click Finish.
 Analytic Solver inserts two new worksheets (PCA_Output
and PCA_Scores). Figure 8.14c shows a portion of the
principal component scores for the 38 countries from the
PCA_Scores worksheet. Here, Record 1 refers to Argentina,
Record 2 refers to Austria, and so on. Recall that the principal
component scores are the weighted linear combinations of the
original variables. Figures 8.14a and 8.14b summarize the
results from the PCA_Output worksheet, showing the weights
used to calculate the principal component scores and the
amount of variance accounted for by each principal
component. We see that the first principal component accounts
for a very large portion, 46.6189%, of the total variance in the
data. Among the original variables, Life Exp has the most
weight, in absolute terms, for the first principal component,
followed by Birth Rate. Cumulatively, the first four principal
components account for 93.3486% of the total variance in the
data, suggesting that if we reduce the dimension of the data
from the original 12 variables to the four newly calculated
principal components, we would still be able to retain almost all
the information in the data.

FIGURE 8.14 PCA results using Analytic Solver

 It is important to note that different versions of Analytic
Solver and Excel may yield slightly different results, especially
for a very small value such as the principal component scores,
weights, and variances for PC12. The results for the other
principal components (i.e., the first four components in the
current example) should remain consistent across versions.

page 358

Using R
The following instructions are based on R version 3.5.3. Also,
the results from R may be slightly different from the Analytic
Solver results due to the differences in the algorithms
employed by the two software. In fact, the sign of some of the
PC scores in R may be the opposite of the ones computed in
Analytic Solver. These sign differences, however, will have no
impact on the resulting data mining techniques.
a. Import the Health data into a data frame (table) and label it
myData.
b. Display a few observations in the myData data frame to
inspect the data using the head function. Enter:

> head(myData)
c. We exclude the Country Name variable (1st variable) from the
analysis, standardize the rest of the variables, and store their
standardized values in a data frame called myData.st. Use the
scale function to standardize the data. Enter:

> myData.st <- scale(myData[ , -1])

d. We use the prcomp function to perform PCA and use the
summary function to display the PCA results. Enter:

> pca <- prcomp(myData.st)

> summary(pca)

Figure 8.15a shows a portion of the PCA results. As you can

see from the cumulative proportion row in the table, the first
four principal components account for 93.35% of the total
variance in the data. This is consistent with the result from
Analytic Solver even though the exact cumulative proportion of
variance is slightly different.

FIGURE 8.15 PCA results using R

e. To review the weights for computing the principal components,
we display the rotation property of the PCA results. Enter:

> pca$rotation
Figure 8.15b shows a portion of the weights. Again, the
weights are very similar to the ones produced by Analytic
Solver.
page 359
f. To review the principal component scores, we
display the x property of the PCA results. Enter:

> pca$x
Figure 8.15c displays a portion of the principal component
scores of the observations. In general, the results are
consistent with those produced by Analytic Solver.
g. In order to list countries (1st variable) along with the principle
component scores, we first use the data.frame statement to
combine the myData data frame with the principal component
scores. We then remove columns 2 to 13, which contain the
 original variables in the myData data frame. We display a few
observations in the new data frame using the head function.
Enter:

> newData <- data.frame(myData, pca$x)

> newData <- newData[ , -(2:13)]
> head(newData)

As mentioned earlier, PCA allows us to significantly reduce the

dimension of the data by replacing a large number of original
variables with a few uncorrelated principal components that
retain a very large portion of the information in the data.
However, these advantages come at a price. As principal
components are simply the weighted linear combinations of the
original variables, they are abstract and cannot be easily
interpreted. In addition, PCA only works with numerical data.

EXERCISES 8.4
Note: and indicate that the exercises must be solved using Analytic Solver

and R, respectively.

Mechanics

36. FILE Exercise_8.36. Perform principal component analysis on the

accompanying data set.

a. Use the data with the covariance matrix to compute the first two principal
components. What are the weights for computing the first principal component
scores? What percent of the total variability is accounted for by the first principal
component?
b. Use the data with the correlation matrix to compute the first two components.
What are the weights for computing the first principal component scores? What
percent of the total variability is accounted for by the first principal component?
c. Comment on the differences in the results of parts a and b.
37. FILE Exercise_8.37. Perform principal component analysis on the

accompanying data set.

a. Use the data with the covariance matrix and choose Smallest # components
explaining at least: 85% of variance. How many principal components were
created? What percent of total variance is accounted for by the
calculated principal components? What are the weights for
page 360
computing the first principal component scores?
b. Use the data with the correlation matrix to perform PCA. Keep all the other
options the same as in part a. How many principal components were created?
What percent of total variance is accounted for by the calculated principal
components? What are the weights for computing the first principal component
scores?

38. FILE Exercise_8.38. Perform principal component analysis on the

accompanying data set.

a. Use the data with the covariance matrix to compute the seven principal
components. How many principal components do you need to retain in order to
account for 80% of the total variance? What are the weights for computing the
first principal component scores?
b. Use the data with the correlation matrix to compute the seven principal
components. How many principal components do you need to retain in order to
account for 80% of the total variance? What are the weights for computing the
first principal component scores?
c. Comment on the differences in the results of part a and b.

39. FILE Exercise_8.39. Standardize the accompanying data set, and then

perform principal component analysis.

a. What percent of total variance is accounted for by the first three principal
components?
b. What are the weights for computing the first principal component scores?
c. What is the first principal component score for the first observation?

40. FILE Exercise_8.40. Standardize the accompanying data set, and then

perform principal component analysis.

a. How many principal components do you need to account for at least 80% of the
total variance?
b. What are the weights for computing the first principal component scores?
c. What is the second principal component score for the first observation?
41. FILE Exercise_8.41. Perform principal component analysis on the

accompanying data set.

a. Use standardized data to compute the principal components. Select the number
of principal components that account for at least 80% of the total variance. How
many principal components are displayed by Analytic Solver?
b. What are the weights for computing the first principal component scores? What
is the second principal component score for the first observation?
42. The following table displays the weights for computing the principal components
and the data for two observations.

Weight PC1 PC2

x1 −0.892  0.452

x2 −0.452 −0.892

x1 x2

Observation 1 4.04 12.77

Observation 2 4.10 12.18

a. The mean and standard deviation for x1 are 3.5 and 1.2, respectively. The
mean and standard deviation for x2 are 11.75 and 3.8, respectively. Compute
the z-scores for the x1 and x2 values for the two observations.
b. Compute the first principal component score for observation 1.
c. Compute the second principal component score for observation 2.
43. The following table displays the weights for computing the principal components
and the data for two observations.

Weight PC1 PC2

x1 −0.81  0.59

x2 −0.59 −0.81

x1 x2
 Weight PC1 PC2

Observation 1 7.30 342.90

Observation 2 3.20 258.60

a. The mean and standard deviation for x1 are 5.8 and 2.4, respectively. The
mean and standard deviation for x2 are 380.5 and 123.4, respectively. Compute
the z-scores for the x1 and x2 values for the two observations.
b. Compute the first principal component score for observation 1.
c. Compute the second principal component score for observation 2.
44. The following table displays the weights for computing the principal components
and the standardized data (z-scores) for three observations.

Weight PC1 PC2 PC3

x1 −0.57 0.62  0.55

x2 −0.64 0.09 −0.76

x3  0.52 0.78 −0.34

z1 z2 z3

Observation 1  0.71  0.33 −0.36

Observation 2  2.81  2.44 −1.41

Observation 3 −0.76 −1.18  1.37

a. Compute the first principal component score for observation 1.

page 361
b. Compute the second principal component score for
observation 2.
c. Compute the third principal component score for observation 3.
45. The following Analytic Solver results were generated by a principal component
analysis.
 a. How many original variables are in the data set?
b. What percent of variance does the third principal component account for?
c. How many principal components need to be retained in order to account for at
least 95% of the total variance in the data?
46. The following R results were generated by a principal component analysis.

a. How many original variables are in the data set?

b. What percent of variance does the third principal component account for?
c. How many principal components need to be retained in order to account for at
least 85% of the total variance in the data?

Applications
47. FILE Development. The accompanying data set contains some economic
indicators for 11 African countries collected by the World Bank in 2015. The
economic indicators include annual % growth in agriculture (Agriculture), annual
% growth in exports (Exports), annual % growth in final consumption (Final
consumption), annual % growth in GDP (GDP growth), and annual % growth in
GDP per capita (GDP per capita growth). A portion of the data set is shown in the
accompanying table.

a. Perform correlation analysis on the data. Discuss the correlation between the
variables in the data set. Which variable pair has the highest correlation?
 b. Conduct principal component analysis on all the variables in the data set.
Standardize the data prior to the analysis or use the correlation matrix if you are
using Analytic Solver.
c. Allow the maximum number of principal components to be calculated by the
software. How many principal components are computed? What percent of the
total variability is accounted for by the first three principal components? How
many principal components must be retained in order to account for at least
80% of the total variance in the data?
d. Which original variable is given the highest weight to compute the first principal
component? Which original variable is given the highest weight to compute the
second principal component?
e. Create a new data set that contains the country names and the principal
components that account for at least 80% of the total variance in the data.

48. FILE Football_Players. Beza Gordon-Smith is a high school senior in

northern California who loves watching football. She keeps track of football results
and statistics of the quarterbacks of each high school team. The accompanying
table shows a portion of the data that Beza has recorded, with the following
variables: the player’s number that Beza assigns to each quarterback (Player),
team’s mascot (Team), completed passes (Comp), attempted pass (Att),
completion percentage (Pct), total yards thrown (Yds), average yards per attempt
(Avg), yards thrown per game (Yds/G), number of touchdowns (TD), and number
of interceptions (Int).

a. Conduct principal component analysis on all variables, except the players’

number, in the data set. Should you use the covariance matrix or the correlation
matrix for the analysis in this case? Explain.
b. Allow the maximum number of principal components to be calculated by the
software. How many principal components are computed? What percent of the
total variability is accounted for by the first three principal components? How
many principal components must be retained in order to account for at least
80% of the total variance in the data?

page 362
c. Which original variable is given the highest weight to compute
the first principal component? Which original variable is given the highest
weight to compute the second principal component?
 d. What is the principal component 1 score for the first record (Player 1)?

49. FILE Baseball_Players. Ben Derby is a scout for a college baseball team.

He attends many high school games and practices each week in order to evaluate
potential players to recruit during each college recruitment season. He also keeps
detailed records about each prospective player. His college team is in particular
need to add another hitter to its roster next season. Luckily, Ben has information
on 144 high school players who have played at least 100 games during the
previous season. portion of the data is shown in the accompanying table that
includes the following variables: Player’s ID (Player), the number of games played
(G), at bats (AB), runs (R), hits (H), homeruns (HR), runs batted in (RBI), batting
average (AVG), on base percentage (OBP), and slugging percentage (SLG).

a. Conduct principal component analysis on all the variables, except the player’s
ID, in the data set. Standardize the data prior to the analysis.
b. How many principal components are computed? How many principal
components must be retained to account for at least 80% of the total variance in
the data?
c. Display the weights used to compute the first principal component scores.
d. Create a new data set that contains the players’ names and the principal
components that account for at least 80% of the total variance in the data.
50. FILE Internet_Addiction. Internet addiction has been found to be a widespread
problem among university students. A small liberal arts college in Colorado
conducted a survey of Internet addiction among its students using the Internet
Addiction Test (IAT) developed by Dr. Kimberly Young. The IAT contains 20
questions that measure various aspects of Internet addiction. Individuals who
score high on the IAT are more likely to have problematic Internet use. Also
included in the data are the sex of the students (Sex, 1 for female and 0 for male)
and whether or not they study at the graduate level (1 for Graduate and 0 for
Undergraduate). A portion of the data set is shown in the accompanying table.
 a. Conduct principal component analysis on all the variables except the Record
Number. Should you standardize the data? Explain.
b. How many principal components are computed? What percent of the total
variability is accounted for by the first principal component? How many principal
components must be retained in order to account for at least 80% of the total
variance in the data?
c. Which original variable is given the highest weight to compute the first principal
component? Which original variable is given the highest weight to compute the
second principal component?
d. What is the principal component 1 score for the first record?

51. FILE Nutritional_Facts. Jenny, a first-year Nutrition Studies student at

Hillside College, is conducting research on various common food items and their
nutritional facts. She compiled a data set that contains the nutrition facts on 30
common food items. Jenny feels that many of the nutritional facts of these food
items are highly correlated and wonders if there is a way to reduce the
dimensionality of the data set by creating principal components. A portion of the
data set is shown in the accompanying table. (Values are based on 100 grams of
the food items.)

a. Conduct principal component analysis on all the variables except the Name
variable. Should you standardize the data? Explain.
b. How many principal components are computed? What percent of the total
variability is accounted for by the first principal component? How many principal
components must be retained in order to account for at least 80% of the total
variance in the data?
c. Display the weights used to compute the first principal component scores.
 d. Create a new data set that contains the names of the food items and the
principal components that account for at least 80% of the total variance in the
data.

page 363
52. FILE Happiness. Since 2014, the United Nations has
conducted annual studies that measure the level of happiness among its member
countries. Experts in social science and psychology are commissioned to collect
relevant data and define measurements related to happiness. Happiness
measurements are based on survey questions such as how people feel about
their life (i.e., life ladder), levels of positive and negative emotion, freedom to
make choices (Life choices), and aggregate indicators such as social support, life
expectancy, and relative household income. These data are converted into
numerical scores for each member country. The accompanying table shows a
portion of the United Nation’s happiness data.

a. Conduct principal component analysis on all the variables except the Country
variable. Should you standardize the data? Explain.
b. What percent of the total variability is accounted for by the first principal
component? How many principal components must be retained in order to
account for at least 80% of the total variance in the data?
c. Display the weights used to compute the first principal component scores.
Which original variable is given the highest weight to compute the second
principal component?
d. What is the principal component 1 score for the first record (Albania)?

53. FILE Tennis. In tennis, how well a player serves and returns serves often

determine the outcome of the game. Coaches and players track these numbers
and work tirelessly to make improvement. The accompanying table shows a
sample of data that includes local youth tennis players. The relevant variables
include the number of aces (Aces), number of double faults (DF), first serve
percentage (1ST SRV), first serve win percentage (1ST SRV WIN), break point
saved percentage (BP SVD), percentage of service games won (SRV WIN), first
serve return win percentage (1ST RTN WIN), percentage of return games won
(RTN WON), and break point conversion percentage (BP CONV).
 a. Conduct principal component analysis on all variables except the Player
variable and display the weights used to compute the first principal component
scores.
b. Which original variable is given the highest weight to compute the first principal
component? Which original variable is given the highest weight to compute the
second principal component?
c. Create a new data set that contains the Player column and the principal
components that account for at least 90% of the total variance in the data.
54. FILE Stocks. Investors usually consider a variety of information to make
investment decisions. The accompanying table displays a sample of large publicly
traded corporations and their financial information. Relevant information includes
stock price (Price), dividend as a percentage of share price (Dividend), price to
earnings ratio (PE), earnings per share (EPS), book value, lowest and highest
share prices within the past 52 weeks (52 wk low and 52 wk high), market value
of the company’s shares (Market cap), and earnings before interest, taxes,
depreciation, and amortization (EBITDA in $billions).

a. Conduct principal component analysis on all the variables except the Name
variable. Should you standardize the data? Explain.
b. What percent of the total variability is accounted for by the first principal
component? How many principal components must be retained in order to
account for at least 80% of the total variance in the data?
c. Which original variable is given the highest weight to compute the first principal
component? Which original variable is given the highest weight to compute the
second principal component?
d. What is the principal component 1 score for the first record (3M)?
 page 364

8.5 WRITING WITH BIG DATA

An inherent problem that comes with big data is that data sets may be complex and
contain many variables (dimensions). To reduce information redundancy and improve
the stability of the resulting models, it is recommended that dimension reduction
techniques be implemented to result in a more manageable number of variables that
capture most, if not all, of the information in the original data. Dimension reduction
techniques, such as principal component analysis, are effective in achieving this goal
and are often performed prior to model building activities.

Case Study
FILE
NBA

Merrick Stevens is a sports analyst working for ACE Sports Management, a sports
agency that represents over 200 athletes. Merrick is tasked with analyzing sports-
related data and developing a predictive model for the National Basketball
Association (NBA). He uses the NBA data set that contains information on 30
competing NBA teams and 455 players. The player statistics are for several seasons
as well as for their career. Because a player’s salary is based on his performance
over multiple seasons, Merrick decides to only look at the career regular season data
rather than data of a particular season.
Given the large number of predictor variables that may explain a player’s salary,
Merrick decides to investigate whether principal component analysis may be
advantageous as a first step in model building.

Sample Report – Dimension Reduction of

the NBA Player Data
An NBA player’s salary is determined by a wide range of variables. Prior to
constructing a model that can be used to predict an NBA player’s salary, data
are collected for 23 possible predictor variables for 455 NBA players. These
variables include physicality variables, such as a player’s age, height, and
weight, as well as performance variables, such as the number of games
played, the number of baskets made, the number of three-pointers made, and
so forth.

Marcio Sanchez/Ap/Pool/EPA/Shutterstock

High correlations between many of the predictor variables suggest that

information redundancy exists in the data. In order to eliminate potential
multicollinearity problems and improve model stability in the resulting salary
model, dimension reduction using principal component analysis (PCA) is
performed on the 23 predictor variables.
Prior to performing the PCA analysis, the data are standardized in order to
remove the impact of data scales. The PCA analysis is not restricted with
respect to the number of principle components to produce; thus, given that
the analysis used 23 predictor variables, 23 principal components are
estimated. Table 8.18 shows a portion of the PCA output with respect to
weights and variances. The first principal component accounts for 44.1218%
of the total variance, and Points (Average points per game) has the largest
weight for the first principal component. Referencing the Cumulative Variance
% row in Table 8.18, it is found that the first seven principal components
account for almost 90% of the total variance in the original data.

page 365

TABLE 8.18 Principal Component Weights and Variances

 In future analyses, it is advisable to use the first seven principal
components as the predictor variables for building models that predict an
NBA player’s salary. A portion of the principal component scores for the first
seven principal components are displayed in Table 8.19. By using the seven
principle components instead of the original 23 predictor variables,
information redundancy has been removed. In addition, a large number of
highly correlated predictor variables have been replaced with a smaller set of
uncorrelated principal components that retain at least 90% of the information
in the original data.

TABLE 8.19 Principal Component Scores

Suggested Case Studies

Principal component analysis can be applied to many situations and different data
sets, especially the big data that accompany this text. Here are some suggestions.

Report 8.1 FILE Longitudinal_Information. Subset the data to select only those
individuals who lived in an urban area. Reduce the dimensionality of the data by
converting numerical variables such as age, height, weight, number of years of
education, number of siblings, family size, number of weeks employed, self-esteem
scale, and income into a smaller set of principal components that retain at least 90%
of the information in the original data. Note: You will need to standardize the data
prior to PCA as the scales of the variables are different.

Report 8.2 FILE College_Admissions. Select the data for one of the three
colleges. Reduce the dimensionality of the data by converting numerical variables
such as high school GPAs, MCA GPAs, SAT scores, and ACT scores into a smaller
set of principal components that retain at least 90% of the information in the original
data. Then use the principal components as predictor variables for building models
for predicting admission.

Report 8.3 FILE House_Price. Choose one of the college towns in the data set.
Reduce the dimensionality of the data by converting numerical variables such as
number of bedrooms, number of bathrooms, home square footage, lot square
footage, and age of the house into a smaller set of principal components that retain
at least 90% of the information in the original data. Then use the principal
components as predictor variables for building models for predicting sale prices of
houses.

Report 8.4 FILE TechSales_Reps. Subset the data to include only the employees
who work in the hardware division and have a college degree. Reduce the
dimensionality of the data by converting numerical variables such as age, years with
the company, number of certifications, feedback scores, and salary into a smaller set
of principal components that retain at least 85% of the information in the original
data. Then use the principal components as predictor variables for building models
for predicting the net promoter score (NPS).
 page 366

9 Supervised Data Mining: k-

Nearest Neighbors and
Naïve Bayes

LEARNING OBJECTIVES
After reading this chapter, you should be able to:

LO 9.1  Explain key differences among supervised

data mining techniques.
LO 9.2  Apply the k-nearest neighbors method to
classify new records.
LO 9.3  Apply the naïve Bayes method to classify
new records.

n Chapters 9 and 10, we discuss supervised data mining, also

I referred to as predictive or directed data mining. These models are
used to predict the outcome of a target variable based on several
predictor variables. As mentioned in Chapter 8, supervised data
mining includes classification and prediction models. In a
classification model, the target variable is categorical, and the
objective is to classify class memberships of new cases. Examples
include classifying a list of prospective customers as buyers and
nonbuyers or classifying a financial asset as buy, hold, and sell. In a
prediction model, the target variable is numerical, and the objective
is to predict its numerical value. Examples include predicting
consumer spending, the selling price of a house, or the time it takes
to recover from a medical procedure.
Linear and logistic regression models, discussed in Chapter 6
and 7, are important prediction and classification models,
respectively. In this chapter, we explore two widely used supervised
data mining techniques: k-nearest neighbors and naïve Bayes.
Another popular technique called decision trees will be discussed in
Chapter 10. The k-nearest neighbors method is a data-driven
technique that classifies or predicts the new case by identifying the k
nearest neighbors of the new case based on the predictor variables.
The naïve Bayes method, a classification technique, computes the
posterior probability that a new case will belong to a target class
based on its prior probability and the conditional probabilities of the
predictor variables. These data mining techniques are frequently
used to build predictive models in a wide range of business
scenarios.
©Halfpoint/Shutterstock

page 367

INTRODUCTORY CASE
24/7 Fitness Center Annual Membership
24/7 Fitness Center is a high-end full-service gym and recruits
its members through advertisements and monthly open house
events. Each open house attendee is given a tour and a one-
day pass. Potential members register for the open house event
by answering a few questions about themselves and their
exercise routine. The fitness center staff places a follow-up
phone call with the potential member and sends information to
open house attendees by mail in the hopes of signing the
potential member up for an annual membership.
Janet Williams, a manager at 24/7 Fitness Center, wants to
develop a data-driven strategy for selecting which new open
house attendees to contact. She has compiled information from
1,000 past open house attendees in the Gym_Data worksheet
of the Gym data file. The data include whether or not the
attendee purchases a club membership (Enroll equals 1 if
purchase, 0 otherwise), the age and the annual income of the
attendee, and the average number of hours that the attendee
exercises per week. Janet also collects the age, income, and
number of hours spent on weekly exercise from 23 new open
house attendees and maintains a separate worksheet called
Gym_Score in the Gym data file. Because these are new open
house attendees, there is no enrollment information on this
worksheet. A portion of the two worksheets is shown in Table
9.1.
FILE
Gym

TABLE 9.1 24/7 Fitness Data

a. The Gym_Data Worksheet

Enroll Age Income Hours

1 26 18000 14

0 43 13000 9

⋮ ⋮ ⋮ ⋮

0 48 67000 18
 b. The Gym_Score Worksheet

Age Income Hours

22 33000 5

23 65000 9

⋮ ⋮ ⋮

51 88000 6

Janet would like to use the data to accomplish the following

tasks.
1. Develop a data-driven classification model for predicting
whether or not a potential gym member will purchase a gym
membership.
2. Identify which of the 23 new open house attendees are likely
to purchase a gym membership.
A synopsis of this case is provided in Section 9.2.

page 368

9.1 INTRODUCTION TO SUPERVISED

DATA MINING
Supervised data mining techniques include classification models
where the target variable is categorical and prediction models where
the target variable is numerical. Examples of classification models
include predicting whether or not a consumer will make a purchase,
a mortgage will be approved, a patient will have a certain illness, and
an e-mail will be a spam. Similarly, examples of prediction models
include predicting the sale price of a house, the salary of a business
school graduate, the total sales of a firm, and the debt payment of a
consumer.
Linear and logistic regression models, discussed in Chapter 6
and 7, are important prediction and classification models,
respectively. In this chapter, we extend the analysis by including two
widely used data mining techniques: k-nearest neighbors and naïve
Bayes. Chapter 10 provides extensive coverage of classification and
regression trees and ensemble trees.
In supervised data mining, it is important to be able to classify or
predict the outcome of a new case; as mentioned in earlier chapters,
we use the terms case, record, and observation interchangeably in
this text. For example, in the introductory case, the manager of 24/7
Fitness Center is interested in whether or not an open house
attendee is likely to purchase a gym membership. The classification
is based on given values of predictor variables and is called scoring
a record. Scoring is straightforward in linear and logistic regression
models where the estimated equation allows us to obtain the
predicted values of a target variable for given predictor values.
Because such an estimated equation does not exist in supervised
data mining techniques, we rely on built-in algorithms in Analytic
Solver and R to score new records. Throughout Chapters 9 and 10,
we will provide two worksheets for each solved example where the
“data” worksheet allows us to find the preferred data mining
technique that we use to predict new records in the “score”
worksheet.

SCORING A RECORD
A process for classifying or predicting the value of the target
variable of a new record for given values of predictor variables.
 As discussed in Sections 7.4 and 8.3, the problem of overfitting
occurs when a predictive model is made overly complex to fit the
quirks of given sample data. It is possible for a model to perform
really well with the data set used for estimation, but then perform
miserably once a new data set is used. To address this issue, we
partition the data into training, validation, and optional test data sets.
The training data set is used to generate one or more models. The
validation data set is used to evaluate the performance of the models
and select the best model. The optional test data set is used to
assess the performance of the final model on a new data set. Hands-
on examples for data partitioning and cross-validation in Analytic
Solver and R will be provided later in this chapter.
We also discussed several performance measures to assess
predictive models in Section 8.3. For classification, a confusion
matrix is used to summarize all outcomes obtained from the
validation or test data set. These outcomes include true positive (TP)
and true negative (TN), where the observations are classified
correctly by the model. Similarly, false positive (FP) and false
negative (FN) outcomes imply that the observation is incorrectly
classified by the model. Based on these outcomes, performance
measures such as the accuracy rate, the sensitivity, and the
specificity are computed. We also use graphical representations
such as the cumulative lift chart, the decile-wise lift chart, and the
receiver operating characteristic (ROC) curve to assess model
performance. For prediction models, we considered the prediction
error, defined as the difference between the actual and the predicted
value of the numerical target variable. We discussed several
performance measures that capture the magnitude of the prediction
errors, including the root mean square error (RMSE), the mean error
(ME), the mean absolute deviation (MAD) or mean absolute error
(MAE), the mean percentage error (MPE), and the mean absolute
percentage error (MAPE). In this chapter, we will simply interpret
predictive models and assess their performance using Analytic
Solver and R.
 page 369

Comparison of Supervised Data Mining

Techniques
LO 9.1
Explain key differences among supervised data mining techniques.

Each data mining technique has its advantages and limitations. As

discussed in Chapter 8, in practice, people tend to develop several
predictive models based on different techniques and choose the best
model based on performance and other practical considerations.
Here, we summarize key differences among supervised data mining
techniques.
Linear and logistic regression models, discussed in Chapters 6
and 7, are the most ubiquitous data mining techniques with a wide
range of practical applications. Regression results describe the
relationship between predictor and target variables in a manner that
is easily interpretable. Regression techniques have been extensively
studied and were used as early as in the 1800s in a study to predict
locations of asteroids. Since then scientists and practitioners have
applied this concept in a countless number of studies, applications,
and research.
The k-nearest neighbors (KNN) method is a relatively simple data
mining tool whose underlying concepts are easy to understand. It
has been widely used for developing recommendation systems
among online companies. Consumer information such as book
purchases and movie or TV viewing habits is used for training KNN
algorithms to offer personalized recommendations for new books,
movies, or TV shows. Online companies that sell or offer these
products can develop a data collection system that efficiently
captures such information. The KNN method has been found to be
very effective as a classifier, such as producing accurate product
recommendations; however, it is not as popular for prediction
problems. In this chapter, we use KNN only for classification.
Naïve Bayes classifiers are relatively fast, even when applied to
big data. The underlying concept is also simple and relatively easy to
understand. It is often used in commercial spam e-mail filters; in fact,
it is one of the oldest and most effective spam filtering techniques.
Unlike other techniques such as keyword filtering, blacklists, and
whitelists, naïve Bayes classifiers can “learn” and recognize new
spam e-mails as well as allow individual users to train the naïve
Bayes algorithm by personally classifying e-mails as spam or
nonspam. In addition, the e-mail information can usually be captured
and reduced into a data set that consists solely of categorical
variables, which is a requirement for traditional naïve Bayes
methods.
Classification and regression trees (CART), discussed later in
Chapter 10, have many advantages. The tree representations are
very simple to understand, and the rule-based results are easy to
interpret. CART algorithms also require a relatively small amount of
data and can handle both numerical and categorical variables at the
same time. Because it is a nonparametric method, it doesn’t rely on
the data belonging to any underlying distributions, which is pertinent
to many real-world situations. As a result, it has been used in a wide
variety of applications such as fraud detection, diagnosis of medical
conditions, stock trading decisions, and prediction of student
success in college.
While decision trees are highly popular, they are sensitive to
changes in the data. Even small changes in the data may render
drastically different trees and results. Ensemble tree models, which
combine multiple single-tree models into an ensemble model, are
often credited to produce better and more stable predictive
performance than a single-tree model.
Two other techniques that are beyond the scope of this text—
discriminant analysis and neural networks—are also included in the
discussion here as reference. Discriminant analysis is a robust
classification technique that, like decision trees, provides information
about predictor importance. However, it assumes that predictor
variables are normally distributed with equal variance and, thus, is
only suitable for numerical predictor variables.
Neural networks can be used for both classification and
prediction. They mimic the neural structure of the brain using
learning, memory, and generalization. As a result, neural networks
can capture highly complex relationships in the data and are used as
a building block of sophisticated machine learning systems (often
called deep learning). A much-publicized neural network application,
Google’s self-driving car project called Waymo, claims to have used
millions of data points to train the car to recognize, adapt, and safely
respond to pedestrians and objects on public roads. Despite their
sophistication, neural networks do not provide page 370
insight into the rationale used for classifications or
predictions and lack interpretability. Many practitioners liken neural
networks to a “black box” and find it difficult to justify the results to a
nontechnical audience.
A summary of advantages and limitations of popular supervised
data mining techniques is presented in Table 9.2. The nature and
some of the prerequisites of these techniques lend themselves to
certain types of analysis and applications.

TABLE 9.2 Comparison of Supervised Techniques

Technique Advantages Limitations

Technique Advantages Limitations

Linear Widely used; computationally Only for

regression efficient; yields “best” prediction; can
(Chapter 6 mathematical function with be affected by
and Sections interpretable parameters and nonlinearity,
7.1 and 7.2) predictor importance information collinearity
between
predictor
variables, and
other violations
of the
assumptions

Logistic Computationally efficient; Only for

regression provides predictor importance classification;
(Section 7.3) information and probabilistic can be affected
interpretation by collinearity
between
predictor
variables

k-nearest Used both for classification and Provides no

neighbors prediction but is particularly information on
(Section 9.2) effective as a classifier; easy to predictor
understand; does not assume importance;
any underlying distribution in the searches the
data entire training
data set, and
therefore, is
computationally
expensive
Technique Advantages Limitations

Naïve Bayes Simple, efficient, and effective Only for

(Section 9.3) classification;
assumes
independence
among
predictors;
predictors must
be categorical
or transformed
into categories
(binning)

Classification Used both for classification and Cannot capture

and prediction; easy to understand interactions
regression and interpret; easily explained among
trees (CART) and justified to non-technical variables;
(Sections decision makers; provides sensitive to
10.2 and predictor importance changes in
10.3) information; does not assume data; single
any underlying distribution in the trees are prone
data; robust when dealing with to overfitting
outliers and missing values

Ensemble Used both for classification and Do not provide

tree models prediction; addresses the the tree
(Section overfitting issue of single tree diagram, thus
10.4) models lack
interpretability
 Technique Advantages Limitations

Discriminant Provides predictor importance Only for

analysis (Not information; accounts for classification;
covered in interactions between predictor assumes
text) variables predictor
variables are
normally
distributed with
equal variance;
uses only
numerical
predictor
variables

Neural Used both for classification and Provides little

networks prediction; flexible; captures insight into
(Not covered complex relationships relationships or
in text) results; hard to
explain or
interpret; lacks
interpretability;
computationally
burdensome

9.2 THE k-NEAREST NEIGHBORS

METHOD
LO 9.2
Apply the k-nearest neighbors method to classify new records.
In this section, we explore a supervised data mining technique called
the k-nearest neighbors method, often abbreviated KNN. The KNN
method has been used in numerous applications including consumer
credit rating assessment, loan application evaluation, fraud
detection, and medical diagnosis. It is one of the simplest, yet most
frequently used data mining techniques for classification when the
response variable is categorical and for prediction when the
response variable is numerical. Also, traditional KNN algorithms
require predictor variables to be numerical; therefore, data
transformation may need to be performed to convert categorical
predictor variables into numerical values. For a discussion on
transforming categorical data, refer to Section 2.5 of this text.

page 371

The KNN method is known to be a very effective classifier;

however, it is much less popular as a predictor. In this section, we
focus primarily on classification problems where the target variable is
binary.
KNN belongs to a category of data mining techniques called
memory-based reasoning. The premise of memory-based reasoning
is that we tend to guide our decisions using the memories of similar
situations we experienced in the past. A credit card company can
use KNN to evaluate a new credit card application based on
information from similar customers. The KNN algorithm considers
the existing credit card holders that are most similar to the new credit
card applicant as its “nearest neighbors” and uses the existing
information to decide on whether or not to approve the new
application. In the medical field, doctors and nurses can use the
KNN algorithm to compare past patients to help determine whether
or not a new patient has a similar disease. Past patients with similar
symptoms are considered “nearest neighbors” of the new patient,
and their information can be used to help diagnose new patients.
 The k-Nearest Neighbors (KNN) Method
A memory-based reasoning technique that uses observations
from the past that are most similar to the new observation to
classify or predict the value of the target variable.

To better understand how the KNN algorithm works, consider a

mortgage company trying to determine whether or not a new
applicant is likely to default on his or her home loan. The new
applicant is 36 years old and has an annual income of $64,000. The
mortgage company extracts information from 10 previous customers
with similar income (in $1,000s) and age characteristics, as well as
whether or not the customer defaulted on the loan. Table 9.3 shows
the data.

TABLE 9.3 Characteristics of Home Loan Customers

Customer ID Income Age Default

001 65 35 Yes

002 59 34 No

003 71 40 No

004 67 32 No

005 60 31 Yes

006 59 37 No

007 72 42 No
 Customer ID Income Age Default

008 63 34 No

009 64 38 No

010 71 34 Yes

The KNN algorithm works by identifying a number of existing

observations (i.e., past loan applicants) that are most similar to the
new observation (i.e., the new applicant) in order to classify the new
observation. Thus, central components of the KNN algorithm are (1)
the chosen similarity measure and (2) the number of nearest
neighbors (k) to consider. In practice, the KNN algorithm often uses
the Euclidean distance to measure the similarity between
observations. In the mortgage approval example, given the
differences in scale of the Age and Income variables, we would
standardize these variables before calculating the Euclidean
distance between observations. Similarity measures, the Euclidean
distance, and the standardization of variables are discussed in
Section 8.2.
To demonstrate this process, Figure 9.1 shows a scatterplot of
Age against Income for the 10 past customers, where customers
who defaulted on the loan are labeled with a red marker and
customers who did not default are labeled with a blue marker. We
add the new applicant to the scatterplot as a green triangle and
identify the three “nearest neighbors” in the green circle. In other
words, we choose the value of k = 3 in the KNN algorithm. The
nearest neighbors can also be determined page 372
mathematically using the Euclidean distance
between the new applicant and other applicants in the data set. In
this scenario, only one of the three nearest neighbors defaulted on
his or her home loan (one red marker among the three nearest
neighbors in the green circle). As a result, we estimate that the
probability that the new applicant will default is 1/3 = 0.33. The
mortgage company can then compare the probability to a
predetermined cutoff point in order to decide whether to approve or
reject the new applicant. If the company uses 0.5 as the default
probability cutoff point, the new loan applicant will be classified as
being unlikely to default and will receive an approval on the loan
application. However, if the cutoff is 0.25, the new loan applicant will
be classified as being likely to default; we discussed cutoff values
and their influence on the classification in Section 8.3.

FIGURE 9.1 Age and Income of home loan applicants

As you can see from this example, the choice of k will result in
different classification results. For example, if k = 1, then we use the
nearest neighbor (the red marker) to determine the class
membership of the new applicant. Now, the probability that the new
applicant will default becomes 1/1 = 1.
Because the KNN method is a supervised data mining technique,
data partitioning is used to determine its performance and to
optimize model complexity. The nearest neighbor observations are
identified in the training data set. Most practitioners try to minimize
the error rate in the KNN method by experimenting with a range of k
values and their corresponding performance measures from the
validation data set in order to determine the optimal k value. In some
software applications such as Analytic Solver, the test data set is
used to assess the performance of the final predictive model and the
optimal k value. The KNN algorithm can also classify new
observations based on the existing observations that are most
similar to the new observation. As mentioned earlier, this process is
called scoring a new record based on the values of the predictor
variables. As presented in Figure 9.1, the KNN algorithm first
searches the entire training data set for the most similar
observations, and then provides a score for a new observation.
Unlike a traditional analysis with linear and logistic regression
models, the KNN algorithm does not assume any underlying
distribution in the data for making predictions. This nonparametric
approach makes the KNN method very useful in a practical situation
because real-world data do not always follow theoretical
assumptions. Despite its simplicity, the KNN method often produces
highly accurate results. Because the KNN algorithm searches the
entire training data set for close neighbors, the process can be quite
time-consuming and computationally intensive, especially with a
large amount of data. Fortunately, several computer software
programs include the KNN algorithm. Example 9.1 illustrates how to
perform KNN classification using Analytic Solver and R.
FILE
Gym

page 373

EXAMPLE 9.1
 Recall from the introductory case that Janet Williams, the
membership manager at 24/7 Fitness Center, has collected
data from 1,000 past open house attendees (Gym_Data
worksheet) and 23 new open house attendees (Gym_Score
worksheet). The data include the age and the annual income of
the attendee, the average number of hours that the attendee
exercises per week, and for the Gym_Data worksheet, whether
or not the attendee purchases a membership. Janet is
interested in a model that will help her identify attendees who
are likely to purchase a membership. Perform KNN analysis to
estimate a classification model and evaluate its performance.
Report the relevant performance measures and score new
records.
SOLUTION:
Using Analytic Solver
As discussed in Chapters 7 and 8, to develop and evaluate a
classification model, we generally perform cross-validation
using either the holdout method or the k-fold method. Analytic
Solver provides a procedure for the holdout method and allows
for partitioning a data set into training, validation, and test data
sets. In this chapter, we partition our data set in Analytic Solver
as follows: 50% for training, 30% for validation, and 20% for
test. Here, an independent assessment of the predictive
performance of the KNN model is conducted with the test data
set that is not used in the model development.
a. Open the Gym data file and go to the Gym_Data worksheet.
b. Choose Data Mining > Partition > Standard Partition.
page 374
c. See Figure 9.2. Click on the ellipsis button next to
Data Range and highlight cells $A$1:$D$1001. Make sure that
the box First Row Contains Headers is checked. The
Variables in Input Data box will populate. Select and move all
the variables to the Selected Variables box. Choose Specify
 percentages with 50% for Training Set, 30% for Validation Set,
and 20% for Test Set. Accept other defaults. Click OK.

FIGURE 9.2 Analytic Solver’s standard data partition

Source: Microsoft Excel

d. Analytic Solver will create a new worksheet called

STDPartition with the partitioned data. In the Partition
Summary table of this worksheet, you should see the following
number of observations in each partition: 500 for Training, 300
for Validation, and 200 for Testing.
e. Choose Data Mining > Classify > k-Nearest Neighbors. See
Figure 9.3. Make sure that the STDPartition worksheet is
active. In the Data tab, select and move Age, Income, and
Hours to the Selected Variables box. Select Enroll as the
Output Variable. Accept the other defaults and click Next.

FIGURE 9.3 The KNN classification

Source: Microsoft Excel

f. See Figure 9.4. In the Parameters tab, click Rescale Data

and check the Rescale Data checkbox. Select Standardization
and click Done. In the # Neighbors (K) box, input 10, and
under Nearest Neighbors: Search, select Search 1..K. Analytic
Solver will repeat the KNN algorithm with the values of k from
1 to 10, but only the k value that yields the best result will be
used in subsequent steps. Click Next.
 FIGURE 9.4 The KNN parameters options

page 375
g. See Figure 9.5. In the Scoring tab, check the
Summary Report boxes for Score Training Data, Score
Validation Data, and Score Test Data. Also check the Lift
Charts box for Score Test Data. Check the In Worksheet
checkbox under the Score New Data section. Under the New
Data (WS) tab, select Gym_Score in the Worksheet box and
enter $A$1:$C$24 in the Data Range. These 23 records of
potential members will be scored by the KNN classification
algorithm. Make sure the First Row Contains Headers box is
checked and click the Match By Name button. Click Finish.

FIGURE 9.5 Analytic Solver’s scoring options

Source: Microsoft Excel

Analytic Solver will create seven new worksheets. The

KNNC_Output worksheet displays the misclassification
percentage for each value of k using the validation data set.
Recall from Chapter 8 that the misclassification rate is the
overall proportion of cases that are classified incorrectly (false
positives and false negatives combined). Figure 9.6 shows that
the lowest misclassification rate is reached at the k value of 3,
which is 8%. Therefore, k = 3 is used in the subsequent
calculations.

FIGURE 9.6 Misclassification rates

Source: Microsoft Excel

page 376

The KNNC_TrainingScore, KNNC_ValidationScore, and

KNNC_TestScore worksheets display the classification results
for the training, validation, and test data sets, with k = 3. Figure
9.7 summarizes the main results from the test data set. Here,
we focus on the most notable performance measures for this
application. The accuracy rate indicates that 89.5% of the
observations are correctly classified; thus, 10.5% are
incorrectly classified (see the error rate). Normally, the
performance measures from the training data will be better
than the results obtained from the validation or test data.
However, if the performance values from the training data are
considerably better than the values from the validation and the
test data, overfitting might be an issue. In the 24/7 Fitness
Center case, the KNN algorithm can accurately predict the
outcome for 94.6% of the observations in the training data and
92% of the observations in the validation data (performance
measures not shown here), while the accuracy rate is slightly
lower for the test data, (89.5%), which indicates that some
overfitting may exist. However, the differences may not be
large enough to suggest any severe overfitting problems. The
sensitivity of 0.90 is the proportion of buyers that the model is
able to identify correctly, while the specificity of 0.89 is the
proportion of the nonbuyers that the model is able to identify
correctly.

FIGURE 9.7 Classification summary tables

Source: Microsoft Excel

To answer the question of whether the resulting KNN model

has good predictive performance or not, we can compare the
accuracy rate of the model to that of the naïve rule, where all
cases are classified into the predominant class. For example,
because there are 80 target class (Enrolled in the gym) cases
among the 200 test data cases, the naïve rule would assume
that each case has the probability of 80/200 or 0.4 of belonging
to the target class. Using the cutoff rate of 0.5, the naïve rule
will classify all cases into the non-target class (Do not enroll in
the gym) and produce an accuracy rate of (200 – 80) ÷ 200 =
0.60, sensitivity of 0, and specificity of 1. Therefore, we
conclude that the KNN model has better predictive
performance than the naïve rule does, especially if the goal is
to identify target class cases.
As illustrated in Example 8.6 in Chapter 8, the choice of the
cutoff value can influence the confusion matrix and the
resulting performance measures. By default, Analytic Solver
uses 0.5 as the cutoff value for binary choice models. The
analyst may choose to increase or decrease the cutoff value to
classify fewer or more observations into the target class due to
asymmetric misclassification costs or uneven class
distributions.
In this example, the opportunity cost of missing a potential
gym member who would have purchased a gym membership is
likely to be higher than the cost of reaching out to a potential
gym member who ends up not purchasing a gym membership.
Furthermore, among the 1,000 past open house attendees in
the data set, 403 purchased gym memberships page 377
suggesting that the probability of an open
house attendee purchasing the gym membership is 403 ÷
1,000 = 0.403 instead of the default probability of 0.5 for a
binary choice situation with equal class distribution.
In this case, the manager of the fitness center may reduce
the cutoff value to a lower value than 0.5 in order to classify
more cases into the target class in order to achieve a higher
sensitivity value. To assess predictive performance using a
different cutoff value in Analytic Solver, input the desired cutoff
value, for example, 0.25, in the Success Probability Cutoff: box
in the Data tab in step e. You may want to experiment with a
number of different cutoff values to examine their impact on the
performance measures.
Figure 9.8 summarizes the output found on the
KNNC_TestLiftChart worksheet. Because Analytic Solver uses
the training and validation data to determine the best k value, it
uses the test data and the charts displayed on the
KNNC_TestLiftChart worksheet to provide an independent
assessment of the KNN classification.

FIGURE 9.8 Performance charts for KNN

 In the cumulative lift chart, the blue line represents the
results from the KNN model, and the diagonal red line
represents a model that randomly classifies potential gym
members (baseline model). In this case, the blue line from the
KNN model lies well above the red diagonal line, indicating that
the KNN model performs considerably better in page 378
terms of predicting whether or not potential gym members will
purchase a gym membership as compared to the baseline
model.
The decile-wise lift chart shows that the first 10% of
potential members selected by the KNN model include about
2.375 times as many individuals who would actually purchase a
gym membership than a randomly selected 10% of potential
members. (The value 2.375 is obtained from the decile table,
which is not shown.) Moreover, the decile-wise lift chart shows
that almost all of the Class 1 cases (that is, purchased a gym
membership) have been identified within the first five deciles of
cases with the highest predicted probability of belonging to
Class 1. Similarly, the blue line from the KNN model in the
ROC curve lies well above the red diagonal line. The area
under the ROC curve, or AUC, is very high (0.9273), indicating
that the KNN model performs extremely well in terms of
sensitivity and specificity across all possible cutoff values.
On the KNNC_NewScore worksheet, the 23 records of
potential members are scored. The Prediction: Enroll column
displays the classification result for each record. The KNN
model predicts that six of the 23 attendees will purchase a gym
membership. We combine the information from the
KNNC_NewScore worksheet with the information that Janet
collected for the 23 attendees at the most recent open house.
Table 9.4 shows a portion of the scoring results. From this
grouping, Janet and her staff should target the second
attendee—a 23-year-old, with an annual income of $65,000,
who exercises 9 hours per week—as well as the last attendee
—a 51-year-old, with an annual income of $88,000, who
exercises 6 hours per week. Most of the open house attendees
that the KNN model classified as more likely to purchase a gym
membership are individuals older than 50 years who either
have a relatively high annual income or spend at least 9 hours
on weekly exercise.
 TABLE 9.4 Analytic Solver’s scoring results for KNN

Using R
As mentioned earlier, in order to evaluate the accuracy of the
KNN algorithm as well as determine the appropriate value of k,
we generally perform cross-validation using either the holdout
method or the k-fold method, both discussed in Chapter 7.
Most data mining packages in R allow for the k-fold cross-
validation method, which, when compared with the holdout
method, is less sensitive to how the data are partitioned.
Therefore, in R we only need to divide the data set into two
partitions, training and validation, and implement the k-fold
cross-validation. Depending on the amount of data available,
when cross-validation is performed, most practitioners partition
data into either 60% training/40% validation or 70%
training/30% validation. We divide the data in the Gym_Data
worksheet into two partitions, 60% for training and 40% for
validation, and then implement a 10-fold cross-validation
technique.
a. Import the data from the Gym_Data worksheet of the Gym
data file into a data frame (table) and label it myData.
The following instructions and results are based on R version
3.5.3. To replicate the results with newer versions of R, enter:
> suppressWarnings(RNGversion(“3.5.3”))
b. For KNN estimation and the resulting performance measures
and diagrams, install and load the caret, gains, and pROC
packages. Enter:
> install.packages(c(“caret”, “gains”, “pROC”))
 > library(caret)
> library(gains)
> library(pROC)
page 379
On some computers, you might also need to
install other packages that support the caret package using
the command > install.packages(”caret”, dependencies =
c(”Depends”, ”Suggests”)). Also, if prompted by R Studio,
install and load the car package.
c. We use the scale function to standardize the Age, Income,
and Hours variables; store the standardized values in a new
data frame called myData1; and append the original Enroll
variable back to myData1. We use the as.factor function to
convert the target variable (Enroll) into a categorical data type.
To simplify the R code, we use the colnames function to
rename myData1$myData.Enroll (in column 4) to
myData1$Enroll. Enter:
> myData1 <- scale(myData[2:4])
> myData1 <- data.frame(myData1, myData$Enroll)
> colnames(myData1)[4] <- ‘Enroll’
> myData1$Enroll <- as.factor(myData1$Enroll)
d. To partition the data into 60% training and 40% validation sets,
we use the createDataPartition function and specify Enroll as
the target variable. To ensure consistency, we use the
set.seed function to set the random seed to 1. Enter:
> set.seed(1)
> myIndex <- createDataPartition(myData1$Enroll, p=0.6, list
= FALSE)
> trainSet <- myData1[myIndex,]
> validationSet <- myData1[-myIndex,]
The data set is partitioned into 60% for training and 40% for
validation with the option p = 0.6. The 60% training set is
 assigned to an object called trainSet and the other 40% is
assigned to the validationSet object. In order to maintain the
same ratio of target and nontarget class cases for both the
training and validation data sets, R partitions 601 cases into
the training data set and 399 cases into the validation set.
e. We use the trainControl function to implement a 10-fold
cross-validation by setting the option method equal to “cv” and
the option number equal to 10. Enter:
> myCtrl <- trainControl(method = ”cv”, number = 10)
f. We use the expand.grid function to specify possible k values
from 1 to 10 and store the results in an object called myGrid.
The optimal k value is determined based on accuracy. The
possible range of k values may vary; you may experiment with
a different range by changing the numbers in the statement.
Enter:
> myGrid <- expand.grid(.k=c(1:10))
g. To implement the KNN method with the training data set with
option values specified in steps e and f, we use the train
function and store the results in an object called KNN_fit. To
ensure consistency of the cross-validation results, we again
use the set.seed function to fix a random seed. Enter:
> set.seed(1)
> KNN_fit <- train(Enroll ~., data = trainSet, method = ”knn”,
trControl=myCtrl, tuneGrid = myGrid)
> KNN_fit
Note that the Enroll variable is specified as a target variable
and “knn” is specified as the classification method. The KNN
results are shown in Figure 9.9. The value k = 6 page 380
yields the highest accuracy rate (0.9102695) and
will be used in subsequent steps.

FIGURE 9.9 KNN training results

h. We apply the KNN model, which uses k = 6, on the validation
data set using the predict function, and then we use the
confusionMatrix function to create a confusion matrix. Enter:
> KNN_Class <- predict(KNN_fit, newdata = validationSet)
> confusionMatrix(KNN_Class,validationSet$Enroll, positive =
‘1’)
Note that we identify Class 1 (Enrolled in the gym) as the
positive or success class. In other situations, the positive class
may be different. Also, the Enroll variable in the validation set
is identified as the target variable. Relevant output is shown in
Figure 9.10. The confusion matrix results from the validation
set indicate a high classification accuracy (0.9148), sensitivity
(0.8944), and specificity (0.9286). Refer to Section 8.3 for
additional details of these performance measures.

FIGURE 9.10 Confusion matrix for KNN

 To evaluate the predictive performance of the KNN model,
we can compare its performance measures to those of the
naïve rule, where all cases are classified into the predominant
class. Because there are 161 target class (Enrolled in the gym)
cases among the 399 validation data cases, page 381
the naïve rule would classify all cases into the
non-target class (Do not enroll in the gym) and produce an
accuracy rate of (399 − 161) ÷ 399 = 0.5965, sensitivity of 0,
and specificity of 1. Therefore, the KNN model shows better
predictive performance than the naïve rule does, especially if
the goal is to identify target class cases.
As illustrated in Example 8.6 in Chapter 8, the choice of the
cutoff value can influence the confusion matrix and the
resulting performance measures. By default, R uses 0.5 as the
cutoff value for binary choice models. The analyst may choose
 to increase or decrease the cutoff value to classify fewer or
more observations into the target class due to asymmetric
misclassification costs or uneven class distributions. In this
example, the opportunity cost of missing a potential gym
member who would have purchased a gym membership is
likely to be higher than the cost of reaching out to a potential
gym member who ends up not purchasing a gym membership.
Furthermore, among the 1,000 past open house attendees in
the data set, 403 purchased gym memberships suggesting that
the probability of an open house attendee purchasing the gym
membership is 403 ÷ 1,000 = 0.403 instead of the default
probability of 0.5 for a binary choice situation with equal class
distribution. In this case, the manager of the fitness center may
reduce the cutoff value to a lower value than 0.5 in order to
classify more cases into the target class in order to achieve a
higher sensitivity value. This is illustrated in step i.
i. To evaluate the predictive performance of the KNN model
using a different cutoff value in R, we first compute the
probability of each case belonging to the target class instead
of its class membership. In the predict function, we set the
type option equal to ‘prob’ to predict the probability values.
Enter:
> KNN_Class_prob <- predict(KNN_fit, newdata =
validationSet, type =‘prob’)
> KNN_Class_prob
Figure 9.11 shows the probabilities associated with the first
few cases. As you can see, the first column lists the
probabilities of the cases belonging to Class 0 (Not enrolled in
the gym), while the second column lists the probabilities of the
cases belonging to Class 1 (Enrolled in the gym). To
determine the class memberships of cases using a cutoff
value other than the default probability of 0.5 (for example,
0.25, in order to accurately classify more Class 1 cases), we
 compare the values in the second column to the new cutoff
value.

FIGURE 9.11 Predicted probabilities

j. To construct a confusion matrix using the new cutoff value of

0.25, we use the ifelse function to determine the class
memberships. We use the as.factor function to convert the
class membership to factor, which is the same data type as
the target variable, Enroll. Enter:
> confusionMatrix(as.factor(ifelse(KNN_Class_prob[,2]>0.25,
‘1’, ‘0’)), validationSet$Enroll, positive = ‘1’)
page 382
The resulting confusion matrix (not shown here)
provides the performance measures of the KNN model using
the cutoff value of 0.25. Please verify that in the new
confusion matrix, accuracy, sensitivity, and specificity are
0.8797, 0.9627, and 0.8235, respectively. Although the overall
accuracy is lower than before, the new cutoff value allows the
manager of the fitness club to identify more target class cases
correctly as signified by the higher sensitivity value.
k. To create a cumulative gain table and a cumulative lift chart,
we need to convert the Enroll variable back as a numerical
data type as required by the gains package. Enter:
> validationSet$Enroll <-
as.numeric(as.character(validationSet$Enroll))
 l. We generate the cumulative gains lift table using the gains
function. The gains function requires two arguments: actual
class memberships and predicted target class probabilities.
Enter:
> gains_table <- gains(validationSet$Enroll,
KNN_Class_prob[,2])> gains_table
Figure 9.12 shows the relevant results. By default, the
cumulative gains table divides the cases into 10 groups based
on their probabilities of belonging to the target class. The last
column of Figure 9.12 shows the average probability for each
group. (Note: These values are rounded to two digits after the
decimal point.) In R, in situations where there are less than ten
possible probability values, the cumulative gains table and
decile-wise lift chart will show less than 10 groups.

FIGURE 9.12 Cumulative gains table

m. We use the plot function to create the cumulative lift chart as

shown in Figure 9.13. The first command plots the lift curve for
the KNN model, and the second command draws a dashed
diagonal line to indicate the baseline model. Enter:
page 383
FIGURE 9.13 R’s cumulative lift chart
 > plot(c(0,
gains_table$cume.pct.of.total*sum(validationSet$Enroll)) ~
c(0, gains_table$cume.obs),xlab = ”# of cases”,
ylab=”Cumulative”, main=”Cumulative Lift Chart”, type = ”l”)
> lines(c(0, sum(validationSet$Enroll)) ~ c(0,
dim(validationSet)[1]), col=”red”, lty=2)
n. To plot the decile-wise lift chart, we use the barplot function to
plot a bar chart. Enter:
> barplot(gains_table$mean.resp/mean(validationSet$Enroll),
names.arg=gains_table$depth, xlab=”Percentile”, ylab=”Lift”,
ylim = c(0,3), main = ”Decile-Wise Lift Chart”)
The decile-wise lift chart is displayed in Figure 9.14. The
decile-wise lift chart shows that if the 24/7 Fitness Center
contacts the top 15% of the potential gym members with the
highest predicted probability of enrolling, the gym would be
able to capture 2.31 times as many potential members who
actually purchase the gym membership, compared to a
scenario where the 24/7 Fitness Center randomly selects 15%
of the potential members to contact. The KNN model shows
superior predictive power compared to the baseline model.
 FIGURE 9.14 R’s decile-wise lift chart

page 384
o. We use the roc function, which produces a roc
object that can be used to plot the ROC curve. We then use
the plot.roc function to plot the ROC curve and the auc
function to retrieve the AUC value. Enter:
> roc_object<- roc(validationSet$Enroll, KNN_Class_prob[,2])
> plot.roc(roc_object)
> auc(roc_object)
The area under the ROC curve, or AUC, is very high (0.9532),
indicating that the KNN model performs extremely well in
predicting the gym enrollment among potential members.
Figure 9.15 displays the ROC curve. The KNN model
performs better than the baseline model (shown as the green
diagonal line) in terms of sensitivity and specificity across all
cutoff values.

FIGURE 9.15 R’s ROC curve, with AUC = 0.9532

p. To score the records of the potential gym members, we import
data from the Gym_Score worksheet of the Gym data file into
a data frame (table) and name it myScoreData. Because KNN
identifies the nearest neighbors among the historical cases,
we use the means and standard deviations calculated from
the historical cases to standardize new data. Here we use the
preProcess function to compute the means and standard
deviations of the predictor variables in myData. The center
method calculates the mean and subtracts it from each value,
and the scale method computes the standard deviation and
divides each value by the standard deviation. The predict
function uses the mean and standard deviation values from
the pre-processing statement to standardize myScoreData
and store the z-scores in a new data frame named
myScoreData1. Finally, we use the predict function to score
the new records. Enter:
> PreProcessing <- preProcess(myData[ , 2:4], method =
c(“center”, “scale”))
> myScoreData1 <- predict(PreProcessing, myScoreData)
 > KNN_Score <- predict(KNN_fit, newdata = myScoreData1)
q. Once the scoring is finished, the KNN_Score object will
contain classification results (whether or not the potential gym
members are likely to enroll: yes = 1 and no = 0). We use the
data.frame function to append the classification results back
to the original data set and the View function to display the
updated data set. Enter:
> myScoreData <- data.frame(myScoreData, KNN_Score)
> View(myScoreData)
A portion of the records is shown in Table 9.5. Because the
built-in algorithms are not the same, Analytic Solver and R
yield different classification outcomes. The R results suggest
that the 24/7 Fitness Center should focus on contacting
potential gym members who are older than 40 years with a
high annual income.
page 385
TABLE 9.5 R’s Scoring Results for KNN

As mentioned earlier, the KNN method is most often used for

classification where the target variable is categorical. However, it can
also be used for predicting a target variable with a numerical value.
To predict the value of a new record, the KNN method uses a similar
approach with the Euclidean distance to identify the nearest
neighbors, but instead uses the average of its k nearest neighbors to
estimate the value of the new record’s target variable. As discussed
in Chapter 8, instead of misclassification or accuracy rates, to
evaluate the performance of the KNN prediction, we use prediction
metrics such as the root mean square error (RMSE) and the mean
absolute deviation (MAD). Functions for KNN prediction can be
found in both Analytic Solver and R. However, because practitioners
use the KNN method mostly for classification purposes, we focus
solely on using the KNN method for classification in this text.

SYNOPSIS OF INTRODUCTORY CASE

Gyms and exercise facilities usually have a high turnover rate among their
members. Like other gyms, 24/7 Fitness Center relies on recruiting new members
on a regular basis in order to sustain its business and financial well-being.
Completely familiar with data analytics techniques, Janet Williams, a manager at
24/7 Fitness Center, uses the KNN method to analyze data from the gym’s past
open house. She wants to gain a better insight into which attendees are likely to
purchase a gym membership after attending this event.
Overall, Janet finds that the KNN analysis provides reasonably high accuracy in
predicting whether or not potential gym members will purchase a membership. The
accuracy, sensitivity, and specificity rates from the test data set are well above 80%.
More importantly, the KNN analysis identifies individual open house attendees who
are likely to purchase a gym membership. For example, the analysis results indicate
that open house attendees who are 50 years or older with a relatively high annual
income and those in the same age group who spend at least nine hours on weekly
exercise are more likely to enroll after attending the open house. With these types of
actionable insights, Janet decides to train her staff to regularly analyze the monthly
open house data in order to help 24/7 Fitness Center grow its membership base.

EXERCISES 9.2
Note: These exercises can be solved using Analytic Solver and/or R. The answers,
however, will depend on the software package used. All answers in R are based on
version 3.5.3. To replicate the results with newer versions of R, execute the following
line of code at the beginning of the R session:
suppressWarnings(RNGversion(“3.5.3”)). For Analytic Solver, partition data sets into
50% training, 30% validation, and 20% test and use 12345 as the default random
seed. For R, partition data sets into 60% training and 40% validation and implement
the 10-fold cross-validation. Use the statement set.seed(1) to specify the random
seed for data partitioning and cross-validation. When searching for the optimal value
of k, search within possible k values from 1 to 10 for both Analytic Solver and R.
Some data files have two worksheets (for example, Exercise_9.3_Data and
Exercise_9.3_Score worksheets) for model development and scoring new records.

Mechanics
1. FILE Exercise_9.1. The accompanying file contains 60 observations with the
binary target variable y along with the predictor variables x1 and x2.
a. Perform KNN analysis. What is the optimal value of k?
b. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).

page 386
2. FILE Exercise_9.2. The accompanying file contains 111
observations with the binary target variable y along with the predictor variables x1,
x2, x3, and x4.
a. Perform KNN analysis on the data set. What is the optimal value of k?
b. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).
c. Comment on the performance of the KNN classification model.
3. FILE Exercise_9.3. The accompanying file contains 80 observations with the
binary target variable y along with the predictor variables x1, x2, x3, and x4.
a. Perform KNN analysis on the data set. What is the optimal value of k?
b. What is the misclassification rate for the optimal k?
c. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).
d. What is the area under the ROC curve (or the AUC value)?
e. Score the new observations in the Exercise_9.3_Score worksheet. What is the
predicted response value of the first observation?
4. FILE Exercise_9.4. The accompanying file contains 200 observations with the
binary target variable y along with the predictor variables x1, x2, x3, x4, and x5.
a. Perform KNN analysis on the data set. What is the optimal value of k?
b. What is the misclassification rate for the optimal k?
c. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).
d. Change the cutoff value to 0.2. Report the accuracy, specificity, sensitivity, and
precision rates for the test data set (for Analytic Solver) or the validation data
set (for R).
e. Comment on the performance of the KNN classification model.
5. FILE Exercise_9.5. The accompanying file contains 1,000 observations with the
binary target variable y along with the predictor variables x1, x2, and x3.
 a. Perform KNN analysis on the data set. What is the optimal value of k?
b. What is the misclassification rate for the optimal k?
c. Report the accuracy, specificity, sensitivity, and precision rates and the AUC
value for the test data set (for Analytic Solver) or validation data set (for R).
d. Comment on the performance of the KNN classification model.
6. FILE Exercise_9.6. The accompanying file contains 2,000 observations with the
binary target variable y along with the predictor variables x1, x2, and x3.
a. Perform KNN analysis on the data set. What is the optimal value of k?
b. Report the accuracy, specificity, sensitivity, and precision rates and the AUC
value for the test data set (for Analytic Solver) or validation data set (for R).
c. Compare the specificity and sensitivity rates and comment on the performance
of the KNN classification model.
d. Display the cumulative lift chart, decile-wise lift chart, and ROC curve.
7. FILE Exercise_9.7. The accompanying file contains 400 observations with the
binary response variable y along with the predictor variables x1, x2, x3, and x4.
a. Perform KNN analysis on the data set. What is the optimal value of k?
b. Report the accuracy, specificity, sensitivity, and precision rates and the AUC
value for the test data set (for Analytic Solver) or validation data set (for R).
c. What is the lift value of the leftmost bar shown in the decile-wise lift chart?
d. Score the new observations in the Exercise_9.7_Score worksheet. What are
the predicted response values of the first two new observations?

Applications
8. FILE Admit. Universities often rely on a high school student’s grade point
average (GPA) and scores on the SAT or ACT for the college admission
decisions. Consider the data for 120 applicants on college admission (Admit
equals 1 if admitted, 0 otherwise) along with the student’s GPA and SAT scores. A
portion of the Admit_Data worksheet is shown in the accompanying table.

Admit GPA SAT

1 3.10 1550

0 2.70 1360

⋮ ⋮ ⋮

1 4.40 1320

a. Perform KNN analysis to estimate a classification model for college admission

decisions in the Admit_Data worksheet and score new applications in the
 Admit_Score worksheet. What is the optimal value of k?
b. Report and interpret the overall accuracy, specificity, sensitivity, and precision
rates for the test data set (for Analytic Solver) or validation data set (for R).
c. What is the area under the ROC curve (or the AUC value)? Comment on the
performance of the KNN classification model.
d. What is the predicted admission outcome for the first new application?

page 387
9. FILE CFA. The Chartered Financial Analyst (CFA) designation
is the de facto professional certification for the financial industry. Employers
encourage their prospective employees to complete the CFA exam. Daniella
Campos, an HR manager at SolidRock Investment, is reviewing 10 job
applications. Given the low pass rate for CFA Level 1, Daniella wants to know
whether or not the 10 prospective employees will be able to pass the CFA Level 1
exam. Historically, the pass rate is higher for those with work experience and a
good GPA in college. With this insight, she compiles the information on 63 current
employees who took the CFA Level I exam last year, including the employee’s
success on the exam (1 for pass, 0 for fail), the employee’s college GPA, and
years of work experience. A portion of the CFA_Data worksheet is shown in the
accompanying table.

Pass GPA Experience

1 3.20 11

1 3.75 15

⋮ ⋮ ⋮

1 3.21 6

a. Perform KNN analysis to estimate a classification model for the CFA Level 1
exam using the CFA_Data worksheet and score the 10 job applicants in the
CFA_Score worksheet. What are the misclassification rates for k = 3, 4, and 5?
b. What is the optimal value of k? Report the overall accuracy, specificity,
sensitivity, and precision rates for the test data set (for Analytic Solver) or
validation data set (for R).
c. What is the predicted CFA Level 1 outcome for the first job applicant?
10. FILE SocialMedia. A social media marketing company is conducting consumer
research to see how the income level and age might correspond to whether or not
consumers respond positively to a social media campaign. Aliyah Turner, a new
college intern, is assigned to collect data from the past marketing campaigns. She
compiled data on 284 consumers who participated in the marketing campaigns in
the past, including income (in $1,000s), age, and whether or not each individual
 responded to the campaign (1 if yes, 0 otherwise). A portion of the
SocialMedia_Data worksheet is shown in the accompanying table.

Social Media Income Age

0 103.3 67

0  61.4 34

⋮ ⋮ ⋮

1  91.3 40

a. Perform KNN analysis to estimate a classification model for the social media
campaign using the SocialMedia_Data worksheet and score new consumer
records in the SocialMedia_Score worksheet. What is the optimal value of k?
b. Report and interpret the overall accuracy, specificity, sensitivity, and precision
rates for the test data set (for Analytic Solver) or validation data set (for R).
c. What is the area under the ROC curve (or the AUC value)? Comment on the
performance of the KNN classification model.
d. What is the predicted outcome for the first new consumer record?
e. Change the cutoff value to 0.3. Report the accuracy, specificity, sensitivity, and
precision rates for the test data set (for Analytic Solver) or validation data set
(for R).
11. FILE Spam. Peter Derby works as a cyber security analyst at a private equity
firm. His colleagues at the firm have been inundated by a large number of spam
e-mails. Peter has been asked to implement a spam detection system on the
company’s e-mail server. He reviewed a sample of 500 spam and legitimate e-
mails with relevant variables: spam (1 if spam, 0 otherwise), the number of
recipients, the number of hyperlinks, and the number of characters in the
message. A portion of the Spam_Data worksheet is shown in the accompanying
table.

Spam Recipients Hyperlinks Characters

0 19 1 47

0 15 1 58

⋮ ⋮ ⋮ ⋮

1 13 2 32
 a. Perform KNN analysis to estimate a classification model for spam detection
using the Spam_Data worksheet and score new e-mails in the Spam_Score
worksheet. What is the optimal value of k?
b. Report the overall accuracy, specificity, sensitivity, and precision rates for the
test data set (for Analytic Solver) or validation data set (for R).
c. What is the area under the ROC curve (or AUC value)?
d. What is the predicted outcome for the first new e-mail?
12. FILE Security. Law enforcement agencies monitor social media sites on a
regular basis, as a way to identify and assess potential crimes and terrorism
activities. For example, certain keywords on Facebook pages are tracked, and the
data are compiled into a data mining model to determine whether or not the
Facebook page is a potential threat. Officer Matthew Osorio is assigned to explore
data mining techniques that can be used for this purpose. He starts by
experimenting with KNN algorithms to monitor and assess social media sites with
war-related terms such as “warfare,” “bomb,” and “attack” as well as suspicious
keywords such as “extremist,” “radical,” and “conspiracy.” He collects a data set
with 300 observations, a portion of which is shown in the accompanying table.
Each record in the data set includes the following variables: Threat (1 if yes, 0
otherwise), the number of suspicious words (WarTerms and Keywords), and the
number of hyperlinks to or mentioning of suspicious sites.

page 388
Threat WarTerms Keywords Links

0 6 5 5

0 3 5 8

⋮ ⋮ ⋮ ⋮

1 4 4 2

a. Perform KNN analysis on the data set. What is the optimal value of k?
b. What is the misclassification rate for the optimal k?
c. Report and interpret the accuracy, specificity, sensitivity, and precision rates for
the test data set (for Analytic Solver) or validation data set (for R).
d. Generate the cumulative lift chart. Does the entire lift curve lie above the
baseline?
e. Generate the decile-wise lift chart. What is the lift value of the leftmost bar?
f. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
g. Comment on the performance of the KNN classification model.
13. FILE HR. Daniel Lara, a human resources manager at a large tech consulting
firm, has been reading about using analytics to predict the success of new
employees. With the fast-changing nature of the tech industry, some employees
have had difficulties staying current in their field and have missed the opportunity
to be promoted into a management position. Daniel is particularly interested in
whether or not a new employee is likely to be promoted into a management role
after 10 years with the company. He gathers information on 300 current
employees who have worked for the firm for at least 10 years. The information
was based on the job application that the employees provided when they
originally applied for a job at the firm. For each employee, the following variables
are listed: Promoted (1 if promoted within 10 years, 0 otherwise), GPA (college
GPA at graduation), Sports (number of athletic activities during college), and
Leadership (number of leadership roles in student organizations). A portion of the
HR_Data worksheet is shown in the accompanying table.

Promoted GPA Sports Leadership

0 3.28 0 2

1 3.93 6 3

⋮ ⋮ ⋮ ⋮

0 3.54 5 0

a. Use the HR_Data worksheet to help Daniel perform KNN analysis to determine
whether or not an employee is likely to be promoted into a management role
after 10 years with the company. Score the records of the 10 new employees in
the HR_Score worksheet. What is the optimal k? What is the predicted outcome
for the first new employee?
b. What is the misclassification rate for the optimal k?
c. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).
d. Display the cumulative lift chart, decile-wise chart, and ROC curve.
e. Comment on the performance of the KNN classification model. Is the KNN
classification an effective way to predict an employee’s success?
14. FILE Heart. In recent years, medical research has incorporated the use of data
analytics to find new ways to detect heart disease in its early stage. Medical
doctors are particularly interested in accurately identifying high-risk patients so
that preventive care and intervention can be administered in a timely manner. A
readily available set of information such as the patient’s age (Age), blood pressure
(BP Systolic and BP Diastolic), and BMI, along with an indicator whether or not
the patient has heart disease (Disease = 1 if heart disease, 0 otherwise). A
portion of the data set is provided in the accompanying table.
 a. Perform KNN analysis on the data set. What is the optimal value of k?
b. What is the misclassification rate for the optimal k?
c. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).
d. Display the cumulative lift chart, decile-wise chart, and ROC curve.
e. Comment on the performance of the KNN classification model. Is the KNN
classification an effective way to predict heart disease?
15. FILE Retail. Online retailers often use a recommendation system to suggest new
products to consumers. Consumers are compared to others with similar
characteristics such as past purchases, age, income, and education level. A data
set, such as the one shown in the accompanying table, is often used as part of a
product recommendation system in the retail industry. The variables used in the
system include whether or not the consumer eventually purchases the suggested
item (Purchase = 1 if purchased, 0 otherwise), the consumer’s age (Age in years),
income (Income, in $1,000s), and number of similar items previously purchased
(PastPurchase).

page 389
Purchase Age Income PastPurchase

1 48 99 21

1 47 32 0

⋮ ⋮ ⋮ ⋮

0 34 110 2

a. Perform KNN analysis on the Retail_Data worksheet to determine whether or

not a consumer is likely to make a purchase. Score the records of the 10 new
consumers in the Retail_Score worksheet. What is the optimal k? What is the
predicted outcome for the first new consumer?
b. What is the misclassification rate for the optimal k?
c. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).
 d. Generate the decile-wise lift chart. What is the lift value of the leftmost bar?
e. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
f. Comment on the performance of the KNN classification model. Is the KNN
classification an effective way to develop a recommendation system?
16. FILE Solar. New Age Solar sells and installs solar panels for residential homes.
The company’s sales representatives contact and pay a personal visit to potential
customers to present the benefits of installing solar panels. This high-touch
approach works well as the customers feel that they receive personal services
that meet their individual needs, but it is more expensive than other, mass-
marketing approaches. The company wants to be very strategic about visiting
potential customers who are more likely to install solar panels. The company has
compiled a data set of past home visits by sales reps. The data include the age
and annual income (in $1,000s) of the potential customer and whether or not the
customer purchases the solar panels (Install: Y/N). A portion of the Solar_Data
worksheet is shown in the accompanying table.

Income Age Install

115 45 N

68 31 Y

⋮ ⋮ ⋮

73 34 N

a. Perform KNN analysis on the Solar_Data worksheet to determine whether or

not a potential customer is likely to purchase solar panels. Score the records of
the 11 new potential customers in the Solar_Score worksheet. What is the
optimal k? What is the predicted outcome for the first new potential customer?
b. What is the misclassification rate for the optimal k?
c. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).
d. Display the cumulative lift chart, decile-wise chart, and ROC curve.
e. Based on your answers in part c and d, is KNN an effective way to classify
potential customers?
17. FILE Machine. Being able to predict machine failures before they happen can
save millions of dollars for manufacturing companies. Manufacturers want to be
able to perform preventive maintenance or repairs in advance to minimize
machine downtime and often install electronic sensors to monitor the machines
and their surrounding environment. However, the more sophisticated the machine,
the more difficult it is to diagnose and predict the failure rate. Data mining has
been used to analyze environmental factors to predict whether or not complex
 machines such as nanotechnology equipment will fail from one production period
to another. The accompanying data set contains 480 observations and three
environmental variables: level of humidity in the room where the equipment is
located (Humid, in percentage); overall temperature in the room (Temp, in
Fahrenheit); and a target variable that indicates whether or not the equipment
broke down during the next production period (Breakdown = 1 if breakdown, 0
otherwise).

Humid Temp Breakdown

9.36 79.99 0

10.37 61.89 0

⋮ ⋮ ⋮

15.02 38.02 1

a. Perform KNN analysis on the data set. What is the optimal value of k?
b. What is the misclassification rate for the optimal k?
c. Report the accuracy, specificity, sensitivity, and precision rates for the test data
set (for Analytic Solver) or validation data set (for R).
d. What is the area under the ROC curve (or the AUC value)?
e. Based on your answers in part c and d, is KNN an effective way to classify
potential customers?
f. Change the cutoff value to 0.3. Report the accuracy, specificity, sensitivity, and
precision rates for the test data set (for Analytic Solver) or validation data set
(for R).

page 390

9.3 THE NAÏVE BAYES METHOD

LO 9.3
Apply the naïve Bayes method for classification.
The naïve Bayes method represents a family of classification models
that allows us to predict the probability that a record belongs to a
particular class. It is based on Bayes’ theorem discussed in Chapter
4. Recall that Bayes’ theorem uses new information to update a prior
probability to form a posterior probability.
Suppose the target variable is defined by two classes: a potential
customer will make a purchase and a potential customer will not
make a purchase. Here, we can estimate the prior probability using
the proportion of customers who made a purchase. If we use a
customer’s sex as a predictor variable, the posterior probability is
estimated conditional on the sex of the potential customer, male or
female. Similarly, other predictor variables, which must be
categorical or transformed into categories, are used to estimate
posterior probabilities. It is important to note that estimating posterior
probabilities can be cumbersome when the number of predictor
variables is large.
For computational simplicity, the naïve Bayes classifier assumes
that the predictor variables are independent. This independence
assumption is often not realistic, hence the label “naïve.” While the
assumption of variable independence does not capture the possible
interaction between predictor variables, this simple method can
perform well in many real-world situations, compared to a more
sophisticated classification method, especially in situations where
ranking cases in terms of their probabilities of belonging to the target
class is the primary objective. The underlying concept of naïve
Bayes is quite simple to understand, easy to implement, and
especially attractive when applied to applications with large data sets
such as target marketing, medical diagnosis, spam detection, textual
document classification, and face recognition.

THE NAÏVE BAYES METHOD

The naïve Bayes method represents a family of classification
models based on Bayes’ theorem. Even though it makes a
strong independence assumption between categorical predictor
 variables, it tends to perform well in many real-world
applications.

In Chapter 4, we used Bayes’ theorem, defined as ,

to update the prior probability P(B) to the posterior probability P(B|A).

Consider a simple classification model where the target variable y
and the predictor variable x are both binary. Here, we use Bayes’

theorem to get . Continuing with the customer purchase

example, let y equal 1 if a potential customer makes a purchase, 0

otherwise, and let x equal 1 if male, 0 otherwise. Table 9.6 shows the
number of observations in each class of the training data set.

TABLE 9.6 Counts in the Training Data Set

Male (x = 1) Female (x = 0)

Purchase (y = 1) 30  50

No Purchase (y = 0) 20 100

We use the relevant proportions to compute the conditional

probability of making (not making) a purchase given that the
potential customer is a male (female), as:

If the potential customer is female, we derive the conditional

probability as:
 In this simple example, we assign a male customer to the
“purchase” class because 0.60 > 0.40 and a female customer to the
“no purchase” class because 0.333 < 0.667. Two page 391
things are noteworthy in this example. First, it uses a
categorical predictor variable, a customer’s sex, with multiple records
for both male and female customers. If we had used a numerical
variable like income as a predictor variable, it would have been
highly unlikely to find multiple records at all levels of income. In
practice, it is common to use numerical predictor variables but first
transform them into categories. Second, it is highly unlikely to find a
classification model that employs only one predictor variable. Often,
several predictor variables are used for classification. We will now
address both of these issues.
Continuing with the customer purchase example, let y represent
whether or not a potential customer makes a purchase. Let x1, x2, . .
. , xk represent the k predictor variables, which now may include a
customer’s sex, income, age, and college education. The extended
Bayes’ theorem describes the probability of y given values of x1
through xk as:

Assuming that x1 through xk are independent of each other, the

probability of y given values of x1 through xk can be simplified as:
 This simplification significantly improves the efficiency of the
naïve Bayes classifiers over other classification methods, especially
when performed on a large data set and/or with a large number of
predictor variables. If a retailer wants to classify potential customers
as y = 1 for purchase, 0 otherwise, we would compute and compare
the following two probability values:

If P(y = 1| x1, x2, . . . , xk) is higher, then we conclude that the

potential customer will make a purchase; i.e., = 1. Similarly, if P(y =

0| x1, x2, . . . , xk) is higher, then we conclude that the potential

customer will not make a purchase; i.e., = 0. In case of ties, various

methods such as a random tiebreaker or a selection that favors the

most frequent outcome, or an alphabetical order might be used.
As mentioned earlier, a naïve Bayes classifier is performed on a
data set where all predictor variables are categorical. To capture the
influence of continuous variables, they must first be converted into
discrete categories, a process called “binning,” which was detailed in
Chapter 2. For example, incomes could easily be converted into
discrete values or bins as 1 = [0, $50,000), 2 = [$50,000, $100,000),
and 3 = [$100,000, $500,000), and so on. There is also a special
family of naïve Bayes classifiers called the Gaussian naïve Bayes
where continuous variables are assumed to follow a normal
distribution, and the density functions are used to estimate
conditional probabilities for the naïve Bayes classifier; this technique,
however, is beyond the scope of this text.
 Another potential issue with the naïve Bayes classifier is a
situation that includes a rare outcome such as classifying patients
with a rare disease. The problem occurs when this rare outcome is
not present in the training data set. According to the formulas shown
earlier, the probability of this rare outcome would be estimated as 0,
and none of the new records would be classified or assigned to this
category. To overcome this problem, most naïve Bayes algorithms
allow a replacement of zero probability values with a nonzero value
(usually with a value of 1). This technique is called smoothing or
Laplace smoothing.

page 392

Because the naïve Bayes method is a supervised data mining

technique, data partitioning is used to assess model performance.
Unlike the KNN method, the naïve Bayes method does not use the
validation data set to optimize model complexity (e.g., finding the
optimal value for k); therefore, it is customary to partition the data
into training and validation sets only. The training data are used to
compute the conditional probabilities, whereas the validation data
are used to assess the performance of the naïve Bayes model on a
previously unseen data set.
The naïve Bayes classifier essentially uses relevant proportions
to estimate conditional probabilities for each class of the target
variable and for all combinations of the predictor variable values in
the training data set. Although simple in principle, this approach can
get quite cumbersome when performed with a large number of
predictor variables. Fortunately, several software packages routinely
implement the naïve Bayes procedure. In Example 9.2, we will
illustrate this procedure with Analytics Solver and R.
FILE
Census

EXAMPLE 9.2
An institute for public policy in Washington, D.C., hires a
number of college interns every summer. This year, Sara
Anderson, a third-year Economics major from Massachusetts,
is selected as one of the research interns. Her first assignment
is to conduct data analysis to help congressional offices gain a
better understanding of U.S. residents whose incomes are
below the poverty level. To complete her assignment, Sara
extracts a relevant data set maintained by the U.S. Census
Bureau. The data set has 9,980 observations and is stored in
the Census_Data worksheet of the Census data file. Each
observation contains an individual’s marital status (Married;
yes/no), sex (Female; yes/no), ethnicity (White; yes/no), age
groups (Age: 1 for [18, 25), 2 for [25, 35), 3 for [35, 45), 4 for
[45, 55), and 5 for 55 years and older), whether or not the
individual receives college-level education (Edu; yes/no), and
whether or not the individual’s income is below the poverty
level (Poverty = 1 if living in poverty, 0 otherwise). In addition,
she keeps records of 66 new individuals with the predictor
variables in the Census_Score worksheet of the Census data
file for scoring based on the naïve Bayes classifier. A portion of
the two worksheets is shown in Table 9.7.

TABLE 9.7 A Portion of Census Data

Help Sara develop a naïve Bayes classification model to

classify whether or not an individual is likely to live in poverty
using all five predictor variables. Report the relevant
performance measures and score new observations.
 SOLUTION:
Using Analytic Solver

a. Open the Census data file and go to the Census_Data

worksheet.
b. Choose Data Mining > Partition > Standard Partition.
page 393
c. For Data Range, highlight cells $A$1:$F$9981
and check the box First Row Contains Headers. The Variables
in Input Data box will populate. Choose Automatic
percentages with 60% for Training Set and 40% for Validation
Set. Accept other defaults and click OK.
d. Analytic Solver will create a new worksheet called
STDPartition with the partitioned data. Make sure that the
STDPartition worksheet is active. Choose Data Mining >
Classify > Naive Bayes.
e. In the Data tab, select and move Married, Female, White,
Age, and Edu to Selected Variables box. Select Poverty as the
Output Variable. Accept the other defaults and click Next.
Note: If a cutoff value other than the default 0.5 is used, input
the desired cutoff value (e.g. 0.75) in the Success Probability
Cutoff: box in this step.
f. In the Parameters tab, note that Analytic Solver provides an
option for Laplace smoothing. In this case, there are no
missing values for any outcome categories in the training data
set, and therefore, there is no need for smoothing. Uncheck
the box next to Laplace Smoothing. Click on the Scoring tab.
g. Check the Summary Report boxes for Score Training Data
and Score Validation Data. Also check the Lift Charts box for
Score Validation Data. Check the In Worksheet box under the
Score New Data section. A new tab called New Data (WS) will
open. In the Worksheet dropdown box, select Census_Score,
select the data range $A$1:SES67, click Match By Name, and
Finish.
 Analytic Solver will create six new worksheets. The
NB_Output worksheet summarizes the naïve Bayes input
parameters that we specified earlier. The NB_TrainingScore
and NB_ValidationScore worksheets display the confusion
matrix and model performance metrics for the training and
validation data sets, respectively. The confusion matrix and
accuracy statistics for the validation data set are shown in
Figure 9.16.

FIGURE 9.16 Analytic Solver’s classification summary tables for naïve Bayes

Source: Microsoft Excel

While the overall accuracy is 79.83%, there is a

considerable difference between specificity (51.36%) and
sensitivity (87.89%). This indicates that, by using a default
cutoff value of 0.5, the model is more accurate in classifying
people in poverty (Class 1) than classifying people who are not
in poverty (Class 0).
The performance of the naïve Bayes classifier can also be
evaluated graphically using the cumulative lift chart, the decile-
wise chart, and the ROC curve for the test data, as shown in
Figure 9.17. The lift chart and the decile-wise lift chart show
that the naïve Bayes classifier performs better than the
baseline model. The ROC curve also shows that the naïve
Bayes classifier performs better than the baseline model in
terms of sensitivity and specificity across all cutoff values. The
area under curve (AUC) is 0.8465, which is closer to the
optimum classifier (AUC = 1) than to the random classifier
(AUC = 0.5).

page 394

FIGURE 9.17 Analytic Solver’s performance charts for naïve Bayes

The NB_NewScore worksheet shows the classification

predictions for the new records from the Census_Score
worksheet. For each observation, the naïve Bayes classifier
displays the predicted value (poverty = 1 or 0, and the
probabilities for both outcomes). Table 9.8 shows a portion of
 the classification outcomes and the original predictor variables.
The naïve Bayes model classified individuals who are married
male with a college education as unlikely to live in poverty.

TABLE 9.8 Naïve Bayes Scoring Results

Using R
As before, we will use the caret package for classification with
k-fold cross-validation. So, we will divide the data into two
partitions for training and validation.
a. Import the data from the Census_Data worksheet of the
Census data file into a data frame (table) and label it myData.
The following instructions and results are based on R version
3.5.3. To replicate the results with newer versions of R, enter:
> suppressWarnings(RNGversion(“3.5.3”))
page 395
b. The caret and klaR packages contain necessary
functions for partitioning the data, k-fold cross-validation, and
naïve Bayes classification. Also, similar to the KNN example,
a cumulative lift chart, decile-wise chart, and ROC curve can
be generated for visual inspection of the model performance
using the gains and pROC packages. Install and load all
packages, if you have not already done so. Enter:
> install.packages(”caret”)
> install.packages(”klaR”)
> install.packages(”gains”)
 > install.packages(”pROC”)
> library(caret)
> library(klaR)
> library(gains)
> library(pROC)
c. We use the as.factor command to convert the poverty
variable into a categorical type. Enter:
> myData$Poverty <- as.factor(myData$Poverty)
d. As in the KNN example, we use the set.seed command to set
the random seed to one for consistency. We use the
createDataPartition function to partition the data into training
(60%) and validation (40%). Enter:
> set.seed(1)
> myIndex <- createDataPartition(myData$Poverty, p=0.6, list
= FALSE)
> trainSet <- myData[myIndex,]
> validationSet <- myData[-myIndex,]
e. We use the trainControl function to specify a 10-fold cross-
validation process. On the training data set, we use the train
function and set the method option equal to “nb”, which stands
for naïve Bayes. The Poverty variable is identified as the
target variable. To ensure consistency of the cross-validation
results, we again use the set.seed function to fix a random
seed. Enter:
> myCtrl <- trainControl(method=’cv’, number=10)
> set.seed(1)
> nb_fit <- train(Poverty ~., data = trainSet, method = ”nb”,
trControl=myCtrl)
> nb_fit
Figure 9.18 displays the naïve Bayes results. The naïve
Bayes classifier generates two accuracy rates based on the
 underlying distribution of the target variables and uses the
option that yields a higher accuracy (77.29% in this case) in
the final model.

FIGURE 9.18 Naïve Bayes results

page 396
f. We use the predict and confusionMatrix
functions to validate the model on the validation data set and
produce a confusion matrix. Enter:
> nb_class <- predict(nb_fit, newdata = validationSet)
> confusionMatrix(nb_class, validationSet$Poverty, positive =
’1’)
Note that in the confusionMatrix statement, we specify
the value ‘1’ in the poverty variable as a positive or success
class. The accuracy value (75.65%) remains consistent with
results from the training data set. Unlike in Analytic Solver, the
naïve Bayes model in R yields sensitivity and specificity
values that are close to each other (76.73% and 71.80%,
respectively). Again, the algorithms in the two software are not
the same and yield different results. A portion of the R results
is shown in Figure 9.19.

FIGURE 9.19 Confusion matrix

 We can also use other cutoff values, instead of the default
value of 0.5, to create the confusion matrix. For example, use
the following commands to specify a cutoff value of 0.75:
>nb_class_prob <- predict(nb_fit, newdata=validationSet,
type= ’prob’)
>confusionMatrix(as.factor(ifelse(nb_class_prob[,2]>0.75, ’1’,
’0’)), validationSet$Poverty, positive = ’1’)
g. To create the two lift charts and ROC curve, we first need to
compute a cumulative gains table by using the predict and
gains functions. Note that the poverty variable was converted
into a categorical type in a previous step. The gains function
requires the variable to be converted back to a numerical type.
Enter:
> nb_class_prob <- predict(nb_fit, newdata=validationSet,
type=’prob’)
> validationSet$Poverty <-
as.numeric(as.character(validationSet$Poverty))
> gains_table <- gains(validationSet$Poverty,
nb_class_prob[,2])
> gains_table
Figure 9.20 shows the gains table.

FIGURE 9.20 Gains table

 page 397
h. We use the plot and line functions to create the
cumulative lift chart as shown in Figure 9.21. Enter:

FIGURE 9.21 The cumulative lift chart

> plot(c(0,
gains_table$cume.pct.of.total*sum(validationSet$Poverty)) ~
c(0, gains_table$cume.obs),xlab = ”# of cases”,
ylab=”Cumulative”, type =”l”)
 > lines(c(0, sum(validationSet$Poverty)) ~ c(0,
dim(validationSet)[1]), col=”red”, lty=2)
i. We use the barplot function to create a decile-wise chart as
shown in Figure 9.22. Enter:

FIGURE 9.22 The decile-wise lift chart

>
barplot(gains_table$mean.resp/mean(validationSet$Poverty),
names.arg=gains_table$depth, xlab=”Percentile”, ylab=”Lift”,
ylim = c(0,1.5), main = ”Decile-Wise Lift Chart”)
j. We use the roc, plot.roc, and auc functions to create the
ROC curve and compute the area under the curve. Enter:
> roc_object<- roc(validationSet$Poverty, nb_class_prob[,2])
> plot.roc(roc_object)
> auc(roc_object)
The area under the ROC curve, or AUC, is 0.8437, indicating
that the naïve Bayes model performs well in predicting
whether an individual is in poverty or not. Figure 9.23 displays
 the ROC curve, which shows that model performs better than
the baseline model (shown as the green diagonal line) in
terms of sensitivity and specificity across all cutoff values.

FIGURE 9.23 The ROC curve

page 398
k. Import the new observations from the
Census_Score worksheet of the Census data file into a data
frame (table) and label it myScoreData. We then use the
predict function to score the 66 new records and append the
classification results back to the original data. Enter:
> nb_class_score <- predict(nb_fit, newdata = myScoreData)
> myScoreData <- data.frame(myScoreData, nb_class_score)
Table 9.9 shows R’s scoring results. Similar to the Analytic
Solver model, the KNN model in R tends to classify new
records who are married male with a college education as
unlikely to live in poverty.

TABLE 9.9 R’s Scoring Results for Naïve Baye

Transforming Numerical into Categorical
Values
As mentioned earlier, in order to perform a naïve Bayes analysis, all
predictor variables must be categorical. Here, we will briefly
demonstrate how to “bin” numerical values into discrete categories;
for additional information, students are encouraged to review related
sections in Chapter 2.
In the introductory case, Janet Williams, the membership
manager at a 24/7 Fitness Center, is holding an open house event
and collecting information from potential members. She wants to
explore naïve Bayes for determining whether or not potential
members are likely to purchase a gym membership. However,
because the predictor variables, Age, Income, and Hours, are
numerical, she has to first convert them into discrete categories.
FILE
Gym

Using the data from the Gym_Data worksheet, we will outline the
procedure for binning the Age variable; other variables can be
binned similarly. Note that the values of the Age variable range from
21 to 68. Suppose we want to bin Age into five categories.

page 399
 In Analytic Solver, go to Data Mining > Data Analysis >
Transform > Transform Continuous Data > Bin. Click on the
ellipsis next to Data range and highlight cells $A$1:$D$1001.

Select Age to convert it into a binned variable called Binned_Age

and select 5 for #bins for variable. Select Equal interval. Click on
Apply to Selected Variable and click Finish. The Bin_Transform
worksheet will show the binned Age variable ranging from one to
five. The new variable can be copied and pasted back to the original
Gym_Data worksheet and used in the naïve Bayes method and
other analysis techniques that require categorical variables.
In R, we first import the data from the Gym_Data worksheet into
R and call the resulting data frame Binning. The following statement
forces the first bin to start at 20 years and cap the last bin at 70
years with each interval value set at 10 years.
> age_bins <- cut(Binning$Age, breaks = seq(20, 70, by = 10))

The age_bins variable can now be appended back to the original

data frame and used in the analysis that requires categorical
variables.
Generally, we adjust the binning options according to the data or
conceptual design. For example, if we want to compare low- versus
high-income individuals, then we might want to bin the Income
variable into two categories. Additional discussions on this topic can
be found in Chapter 2.

EXERCISES 9.3
Note: These exercises can be solved using Analytic Solver
and/or R. The answers, however, will depend on the software
package used. All answers in R are based on version 3.5.3. To
replicate the results with newer versions of R, execute the
following line of code at the beginning of the R session:
suppressWarnings(RNGversion(“3.5.3”)). For Analytic Solver,
 partition the data into 60% training and 40% validation and use
12345 as the default random seed. For R, partition the data into
60% training and 40% validation and implement the 10-fold
cross-validation. Use the statement set.seed(1) to specify the
random seed one for data partitioning and cross-validation.
Some data files have two worksheets (e.g.,
Exercise_9.18_Data and Exercise_9.18_Score worksheets) for
model development and scoring new records.

Mechanics
18. FILE Exercise_9.18. The accompanying data set contains three predictor
variables (x1, x2, and x3) and the target variable (y). Partition the data in the
Exercise_9.18_Data worksheet to develop a naïve Bayes classification model
where “Yes” denotes the positive or success class for y. Score the 10 new
observations on the Exercise_9.18_Score worksheet.
a. Report the accuracy, sensitivity, and specificity rates for the validation data set.
b. Generate the decile-wise lift chart. What is the lift value of the leftmost bar?
What does this value imply?

page 400
c. Generate the ROC curve. What is the area under the ROC
curve (or the AUC value)?
d. Report the scoring results for the first three new observations.
19. FILE Exercise_9.19. The accompanying data set contains three predictor
variables (x1, x2, and x3) and the target variable (y). Partition the data in the
Exercise_9.19_Data worksheet to develop a naïve Bayes classification model
where “Y” denotes the positive or success class for y. Score the five new
observations on the Exercise_9.19_Score worksheet.
a. Report the accuracy, sensitivity, and specificity rates for the validation data set.
b. What is the area under the ROC curve (or AUC value)?
c. Report the scoring results for the five new observations.
20. FILE Exercise_9.20. The accompanying data set contains four predictor
variables (x1, x2, x3, and x4) and the target variable (y). Partition the data in the
Exercise_9.20_Data worksheet to develop a naïve Bayes classification model
where “1” denotes the positive or success class for y. Score the five new
observations on the Exercise_9.20_Score worksheet.
a. Report the accuracy, sensitivity, and specificity rates for the validation data set.
b. Generate the cumulative lift chart. Does the entire lift curve lie above the
baseline?
 c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
d. Report the scoring results for the five new observations.
e. Develop the naïve Bayes model with only x1, x2, and y in the naïve Bayes
model. Repeat parts a through c and compare the results.

page 401
21. FILE Exercise_9.21. The accompanying data set contains
three predictor variables (x1, x2, and x3) and the target variable (y). Partition the
data to develop a naïve Bayes classification model where “1” denotes the positive
or success class for y.
a. Report the accuracy, sensitivity, and specificity rates for the validation data set.
b. Generate the decile-wise lift chart. What is the lift value of the leftmost bar?
What does this value imply?
c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
d. Can the naïve Bayes model be used to effectively classify the data? Explain
your answer.
22. FILE Exercise_9.22. The accompanying data set contains three predictor
variables (x1, x2, and x3) and the target variable (y). Partition the data to develop
a naïve Bayes classification model where “1” denotes the positive or success
class for y.
a. Report the accuracy, sensitivity, specificity, and precision rates for the validation
data set.
b. Generate the cumulative lift chart. Does the entire lift curve lie above the
baseline?
c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
d. Can the naïve Bayes model be used to effectively classify the data? Explain
your answer.
23. FILE Exercise_9.23. The accompanying data set contains two predictor
variables (x1 and x2 ) and the target variable (y). Partition the data to develop a
naïve Bayes classification model where “1” denotes the positive or success class
for y.
a. Report the accuracy, sensitivity, specificity, and precision rates for the validation
data set.
b. Generate the decile-wise lift chart. What is the lift value of the leftmost bar?
c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
d. Can the naïve Bayes model be used to effectively classify the data? Explain.
24. FILE Exercise_9.24. The accompanying data set contains three predictor
variables (x1, x2, and x3) and the target variable (y).
 a. Bin predictor variables x1, x2, and x3. For Analytic Solver, choose the Equal
count option and three bins for each of the three variables. For R, bin x1 into [0,
6), [6, 14), and [14, 30); x2 into [0, 10), [10, 20), and [20, 61); and x3 into [0, 3),
[3, 5), and [5, 10). What are the bin numbers for the variables of the first two
observations?
b. Partition the transformed data to develop a naïve Bayes classification model
where “1” denotes the positive or success class for y. Report the accuracy,
sensitivity, specificity, and precision rates for the validation data set.
c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
d. Change the cutoff value to 0.2. Report the accuracy, sensitivity, specificity, and
precision rates for the validation data set.
25. FILE Exercise_9.25. The accompanying data set contains three predictor
variables (x1, x2, and x3) and the target variable (y).
a. Bin predictor variables x1 and x2. For Analytic Solver, choose the Equal count
option and three bins for each of the three variables. For R, bin x1 into [0, 60),
[60, 400), and [400, 30000); and x2 into [0, 160), [160, 400), and [400, 800).
What are the bin numbers for the variables of the first two observations?
b. Partition the transformed data to develop a naïve Bayes classification model
where “1” denotes the positive or success class for y. Report the accuracy,
sensitivity, specificity, and precision rates for the validation data set.
c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
26. FILE Exercise_9.26. The accompanying data set contains three predictor
variables (x1, x2, and x3) and the target variable (y).
a. Bin predictor variables x1, x2, and x3. For Analytic Solver, choose the Equal
interval option and 2 bins for each of the three variables. For R, bin x1 into [0,
125) and [125, 250); x2 into [0, 30) and [30, 60); and x3 into [0, 30) and [30,
60). What are the bin numbers for the variables of the first two observations?
b. Partition the transformed data to develop a naïve Bayes classification model
where “1” denotes the positive or success class for y. Report the accuracy,
sensitivity, specificity, and precision rates for the validation data set.
c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
d. Change the cutoff value to 0.4. Report the accuracy, sensitivity, specificity, and
precision rates for the validation data set.
27. FILE Exercise_9.27. The accompanying data set contains four predictor
variables (x1, x2, x3, and x4) and the target variable (y).
a. Bin predictor variables x1, x2, x3, and x4. For Analytic Solver, choose the Equal
interval option and 2 bins for each of the four variables. For R, bin x1 into [0,
 40000) and [40000, 80000); x2 into [0, 50) and [50, 100); x3 into [50, 75) and
[75, 100); and x4 into [0, 20000) and [20000, 40000). What are the bin numbers
for the variables of the first two observations?
b. Partition the transformed data to develop a naïve Bayes classification model
where “1” denotes the positive or success class for y. Report the accuracy,
sensitivity, specificity, and precision rates for the validation data set.
c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
d. Can the naïve Bayes model be used to effectively classify the data? Explain
your answer.

Applications
28. FILE International. Every year, hundreds of thousands of international students
apply to graduate programs in the United States. Two of the most important
admissions criteria are undergraduate GPAs and TOEFL scores. An English
language preparation school in Santiago, Chile, wants to examine the acceptance
records of its former students who had applied to graduate school in the United
States during the past two years. The results will be used to help advise new
students about their chance of acceptance to their first choice of graduate
programs. A portion of the data set is shown in the accompanying table with the
following variables: Accept (1 if accepted, 0 otherwise); GPA (1 for below 3.00, 2
for 3.00–3.49, 3 for 3.50 and above); and TOEFL (H for 80 or above, L for below
80).

Accept GPA TOEFL

1 1 H

0 1 H

⋮ ⋮ ⋮

0 1 L

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
c. Can the naïve Bayes model be used to effectively classify the data? Explain
your answer.
29. FILE OnlineRetail. An online retailer is offering a new line of running shoes. The
retailer plans to send out an e-mail with a discount offer to some of its existing
customers and wants to know if it can use data mining analysis to predict whether
 or not a customer might respond to its e-mail offer. The retailer prepares a data
set of 170 existing customers who had received online promotions in the past,
which include the following variables: Purchase (1 if purchase, 0 otherwise); Age
(1 for 20 years and younger, 2 for 21 to 30 years, 3 for 31 to 40 years, 4 for 41 to
50 years, and 5 for 51 and older); Income (1 for $0 to $50K, 2 for $51K to $80K, 3
for $81K to $100K, 4 for $100K+); and PastPurchase (1 for no past purchase, 2
for 1 or 2 past purchases, 3 for 3 to 6 past purchases, 4 for 7 or more past
purchases). A portion of the data set is shown in the accompanying table.

Purchase Age Income PastPurchase

1 4 3 4

1 4 1 1

⋮ ⋮ ⋮ ⋮

1 3 4 3

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Generate the decile-wise lift chart. What is the lift value of the leftmost bar?
What does this value imply?
c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Can the naïve Bayes model be used to effectively classify the data? Explain
your answer.
30. FILE MedSchool. Admission to medical schools in the United States is highly
competitive. The acceptance rate to the top medical schools could be as low as
2% or 3%. With such a low acceptance rate, medical school admissions
consulting has become a growing business in many cities. In order to better serve
his clients, Paul Foster, a medical school admissions consultant, wants to build a
data-driven model to predict whether or not a new applicant is likely to get
accepted into one of the top 10 medical schools. He collected a database of 1,992
past applicants to the top 10 medical schools with the following information: Sex
(F = female, M = male), CollegeParent (1 if parents with college degrees, 0
otherwise), GPA (1 if undergraduate GPA of 3.50 or higher, 0 otherwise), Med (1 if
accepted to the top 10 medical school, 0 otherwise). A portion of the data set is
shown in the accompanying table.

Sex CollegeParent GPA Med

F 1 1 1
 Sex CollegeParent GPA Med

M 1 0 1

⋮ ⋮ ⋮ ⋮

M 0 0 0

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Generate the decile-wise lift chart. What is the lift value of the leftmost bar?
What does this value imply?
c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Can the naïve Bayes model be used to effectively classify the medical school
applicant data? Explain your answer.
31. FILE CreditCard. A home improvement retail store is offering its customers
store-branded credit cards that come with a deep discount when used to
purchase in-store home improvement products. To maintain the profitability of this
marketing campaign, the store manager would like to make these offers only to
the customers who are likely to carry a high monthly balance on the credit card. A
data set obtained from a nationwide association of home improvement stores
contains records of 500 consumers who carry similar credit cards offered by other
home improvement stores. Relevant variables include Sex (Female or Male),
Education (1 if did not finish college, 2 if undergraduate degree, 3 if graduate
degree), Children (1 if have children, 0 otherwise), Age (1 if below 20 years old, 2
if 20–29 years, 3 if 30–39 years, 4 if 40–49 years, 5 if 50–59 years, 6 if 60 years
and older). Balance is the target variable where one indicates the customer
maintains a high monthly balance, and 0 otherwise. A portion of the
CreditCard_Data worksheet is shown in the accompanying table.

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
 c. Interpret the performance measures and evaluate the
effectiveness of the naïve Bayes model.
page 402
d. Score the new customer records in the CreditCard_Score worksheet. What is
the scoring result of the first customer record?
e. Change the cutoff value to 0.3. Report the accuracy, sensitivity, specificity, and
precision rates for the validation data set.
32. FILE Volunteer. A community center is launching a campaign to recruit local
residents to help maintain a protected nature preserve area that encompasses
extensive walking trails, bird watching blinds, wild flowers, and animals. The
community center wants to send out a mail invitation to selected residents and
invite them to volunteer their time to help but does not have the financial
resources to launch a large mailing campaign. As a result, they solicit help from
the town mayor to analyze a data set of 5,000 local residents and their past
volunteer activity, stored in the Volunteer_Data worksheet. The data include Sex
(F/M), Married (Y = married, N = not married), College (1 if college degree, 0
otherwise), Income (1 if annual income of $50K and above, 0 otherwise), and
Volunteer (1 if participated in volunteer activities, 0 otherwise). They want to use
the analysis results to help select potential residents who are likely to accept the
invitation to volunteer. A portion of the data set from the mayor is shown in the
accompanying table.

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Generate the ROC curve and decile-wise lift chart. What is the area under the
ROC curve (or AUC value)? What is the lift of the leftmost bar of the decile-wise
lift chart?
c. Score the new volunteer records in the Volunteer_Score worksheet. What is the
scoring result of the first customer record?
33. FILE Vacation. Nora Jackson owns a number of vacation homes on a beach.
She works with a consortium of rental home owners to gather a data set to build a
classification model to predict the likelihood of potential customers renting a
beachfront home during holidays. A portion of the data set is shown in the
accompanying table with the following variables: whether the potential customer
owns a home (Own = 1 if yes, 0 otherwise), whether the customer has children
(Children = 1 if yes, 0 otherwise), the customer’s age in years (Age), annual
income (Income), and whether or not the customer has previously rented a
beachfront house (Rental = 1 if yes , 0 otherwise).
 a. Bin the Age and Income variables as follows. For Analytic Solver, choose the
Equal count option and two bins for each of the two variables. For R, bin Age
into [22, 45) and [45, 85) and Income into [0, 85000) and [85000, 300000).
What are the bin numbers for Age and Income of the first two observations?

page 403
b. Partition the transformed data to develop a naïve Bayes
classification model. Report the accuracy, sensitivity, specificity, and precision
rates for the validation data set.
c. Generate the decile-wise lift chart. What is the lift value of the leftmost bar?
d. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
e. Interpret the performance measures and evaluate the effectiveness of the naïve
Bayes model.
34. FILE InGame. A mobile gaming company wants to study a group of its existing
customers about their in-game purchases. A data set, a portion of which is shown
in the accompanying table, is extracted and includes how old the customer is
(Age), Sex (1 if female, 0 otherwise), the amount of weekly play time in hours
(Hours), whether or not the customer’s mobile phone is linked to a Facebook
account (Facebook = 1 if yes, 0 otherwise), and whether or not the customer has
made an in-game purchase (Buy = 1 if yes, 0 otherwise).

a. Bin the Age and Hours variables as follows. For Analytic Solver, choose the
Equal interval option and two bins for each of the two variables. For R, bin Age
into [15, 40) and [40, 65) and Hours into [0, 20) and [20, 40). What are the bin
numbers for Age and Hours for the first two observations?
b. Partition the transformed data to develop a naïve Bayes classification model.
Report the accuracy, sensitivity, specificity, and precision rates for the validation
data set.
 c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Interpret the performance measures and evaluate the effectiveness of the naïve
Bayes model.
35. FILE Grit. Forbes magazine published an article that studied career
accomplishments and factors that might contribute to career success (August 30,
2018). It turns out that career success has less to do with talents and is not
necessarily influenced by test scores or IQ scores. Rather, “grit,” or a combination
of persistence and passion, was found to be a good indicator of a person’s career
success. Tom Weyerhaeuser, an HR manager at an investment bank, is
conducting a campus recruitment and wants to know how he might be able to
measure “grit.” He thought that he could ask each job candidate about his or her
GPA, athletic activities, leadership roles in college, study abroad experience, and
employment during college as a way to gauge whether or not each candidate has
the persistence and passion to succeed in the investment banking industry. He
extracts data from the corporate HR database on 157 current employees with
variables indicating whether or not they currently hold an upper-management
position (Success = 1 if yes, 0 otherwise), had a part-time job in college (Job = 1 if
yes, 0 othwerwise), graduated with a GPA higher than 3.5 (GPA = 1 if yes, 0
otherwise), represented their university in athletic activities (Sports = 1 if yes, 0
otherwise), and had a leadership role in college organizations (Leadership = 1 if
yes, 0 otherwise). A portion of the Grit_Data worksheet is shown in the
accompanying table.

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Generate and display the cumulative lift chart, the decile-wise lift chart, and the
ROC curve.
c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Interpret the performance measures and evaluate the effectiveness of the naïve
Bayes model.
e. Score the new job applicants in the Grit_Score worksheet. What is the predicted
outcome of the first applicant?
36. FILE Graduation. Predicting whether or not an entering freshman student will
drop out of college has been a challenge for many higher education institutions.
Nelson Touré, a senior student success adviser at an ivy-league university, has
 been asked to investigate possible indicators that might allow the university to be
more proactive to provide support for at-risk students. Nelson reviews a data set
of 200 former students and selects the following variables to include in his study:
Graduate (Graduate = 1 if graduated, 0 otherwise); whether or not the student
received a passing grade in his or her first calculus, statistics, or math course
(Math = 1 if yes, 0 otherwise); whether or not the student received a passing
grade in his or her first English or communications course (Language = 1 if yes, 0
otherwise); whether or not the student had any contact with the advising center
during his or her first semester (Advise = 1 if yes, 0 otherwise); and whether or not
the student lived on campus during his or her first year at college (Dorm = 1 if yes,
0 otherwise). A portion of the data is shown in the accompanying table.

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Generate the cumulative lift chart. Does the entire lift curve lie above the
baseline?
c. Generate the ROC curve. What is the area under the ROC curve (or the AUC
value)?
d. Interpret the results and evaluate the effectiveness of the naïve Bayes model.
37. FILE Fraud. Credit card fraud is becoming a serious problem for the financial
industry and can pose a considerable cost to banks, credit card issuers, and
consumers. Fraud detection using data mining techniques has become an
indispensable tool for banks and credit card companies to combat fraudulent
transactions. A sample credit card data set contains the following variables: Fraud
(1 if fraudulent activities, 0 otherwise), Amount (1 if low, 2 if medium, 3 if high),
Online (1 if online transactions, 0 otherwise), and Prior (1 if products that the card
holder previously purchased, 0 otherwise). A portion of the data set is shown in
the accompanying table.

Fraud Amount Online Prior

0 2 0 1

0 3 0 0

⋮ ⋮ ⋮ ⋮
 Fraud Amount Online Prior

0 2 0 1

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Generate the decile-wise lift chart. What is the lift value of the page 404
leftmost bar? What does this value imply?
c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Interpret the performance measures and evaluate the effectiveness of the naïve
Bayes model.
e. Change the cutoff value to 0.1. Report the accuracy, sensitivity, specificity, and
precision rates for the validation data set.
38. FILE Insurance. Insurance companies use a number of factors to help
determine the premium amount for car insurance coverage. Discounts or a lower
premium may be given based on factors including credit scores, history of at-fault
accidents, age, and sex. Consider the insurance discount data set from 200
existing drivers. The following variables are included in the data set: Discount (1 if
yes, 0 otherwise), Female (1 if female, 0 otherwise), Credit (1 if low, 2 if medium,
3 if high scores), AtFault (1 if history of at-fault accidents, 0 otherwise), and Age
(1 if 25 years and older, 0 otherwise). A portion of the data is shown in the
accompanying table.

a. Partition the data to develop a naïve Bayes classification model. Report the
accuracy, sensitivity, specificity, and precision rates for the validation data set.
b. Display the cumulative lift chart, the decile-wise lift chart, and the ROC curve.
c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Interpret the performance measures and evaluate the effectiveness of the naïve
Bayes model.
39. FILE Solar. Refer to Exercise 9.16 for the description of a solar panel company
called New Age Solar and the Solar_Data worksheet.
a. Bin the Age and Income variables in the Solar_Data worksheet as follows. For
Analytic Solver, choose the Equal count option and two bins for each of the two
 variables. For R, bin Age into [30, 50) and [50, 90) and Income into [30, 85) and
[85, 140). What are the bin numbers for Age and Income of the first two
observations?
b. Partition the transformed data and develop a naïve Bayes classification model.
Report the accuracy, sensitivity, specificity, and precision rates for the validation
data set.
c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Interpret the results and evaluate the effectiveness of the naïve Bayes model.
40. FILE Depression. Michelle McGrath is a college student working to complete an
undergraduate research project to fulfill her psychology degree requirements. She
is interested in how physical and behavioral factors might be used to predict an
individual’s risk of having depression. After receiving an approval from her
adviser, she sends out a survey to local residents asking for their age (Age, in
years), years of education (Education), the number of hours per month they
engaged in moderate or vigorous physical activities (Hours), and whether or not
they have experienced depression (Depression: Y/N). A portion of the data from
261 respondents is shown in the accompanying table.

Age Education Hours Depression

44 12 20 Y

49 9 30 Y

⋮ ⋮ ⋮ ⋮

69 15 34 Y

a. Bin the Age, Education, and Hours variables as follows. For Analytic Solver,
choose the Equal interval option and three bins for each of the variables. For R,
bin Age into [20, 40), [40,60), and [60, 81), Education into [9, 12), [12, 16), and
[16, 20), and Hours into [0, 55), [55, 105), and [105, 150). What are the bin
numbers for the variables of the first two observations?
b. Partition the transformed data and develop a naïve Bayes classification model.
Report the accuracy, sensitivity, specificity, and precision rates for the validation
data set.
c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Interpret the results and evaluate the effectiveness of the naïve Bayes model.
9.4 Writing With Big Data
FILE
College_Admission

In this chapter, we discussed two well-known supervised data mining techniques for
classification problems: the KNN method and the naïve Bayes method. In the
introductory case, we used the KNN method to analyze gym membership data that
have numerical predictor variables to predict whether or not a potential member will
purchase a gym membership. The naïve Bayes method, however, requires
categorical predictor variables. In the Big Data case presented next, we will
transform numerical predictor variables into categorical variables and use the naïve
Bayes method to predict the college admissions of high school students. Details on
data wrangling and transformation are discussed in Chapter 2.

page 405

Case Study
Every year, millions of high school students apply and vie for acceptance to a college
of their choice. For many students and their parents, this requires years of
preparation, especially for those wishing to attend a top-ranked college. In high
schools, students usually work with college advisors to research different colleges
and navigate the admissions process.
Elena Sheridan, a college counselor at Beachside High School, is working with 14
students who are interested in applying to the same selective four-year college. She
is asked by her school principal to prepare a report that analyzes the chances of the
14 students getting accepted into one of the three academic programs. In a database
of past college applicants available to counselors at Beachside High, predictor
variables include the student’s high school GPA, SAT score, and the Male, White,
and Asian dummy variables that capture the student’s sex and ethnicity. Elena also
wants to know whether or not the parents’ education can be a predictor of a student’s
college acceptance and plans to include the education level of both parents in her
analysis.
Based on her conversation with college counselors at other high schools, she
believes that high school students with a GPA of 3.5 or above have a much higher
chance of getting accepted into a selective college. She also thinks that SAT scores
of at least 1,200 substantially increase the chance of acceptance. To test these
anecdotal assumptions, Elena wants to convert the GPAs and SAT scores into the
categories corresponding to these thresholds. In addition, the database has a target
variable indicating whether or not the past applicant was accepted to the college.
 Develop the naïve Bayes classification model and create a report that presents
an analysis of the factors that may influence whether or not a high school student is
admitted to a selective four-year college. Predictor variables should include the
applicant’s sex, ethnicity, parents’ education levels, GPA, and SAT scores. Transform
the GPAs and SAT scores into appropriate categorical variables. Make predictions
whether or not each of the 14 high school students at Beachside High in the
College_Admission_Score worksheet will be admitted.

Sample Report – College Admissions

High school students work hard to excel academically and set themselves
apart with extracurricular achievements. Getting into the right college and
choosing the right major can help start them out on a successful professional
career.
The college admissions data set that is available to Beachside High
School counselors includes records of past students who had applied to a
selective four-year college that has three academic units: School of Arts and
Letters, School of Business and Economics, and School of Mathematics and
Sciences. The data set is used to develop classification models based on a
naïve Bayes algorithm to predict whether or not the 14 high-achieving
students at Beachside High will be admitted to any of the three academic
units at the college.
The 14 current students are the top students in their graduating class.
However, because the admission process can be highly competitive, some of
these students might not get admitted to the college they wish to attend.
Moreover, different academic programs may have different admission criteria.
As a result, a naïve Bayes classification model is developed for each of the
three academic schools based on the following variables: parents’ education,
high school GPAs, SAT scores, and the male, white, and
Asian indicators. Because the naïve Bayes algorithm
page 406
requires that all predictor variables are categorical, GPAs and SAT scores are
converted into binary values where GPAs that are at least 3.50 are denoted
as 1, 0 otherwise, and SAT scores that are at least 1,200 are denoted as 1, 0
otherwise. A summary of the students’ demographic and academic
information is as follows.
My Life Graphic/Shutterstock

Of the 14 students, eight of them are female. Four

students are of Asian descent, and six are nonwhite.
The average high school GPA and SAT score of the 14
students are 3.64 and 1,261, respectively.
Nine students have a current GPA of 3.50 or higher, and
10 students scored at least 1,200 on the SAT exam.
All but three students have at least one parent who
completed a four-year college degree.
Even though the current students have not decided which academic
program they want to pursue, most of them express an interest in the School
of Arts and Letters. The data set has 6,964 records of past applicants to this
program, which are partitioned into training, validation, and test data sets.
Based on the test data set, the naïve Bayes model for the School of Arts and
Letters has an overall accuracy rate of 75.81%. The specificity and sensitivity
rates are 84.65% and 48.83%, respectively.
A summary of performance measures of the classification models for the
three academic schools is shown in Table 9.10. The naïve Bayes model
predicts that only four of our 14 top students will be admitted to the School of
Arts and Letters. This program appears to be the most selective of the three
academic units. The scoring results of the 14 students for each of the three
academic schools are presented in Table 9.11.

TABLE 9.10 Performance Measures of Naïve Bayes Classifiers

TABLE 9.11 Prediction Results for the 14 Students

The School of Business and Economics is also a popular choice among

the current students. The data set has 4,103 records of past applicants to this
program, which are partitioned into training, validation, and test data sets.
The accuracy, specificity, and sensitivity rates based on the test data are
77.44%, 83.20%, and 67.86%, respectively. The model predicts that six out of
the 14 students will be admitted into this program.
A similar naïve Bayes classifier is developed based on the 6,272 records
of past applicants to the School of Mathematics and Sciences. As presented
in Table 9.10, the accuracy, specificity, and sensitivity rates based on the test
data are 76.24%, 79.81%, and 69.19%, respectively. Based on the scoring
results, six out of the 14 students are likely to be admitted into this program.
These are the same six students that the previous model classifies as likely to
be admitted into the School of Business and Economics.
 Based on the overall accuracy rate, the naïve Bayes classifiers perform
reasonably well. The lift ratios and the decile-wise lift chart also indicate that
the naïve Bayes classifiers are more effective than a baseline random model.
As shown in Table 9.10, the lift values of the first decile of the three models
are above 2.0, and the AUC values are around 80%. However, compared to
other performance measures, the sensitivity rate of the models is relatively
low, especially for the School of Arts and Letters. This may be because the
schools also use qualitative information that is not captured in the database,
making it difficult to correctly classify all of the past applicants who were
admitted to the college (i.e., identifying true positive cases). The qualitative
factors that are relevant to the admission process include letters of
recommendations, written essays, and, for the School of Arts and Letters, a
student’s artwork.

page 407
Table 9.11 presents the scoring results of the 14 current students. In
general, the only students who are likely to be admitted into any of the three
academic programs are those who maintain a GPA of at least 3.50 and score
1,200 or above on the SAT exam. Out of the 14 students, only six of them
meet both criteria. A high GPA or a high SAT score alone is not likely to result
in an acceptance to the college.
These data-driven results confirm the anecdotal intuition that a GPA of
3.50 and an SAT score of 1,200 are the minimum thresholds that students at
Beachside High need to achieve in order to be admitted into a more selective
college or university. Moreover, the School of Arts and Letters appears to be
more selective than the other two schools. Only four of the six students with a
GPA above 3.50 and an SAT score above 1,200 are likely to get accepted
into this program. With this information, it is advised that some of the 14
students wishing to attend the School of Arts and Letters apply to additional
colleges and universities with a similar degree program as a back-up plan.

Suggested Case Studies

Supervised data mining techniques can be applied to many situations and different
data sets. The following are some suggestions based on the Big Data sets that are
provided with this text.

Report 9.1 FILE Longitudinal_Survey. Subset the data to include only those
individuals who lived in an urban area. Predict whether or not an individual’s
marriage will end up in a divorce, a separation, or a remarriage using predictor
variables such as sex, parents’ education, height, weight, number of years of
education, self-esteem scale, and whether the person is outgoing as a kid and/or
adult. Note: You may need to remove observations with missing values prior to the
analysis.
Report 9.2 FILE Longitudinal_Survey. Subset the data to include only those
individuals who lived in a non-urban area. Predict whether or not an individual will
have full-time employment (i.e., work for 52 weeks) using predictor variables such as
parents’ education, race, self-esteem score, and whether the person is outgoing as a
kid. Note: You may need to remove observations with missing values prior to the
analysis.

Report 9.3 FILE TechSales_Reps. Subset the data set to include only college-
educated sales professionals in the software product group. Predict whether the
sales professional will receive a high (9 or 10) net promoter score
(NPS) or not using predictor variables such as age, sex, tenure with
page 408
the company, number of professional certificates acquired, annual evaluation score,
and personality type. Note: You may need to perform data transformation in order to
meet the requirements of the analytics technique.

Report 9.4 FILE College_Admissions. Subset the data to include only one of the
colleges. Predict whether an admitted applicant will eventually decide to enroll at the
college using predictor variables such as gender, race, high school GPA, SAT/ACT
score, and parents’ education level.

Report 9.5 FILE Car_Crash. Subset the data to include only the accidents that
occured in one city or during one month. Predict whether or not a car crash will result
in fatality or severe injuries by using predictor variables such as the weather
condition, amount of daylight, and whether or not the accident takes place on a
highway.

page 409
 page 410

10 Supervised Data Mining:

Decision Trees

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 10.1  Apply classification trees to classify new
records.
LO 10.2  Apply regression trees to predict new
records.
LO 10.3  Apply ensemble tree models to classify
new records.

n the last chapter, we examined two widely used supervised data

I mining techniques: the k-nearest neighbors (KNN) method and the
naïve Bayes method. Recall that the KNN method is a data-driven
technique that classifies a new case by identifying the k-nearest
neighbors of the new case based on numerical predictor variables.
The naïve Bayes method, another classification technique, computes
the posterior probability that a new case will belong to a target class
based on its prior probability and the conditional probabilities of
categorical predictor variables.
In this chapter, we focus on another popular supervised data
mining technique called decision trees. Decision trees can be
constructed using both numerical and categorical predictor variables
and are used for both classification and prediction problems. We
start with a detailed discussion of the classification and regression
trees (CART) method, which is considered one of the most
transparent and easily interpretable data mining techniques. At the
end of the chapter, we also discuss a methodology called ensemble
trees, where multiple decision trees are combined to improve the
predictive performance of the model and reduce overfitting.

page 411

©Andy Dean Photography/Shutterstock

INTRODUCTORY CASE
Home Equity Line of Credit from Sunnyville
Bank
Established in 1990, Sunnyville Bank is a local community
bank serving businesses, individuals, and nonprofit
organizations throughout southern California. The bank is
committed to offering personalized services and a wide range
of banking products and services that match those of larger
financial institutions. For the last five years, Sunnyville Bank
has been named “Best Local Bank” by residents of southern
California.
Hayden Sellar, the retail banking manager of Sunnyville
Bank, was taken by surprise when he heard of a recent decline
in the bank’s home equity line of credit (HELOC) business.
Hayden does not believe that the decline is due to the
competitiveness of the bank’s interest rates. After all,
customers choose to bank with Sunnyville for the superior
service, not just for low prices. Hayden’s intuition suggest that
Sunnyville Bank has not been targeting the right customers to
market its HELOC products.
As an experiment, Hayden compiles data on 500 past
customers to whom Sunnyville Bank marketed its HELOC
products. The data include the age, sex, income, and whether
or not the customer responded to the HELOC offer (HELOC
equals 1 if responded, 0 otherwise). Table 10.1a shows a
portion of this data, labeled the HELOC_Data worksheet.
Hayden also collects the age, sex and income from 20 new
bank customers. Table 10.1b shows a portion of these data,
labeled the HELOC_Score worksheet. Hayden plans to use the
predictive model generated from historical data to predict
whether these new customers will respond to a HELOC offer.
Both worksheets are located in the HELOC data file.
 FILE
HELOC

TABLE 10.1 Bank Customers and Their Response to a HELOC Offer

a. The HELOC_Data worksheet

Age Sex Income HELOC

30 Female 101000 0

25 Male 86000 0

⋮ ⋮ ⋮ ⋮

47 Male 33000 1

b. The HELOC_Score worksheet

Age Sex Income

25 Female 45000

23 Male 22000

⋮ ⋮ ⋮

51 Male 43000

Hayden would like to use the data to accomplish the following

tasks.
1. Develop a classification tree for predicting whether or not a
bank customer will respond to a HELOC offer.
2. Identify which of the 20 new customers are likely to respond to
a HELOC offer.
A synopsis of this case is provided in Section 10.2.

page 412

10.1 INTRODUCTION TO
CLASSIFICATION AND
REGRESSION TREES (CART)
Decision trees are a popular supervised data mining technique with
a wide range of applications. Their appeal lies in the fact that the
output is displayed as one or more upside-down trees that are easy
to interpret. In general, a decision tree represents a set of If-Then
rules, whereby answers to these statements will eventually lead to a
solution to the application at hand.
Imagine a business decision such as the one in the introductory
case where a bank manager tries to decide which customers are
likely to respond to the bank’s home equity line of credit (HELOC)
offer. Figure 10.1 shows a hypothetical decision tree applied to the
HELOC example. The top node of the decision tree is called the root
node. The root node is the first variable to which a split value is
applied. The left branch from the root node represents the cases
whose variable values are less than (<) the split value, and the right
branch represents the cases whose variable values are greater than
or equal to (≥) the split value. Figure 10.1 shows that the root node is
the Age variable. Customers who are younger than 50 years old are
placed in the left branch, and customers who are at least 50 years
old are placed in the right branch. These branches often lead to
interior nodes where more decision rules are applied.
FIGURE 10.1 A simplified decision tree

In this simplified example, the Income variable is an interior node.

The split value for a variable can be different if it originates from a
different branch. Here, the split value for Income is $100,000 for
those customers younger than 50 years old, whereas it is $50,000
for those customers at least 50 years old. The bottom nodes of the
decision tree are called leaf nodes or terminal nodes, and this is
where the classification or prediction outcomes are given. In the
HELOC example, the leaf node indicates whether or not the
customer will respond to the HELOC offer.
The following four rules can be derived from Figure 10.1:
If Age < 50 and Income < $100,000, then the customer will not
respond.
If Age < 50 and Income ≥ $100,000, then the customer will
respond.
If Age ≥ 50 and Income < $50,000, then the customer will not
respond.
If Age ≥ 50 and Income ≥ $50,000, then the customer will respond.
The bank manager can then apply these rules to future customers
when marketing the bank’s HELOC products.
Classification and Regression Trees page 413
(CART)
As the HELOC example shows, a decision tree sorts or splits the
cases based on the values of predictor variables (such as Age < 50
versus Age ≥ 50) into subsets of cases. At each node, potential
customers are sorted or split into groups depending on their age or
income. This process is repeated at each level of the decision tree
until each subset contains only the cases with similar values for the
target variable. The process is recursive because every split is
dependent on the previous splits. In Figure 10.1, for example, the
split on the Income variable is dependent on the previous split on the
Age variable. Each of the final four subsets (the green rectangular
boxes) contains cases with the same value of the target variable,
that is, whether or not the customer will respond. A leaf node such
as these four subsets is called a pure subset where each leaf node
contains cases with the same value of the target variable and,
therefore, there is no need to further split it.
The decision tree in Figure 10.1 is kept simple intentionally to
help illustrate the relevant concepts. It is important to understand
that for real-world, large, and complex data sets, it will require many
interior nodes to reach pure subsets. Sometimes, it may not even be
possible to reach purity in the case of identical observations with
different values of the target variable.

PURE SUBSETS
Pure subsets consist of leaf nodes that contain cases with the
same value of the target variable. There is no need to further
split pure subsets.

This successive split or partitioning of the cases into pure subsets

is the primary building block of the classification and regression
trees (CART) algorithm. Decision trees produced by the CART
algorithm are binary, meaning that there are two branches for each
decision node. When the target variable is categorical, the CART
algorithm produces a classification tree to predict the class
memberships of new cases. When the target variable is numerical,
the CART algorithm produces a regression tree to estimate the
numeric values of the target variable of new cases.

CLASSIFICATION AND REGRESSION TREES (CART)

The CART algorithm successively splits the cases into
increasingly homogeneous subsets based on the values of the
predictor variables.
When the target variable is categorical, the CART algorithm
produces a classification tree to predict the class
memberships of future cases.
When the target variable is numerical, the CART algorithm
produces a regression tree to predict the value of the outcome
variable of future cases.

Like other supervised data mining methods discussed in Chapter

9, a decision tree is usually built using partitioned data sets. It is
common practice to partition data into three sets: training, validation,
and test data sets. The CART algorithm will first grow the tree using
the training data set by selecting the optimal splits of all predictor
variables and all possible split values until no further splits can be
made. This is usually achieved when all the leaf nodes only contain
cases that belong to the same class for classification trees or have
the same outcome value for regression trees. In other words, the leaf
nodes are all pure subsets. The decision tree produced in this step is
called a full tree, also called a full-grown tree or maximal tree, as it
represents the maximum number of splits the CART algorithm can
make to identify pure subsets.
A full tree, however, tends to overfit the data, resulting in low
predictive performance when applied to new cases. This is
especially true for the decision nodes and branches page 414
close to the end of the tree as these splits are based
on a very small number of cases and tend to fit the noise rather than
the underlying patterns of the data. A validation, and sometimes
even a test, data set is used to assess the performance of a decision
tree and the possible issue of overfitting.
In CART, the validation data set is used to optimize the
complexity of the tree by “pruning” the full tree to a simpler tree that
generalizes better to new data. Pruning is the process of removing
branches of the decision tree to improve its ability to predict new
cases more accurately. The pruning process results in a set of
candidate decision trees with different levels of complexity.
Performance measures such as the misclassification rate for the
classification tree or the root mean square error (RMSE) for the
regression tree from the validation data set are then compared to
identify the tree that produces the lowest error rate, called the
minimum error tree. The minimum error tree is the smallest (least
complex) tree that produces the smallest classification or prediction
error on the validation data set among all the candidate trees.
Alternatively, the best-pruned tree, which is the smallest (least
complex) tree with a validation error within one standard error of the
minimum error tree, can be used as the final decision tree. Because
the validation data set is a sample from the original data, relaxing the
tree selection criterion by one standard error accounts for the
sampling error. Finally, an independent assessment of the predictive
performance of the tree is sometimes conducted by applying the final
tree to the test data set.
While the underlying concept remains the same, different
software packages tend to perform data mining techniques
differently. As illustrated in the KNN example in Chapter 9, Analytic
Solver uses the validation data set to determine the optimal model
complexity, whereas R uses the cross-validation technique. Similarly
for the decision tree method, while Analytic Solver uses three-way
partitioning (i.e., train, validation, and test data) for building and
pruning decision trees, R implements a k-fold cross-validation
process for pruning decision trees and therefore uses two-way
partitioning (i.e., only train and validation data). For a detailed
disucssion on the cross-validation method, refer to section 7.4 of this
text.

VARIOUS DECISION TREES

The full tree is a fully grown decision tree that represents the
maximum number of splits the CART algorithm is willing to
make in order to identify pure subsets. As the full tree tends to
overfit and does not generalize well to new data, the full tree is
often pruned by removing the weaker branches to create the
minimum error tree or the best-pruned tree.
The minimum error tree is the least complex tree with the
smallest validation error.
The best-pruned tree is the least complex tree with a
validation error within one standard error of the minimum error
tree.

The decision tree methodology has a number of characteristics

that make it highly popular. First, the technique is versatile and can
deal with both categorical and numerical target variables. Second,
the tree diagram and the If-Then rules are easy to understand,
interpret, and implement in real-world situations even for a non-
technical audience. Third, the selection of the most discriminatory
predictor variables is automatic through the recursive partitioning
and pruning processes. The most important predictors appear at the
top of the tree diagram. The variable importance information can
help refine future data collection efforts. Fourth, because the
methodology is data-driven, it does not assume a particular type of
relationship between the predictor and the target variables as
assumed in linear and logistic regression models. Moreover, the
methodology can handle incomplete data and is not affected by
outliers, and, therefore, requires relatively less effort for data
preparation.

page 415

At the same time, however, it does require a relatively large

training data set to build a reasonably accurate model, and the
computational cost increases exponentially as the number of
predictor variables and observations increase. Furthermore, due to
the recursive partitioning process used in the CART algorithm, small
changes to the training data set will result in drastically different
trees. Later in this chapter, we will discuss ensemble trees, which
are designed to improve model stability.

10.2 CLASSIFICATION TREES

LO 10.1
Apply classification trees to classify new records.

In this section, we focus on applications of decision trees where the

outcome variable is binary. The decision trees generated in these
applications are called classification trees. As mentioned earlier, the
CART algorithm uses a recursive partitioning process to split the
cases into increasingly pure subsets to build the tree. In a
classification problem with a binary target variable, a subset is
considered “pure” when it contains cases that all belong to the same
class (100% and 0% split between classes). A subset with the
highest degree of impurity is when half of the cases belong to one
class and the other half belong to the other (50% and 50% split
between classes). Impurity is therefore measured by how well the
two classes are separated. In order to build a decision tree, we need
to understand the possible split points on each predictor variable and
how the CART algorithm determines the optimal splits.
Identifying Possible Split Points
Consider the HELOC example from the introductory case. Recall
that the bank manager has collected data from 500 past customers
that include the customer’s age, sex, income, and whether or not the
customer responded to the HELOC offer (HELOC equals 1 if
responded, 0 otherwise). Table 10.2 shows a portion of 20 randomly
selected cases from this data set that are labeled HELOC_Data_20.
These 20 cases as a training data set are used only for illustration to
develop a classification tree with two predictor variables: Age and
Income.
FILE
HELOC_Data_20

TABLE 10.2 The HELOC_Data_20 worksheet

Age Income HELOC

44 99000 0

63 158000 1

⋮ ⋮ ⋮

40 97000 0

The CART algorithm uses the midpoints between consecutive cases

of the predictor variable as possible split points. For example, the
ages of the 20 cases can be sorted in ascending order into the
following data array:
(Note that some values have multiple occurrences, but for
determining splitting points, each value only needs to appear once in
the array.)
Therefore, the first possible split point is calculated as (23 + 24)/
2 = 23.5, the next possible split point is calculated as (24 + 25)/ 2 =
24.5, and the remaining possible split points are found in the same
way. The following data array shows all possible split points for Age:

page 416

EXAMPLE 10.1
What are the possible split points on the Income variable for
the 20 cases in Table 10.2?

SOLUTION: First, we sort the values for Income in ascending

order to produce the following data array:

Next, we find the midpoints between consecutive cases. The

first split value is found as (4000 + 11000)/ 2 = 7500, the next
possible split point is calculated as (11000 + 14000)/ 2 =
12500. The following data array shows all possible split points
for Income:
 After identifying all the possible split points on the two variables,
the next step is to find the optimal partition that minimizes impurity in
the two resulting subsets. Recall that the impurity of a node in a
decision tree is measured by how well the classes are separated
from one another. There are several different ways to measure
impurity, including the Gini impurity index, entropy index, and Chi-
Square test; in this chapter, we use the Gini impurity index to
illustrate how to identify the optimal splits for producing a
classification tree.

The Gini Impurity Index

The Gini impurity index, or simply the Gini index, measures the
degree of impurity of a set of cases in a multiclass classification
context. Using m as the number of classes of the target variable, the
Gini index is computed as

where

For a binary classification problem (m = 2), the Gini index of a set

of cases ranges from 0 (highest purity) to 0.5 (highest impurity). For
example, if all the cases in the set belong to Class 0 and none
belongs to Class 1, then the Gini index is 1 − (12 + 02) = 0. On the
other hand, if half of the cases belong to Class 0 and the other half
of the cases belong to Class 1, then the Gini index is 1 − (0.52 +
0.52) = 0.5. As mentioned earlier, we have the highest degree of
impurity when half of the cases belong to one class and the other
half belong to the other. With any other split, the Gini index will be
less than 0.5. For example, if 90% of the cases belong to Class 0
and 10% of the cases belong to Class 1, then the Gini index is 1 −
(0.902 + 0.102) = 0.18, implying that it is relatively pure as its value is
closer to 0 than to 0.5. During the recursive partitioning process, the
CART algorithm determines the optimal split at each step by
comparing the Gini indexes for all possible splits across all variables.
The optimal split is the one that produces the smallest Gini index. In
Examples 10.2, 10.3, and 10.4, we illustrate the steps in the
recursive partitioning process to determine the optimal splits using
the Gini index.

page 417

EXAMPLE 10.2
Let’s return to the HELOC example with the 20 randomly
selected customers. Figure 10.2 shows a plot of Income
against Age for these 20 cases. The 15 blue dots represent
customers who did not respond to the HELOC offer (Class 0),
whereas the five red dots represent customers who responded
to the HELOC offer (Class 1).

FIGURE 10.2 Plot of Income against Age for n = 20

 Using the information in Figure 10.2, find the overall Gini index.

SOLUTION: Among the 20 cases, five cases belong to Class 1

while the rest of the cases belong to Class 0. The Gini index for
the 20 cases can be computed as follows:

This group of 20 cases is an impure subset as 15 cases belong

to Class 0 and five cases belong to Class 1 (75% and 25%
split).

Example 10.3 demonstrates the next step in the recursive

partitioning process that reduces impurity to construct a classification
tree.

EXAMPLE 10.3
We continue with the HELOC example as applied to the 20
randomly selected cases from Table 10.2. Earlier, we
calculated the possible split points for the Age variable. One of
the possible split points is when Age = 49. Figure 10.3 shows a
black vertical line at Age = 49. The subset on the left includes
all the cases whose ages are less than 49, and the subset on
the right contains all the cases whose ages are greater than or
equal to 49. Use Figure 10.3 to compute the Gini index given
an Age split of 49.

FIGURE 10.3 Split at Age = 49

SOLUTION: For the 16 cases in the left subset (Age < 49), two
belong to Class 1 and 14 belong to Class 0. For the four cases
in the right subset (Age ≥ 49), one belongs to Class 1 and three
belong to Class 0. Therefore, the Gini indexes for the two
subsets are:

page 418
 To compute the overall Gini index for the split, we use the
weighted combination of the Gini indexes using the percentage
of cases in each partition as the weight:

Partitioning the cases using an Age split value of 49 has

reduced the Gini index from 0.375 to 0.25. The best split on a
predictor is chosen by comparing the Gini indexes across all
possible splits for the predictor. In our example, if you compute
the Gini indexes for all possible splits for Age, you will find that
the split value of 49 produces the smallest Gini index;
therefore, this split is chosen as the best split for Age.

Because we have not produced a pure subset from the 20 cases in

the HELOC example, we will continue the recursive partitioning
process using the Income variable. Consider Example 10.4.

EXAMPLE 10.4
Recall from Example 10.1 that one of the possible split points
on the Income variable is when Income = 64,500. Figure 10.4
shows a horizontal black line at Income = 64,500. This split
creates top and bottom partitions. Compute the overall Gini
index given an Income split of 64,500.

FIGURE 10.4 Split at Income = $64,500

 SOLUTION: For the 11 cases in the top subset (Income ≥ 64,500),
five belong to Class 1 and six belong to Class 0. For the nine
cases in the bottom subset (Income < 64,500), none belongs to
Class 1 and nine belong to Class 0. Therefore, the Gini indexes
for the two subsets are

The Gini index for this split is computed as

page 419

It can be shown that the split with Income = 64,500 produces the
smallest Gini index across all possible Income splits. Note that the
splits at Age = 49 and Income = 64,500 (Examples 10.3 and 10.4)
produce a Gini index that is better than the original Gini index
computed in Example 10.2. Next, we will start constructing a
decision tree by considering the Gini index values from the two Age
and Income splits.

Building the Decision Tree

In the HELOC example with the Age and Income predictor variables,
we compare the Gini indexes of the best splits for the two predictors.
The split using Age = 49 produces the most reduction in impurity
across all possible splits for the two predictors because its Gini index
is the smallest (0.25 < 0.2727). Therefore, Age is used as the root
node on the decision tree with a split of 49. Figure 10.5 shows this
result.

FIGURE 10.5 Root node partition

Next, we find the probability that a customer responds to the

HELOC offer given the customer’s age. In general, this probability is
found by finding the proportion of cases in a partition that belongs to
the target class. In the HELOC example, the probability that a
customer is younger than 49 years old and responds to the HELOC
offer (the target class or Class 1) is 2/16 = 0.125 (see Figure 10.3).
Similarly, the probability that a customer is 49 years old or older and
responds to the HELOC offer is 3/4 = 0.75.
To grow the decision tree, the process is repeated for all the
partitions until no splits can further reduce impurity. In other words,
the recursive process terminates when it can no longer produce a
lower Gini index or when all the partitions become pure subsets (i.e.,
the Gini index = 0). Figure 10.6 shows the result of the entire
recursive partitioning process. Note that Age is split at 23.6 and 49
and Income is split at 39,500; 59,000; 102,000; and 107,000. In this
case, all partitions are pure subsets.

FIGURE 10.6 Recursive partitioning results

page 420

The recursive partitioning process produces the full-grown tree in

Figure 10.7. Within the older age group (49 years and older), those
who earn at least $39,500 in annual income are likely to respond to
the HELOC offer. Among the younger age groups, individuals who
are younger than 23.5 years with an annual income of at least
$59,000 are likely to respond to the offer. Individuals who are
between 24 and 49 years old and earn an annual income between
$102,000 and $107,000 are also likely to respond to the offer.

FIGURE 10.7 Full-grown classification tree

 page 421

While the classification tree produced in this example is relatively

simple due to the fact that we use a small data set with only 20
observations and two predictor variables, a large data set with many
predictor variables will likely generate a very complex tree with many
levels of decision nodes. As the number of partitions increases, the
misclassification rate from the training data set will decrease and
eventually reach 0 when all partitions include only cases that belong
to the same class. However, for a new data set (e.g., validation
data), the misclassification rate will likely decrease initially but
increase after a certain point due to overfitting. Therefore, pruning is
an important step to optimize the complexity of the final decision
tree.

Pruning the Decision Tree

Table 10.3 shows five additional randomly selected observations
from the HELOC_Data worksheet that we use as validation data to
prune the fully grown classification tree in Figure 10.7.

TABLE 10.3 Validation Data

Age Sex Income HELOC

50 Female 35000 0

23 Male 60000 0

56 Male 87000 1

42 Female 90000 0

62 Male 71000 1

A simple approach to prune a classification tree is to reduce the

misclassification rate in the validation data set by replacing a branch
of the tree with a leaf node; soon we will show how we build and
prune a classification tree with Analytic Solver and R. For example,
Figure 10.8 shows a pruned tree where what used to be an interior
node with a split point where Age = 23.5 is now replaced with a leaf
node. Because the majority of the cases that reach this node belong
to Class 0, the leaf node predicts that customers in this subset will
not respond to the HELOC offer. This simpler tree yields the best
classification performance on the validation data set. All five
observations in the validation data are correctly classified. If we
score the validation data using the full-grown tree in Figure 10.7, we
would get an accuracy rate of 80% as the second observation is
misclassified.

FIGURE 10.8 Resulting decision tree

Unlike the full-grown tree, the tree obtained from the pruning
process is much simpler. The following three rules can be derived
from the optimized tree:
If Age < 49, then the customer will not respond.
If Age ≥ 49 and Income < $39,500, then the customer will not
respond.
If Age ≥ 49 and Income ≥ $39,500, then the customer will respond.
The optimized tree shows that individuals whose age is 49 or older
and income is $39,500 or higher are likely to respond to a HELOC
offer from the bank. As this example demonstrates, pruning is
effective in determining the optimal complexity of a decision tree and
reducing the chance of overfitting.

page 422
Using Analytic Solver and R to Develop a
Classification Tree
We now demonstrate how to use Analytic Solver and R to build
classification trees, using the complete data set. Consider Example
10.5.
FILE
HELOC

EXAMPLE 10.5
Recall from the introductory case that Hayden Sellar, the bank
manager of Sunnyville Bank, wants to use historical data from
500 bank customers to develop a classification model for
determining whether or not new bank customers will respond to
a HELOC offer. He also thinks that Sunnyville Bank is not
targeting the right customers when marketing its HELOC
products and hopes that the classification model will offer
actionable insights for his company. Hayden plans to assess
the performance of the model and then classify 20 new bank
customers as likely or unlikely to respond to a HELOC offer.
Use Analytic Solver and R to build the preferred classification
tree, and then use this tree to score the new cases.
SOLUTION:
Using Analytic Solver

a. Open the HELOC_Data worksheet of the HELOC data file.

b. Choose Data Mining > Partition (under the Data Mining
group) > Standard Partition.
c. In the Standard Data Partition dialog box, we select the data
range $A$1:$D$501 and move Age, Sex, Income, and
HELOC to the Selected Variables box. Select Pick up rows
randomly and check the Set seed checkbox to use the default
random seed of 12345. Choose Specify percentages to
 allocate 50%, 30%, and 20% of the data to training, validation,
and test sets, respectively. Click OK.
d. Analytic Solver will create a new worksheet called
STDPartition with the partitioned data. With the STDPartition
worksheet active, choose Data Mining > Classify >
Classification Tree to open the Classification Tree dialog box.
In the Data tab, select and move the predictor variables Age
and Income to the Selected Variables box and Sex to the
Categorical Variables box. Select HELOC as the Output
Variable. Accept other defaults. Click Next.
e. In the Parameters tab, check the Prune (Using Validation Set)
checkbox and click on the Tree for Scoring button. Select Best
Pruned. Click Done. This selection tells Analytic Solver to
score cases using the best-pruned tree to evaluate the
performance of the classification tree and score new cases.
Check the Show Feature Importance checkbox and click on
the Trees to Display button. Check the boxes next to Fully
Grown, Best Pruned, and Minimum Error. Click Done. This
tells Analytic Solver to display the full, best-pruned, and
minimum error trees in the results. Accept other defaults. Click
Next.
f. In the Scoring tab, check the Summary Report boxes for
Score Training Data, Score Validation Data, and Score Test
Data. Also check the Lift Charts box for Score Test Data.
Check the In Worksheet check box under the Score New Data
section. Under the New Data (WS) tab, select HELOC_Score
in the Worksheet box. These 20 cases of new bank customers
in Data Range $A$1:$C$21 will be scored by the classification
tree algorithm. Make sure the First Row Contains Headers box
is checked and click the Match By Name button. Accept other
defaults. Click Finish.

page 423
 The CT_Output worksheet shows a prune log that reports
the error rate depending on the number of decision nodes.
Table 10.4 shows a portion of the prune log listing error rates
against decision nodes. The simplest tree with the minimum
validation error rate of 0.16 occurs at three decision nodes.

TABLE 10.4 Analytic Solver’s Prune Log

# Decision Nodes Error Rate

0 0.2267

1 0.2267

2 0.2267

3 0.1600

4 0.2000

5 0.2000

6 0.2000

Is there a simpler tree with a validation error rate that is

within one standard error of the minimum validation error rate?
To answer this question, we have to inspect the tree diagrams.
The CT_FullTree worksheet shows the full tree. The full tree is
quite complex with many decision nodes. Due to its size, the
full tree is not displayed here. The CT_BestTree and
CT_MinErrorTree worksheets show the best-pruned and
minimum error trees. When comparing these two tree
diagrams, we find the best-pruned and minimum error trees are
the same. In this example, therefore, no simpler tree with a
validation error rate that is within one standard error of the
minimum validation error rate can be found.
The best-pruned and minimum error tree is displayed in
Figure 10.9. The tree is relatively simple with three decision
nodes and seven total nodes if the terminal nodes are included.
The predictor variable in the root node is Sex. Because Sex is
a categorical variable, Analytic Solver uses the generic value
“Set” as the split value. To find out what it page 424
means, we place our mouse pointer over the
root node, and the following text would appear above the tree:

FIGURE 10.9 Analytic Solver’s best-pruned classification tree

Decision Node:
Go left if Sex is from set {Female}
Go right if Sex is not from set {Female}
 A number of If-Then rules can be derived from the
classification tree. For example, one rule is that if the customer
is a male who is younger than 32.5 years old, then he will not
respond to a HELOC offer. Another rule suggests that if a
customer is a male who is 32.5 years or older with an income
greater than $26,000, then he will respond to a HELOC offer.
Finally, and perhaps more importantly, the classification tree
indicates that female customers are not likely to respond to the
HELOC offer, confirming the bank manager’s intuition that the
current marketing campaign is not targeting all the right
customers.
Because the validation data were used for pruning, the
performance of the classification tree on new data should be
assessed using the test data set. To assess the performance of
the resulting classification tree, we turn to Testing:
Classification Summary in the CT_TestScore worksheet, a
portion of which is shown in Figure 10.10.

FIGURE 10.10 Analytic Solver’s summary measures

Source: Microsoft Excel

The performance measures show that the model has an

overall error rate of 23%, sensitivity of 0.5385, and specificity of
0.8514. Note that these performance measures are highly
sensitive to the cutoff value used. In this example, the default
cutoff value is 0.5. If we lower the cutoff value, more cases will
be classified into the target class, resulting in different values
for the performance measures. To evaluate the performance of
the classification tree using a different cutoff value, in Analytic
Solver, input the desired cutoff value in the Success Probability
Cutoff: box in the Data tab in step d. To evaluate model
performance independent of the cutoff value, we now examine
the cumulative lift chart, the decile-wise lift chart, and the ROC
curve.
Figure 10.11a shows the cumulative lift chart. The blue line
lies well above the red diagonal line, indicating that the
classification tree performs considerably better in terms of
predicting whether or not a potential customer will respond to
the HELOC offer, compared to the baseline model. The lift
chart shows that if we send a HELOC offer to the 50 bank
customers (half of the test data sample size) who have the
highest predicted probability of responding to the offer, we will
be able to reach 22 out of 26 (84.6%) individuals who
responded to a HELOC offer. Alternately, if we send the offer to
50 randomly chosen customers, we would reach only 13 of the
customers who responded to the offer.

FIGURE 10.11 Analytic Solver’s performance charts

 Figure 10.11b shows the decile-wise lift chart. The first 10%
of potential customers selected by the classification tree, for
example, include 1.92 times as many customers who would
respond to the HELOC offer, compared to a random selection
of 10% of potential customers.

page 425

Finally, the ROC curve (see Figure 10.11c) shows that the
classification tree performs much better than the baseline
 model in terms of sensitivity and specificity across all cutoff
values. The area under the curve (AUC) is 0.7710, which is
closer to the optimum classifier (AUC = 1) as compared to the
random classifier (AUC = 0.5).
The CT_NewScore worksheet shows the classification
predictions for the 20 records of new customers. The results
are summarized in Table 10.5. As shown in the page 426
Prediction: HELOC column, the first two
potential customers are classified as not likely to respond to a
HELOC offer, while the last customer is likely to respond. The
PostProb:0 and PostProb:1 columns provide the predicted
probabilities of the customer belonging to Class 0 and Class 1,
respectively.

TABLE 10.5 Analytic Solver’s Scoring Results

Using R
The most popular algorithms for building decision trees in R
can be found in the rpart package. As in the case of the KNN
method discussed in Chapter 9, while Analytic Solver uses
three-way partitioning (i.e., train, validation, and test data) for
building decision trees, R implements a k-fold cross-validation
process for pruning decision trees using two-way partitioning
(i.e., only train and validation data).
a. Import the data from the HELOC_Data worksheet of the
HELOC data file into a data frame (table) and label it myData.
The following instructions and results are based on R version
3.5.3. To replicate the results with newer versions of R, enter:
 >suppressWarnings(RNGversion(“3.5.3”))
b. Install and load the caret, gains, rpart, rpart.plot, and pROC
packages using the following commands. Enter:
>install.packages(“caret”, dependencies = c(“Depends”,
“Suggests”))
>install.packages(“gains”)
>install.packages(“rpart”)
>install.packages(“rpart.plot”)
>install.packages(“pROC”)
>library(caret)
>library(gains)
>library(rpart)
>library(rpart.plot)
>library(pROC)
c. For constructing a classification tree model, R requires that
the target variable, HELOC, be a factor variable, a categorical
data type. We use the as.factor command to convert the
HELOC variable into a categorical type. Enter:
>myData$HELOC <- as.factor(myData$HELOC)
d. We use the set.seed command to set the random seed to 1,
thus generating the same partitions as in this example. We
use the createDataPartition function to partition the data into
training (70%) and validation (30%). Enter:
>set.seed(1)
>myIndex <- createDataPartition(myData$HELOC, p=0.7, list
= FALSE)
>trainSet <- myData[myIndex,]
>validationSet <- myData[-myIndex,]
e. We use the rpart function to generate the default classification
tree, default_tree. Within the rpart function, we specify the
model structure, data source, and method. The method option
 is set to “class” for developing a classification tree. To view the
details about the default tree, use the summary function.
Because R uses the cross-validation method for pruning the
tree, to ensure consistency of the cross-validation results, we
use the set.seed function to set a random seed of 1. Enter:
>set.seed(1)
>default_tree <- rpart(HELOC ~ ., data=trainSet,
method=“class”)
>summary(default_tree)
f. To view the classification tree visually, use the prp function.
The type option is set equal to 1 so that all nodes except the
leaf nodes are labeled in the tree diagram. The extra option is
set equal to 1 so that the number of observations that fall into
each node are displayed. The under option is set equal to
TRUE to put the number of cases under each decision node in
the diagram. Enter:
>prp(default_tree, type=1, extra=1, under = TRUE)
page 427
Figure 10.12 shows the default classification tree.
The first decision node is on the Sex variable, followed by Age
and Income splits. Note that R presents decision trees in a
slightly different format as Analytic Solver does. The root node
provides information about how to interpret the tree. For
example, in Figure 10.12, the root node shows that if Sex is
“Female” then go to the left branch, otherwise go to the right
branch. The subsequent decision nodes follow the same
format. For example, the second decision node suggests that
if Age is less than 25 then go to the left branch, otherwise go
to the right branch.

FIGURE 10.12 R’s default classification tree

How are the number of splits determined in the default
classification tree? The rpart function uses the complexity
parameter (cp) to determine when to stop growing the tree. If
the cost of adding another split to the tree exceeds the value
of cp, then the tree growth will not continue. The default cp
value for the rpart function is 0.01. However, in most cases, it
is very difficult to know which cp value will produce the best-
performing tree beforehand. Therefore, a common practice is
to grow the full tree and then prune it to a less-complex tree
based on the classification errors produced by a built-in cross-
validation process of the rpart function. By identifying the
value of cp associated with the smallest cross-validated
classification error, we can create the minimum error tree.
Alternatively, we can produce the best-pruned tree, which is
the smallest tree with an error rate that is within one standard
 error of the minimum error rate. Next, we will demonstrate the
pruning process to optimize the complexity of the tree.
g. We first grow the full tree by using the rpart function. We set
the options cp equal to 0, minsplit equal to 2, and minbucket
equal to 1. The minsplit option specifies the minimum number
of observations in the parent node that can be split further,
and the minbucket option specifies the minimum number of
observations that are allowed in the leaf node. These settings
ensure that the largest possible tree will be produced. We plot
the full tree using the prp function. Again, to page 428
ensure consistency of the cross-validation
results, we specify a random seed of 1 using the set.seed
function. Enter:
>set.seed(1)
>full_tree <- rpart(HELOC ~ ., data= trainSet, method=“class”,
cp=0, minsplit=2, minbucket=1)
>prp(full_tree, type=1, extra=1, under = TRUE)
The full tree in this case is very complex, so it is not displayed
here.
h. To identify the value of cp that is associated with the smallest
cross-validated classification error, we use the printcp
function to display the complexity parameter table. Enter:
>printcp(full_tree)
Figure 10.13 displays the complexity parameter table for the
nine candidate trees with increasing complexity. The nsplit
column shows the number of splits for each tree. The number
of leaf nodes for each tree can be calculated using nsplit + 1.
For example, the last tree has 82 splits, so it has 83 leaf
nodes.

FIGURE 10.13 R’s complexity parameter table

 The rel error column shows the fraction of misclassified
cases for each tree relative to the fraction of misclassified
cases in the root node if all cases are classified into the
predominant class. As you can see, the last tree has a relative
error of 0 because it is the fully grown tree whose leaf nodes
only contain cases that belong to the same class; therefore,
there are no misclassified cases.
The xerror column shows the cross-validation errors
associated with each candidate tree. It is the recommended
measure for identifying the tree that can potentially perform
well on new data sets. As you can see from the xerror column,
the cross-validation errors decrease initially as the
classification tree becomes more complex and then increase
after a certain point. This is common and indicative of the
overfitting problems with complex tree models. The fourth tree,
with six splits, has the lowest cross-validation error (0.83516
relative to the cross-validation error of the root node);
therefore, it is the minimum error tree.
The xstd column can be used to identify the best-pruned
tree, which is the smallest tree with an error that is within one
 standard error of the minimum error tree. In page 429
this case, no simpler tree has a cross-
validation error that meets this criterion. In other words, no
simpler tree has a relative cross-validation error that is less
than 0.83516+0.084763 or 0.919923. Therefore, the best-
pruned tree and minimum error tree are the same tree.
However, in many cases, you may find that the best-pruned
and minimum error trees are two different trees.
Figure 10.14 displays a hypothetical complexity parameter
table from another classification tree. In this case, the third
tree, with five splits, has a relative cross-validation error
(0.87912) that is within one standard error of the minimum
error (0.81319 + 0.083945 = 0.897135); hence, the third tree,
with five splits, is the best-pruned tree. Next, we use the value
of cp associated with the third tree to produce our pruned tree.

FIGURE 10.14 A hypothetical complexity parameter table in

R

i. We use the prune function to create the pruned tree by using

the cp value associated with the fourth tree. Please note that
the cp values provided in R results are rounded to seven digits
after the decimal points. To ensure that we do not use a cp
value that is less than the actual cp value, we use a cp value
that is slightly larger than the cp value displayed in the table
but lower than the cp value for the next smaller tree. In this
case, we will use 0.0164836, a number that is slightly larger
than the cp number associated with the fourth tree
 (0.0164835). We display the pruned tree using the prp
function. Enter:
>pruned_tree <- prune(full_tree, cp=0.0164836)
>prp(pruned_tree, type=1, extra=1, under = TRUE)
Figure 10.15 displays the pruned tree, which has six splits and
seven leaf nodes. Note that this is a much simpler tree with
fewer branches, compared to the default tree in Figure 10.12.
Next, we will evaluate the performance of the pruned tree
using the validation data set.

FIGURE 10.15 R’s pruned tree

j. We predict the class memberships of the observations in the

validation data set using the predict function. We set the type
option equal to “class” so that the class membership is
produced instead of its probability. Enter:
 >predicted_class <- predict(pruned_tree, validationSet,
type=“class”)
k. The confusion matrix can be created by comparing the
predicted class memberships and actual class memberships
of the validation data set. We use the confusionMatrix
function to produce the confusion matrix and various
performance measures. The positive = “1” option specifies 1
as the target class (respond to a HELOC offer). Enter:
>confusionMatrix(predicted_class, valid.data$HELOC,
positive=“1”)
page 430
Figure 10.16 reports the confusion matrix and
relevant performance measures. The model has an overall
accuracy rate of 76.67%, sensitivity of 0.4872, and specificity
of 0.8649. These measures suggest that, while the overall
accuracy rate is relatively high, the model classifies a much
larger portion of the non-target class (customers who do not
respond to HELOC offers) cases than target page 431
class (customers who respond to HELOC offers)
cases correctly using the default cutoff rate of 0.5. As noted
earlier, these performance measures are highly sensitive to
the cutoff value. In this example, the default cutoff value of 0.5
is much higher than the proportion of target class cases in the
data set, which is 0.26. By lowering the cutoff value to be
close to the actual class distribution, we will be able to classify
more cases into the target class and improve the sensitivity
measure. This is illustrated in the following steps.

FIGURE 10.16 R’s summary measures

l. To evaluate the predictive performance of the classification
tree model using a different cutoff value in R, we first compute
the probability of each validation case belonging to the target
class instead of its class membership. In the predict function,
we set the type option equal to ‘prob’ to predict the probability
values. Enter:
>predicted_prob <- predict(pruned_tree, validationSet,
type=“prob”)
>head(predicted_prob)
Figure 10.17 shows the probabilities associated with the first
few cases. The first column lists the probabilities of the cases
belonging to Class 0, while the second column lists the
probabilities of the cases belonging to Class 1. To determine
the class memberships of cases using a cutoff value other
than the default value of 0.5 (for example, 0.26, in order to
accurately classify more Class 1 cases), we compare the
values in the second column to the new cutoff value.
 FIGURE 10.17 R’s predicted probabilities

m. To construct a confusion matrix using the new cutoff value of

0.26, we use the ifelse function to determine the class
memberships. We use the as.factor function to convert the
class membership to factor, which is the same data type as
the target variable, HELOC. Enter:
>confusionMatrix(as.factor(ifelse(predicted_prob[,2]>0.26, ‘1’,
‘0’)), validationSet$HELOC, positive = ‘1’)
The resulting confusion matrix (not shown here) provides the
performance measures of the pruned decision tree using the
cutoff value of 0.26. Please verify that in the new confusion
matrix, accuracy, sensitivity, and specificity are 0.7667,
0.6410, and 0.8108, respectively. The new cutoff value allows
the manager of the bank to identify more target class cases
correctly as signified by the higher sensitivity value.
To evaluate model performance independent of the cutoff
value, we now examine the cumulative lift chart, the decile-
wise lift chart, and the ROC curve.
n. We first convert the target variable (HELOC) to a numerical
data type as required by the gains package using the
as.numeric function. We generate the cumulative lift table
using the gains function. The gains function requires two
inputs: actual class memberships and predicted target class
probabilities. Because in Figure 10.17, ‘1’ is the target class,
we refer to column two for the predicted target class
probabilities. Enter:
>validationSet$HELOC <-
as.numeric(as.character(validationSet$HELOC))
 >gains_table <- gains(validationSet$HELOC,
predicted_prob[,2])
>gains_table
page 432
Figure 10.18 shows the cumulative gains table.
Because there are only six possible target class probabilities
in the pruned tree—0.73, 0.67, 0.27, 0.25, 0.07, and 0.00—
that can be assigned to each case, the gains function
generates only six groups instead of the default 10 groups,
one for each unique probability value. (Note: These values are
rounded to two digits after the decimal point. That is why you
will get the warning: Fewer distinct predicted values than
groups requested.)

FIGURE 10.18 R’s cumulative gains table

o. Because the R instructions for the cumulative lift chart, the

decile-wise lift chart, and the ROC curve have been discussed
in detail in Chapter 9, we group the commands here. Enter:
>plot(c(0,
gains_table$cume.pct.of.total*sum(validationSet$HELOC)) ~
c(0, gains_table$cume.obs), xlab = “# of cases”,
ylab=“Cumulative”, main=“Cumulative Lift Chart”, type =“l”)
>lines(c(0, sum(validationSet$HELOC)) ~ c(0,
dim(validationSet)[1]), col=“red”, lty=2)
>barplot(gains_table$mean.resp/mean(validationSet$HELOC)
, names.arg=gains_table$depth, xlab=“Percentile”, ylab=“Lift”,
ylim=c(0,3), main = “Decile- Wise Lift Chart”)
>roc_object<- roc(validataionSet$HELOC, predicted_prob[,2])
>plot.roc(roc_object)
>auc(roc_object)
Figure 10.19a displays the cumulative lift chart. The chart
shows that the classification model shows superior predictive
power compared to the default model. Sunnyville Bank can
reach a larger portion of potential customers who will actually
respond to the HELOC offer by targeting a smaller percentage
of the potential customers with the highest predicted
probability of responding to the offer. Figure 10.19b displays
the decile-wise lift chart. If Sunnyville Bank sends offers to the
top 14% of the potential customers with the highest predicted
probability of responding to a HELOC offer, it would be able to
capture twice as many individuals who actually respond to the
HELOC offer as compared to if 14% of the customers are
randomly selected. Figure 10.19c displays the ROC curve.
Again, this curve shows that the classification tree model
outperforms the baseline model in terms of sensitivity and
specificity across all cutoff values. The AUC value of 0.805
also reinforces this finding.

FIGURE 10.19 R’s performance charts

p. Finally, to score the 20 new cases, we import the data from
the HELOC_Score worksheet of the HELOC data file into a
data frame (table) called myScoreData and use the predict
function to produce the predicted class memberships and
probabilities for the new cases using our classification tree.
Enter:
>predicted_class_score <- predict(pruned_tree, myScoreData,
type=“class”)
>predicted_class_score
 >predicted_class_prob <- predict(pruned_tree, myScoreData,
type=“prob”)
>predicted_class_prob
page 433
The scoring results are summarized in Table
10.6. The predicted classes for the cases are similar to those
predicted by the classification tree constructed in Analytic
Solver although the probabilities are slightly different due to
the differences in constructing the trees.
TABLE 10.6 R’s Scoring Results

page 434

SYNOPSIS OF INTRODUCTORY CASE

Data-driven marketing decisions can help an organization better target potential
customers and at the same time reduce marketing costs. Hayden Sellar, retail
banking manager of Sunnyville Bank, is excited about using data mining tools to
help his team identify customers who are more likely to be interested in certain
banking products and services.
Experimenting with a relatively small sample of 500 bank customers who
previously received HELOC offers from the bank, Hayden and his team built a
classification tree to predict which new customers are likely to respond to HELOC
offers. The model offers a classification accuracy rate of about 77%, which is a
significant improvement over what Hayden and his team were able to achieve using
only past experience and intuition. With this data-driven model, Sunnyville Bank will
be able to focus on a small proportion of the customers who have high predicted
probabilities of responding to HELOC offers. For example, male customers in their
late 30s with high income are likely to respond to the offer.

Syda Productions/Shutterstock

A closer study of the tree diagram also reveals a rather troubling issue for
Hayden and his team. The decision tree predicts that female customers are not
likely to respond to the HELOC offer. At the same time, Hayden knows that female
customers account for a substantial portion of Sunnyville customers. Could the
alarmingly low percentage of female customers who are interested in HELOC
products be due to the unconscious bias in the bank’s marketing materials? To
answer this question, Hayden and his team decide to review the previous ads of
HELOC products. As they expected, all previous HELOC ads indeed seem to have a
bias toward male customers. This confirms Hayden’s intuition that Sunnyville’s
current marketing strategy is not targeting the right customers. As a result, he has
scheduled to meet and work with the bank’s marketing team to eliminate any
unconscious bias in its marketing materials and focus more on the female
customers.

EXERCISES 10.2

Note: and indicate that the exercise problems must be

solved using Analytic Solver and R, respectively. For Analytic

Solver, partition data sets into 50% training, 30% validation, and
20% test and use 12345 as the default random seed. For R,
partition data sets into 70% training and 30% validation. Use
the statement set.seed(1) to specify the random seed of 1 for
both data partitioning and cross-validation. All answers in R are
based on version 3.5.3. To replicate the results with newer
versions of R, execute the following line of code at the
beginning of the R session:
suppressWarnings(RNGversion(“3.5.3”)). Some data files have
two worksheets (for example, Exercise_10.7_Data and
Exercise_10.7_Score worksheets) for model development and
scoring new records. If the predictor variable values are in the
character format, then treat the predictor variable as a
categorical variable. Otherwise, treat the predictor variable as a
numerical variable.

Mechanics
1. FILE Exercise_10.1. The accompanying data set contains two predictor
variables, age and income, and one binary target variable, newspaper
subscription (subscribe), indicating whether or not the person subscribes to a
newspaper. A media company wants to create a decision tree for predicting
whether or not a person will subscribe to the newspaper.
a. List the possible split values for age in ascending order.
b. List the possible split values for income in ascending order.
2. FILE Exercise_10.2. The accompanying data set contains two predictor
variables, average annual number of sunny days (days) and average annual
precipitation (precipitation), and one numeric target variable, average annual crop
yield in bushels per acre (yield). An agricultural researcher wants to create a
decision tree for predicting the annual crop yield in bushels per acre for various
areas.
a. List the possible split values for days in ascending order.
b. List the possible split values for precipitation in ascending order.
3. After a data set was partitioned, the first partition contains 43 cases that belong to
Class 1 and 12 cases that belong to Class 0, and the second partition contains 24
cases that belong to Class 1 and 121 cases that belong to Class 0.
a. Compute the Gini impurity index for the root node.
b. Compute the Gini impurity index for partition 1.
c. Compute the Gini impurity index for partition 2.
d. Compute the Gini impurity index for the split.
4. After a data set was partitioned using the split value of 45.5 for age. The age <
45.5 partition contains 22 patients with a diabetes diagnosis and 178 patients
without a diabetes diagnosis, and the age ≥ 45.5 partition contains 48 patients
page 435
 with a diabetes diagnosis and 152 patients without a diabetes
diagnosis.
a. Compute the Gini impurity index for the root node.
b. Compute the Gini impurity index for the age < 45.5 partition.
c. Compute the Gini impurity index for the age ≥ 45.4 partition.
d. Compute the Gini impurity index for the split.
e. State the rules generated from this split.
5. FILE Exercise_10.5. Use the accompanying data set to answer the following
questions.
a. Which split value for age would best separate the newspaper subscribers from
nonsubscribers based on the Gini impurity index?
b. Which split value for income would best separate the newspaper subscribers
from nonsubscribers based on the Gini impurity index?
c. Between the best split values for age and income, which one should be chosen
as the first split for a decision tree that predicts whether someone is a
newspaper subscriber or not?
d. State the rules generated from this split.
6. The classification tree below relates type of wine (A, B, or C) to alcohol content,
flavonoids, malic acid, and magnesium. Classify each of the following wines of
unknown class.
a. Wine with alcohol = 13.3, flavanoids = 1.95, malic acid = 1.03, and magnesium
= 114
b. Wine with alcohol = 11.8, flavanoids = 2.44, malic acid = 0.97, and magnesium
= 103
c. Wine with alcohol = 12.9, flavanoids = 2.21, malic acid = 1.22, and magnesium
= 99
7. FILE Exercise_10.7. The accompanying data set in the Exercise_10.7_Data

worksheet contains four predictor variables (x1 to x4) and one binary target
variable (y). Select the best-pruned tree for scoring and display the full-grown,
best-pruned, and minimum error trees.
a. What is the minimum error in the prune log for the validation data? How many
decision nodes are associated with the minimum error?
b. Display the best-pruned tree. How many leaf nodes are in the best-pruned tree?
How many leaf nodes are in the minimum error tree?
c. What are the predictor variable and split value for the first split of the best-
pruned tree?
d. What are the classification accuracy rate, sensitivity, specificity, and precision of
the best-pruned tree on the test data?
e. Generate the decile-wise lift chart of the best-pruned tree on the test data. What
is the lift value of the leftmost bar of the decile-wise lift chart? What does this
value imply?
f. Generate the ROC curve of the best-pruned tree on the test data. What is the
area under the ROC curve (or AUC value)?
g. Score the new observations in the Exercise_10.7_Score worksheet using the
best-pruned tree. What is the predicted response value of the new
observations? What is the Class 1 probability of the first observation? Round
your answers to four decimal places.

8. FILE Exercise_10.8. The accompanying data set in the Exercise_10.8_Data

worksheet contains four predictor variables (x1 to x4) and one binary target
variable (y). Select the best-pruned tree for scoring and display the full-grown,
best-pruned, and minimum error trees.
a. What is the minimum error in the prune log for the validation data? How many
decision nodes are associated with the minimum error?
b. Display the best-pruned tree. How many leaf nodes are in the best-pruned tree?
How many leaf nodes are in the minimum error tree?
c. What are the predictor variable and split value for the first split of the best-
pruned tree?
d. What are the classification accuracy rate, sensitivity, specificity, and precision of
the best-pruned tree on the test data?
e. Generate the lift chart of the best-pruned tree on the test data. Does the entire
lift curve lie above the baseline?
f. Generate the ROC curve of the best-pruned tree on the test data. What is the
area under the ROC curve (or AUC value)?
g. Score the new observations in the Exercise_10.8_Score worksheet using the
best-pruned tree. What is the predicted response value of the new
 observations? What is the Class 1 probability of the first observation? Round
your answers to four decimal places.

9. FILE Exercise_10.9. The accompanying data set contains five predictor

variables (x1 to x5) and one binary target variable (y). Follow the instructions
below to create classification trees using the Exercise_10.9_Data worksheet.

page 436
a. Use the rpart function to build a default classification tree.
Display the default classification tree. How many leaf nodes are in the default
classification tree? What are the predictor variable and split value for the first
split of the default classification tree (root node)?
b. Use the rpart function to build a fully-grown classification tree. Display the cp
table. What is the cp value associated with the lowest cross-validation error?
How many splits are in the minimum error tree?
c. Is there a simpler tree within one standard error of the cross-validation error of
the minimum error tree? If there is, then what is the cp value associated with
the best-pruned tree?
d. Use the prune function to prune the full tree to the minimum error tree. Display
the minimum error tree. How many leaf nodes are in the minimum error tree?
e. Assign Class 1 to be the positive class. What are the accuracy, sensitivity,
specificity, and precision of the minimum error tree on the validation data?
f. Generate the cumulative lift chart and decile-wise lift chart of the minimum error
tree on the validation data. Does the entire lift curve lie above the baseline?
What is the lift value of the leftmost bar of the decile-wise lift chart? What does
this value imply?
g. Generate the ROC curve of the minimum error tree on the validation data. What
is the area under the ROC curve (or AUC value)?
h. Score the new observations in the Exercise_10.9_Score worksheet. What is the
predicted response value of the new observations? What is the Class 1
probability of the first observation? Round your answers to four decimal places.

10. FILE Exercise_10.10. The accompanying data set contains four predictor

variables (x1 to x4) and one binary target variable (y). Follow the instructions
below to create classification trees using the Exercise_10.10_Data worksheet.
a. Use the rpart function to build a default classification tree. Display the default
classification tree. How many leaf nodes are in the default classification tree?
What are the predictor variable and split value for the first split of the default
classification tree (root node)?
b. Use the rpart function to build a fully-grown classification tree. Display the cp
table. What is the cp value associated with the lowest cross-validation error?
How many splits are in the minimum error tree?
 c. Is there a simpler tree within one standard error of the cross-validation error of
the minimum error tree? If there is, then what is the cp value associated with
the best-pruned tree?
d. Use the prune function to prune the full tree to the best-pruned tree or minimum
error tree if the answer to part c is “No.” Display the pruned tree. How many leaf
nodes are in the pruned tree?
e. Assign Class 1 to be the positive class. What are the accuracy, sensitivity,
specificity, and precision of the pruned tree on the validation data?
f. Generate the decile-wise lift chart of the pruned tree on the validation data.
What is the lift value of the leftmost bar of the decile-wise lift chart? What does
this value imply?
g. Generate the ROC curve of the pruned tree on the validation data. What is the
area under the ROC curve (or AUC value)?
h. Score the new observations in the Exercise_10.10_Score worksheet using the
pruned tree. What is the predicted response value of the new observations?
What is the Class 1 probability of the first observation? Round your answers to
four decimal places.

Applications
11. FILE Travel_Plan. Jerry Stevenson is the manager of a travel agency. He

wants to build a model that can predict whether or not a customer will travel within
the next year. He has compiled a data set that contains the following variables:
whether the individual has a college degree (College), whether the individual has
credit card debt (CreditCard), annual household spending on food (FoodSpend),
annual income (Income), and whether the customer has plans to travel within the
next year (TravelPlan, 1 = has travel plans, 0 = does not have travel plans). A
portion of the Travel_Plan_Data worksheet is shown in the accompanying table.
Create a classification tree model for predicting whether or not the customer will
travel within the next year (TravelPlan). Select the best-pruned tree for scoring
and display the full-grown, best-pruned, and minimum error trees.

a. How many leaf nodes are in the best-pruned tree and minimum error tree?
What are the predictor variable and split value for the root node of the best-
pruned tree?
b. What are the classification accuracy rate, sensitivity, and specificity of the best-
pruned tree on the test data?
 c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Score the two new customers in the Travel_Plan_Score worksheet using the
best-pruned tree. What is the probability of the first customer having plans to
travel within the next year? What is the probability for the second customer?

12. FILE Travel_Plan. Refer to the previous exercise for a page 437
description of the problem and data set. Build a default classification tree to
predict whether a customer has plans to travel within the next year. Display the
default classification tree.
a. How many leaf nodes are in the tree? What are the predictor variable and the
split value for the first split of the default classification tree?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error? How many splits are in the minimum error tree?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum cross-validation error? If there is, then which cp value is
associated with the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. How many leaf nodes are in the
pruned tree?
e. Create a confusion matrix and display the various performance measures.
Assign Class 1 to be the positive class. What are the accuracy, sensitivity,
specificity, and precision of the pruned tree on the validation data?
f. Display the cumulative lift chart, the decile-wise lift chart, and the ROC curve of
the minimum error tree on the validation data. Comment on the performance of
the classification tree.
g. Score the two new customers in the Travel_Plan_Score worksheet using the
pruned tree. What is the probability of the first customer having plans to travel
within the next 12 months? What is the probability for the second customer?
Round your answers to four decimal places.

13. FILE Continue_Edu. Samantha Brown is Director of Continuing Education

of a major university. The Continuing Education department offers a wide range of

five-week courses to the community during the summer. Samantha would like to
find out which community members are more likely to enroll in these summer
courses. She has compiled a data set of 2,000 community members that includes
the following variables: age (Age), years of marriage (Marriage), annual income
(Income), education level (Edu), whether the person owns his or her residence
(Own), whether the person owns a pet (Pet), whether the person lives in the city
or suburb (City), and whether the person has enrolled in at least one continuing
education course from the university before (ContinueEdu, 1 = enrolled, 0
otherwise). A portion of the Continue_Edu_Data worksheet is shown in the
accompanying table. Build a default classification tree to predict whether a
 community member is likely to enroll in summer courses. Display the default
classification tree.

a. How many leaf nodes are in the tree? What are the predictor variable and the
split value for the first split of the default classification tree?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error? How many splits are in the minimum error tree?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum cross-validation error? If there is, then which cp value is
associated with the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. How many leaf nodes are in the
pruned tree?
e. Create a confusion matrix and display the various performance measures.
Assign Class 1 to be the positive class. What are the accuracy, sensitivity,
specificity, and precision of the pruned tree on the validation data?
f. Change the cutoff value to 0.1. Report the accuracy, sensitivity, specificity, and
precision rates of the pruned tree on the validation data.
g. Generate the decile-wise lift chart. What is the lift value of the leftmost bar of
the decile-wise lift chart?
h. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
i. Score the two new individuals in the Continue_Edu_Score worksheet using the
pruned tree. What is the probability of the first community member enrolling in
summer courses? What is the probability for the second community member?
Round your answers to four decimal places.

14. FILE Continue_Edu. Refer to the previous exercise for a description of the

problem and data set. Create a classification tree model for predicting whether the
community member is likely to enroll in summer courses (ContinueEdu). Select
the best-pruned tree for scoring and display the full-grown, best-pruned, and
minimum error trees.
a. How many leaf nodes are in the best-pruned tree and minimum error tree?
What are the predictor variable and split value for the root node of the best-
pruned tree?
b. What are the accuracy rate, sensitivity, and specificity of the best-pruned tree
on the test data?
 c. Display the cumulative lift chart, the decile-wise lift chart, and the ROC curve.
Comment on the performance of the classification model?
d. Score the two new cases in the Continue_Edu_Score worksheet using the best-
pruned tree. What is the probability of the first community member enrolling in
summer classes? What is the probability for the second community member?

15. FILE Church. The following data set in the Church_Data worksheet is used

to classify individuals as likely or unlikely to attend church using five predictor

variables: years of education (Educ), annual income (Income in $), age, sex (F =
female, M = male), and marital status (Married, Y = yes, N = no).
The outcome variable is Church (1 = attends, 0 otherwise).
page 438
Create a classification tree model for predicting whether the individual is likely to
attend church. Select the best-pruned tree for scoring and display the full-grown,
best-pruned, and minimum error trees.

a. How many leaf nodes are in the best-pruned tree and minimum error tree?
What are the rules that can be derived from the best-pruned tree?
b. What are the accuracy rate, sensitivity, specificity, and precision of the best-
pruned tree on the test data?
c. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
d. Score the cases in the Church_Score worksheet using the best-pruned tree.
What percentage of the individuals in the score data set are likely to go to
church based on a cutoff probability value of 0.5?

16. FILE Church. Refer to the previous exercise for a description of the problem

and data set. Build a default classification tree to predict whether an individual is
likely to attend church. Display the default classification tree.
a. How many leaf nodes are in the tree? What are the predictor variable and the
split value for the first split of the default classification tree? What are the rules
that can be derived from the root node?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum cross-validation error? If there is, then
page 439
 which cp value is associated with the best-pruned tree? How many splits are in
the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Create a confusion matrix and display the various
performance measures. Assign Class 1 to be the positive class. What are the
accuracy, sensitivity, specificity, and precision of the pruned tree on the
validation data?
e. Display the cumulative lift chart, the decile-wise lift chart, and the ROC curve of
the minimum error tree on the validation data. Comment on the performance of
the classification tree.
f. Score the cases in the Church_Score worksheet using the best-pruned tree.
What percentage of the individuals in the score data set are likely to go to
church based on a cutoff probability value of 0.5?

17. FILE Mobile_Banking. Sunnyville Bank wants to identify customers who

may be interested in its new mobile banking app. The worksheet called
Mobile_Banking_Data contains 500 customer records collected from a previous
marketing campaign for the bank’s mobile banking app. Each observation in the
data set contains the customer’s age (Age), sex (Male/Female), education level
(Edu, ranging from one to three), income (Income in $1,000s), whether the
customer has a certificate of deposit account (CD), and whether the customer
downloaded the mobile banking app (App equals 1 if downloaded, 0 otherwise). A
portion of the data set is shown in the accompanying table. Create a classification
tree model for predicting whether a customer will download the mobile banking
app. Assign one as the success class as we are more interested in identifying
customers who download the app. Select the best-pruned tree for scoring and
display the full-grown, best-pruned, and minimum error trees.

page 440

a. How many leaf nodes are in the best-pruned tree and minimum error tree?
What are the predictor variable and split value for the root node of the best-
pruned tree?
b. What are the accuracy rate, sensitivity, specificity, and precision of the best-
pruned tree on the test data?
c. Generate the decile-wise lift chart. What is the lift value of the leftmost bar of
the decile-wise lift chart? What does it imply?
 d. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
e. Score the 20 customers in the Mobile_Banking_Score worksheet using the
best-pruned tree. How many of the 20 new customers will likely download the
mobile banking app based on your classification model? What is the probability
of the first new customer downloading the app?

18. FILE Mobile_Banking. Refer to the previous exercise for a description of

the problem and data set. Build a default classification tree to predict whether a
customer will download the mobile banking app. Display the default classification
tree.
a. How many leaf nodes are in the tree? What are the predictor variable and split
value for the first split of the default classification tree? State the rule that can
be derived from the first leaf node from the top of the tree diagram.
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum cross-validation error? If there is, then which cp value is
associated with the best-pruned tree? How many splits are in the best-pruned
tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. Create a confusion matrix and display
the various performance measures. Assign Class 1 to be the positive class.
What are the accuracy, sensitivity, specificity, and precision of the pruned tree
on the validation data?
e. Generate the decile-wise lift chart. What is the lift value of the leftmost bar of
the decile-wise lift chart?
f. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
g. Score the 20 new customers in the Mobile_Banking_Score worksheet using the
pruned tree. How many customers will likely download the mobile banking app
based on your classification model? What is the probability of the first customer
to download the mobile banking app? Round your answer to 4 decimal places.

19. FILE Graduate. Dereck Anderson is an institutional researcher at a major

university. The university has set a goal to increase the number of students who
graduate within four years by 20% in five years. Dereck is asked by his boss to
create a model that would flag any student who has a high likelihood of not being
able to graduate within four years. He has compiled a data set of 2,000 previous
students of the university that contains the following variables: sex (M/F), whether
the student is Caucasian (White), high school GPA (HS GPA), SAT score (SAT),
College GPA (GPA), whether the student’s parents are college educated (College
Parent), and whether the student graduated within four years (Grad). A portion of
 the Graduate_Data worksheet is shown in the accompanying table. Build a default
classification tree to predict whether the student will be able to graduate within
four years (Grad). Display the classification tree.

a. What are the predictor variable and the split value for the first split of the default
classification tree? State the rules that can be derived from the root node.
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum cross-validation error? If there is, then which cp value is
associated with the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. Is the pruned tree the same tree as
the default tree created in a?
e. Create a confusion matrix and display the various performance measures.
Assign Class 1 to be the positive class. What are the accuracy, sensitivity,
specificity, and precision of the pruned tree on the validation data?
f. Generate the cumulative lift chart. Does the lift curve lie above the baseline?
What does this mean?
g. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)? What does the AUC value imply?
h. Score the three university students in the Graduate_Score worksheet using the
pruned tree. How many of these three students will be able to graduate within
four years according to your model?

20. FILE Graduate. Refer to the previous exercise for a description of the

problem and data set. Create a classification tree model for predicting whether the
student will be able to graduate within four years (Grad). Assign 0 as the success
class as we are more interested in identifying students who are at the risk of not
being able to graduate within four years. Select the best-pruned tree for scoring
and display the full-grown, best-pruned, and minimum error trees.
a. How many leaf nodes are in the best-pruned tree and minimum error tree?
What are the rules that can be derived from the best-pruned tree?
b. What are the accuracy rate, sensitivity, specificity, and precision of the best-
pruned tree on the test data?
c. Generate the cumulative lift chart. Does the lift curve lie above the baseline?
What does this mean?
 d. Generate the ROC curve. What is the area under the ROC curve (or AUC
value)?
e. Score the three university students in the Graduate_Score worksheet using the
best-pruned tree. Will these three students be able to graduate within four years
according to your model?

21. FILE In_App_Pur. Monstermash, an online game app development

company, wants to be able to predict which gamers are likely to make in-app
purchases. Ranon Weatherby, the company’s data analyst, has compiled a data
set about customers that contains the following variables: customer age (Age),
sex (1 if male, 0 otherwise), household income (Income in $1,000s), the number
of years playing online games (Years), the number of hours playing online games
per week (Hours), whether the customer has a credit card (CreditCard), whether
the customer has a Facebook profile (Facebook), and whether the customer has
made in-app purchases before (Buy). A portion of the In_App_Pur_Data
worksheet is shown in the accompanying table. Create a classification tree model
for predicting whether the customer will make in-app purchases (Buy). Select the
best-pruned tree for scoring and display the full-grown, best-pruned, and
minimum error trees.

a. How many leaf nodes are in the best-pruned tree and minimum error tree?
What are the predictor variable and split value for the root node of the best-
pruned tree?
b. Describe the rules produced by the best-pruned tree.
c. What are the accuracy rate, sensitivity, specificity, and precision of the best-
pruned tree on the test data?
d. Display the cumulative lift chart, the decile-wise lift chart, and the ROC curve,
does the classification model outperform the baseline model? What is the area
under the ROC curve (or AUC value)?
e. Score the two new customers in the In_App_Pur_Score worksheet using the
best-pruned tree. Will the first new customer make in-app purchases according
to your model? What about the second new customer? What are the
probabilities of the two new customers to make in-app purchases?

22. FILE In_App_Pur. Refer to the previous exercise for a description of the

problem and data set. Build a default classification tree to predict whether the
gamer will make in-app purchases. Display the classification tree.
 a. What are the predictor variable and the split value for the first split of the default
classification tree?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum cross-validation error? If there is, then which cp value is
associated with the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Create a confusion matrix and display the various
performance measures. Assign Class 1 to be the positive class. What are the
accuracy, sensitivity, and specificity of the minimum error tree on the validation
data?
e. Display the cumulative lift chart, the decile-wise lift chart, and the ROC curve of
the minimum error tree on the validation data. Comment on the performance of
the classification tree.
f. Score the two gamers in the In_App_Pur_Score worksheet using the pruned
tree. What is the probability of the first gamer making in-app purchases
according to your classification model? What is the probability for the second
gamer?

10.3 Regression Trees

LO 10.2
Apply regression trees to predict new records.

In Section 10.2, we presented classification trees, where the target

variable assumed a categorical value. We now discuss regression
trees, where the target variable assumes a numerical value. Like
classification trees, regression trees can be constructed using the
CART algorithm. The CART algorithm operates similarly for
classification and regression trees, with the following three important
differences.
 First, as the target variable of the regression tree is numerical,
the predicted value is the average value of the previous cases that
belong to the same leaf node. Consider, for example, a regression
tree for predicting customer spending over a 12-month period. Let
one of the leaf nodes consist of four cases in the training data set
with the following customer spending values: 259.00, 412.50,
139.75, and 188.45. When scoring any new customer that falls into
this leaf node, the predicted spending amount for the new customer
is the average of the four values, or .

Second, while the CART algorithm uses the Gini index to

measure impurity in the classification tree, it uses the mean squared
error (MSE) to measure impurity in a regression tree. Recall that the
MSE is calculated as the average sum of the squared differences
between the actual and the predicted values of the target variable in
each leaf node or subset. If the exact values of all cases in a subset
are predicted accurately, then the MSE equals 0, indicating that the
subset is pure. There is no upper limit for the MSE for an impure
subset. The higher the MSE, the more impure the subset is. In
principle, a recursive partitioning process for a regression tree is
similar to the classification tree process presented in Examples 10.1
through 10.5. As such, in Example 10.6, we will only demonstrate
how the CART algorithm computes the MSE to determine the splits
of a regression tree.

page 441

Finally, as a regression tree is a prediction technique, the

performance of the tree is evaluated using the performance
measures for prediction that were discussed in Section 8.3; that is,
we use the root mean square error (RMSE), the mean error (ME),
the mean absolute deviation (MAD) or mean absolute error (MAE),
the mean percentage error (MPE), and the mean absolute
percentage error (MAPE).
 In Section 10.2, we showed business insights gained from the
classification tree that Hayden and his team built for predicting which
new customers are likely to respond to HELOC offers from
Sunnyville Bank. Encouraged by these insights, Hayden has an idea
for another data mining project. His retail banking team is interested
in attracting high-value customers, defined as those who routinely
maintain high balances in their accounts with Sunnyville Bank.
Hayden hopes that a predictive model can help his team identify and
focus on this group of high-value customers. He compiles a data file,
Balance, that has two worksheets, Balance_Data and
Balance_Score. Table 10.7a shows a portion of the Balance_Data
worksheet, which contains the information from 500 current
customers including the customer’s age (Age, in years), sex (Sex),
annual income (Income), and the average monthly account balance
(Balance). Table 10.7b shows a portion of the Balance_Score
worksheet, which includes the age, sex, and income of 20 potential
customers. Ultimately, Hayden wants to build a predictive model
using the data of current customers to predict the average account
balance for the 20 potential customers.
FILE
Balance

TABLE 10.7 The Balance Data

a. The Balance_Data Worksheet

Age Sex Income Balancec

38 Female 60000 3250

20 Male 41000 3546

⋮ ⋮ ⋮ ⋮
 Age Sex Income Balancec

50 Female 17000 2529

b. The Balance_Score Worksheet

Age Sex Income

35 Female 65000

56 Male 160000

⋮ ⋮ ⋮

52 Female 155000

Before building a prediction tree, we first show how to identify

possible split points using a smaller data set.

Identifying Possible Split Points

Similar to the classification tree, one of the first steps in constructing
a regression tree is to identify possible splits of the predictor
variables. In principle, a recursive partitioning process for a
regression tree is similar to the classification tree process presented
in Section 10.2. In Example 10.6, we demonstrate how the CART
algorithm computes the MSE to determine the splits of a regression
tree.
As mentioned earlier, the recursive partitioning process used for
classification trees also applies to regression trees. Using the MSE
as the criterion, a regression tree can be constructed by selecting
the best split at each recursive partitioning step across all predictor
variables. We will now develop a prediction tree with Analytic Solver
and R.
FILE
Balance_Data_20

EXAMPLE 10.6
For illustration, we use only a small data set of 20 randomly
selected bank customers from the Balance_Data worksheet.
Table 10.8 shows a portion of the data, which is stored in the
Balance_Data_20 worksheet. Select the optimal split for the
Age variable based on the MSE impurity measure.

TABLE 10.8 The Balance_Data_20 worksheet

Age Balance

43 1775

34 3675

⋮ ⋮

26 10761

page 442

SOLUTION: Similar to Example 10.1, we identify the possible split

points by finding the midpoints between the consecutive values
of Age. The ages of the 20 potential customers are first sorted
in ascending order in the following data array:
(Note that some values have multiple occurrences, but for
determining splitting points, each value only appears once in
the array.)
Therefore, the first possible split point is calculated as (18 +
20)/ 2 = 19, the next possible split point is calculated as (20 +
24)/ 2 = 22, and the remaining possible split points are found in
a like manner, as shown in the following data array:

Consider two of the possible splits, 30 and 32. We will compute

and compare the MSE for the two split values to find the
preferred split. Recall that the MSE is the average sum of the
squared differences between the actual account balances and
predicted account balances in each partition. A split at Age =
30 creates two partitions: customers whose ages are less than
30 (8 cases) and customers whose ages are equal to or
greater than 30 (12 cases). The average bank account
balances of the two partitions are $7,811.25 and $4,445,
respectively. This suggests that if we create a regression tree
with one decision node that splits Age at the value of 30, all
potential customers younger than 30 years will have a
predicted balance of $7,811.25, whereas all customers 30
years or older will have a predicted balance of $4,445. For the
Age = 30 split, the MSE for each partition is calculated as
To compute the overall MSE for the split, we calculate the
weighted combination of the MSEs using the percentage of
cases in each partition as the weight:

Similarly, a split at Age = 32 creates two partitions as well:

customers whose ages are less than 32 (9 cases) and
customers whose ages are equal to or greater than 32 (11
cases). The average bank account balances for the two
partitions are $7,494.667 and $4,398, respectively. For the Age
= 32 split, the MSE for each partition is calculated as

The overall MSE for the split is calculated as

page 443

The MSE value of the Age = 30 split is slightly smaller than that
of the Age = 32 split, suggesting that the Age = 30 split
generates a lower level of impurity; therefore, Age = 30 is a
better split for constructing the regression tree. In fact, if we
 compute the MSE values for all the possible splits for Age in
our example, we would find that the split at Age = 30 produces
the smallest MSE. Therefore, the Age = 30 split is chosen as
the best split for Age. Figure 10.20 shows the regression tree
when Age is the only predictor variable used in the analysis.

FIGURE 10.20 Regression tree if Age only predictor

Using Analytic Solver and R to Develop a

Prediction Tree
FILE
Balance

EXAMPLE 10.7
Hayden wants to use historical data from 500 bank customers
to develop a prediction model for predicting a customer’s
account balance. He also plans to assess the performance of
the regression tree and then score 20 new customers. Use
Analytic Solver and R to build the preferred prediction tree, and
then use this tree to score the new cases.
SOLUTION:
Using Analytic Solver
a. Open the Balance_Data worksheet of the Balance data file.
b. Choose Data Mining > Partition (under the Data Mining
group) > Standard Partition.
c. In the Standard Data Partition dialog box, we select the data
range $A$1:$D$501 and move Age, Sex, Income, and
Balance to the Selected Variables box. Select Pick up rows
randomly and check the Set seed checkbox to use the default
random seed of 12345. Select Specify percentages to allocate
50%, 30%, and 20% of the data to training, validation, and test
sets, respectively. Click OK.
d. Analytic Solver will create a new worksheet called
STDPartition with the partitioned data. With the STDPartition
worksheet active, choose Data Mining > Predict >
Regression Tree to open the Regression Tree dialog box. In
the Data tab, move the predictor variables Age page 444
and Income to the Selected Variables box and
Sex to the Categorical Variables box. Select and move the
target variable, Balance, to the Output Variable box. Accept
other defaults. Click Next.
e. In the Parameters tab, check the Prune (Using Validation Set)
checkbox and click on the Tree for Scoring button. Select Best
Pruned, then click Done. This selection tells Analytic Solver to
score cases using the best-pruned tree. Click on the Trees to
Display button. Check the checkboxes next to Fully Grown,
Best Pruned, and Minimum Error, then click Done. Accept
other defaults. Click Next.
f. In the Scoring tab, Check the Summary Report boxes for
Score Training Data, Score Validation Data, and Score Test
Data. Check the In Worksheet check box under the Score
New Data section. Under the New Data (WS) tab, select
Balance_Score in the Worksheet box. These 20 records of
new bank customers will be scored by the regression tree.
Make sure the First Row Contains Headers box is checked
and click the Match by Name button. Accept other defaults.
Click Finish.
Analytic Solver creates a number of worksheets. Table 10.9
shows a portion of the prune log, which is found in the
RT_Output worksheet. The prune log shows that the simplest
tree with the minimum validation MSE of 34,625,847.56 occurs
at two decision nodes. Because no smaller tree has a
validation MSE within one standard error of this value, the
minimum error tree and best-pruned tree coincide in this
example.

TABLE 10.9 Analytic Solver’s Prune Log

# Decision Nodes Cost Complexity Validation MSE

0 10010714.87 47757025.42

1 8713455.60 39391178.20

2 7016220.13 34625847.56

3 7091273.75 35733534.52

4 6440499.50 35411119.80

5 5154901.31 35536360.84

6 5764821.26 36632166.92

The RT_FullTree worksheet shows the full tree. The full tree is
quite complex with many decision nodes. Due to its size, the
full tree is not displayed here. The RT_BestTree and
RT_MinErrorTree worksheets show the best-pruned and
minimum error trees, respectively. As mentioned earlier, the
two trees are identical in this example. Figure 10.21 displays
the tree. It is relatively simple with two decision nodes and five
total nodes if the terminal nodes are included. The following If-
Then rules can be derived:

FIGURE 10.21 Analytic Solver’s minimum error and best-pruned regression

tree

If the customer’s annual income is at least $157,500, then

the predicted account balance is $15,448.3.
If the customer’s annual income is at least $72,500 but less
than $157,500, then the predicted account balance is
$7,630.6.
If the customer’s annual income is less than $72,500, then
the predicted balance is $3,200.6.
The regression tree suggests that bank customers who have
high income tend to maintain high balances in their checking
accounts. Age and Sex have no impact on predicting Balance.
 page 445

Because the validation data were used for pruning, the

performance of the regression tree on a new data set should
be assessed using the test data set. A portion of the
performance measures in Testing: Prediction Summary in the
RT_TestScore worksheet is shown in Table 10.10. The RMSE
implies that the standard error of the average checking account
balance prediction is $6,151.07, and the MAD implies that the
predicted balances on average differ from the actual balance
by $4,326.83.

TABLE 10.10 Analytic Solver’s Performance Measures

Performance Measure Value

RMSE 6151.07

MAD 4326.83

Sometimes the prediction errors implied by the performance

measures appear large. This is quite common for regression
trees when the data do not have a very clear and simple
pattern. However, if you rank order the customers based on
their predicted balances from the highest to lowest, you will find
that the regression tree model does a relatively good job in
predicting whether the customer tends to have a high or a low
account balance. Therefore, in many cases, predicting the
values of the target variable may not be the ultimate goal of a
regression tree model. In our example, Hayden will likely use
the regression tree model to identify customers who tend to
have higher account balances instead of trying to predict the
actual balances. The RT_NewScore worksheet shows the
predicted balances of the 20 new customers using our
 regression tree. A portion of the predicted balances are shown
in Table 10.11.

TABLE 10.11 Analytic Solver’s Scoring Results

page 446
Using R
As mentioned earlier in the construction of classification trees,
R’s built-in cross-validation process means that there is no
need to separate the data into three partitions as we did when
using Analytics Solver. Here, we separate the data into two
partitions: training and validation data sets.
a. Import the data from the Balance_Data worksheet of the
Balance data file into a data frame (table) and label it myData.
The following instructions and results are based on R version
3.5.3. To replicate the results with newer versions of R, enter:
>suppressWarnings(RNGversion(“3.5.3”))
b. Install and load the caret, rpart, rpart.plot, and forecast
packages using the following command if you have not
already done so. Enter:
>install.packages(“caret”, dependencies = c(“Depends”,
“Suggests”))
>install.packages(“rpart”)
>install.packages(“rpart.plot”)
>install.packages(“forecast”)
 >library(caret)
>library(rpart)
>library(rpart.plot)
>library(forecast)
c. By setting the random seed to 1, we will generate the same
partitions as shown in this example. As the construction of a
decision tree is a data-driven process and can benefit from a
large training set, we use the createDataPartition function to
randomly allocate 70% of the data into the training data set
and 30% into the validation data set. Enter:
>set.seed(1)
>myIndex <- createDataPartition(myData$Balance, p=0.7, list
= FALSE)
>trainSet <- myData[myIndex,]
>validationSet <- myData[-myIndex,]
d. We use the rpart function to generate the default classification
tree labeled default_tree. Within the rpart function, we specify
the model structure, data source, and method. The method =
“anova” option tells the function to build a regression tree to
estimate a numerical target value. To view the details of the
default tree, use the summary function. To ensure
consistency of the cross-validation results, we use the
set.seed function to fix the random seed to 1. Enter:
>set.seed(1)
>default_tree <- rpart(Balance ~ ., data=trainSet,
method=“anova”)
>summary(default_tree)
e. To view the regression tree visually, we use the prp function.
The type option is set equal to 1 so that all nodes except the
leaf nodes are labeled in the tree diagram. The extra option is
set equal to 1 so that the number of observations that fall into
each node are displayed. The under option is set equal to
TRUE in order to put the number of cases under each
decision node in the diagram. Enter:
>prp(default_tree, type=1, extra=1, under = TRUE)
Figure 10.22 shows the default regression tree for predicting
average balance.

FIGURE 10.22 R's default regression tree

page 447
As discussed earlier, to find the optimal decision
tree, a common practice is to grow the full tree and then prune
it to a less-complex tree based on the prediction errors
 produced by the cross-validation process of the rpart function.
By identifying the value of the complexity parameter (cp)
associated with the smallest cross-validated prediction error,
we can create the minimum error tree. Next, we demonstrate
the pruning process to optimize the complexity of the tree.
f. We first grow the full tree by using the rpart function. We set
the options cp equal to 0, minsplit equal to 2, and minbucket
equal to 1. As discussed in the classification tree section,
these settings ensure that the largest possible tree will be
produced. We plot the full tree using the prp function. Again,
to ensure consistency of the cross-validation results, we set
the random seed to 1. Enter:
>set.seed(1)
>full_tree <- rpart(Balance ~ ., data= trainSet,
method=“anova”, cp=0, minsplit=2, minbucket=1)
>prp(full_tree, type=1, extra=1, under = TRUE)
The full tree in this case is very complex so it is not displayed
here.
g. To identify the value of cp that is associated with page 448
the smallest cross-validated prediction error, we
use the printcp function. Enter:
>printcp(full_tree)
Due to the complexity of the full tree, 273 subtree options are
displayed in the cp table. Figure 10.23 displays a portion of
the complexity parameter table showing the first eight
candidate trees with increasing complexity. The nsplit column
shows the number of splits for each tree. The rel error column
shows the prediction error for each tree, relative to the
prediction error of the root node if all cases are given the
predicted balance that equals the average of all balances. The
prediction performance of the trees needs to be evaluated by
inspecting the cross-validation errors associated with each
tree; see the xerror column. The third tree, with two splits, has
the lowest cross-validation error (xerror = 0.72172); therefore,
 it is the minimum error tree. The xstd column (standard error)
can be used to identify the best-pruned tree, which is the
smallest tree whose cross-validation error falls within one
standard error of the minimum error tree (0.72172 + 0.10521 =
0.82693). In this case, the second tree, with just one split, has
a cross-validation error of 0.78354, which is within the range;
hence, the best-pruned tree is the second tree. Next, we use
the value of cp associated with the second tree to produce our
best-pruned tree.

FIGURE 10.23 R’s complexity parameters

h. We use the prune function to create the best-pruned tree by

using the cp value associated with the second tree. Again, to
avoid issues caused by rounding, we use a cp value that is
slightly larger than the cp value associated with the second
tree but lower than that of the next smaller tree. In this case,
we will use 7.5219e-02. We display the pruned tree using the
prp function. Enter:
>pruned_tree <- prune(full_tree, cp=7.5219e-02)
>prp(pruned_tree, type=1, extra=1, under = TRUE)
 Figure 10.24 displays the tree diagram of the best-pruned
tree. The tree has just one split and two leaf nodes. The
values in the boxes are the predicted balance for customers
that belong to each node. The tree shows that if a customer’s
income is less than $74,000, then the predicted balance is
$2,942; otherwise, the predicted balance is page 449
$9,421. This suggests that high-income earners
tend to have higher account balances. Next, we will evaluate
the performance of the pruned tree using the validation data
set.

FIGURE 10.24 R’s best-pruned tree

i. We predict the average balances of the observations in the

validation data set using the predict function. Enter:
>predicted_value <- predict(pruned_tree, validationSet)
j. To evaluate the performance of our regression tree, we
compute the performance measures for prediction using the
accuracy function. The accuracy function requires two
arguments: predicted values and actual values. Enter:
>accuracy(predicted_value, validationSet$Balance)
Various performance measures are produced, as shown in
Figure 10.25.

FIGURE 10.25 R’s performance measures

k. Finally, to score the 20 new cases, we import the data from
the Balance_Score worksheet of the Balance data file into a
data frame (table) called myScoreData and use the predict
function to produce the predicted average checking account
balance (called predicted_value_score) for the new cases
using our regression tree. Enter:
>predicted_value_score <- predict(pruned_tree,
myScoreData)
>predicted_value_score
The scoring results are summarized in Table 10.12. As you
can see, the annual balances predicted by the regression tree
constructed in R are somewhat different from those predicted
by the regression tree produced in Analytic Solver. This is due
to the different training data sets randomly selected in the two
software products. However, as in the case of Analytic Solver,
the pruned regression tree suggests that while Income plays a
crucial role, Age and Sex have no impact on predicting
Balance.
TABLE 10.12 R’s Scoring Results

page 450

EXERCISES 10 3
 EXERCISES 10.3

Note: and indicate that the exercise problems must be

solved using Analytic Solver and R, respectively. For Analytic

Solver, partition data sets into training, 30% validation, and 20%
and use 12345 as the default random seed. For R, partition
data sets into 70% and 30% validation. Use the statement
set.seed(1) to specify the random seed of 1 for both data
partitioning and cross-validation. All answers in R are based on
version 3.5.3. To replicate the results with newer versions of R,
execute the following line of code at the beginning of the R
session: suppressWarnings(RNGversion(“3.5.3”)). Some data
files have two worksheets (for example, Exercise_10.28_Data
and Exercise_10.28_Score worksheets) for model development
and scoring new records. If the predictor variable values are in
the character format, then treat the predictor variable as a
categorical variable. Otherwise, treat the predictor variable as a
numerical variable.

Mechanics
23. FILE Exercise_10.23. The accompanying data set contains two predictor
variables, x1 and x2, and one numerical target variable, y. A regression tree will
be constructed using the data set.
a. List the possible split values for x1 in ascending order.
b. List the possible split values for x2 in ascending order.
c. Compute the MSE of the partition x1 = 131.
d. Compute the MSE of the partition x2 = 105.
e. Compare and interpret the results from parts c and d.
24. FILE Exercise_10.24. The accompanying data set contains three predictor
variables, x1, x2, and x3, and one numerical target variable, y. A regression tree
will be constructed using the data set.
a. List the possible split values for x1 in ascending order.
b. List the possible split values for x2 in ascending order.
c. List the possible split values for x3 in ascending order.
 d. Compute the MSE of the partition x1 = 252.
e. Compute the MSE of the partition x2 = 92.5.
f. Compute the MSE of the partition x3 = 14.25.
g. Compare and interpret the results from parts d, e, and f.
25. FILE Exercise_10.25. The accompanying data set contains two predictor
variables, x1 and x2, and one numerical target variable, y. A regression tree will
be constructed using the data set.
a. Which split on x1 will generate the smallest MSE?
b. Which split on x2 will generate the smallest MSE?
c. Which variable and split value should be used to create the root node if a
regression tree is constructed using the accompanying data set?
d. State the rules generated from this split.
26. FILE Exercise_10.26. The accompanying data set contains three predictor
variables, x1, x2, and x3, and one numerical target variable, y. A regression tree
will be constructed using the data set.
a. Which split on x1 will generate the smallest MSE?
b. Which split on x2 will generate the smallest MSE?
c. Which split on x3 will generate the smallest MSE?
d. Which variable and split value should be used to create the root node if a
regression tree is constructed using the accompanying data set?
e. State the rules generated from this split.
27. The regression tree below relates credit score to number of defaults (NUM DEF),
revolving balance (REV BAL), and years of credit history (YRS HIST). Predict the
credit score of each of the following individuals.

a. An individual with no defaults, $4,200 revolving balance, and 12 years of credit

history.
 b. An individual with two defaults, $6,500 revolving balance, and seven years of
credit history.
c. An individual with one default, $12,700 revolving balance, and 25 years of credit
history.
d. An individual with two defaults, $8,100 revolving balance, and 16 years of credit
history.

28. FILE Exercise_10.28. Create a regression tree using the accompanying

data set in Exercise_10.28_Data worksheet (predictor variables: x1 to x5; target:

y). Select the best-pruned tree for scoring and display the full-grown, best-pruned,
and minimum error trees.
a. What is the minimum validation MSE in the prune log? How many decision
nodes are associated with the minimum error?
b. How many leaf nodes are in the best-pruned and minimum error trees?
c. Display the best-pruned tree. What are the predictor variable and split value for
the first split (root node) of the best-pruned tree? What are the rules that can be
derived from the root node?
d. What are the RMSE and MAD of the best-pruned tree on the test data?
e. Score the new observations in the Exericse_10.28_Score worksheet using the
best-pruned tree. What is the predicted value of y for the first observation?
f. What are the minimum, maximum, and average values of predicted y?

29. FILE Exercise_10.29. Create a regression tree using the page 451
accompanying data set (predictor variables: x1 to x4; target: y). Select the best-
pruned tree for scoring and display the full-grown, best-pruned and minimum error
trees.
a. What is the minimum validation MSE in the prune log? How many decision
nodes are associated with the minimum error?
b. How many leaf nodes are in the best-pruned and minimum error trees?
c. Display the best-pruned tree. What are the predictor variable and split value for
the first split (root node) of the best-pruned tree? What are the rules that can be
derived from the root node?
d. What are the RMSE and MAD of the best-pruned tree on the test data?
e. According to the best-pruned tree, what is the predicted y for a new observation
with the following values: x1 = 20; x2 = 40; x3 = 36; x4 = 8.3?

30. FILE Exercise_10.30. Create a regression tree using the accompanying

data set in the Exercise_10.30_Data worksheet (predictor variables: x1 to x4;

target: y).
 a. Use the rpart function to build a default regression tree. Display the default
regression tree. How many leaf nodes are in the default regression tree?
b. What are the predictor variable and split value for the first split (root node) of the
default regression tree? What are the rules that can be derived from the root
node?
c. Use the rpart function to build a fully-grown regression tree. Display the cp
table. Which tree has the lowest cross-validation error? How many splits are in
the minimum error tree?
d. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum cross-validation error? What is the cp value associated
with the best-pruned tree?
e. Prune the full tree to the best-pruned tree or minimum error tree if the answer to
part c is “No.” Display the pruned tree. How many leaf nodes are in the pruned
tree?
f. What are the ME, RMSE, MAE, MPE and MAPE measures of the pruned tree
on the validation data?
g. Score the new observations in the Exercise_10.30_Score worksheet. What is
the predicted value of y for the first observation? What are the minimum,
maximum, and average values of predicted y?

31. FILE Exercise_10.31. Create a regression tree using the accompanying

data set in the Exercise_10.31 worksheet (predictor variables: x1 to x4; target: y).
a. Use the rpart function to build a default regression tree. Display the default
regression tree. How many leaf nodes are in the default regression tree? What
are the predictor variable and split value for the first split of the default
regression tree?
b. Use the rpart function to build a fully-grown regression tree. What is the cp
value that is associated with the lowest cross-validation error? How many splits
are in the minimum error tree?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum cross-validation error? What is the cp value associated
with the best-pruned tree? How many splits are in the best-pruned tree?
d. Prune the full tree to the best-pruned tree or minimum error tree if the answer to
part c is “No.” Display the pruned tree. What are the ME, RMSE, MAE, MPE,
and MAPE measures of the pruned tree on the validation data?
e. Comment on the performance of the pruned regression tree.

32. FILE Exercise_10.32. Create a regression tree using the accompanying

data set (predictor variables: x1 to x4; target: y). Select the best-pruned tree for
scoring and display the full-grown, best-pruned, and minimum error trees.
a. What is the minimum validation MSE in the prune log? How many decision
nodes are associated with the minimum error?
 b. Display the best-pruned tree. How many leaf nodes are in the best-pruned and
minimum error trees?
c. What are the predictor variable and split value for the first split (root node) of the
best-pruned tree? What are the rules that can be derived from the root node?
d. What are the RMSE and MAD of the best-pruned tree?

33. FILE Exercise_10.33. Create a regression tree using the accompanying

data set (predictor variables: x1 to x4; target: y).

a. Use the rpart function to build a default regression tree. Display the tree using
the prp function. How many leaf nodes are in the default regression tree?
b. What are the predictor variable and split value for the first split (root node) of the
default regression tree? What are the rules that can be derived from the root
node?
c. Use the rpart function to build a fully grown regression tree. Display the cp
table. Which tree has the lowest cross-validation error? Which cp value is
associated with the minimum error tree?
d. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum error? If there is, then which cp value is associated with
the best-pruned tree?
e. Use the prune function to prune the full tree to the best-pruned tree or the
minimum error tree if the answer to part d is “No.” Display the pruned tree using
the prp function. What are the ME, RMSE, MAE, MPE, and MAPE measures of
the pruned tree on the validation data?

page 452

Applications
34. FILE Travel. Jerry Stevenson is the manager of a travel agency. He wants to

build a model that can predict customers’ annual spending on travel products. He
has compiled a data set that contains the following variables: whether the
individual has a college degree (College), whether the individual has credit card
debt (CreditCard), annual household spending on food (FoodSpend), annual
income (Income), and annual household spending on travel products
(TravelSpend). A portion of the Travel_Data worksheet is shown in the
accompanying table. Create a regression tree model for predicting the customer’s
annual household spending on travel products (TravelSpend). Select the best-
pruned tree for scoring and display the full-grown, best-pruned, and minimum
error trees.
 a. How many leaf nodes are in the best-pruned tree and the minimum error tree?
What are the rules that can be derived from the root node of the best-pruned
tree?
b. What are the RMSE and MAD of the best-pruned tree on the test data?
c. Score the two new customers in the Travel_Score worksheet using the best-
pruned tree. What are their predicted annual spending amounts on travel
products?

35. FILE Travel. Refer to the previous exercise for a description of the data set.

Build a default regression tree to predict the customer’s annual household

spending on travel products (TravelSpend). Display the regression tree.
a. How many leaf nodes are in the default tree? What are the predictor variable
and split value for the first split of the default regression tree?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error? How many splits are in the minimum-error tree? What is the
minimum cross-validation error?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum error? If there is, then which cp value is associated with
the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. How many leaf nodes are in the
pruned tree?
e. What are the ME, RMSE, MAE, MPE, and MAPE of the pruned tree on the
validation data?
f. Score the two new customers in the Travel_Score worksheet using the pruned
tree. What are their predicted annual spending amounts on travel products?

36. FILE Houses. Melissa Hill is a real estate agent in Berkeley, California. She

wants to build a predictive model that can help her price a house more accurately.
Melissa has compiled a data set in the House_Data worksheet that contains the
information about the houses sold in the past year. The data set contains the
following variables: number of bedrooms (BM), number of bathrooms (Bath),
square footage of the property (SQFT), lot size (Lot_Size), type of property
(Type), age of the property (Age), and price sold (Price). A portion of the data set
is shown in the accompanying table. Build a default regression tree to predict
house prices (Price). Display the regression tree.
 a. What are the predictor variable and split value for the first split of the default
regression tree? What are the rules that can be derived from the root node?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error? How many splits are in the minimum-error tree?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum error? If there is, then which cp value is associated with
the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. How many leaf nodes are in the
pruned tree?
e. What are the ME, RMSE, MAE, MPE, and MAPE of the pruned tree on the
validation data? On average, does the regression tree over- or under-predict
prices of houses? Is the regression tree model effective in predicting prices of
houses?
f. Score the two new houses on the market in the Houses_Score page 453
worksheet using the pruned tree. What are their predicted
prices?

37. FILE Houses. Refer to the previous exercise for a description of the data

set. Create a regression tree model for predicting house prices (Price). Select the
best-pruned tree for scoring and display the full-grown, best-pruned, and
minimum error trees.
a. Display the best-pruned tree. How many leaf nodes are in the best-pruned tree?
What are the predictor variable and split value of the root node of the best-
pruned tree?
b. What are the RMSE and MAD of the best-pruned tree on the test data? On
average, does the regression tree over- or under-predict prices of houses? Is
the regression tree model effective in predicting prices of houses?
c. Score the two new houses on the market in the Houses_Score worksheet using
the best-pruned tree. What are their predicted prices according to your model?

38. FILE E_Retailer. An online retail company is trying to predict customer

spending in the first three months of the year. Brian Duffy, the marketing analyst of
the company, has compiled a data set on 200 existing customers that includes
sex (Female: 1 = Female, 0 otherwise), annual income in 1,000s (Income), age
(Age, in years), and total spending in the first three months of the year
 (Spending). A portion of the E-Retailer_Data worksheet is shown in the
accompanying table. Create a regression tree model for predicting customer
spending during the first three months of the year (Spending). Select the best-
pruned tree for scoring and display the full-grown, best-pruned, and minimum
error trees.

Female Income Age Spending

0 87.5 52 156.88

1 66.5 43 275.16

⋮ ⋮ ⋮ ⋮

0 51.9 61 159.51

a. Display the best-pruned tree. How many leaf nodes are in the best-pruned tree?
b. What are the RMSE and MAD of the best-pruned tree on the test data?
c. Score the 10 new customers in the E-Retailer_Score worksheet using the best-
pruned tree. What are the mean and median values of the predicted spending
amounts according to your model?

39. FILE E_Retailer. Refer to the previous exercise for a description of the data

set. Build a default regression tree to predict the customer’s spending during the
first three months of the year (Spending). Display the regression tree.
a. What are the rules that can be derived from the default regression tree?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error? How many leaf nodes are in the minimum-error tree?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum error? If there is, then which cp value is associated with
the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. How many leaf nodes are in the
pruned tree?
e. What are the ME, RMSE, MAE, MPE, and MAPE of the pruned tree on the
validation data?
f. Score the 10 new customers in the E-Retailer_Score worksheet using the
pruned tree. What are the mean and median values of the predicted spending
amounts during the first three months?

40. FILE Electricity. Kyle Robson, an energy researcher for the U.S. Energy

Information Administration, is trying to build a model for predicting annual

 electricity retail sales for states. Kyle has compiled a data set for the 50 states
and the District of Columbia that contains average electricity retail price (Price in
cents/kWh), per capita electricity generation (Generation), median household
income (Income), and per capita electricity retail sales (Price in MWh). A portion
of the data set is shown in the accompanying table. Build a default regression tree
to predict per capita electricity retail sales (Sales). Display the regression tree.

a. What are the predictor variable and split value for the first split of the default
regression tree? What are the rules that can be derived from the default
regression tree?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error? How many leaf nodes are in the minimum-error tree?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum error? If there is, then which cp value is associated with
the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. How many leaf nodes are in the
pruned tree?
e. What are the ME, RMSE, MAE, MPE, and MAPE of the pruned tree on the
validation data?
f. What is the predicted per capita electricity retail sales for a state with the
following values: Price = 11, Generation = 25, and Income = 65,000?

41. FILE Electricity. Refer to the previous exercise for a description of the data

set. Create a regression tree model for predicting per capita electricity retail sales
(Sales). Select the best-pruned tree for scoring and display the full-grown, best-
pruned, and minimum error trees.
a. How many leaf nodes are in the best-pruned tree and minimum error tree?
b. What are the predictor variable and split value for the first split of the best-
pruned tree? What are the rules that can be derived from the root node?
c. What are the RMSE and MAD of the best-pruned tree on the test data?
d. What is the predicted per capita electricity retail sales for a state with the
following values: Price = 11, Generation = 25, and Income = 65,000?

page 454
42. FILE NBA. Merrick Stevens is a sports analyst working for ACE Sports

Management, a sports agency that represents over 200 athletes. He is interested

in understanding the relationship between an NBA player’s salary and his
physicality and performance statistics. Merrick has constructed a data set that
contains information on 30 competing NBA teams and 445 players. A portion of
the NBA_Data worksheet is shown in the table below. A detailed description of
each field in the data set can be found in NBA Information.docx. Create a
regression tree model for predicting an NBA player’s salary (salary). Select the
best-pruned tree for scoring and display the full-grown, best-pruned, and
minimum error trees.

a. How many leaf nodes are in the best-pruned tree and minimum error tree?
b. Display the best-pruned tree. What are the predictor variable and split value for
the first split of the best-pruned tree?
c. What are the RMSE and MAD of the best-pruned tree on the test data?
d. Score the three NBA players Merrick is trying to sign as ACE Sports
Management clients in the NBA_Score worksheet using the pruned tree. What
is the average predicted salary of the three players?

43. FILE NBA. Refer to the previous exercise for a description of the data set.

Build a default regression tree to predict an NBA player’s salary (salary). Display
the regression tree.
a. What are the predictor variable and split value for the first split of the default
regression tree?
b. Build a full-grown tree. Which cp value is associated with the lowest cross-
validation error? How many leaf nodes are in the minimum-error tree?
c. Is there a simpler tree with a cross-validation error that is within one standard
error of the minimum error? If there is, then which cp value is associated with
the best-pruned tree?
d. Prune the full tree to the best-pruned tree or the minimum error tree if the
answer to part c is “No.” Display the tree. What are the rules that can be derived
from the pruned tree?
e. What are the ME, RMSE, MAE, MPE, and MAPE of the pruned tree on the
validation data?
 f. Score the three NBA players Merrick is trying to sign as ACE Sports
Management clients in the NBA_Score worksheet using the pruned tree. What
is the average predicted salary of the three players?

10.4 ENSEMBLE TREE MODELS

LO 10.3
Apply ensemble tree models to classify new records.

While decision trees are a highly popular classification and prediction

technique, they are sensitive to changes in the data. Even small
changes in the data may render drastically different trees and
results. Moreover, a single tree model tends to overfit, meaning that
it fits the current data well but fails to generalize to unseen data sets.
One solution to these problems is to use ensemble tree models,
which combine multiple single-tree models into an ensemble model.
Ensemble trees have gained increasing popularity and have been
proven to produce better and more stable predictive performance
than a single-tree model in many situations.
The principle behind the ensemble trees is that the outcomes of a
collection of single trees are combined into a single classification or
prediction model. In other words, a group of relatively weak single-
tree models (often called weak learners) are combined to form a
stronger ensemble tree model to reduce the variation in prediction
error. For classification, we use the majority rule where the most
frequently predicted class from all the single-tree models is chosen
as the final classification. For prediction, we use the average of the
predictions from all the single-tree models. Sometimes, more
sophisticated techniques such as weighted majority vote or weighted
average can be used by giving more weight to better-performing
trees and may produce more accurate predictions. Unlike a single
tree, an ensemble tree model cannot be displayed visually using a
tree diagram as multiple trees are combined to construct the
ensemble model.
Three common strategies are used when creating ensemble
models: bagging, boosting, and random forest. Bagging uses the
bootstrap aggregation technique to create multiple training data sets
by repeatedly sampling the original data with replacement. Each
training set is then used to construct a single page 455
decision tree. Bagging helps improve the stability of
the tree model by combining the results from single-tree models
produced by different training data.
The second ensemble strategy is called boosting. The boosting
strategy also uses repeated sampling with replacement to generate
multiple single-tree models. However, it is an iterative and sequential
process that pays more attention to cases that are misclassified or
have large prediction errors. In each iteration, cases that are
misclassified or have large prediction errors from the previous round
are given a higher probability (or more weight) of being included in
the next sample. As a result, the individual single-tree models that
initially have high classification or prediction error rates will
eventually adjust to make better classifications or predictions on
these difficult cases, and therefore the model performance will
improve over multiple iterations.
The key difference between the bagging and boosting strategies
is that, with bagging, every case has an equal probability of being
included in the sample, but with boosting, the probability of inclusion
increases for cases that have been previously misclassified or have
large prediction errors. The advantage of the boosting strategy is
that it forces the modeling process to place emphasis on the most
difficult cases, but at the same time, it is more prone to overfitting.
Random forest, also called random trees, is an extension of the
bagging strategy. With the bagging strategy, the single trees are
created from the same pool of predictor variables in the training set
and, therefore, tend to be very similar to each other. As a result,
single trees generated from the bagging method and their
predictions may be highly correlated to each other. Averaging
predictions from highly correlated trees may not result in a
substantial decrease in the variance of prediction errors. To
overcome this issue, the random forest strategy implements not only
repeated sampling of the training data, but also a random selection
of a subset of predictor variables, called features, to construct each
tree.
The random forest technique is particularly useful if the predictor
variables are highly correlated. This approach results in a greater
diversity and less-correlated single-tree models, which can
substantially reduce the variance in the prediction errors. As a
guideline, the square root of the total number of predictor variables is
used as the number of random features to select for each tree. For
example, if there are 16 predictor variables in the data, each tree
would randomly select features to be included in the tree. The

feature importance information allows you to decide which predictor

variables should be excluded from future analyses as they do not
contribute much to the classification or prediction process.

ENSEMBLE TREE MODELS

Ensemble tree models combine multiple single-tree models to
reduce the variation in prediction error. For classification,
majority rule is often used, where the most frequently predicted
class from all the single-tree models is chosen as the final
classification. For prediction, the average of the predictions
from all the single-tree models is often used. Common
ensemble tree strategies include bagging, boosting, and
random forest.
Bagging is an ensemble modeling strategy that uses the
bootstrap aggregation technique to create multiple training
data sets by repeatedly sampling the original data with
replacement.
 Boosting is an ensemble modeling strategy that forces the
model to pay more attention to cases that are misclassified or
have large prediction errors in previous trees through a
weighted sampling process.
Random forest is an ensemble modeling strategy that creates
a collection of trees by repeatedly sampling the original data
with replacement and randomly selecting a subset of predictor
variables, called features, to construct each tree.

page 456

In summary, bagging is easy to implement, and trees can be

developed in parallel. Boosting focuses on cases that are difficult to
classify or predict. However, boosting cannot develop trees in
parallel (weights depend on prior trees); thus, it is more
computationally expensive and may lead to overfitting. Random
forest provides the simplicity and computational efficiency of bagging
and adds more diversity to the process. Finally, because multiple
single-tree models are generated, pruning is usually not necessary
for the ensemble model. As such, we only need to partition the data
into training and validation sets. Without the need for a test data set,
this approach increases the amount of data available for sampling
and training the model. Unfortunately, ensemble trees cannot be
displayed graphically using the tree diagram; therefore, they are
more difficult to interpret than single trees. Whether or not ensemble
tree models perform better than single-tree models will depend on
the data, the predictor variables, and other factors unique to each
problem. In most cases, ensemble models lead to more accurate
and reliable predictions. In Example 10.8, we demonstrate how to
use Analytic Solver and R to develop ensemble classification tree
models; ensemble prediction tree models can be developed similarly.
Using Analytic Solver and R to Develop
Ensemble Classification Tree Models
FILE
HELOC

EXAMPLE 10.8
Use Analytic Solver and R to develop ensemble classification
tree models using the HELOC data set used in Example 10.5.
Compare the model performance of the ensemble tree model
with the single-tree model developed in Example 10.5.

SOLUTION: We will first provide detailed instructions for the

bagging technique. The necessary adjustments to develop the
boosting and the random forest models are given at the end of
the example.
Using Analytic Solver

a. Open the HELOC_Data worksheet of the HELOC data file.

b. Choose Data Mining > Partition (under the Data Mining
group) > Standard Partition.
c. In the Standard Data Partition dialog box, we select the data
range $A$1:$D$501 and move Age, Sex, Income, and
HELOC to the Selected Variables box. As the ensemble tree
modeling process does not involve pruning, only the train and
validation data sets are needed for constructing page 457
the model and evaluating the model’s
performance. Select Pick up rows randomly and check the Set
seed checkbox to use the default random seed of 12345.
Select Specify percentages to allocate 60% and 40% of the
data to training and validation sets, respectively. Click OK.
d. With the STDPartition worksheet active, choose Classify >
Ensemble > Bagging. In the Data tab, move the predictor
 variables Age and Income to the Selected Variables box and
Sex to the Categorical Variables box. Select and move
HELOC to the Output Variable box. Accept other defaults.
Click Next.
e. In the Parameters tab, select Decision Tree from the pull-
down menu for Weak Learner. Accept other defaults and click
Next. Analytic Solver will construct 10 decision trees as the
weak learners and uses the bagging technique for the
ensemble tree model.
f. In the Scoring tab, check the Summary Report boxes for
Score Training Data and Score Validation Data. Also check
the Lift Charts box for Score Validation Data. Check the In
Worksheet check box under the Score New Data section.
Under the New Data (WS) tab, select HELOC_Score in the
Worksheet box. These 20 records of new bank customers in
Data Range $A$1:$C$21 will be scored by the ensemble tree
model. Make sure the First Row Contains Headers box is
checked and click the Match By Name button. Accept other
defaults and click Finish. You will notice that it takes slightly
longer for Analytic Solver to return the output of the ensemble
tree model as it needs to create 10 single classification trees
before constructing the final ensemble tree model.
To assess model performance, we turn to Validation:
Classification Summary in the CBagging_validationScore
worksheet shown in Figure 10.26. The performance measures
show that the ensemble model yields an overall
misclassification rate of 21.5%, sensitivity of 0.6735, and
specificity of 0.8212. Recall that the single classification tree in
Example 10.5 has a misclassification rate of 23%, sensitivity of
0.5385, and specificity of 0.8514. The current ensemble tree
model has slightly better overall predictive performance, a
higher sensitivity rate, and performs better at correctly
identifying Class 1 cases (customers who respond to the
HELOC offer) than the single-classification tree model.
FIGURE 10.26 Analytic Solver’s summary measures

Source: Microsoft Excel

Recall that the performance measures in Figure 10.26 are

based on the cutoff value of 0.5. Lowering the cutoff value will
classify more cases into the target class, resulting in different
performance measurement values. Therefore, in order to
evaluate model performance, it is recommended that we
inspect the cumulative lift chart, the decile-wise lift chart, and
the ROC curve because they are independent of the cutoff
value.
Figure 10.27a shows the cumulative lift chart. Because the
blue line lies well above the red diagonal line, the ensemble
tree performs considerably better in terms of predicting whether
or not a potential customer will respond to the HELOC offer,
compared to the baseline model. For example, if we send
HELOC offers to 100 bank customers (half of the validation
data sample size) who have the highest predicted probability of
responding to the offer, we will be able to reach 42 out of 49
(85.71%) customers who actually respond to the offer.
Alternately, if we send the offer to 100 randomly chosen
customers, we would reach only 50% of the customers who
actually responded to the offer.

FIGURE 10.27 Analytic Solver’s performance charts

 Figure 10.27b shows the decile-wise lift chart. The first 10%
of potential customers selected by the ensemble tree, for
example, include 1.84 times as many customers who would
respond to the HELOC offer, compared to a random selection
of 10% of potential customers.
 Finally, the ROC curve (see Figure 10.27c) shows that the
ensemble tree performs much better than the baseline model in
terms of sensitivity and specificity across all cutoff values. The
area under the curve (AUC) is 0.8008, which is closer to the
optimum classifier (AUC = 1) as compared to the random
classifier (AUC = 0.5). Overall, the ensemble model
demonstrates slightly better predictive page 458
performance than that of the single-tree
classification model produced in Analytic Solver; refer to
Example 10.5 for comparing measures such as sensitivity,
specificity, and AUC.

In the CBagging_NewScore worksheet, the 20 new bank

customers are scored, and the results are summarized in Table
10.13. The Prediction: HELOC column displays the
classification result for each of the 20 records. The first
potential customer on the list is classified as not likely to
respond to the HELOC offer, while the second page 459
potential customer is likely to respond. The
PostProb:0 and PostProb:1 columns provide the predicted
probabilities of the customer belonging to Class 0 and Class 1,
respectively.

TABLE 10.13 Analytic Solver’s Scoring Results

To create an ensemble tree model using the boosting strategy,

repeat the previous procedure while choosing Classify >
Ensemble > Boosting in step d. In the Parameters tab,
Analytic Solver offers three variants of the AdaBoost algorithm
that differ based on how training record weights are updated.
Accept all default settings. In this case, the boosting ensemble
model has slightly worse predictive performance than the
bagging ensemble model and the single classification tree in
Example 10.5, with an overall misclassification rate of 27.5%,
sensitivity of 0.3469, and specificity of 0.8477. The cumulative
lift chart, decile-wise lift chart, and ROC curve suggest the
same information. The AUC value for the boosting ensemble
tree is only 0.7590.
To create an ensemble tree model using the random forest
strategy, repeat the previous procedure while choosing
Classify > Ensemble > Random Trees in step d. In the
Parameters tab, the default number two for the Number of
Randomly Selected Features indicates that two predictor
variables will be randomly selected to construct each tree.
Check the box next to Show Feature Importance. Accept all
other default settings. The performance measures and charts
show that the random forest model performs almost the same
as the bagging tree model with an overall misclassification rate
of 21%, sensitivity of 0.6939, and specificity of 0.8212. The
AUC value for the random forest ensemble model is 0.8100.
The feature importance table shown in Table 10.14 indicates
that Age is the most important predictor of whether a bank
customer will respond to a HELOC offer, followed by Income.

TABLE 10.14 Analytic Solver’s Feature Importance Table

Feature Importance

Age 3.048356023

Income 1.874119349

Sex 0.285117909
 You probably have noticed that this result contradicts with
the single-tree model developed in Example 10.5, which finds
Sex to be the most important variable to distinguish responders
from nonresponders. The reason for the discrepancy is that the
student edition of Analytic Solver only allows you to create 10
weak learners, each of which randomly selects only two
predictor variables to construct the tree. The predictor variable
Sex may not have been sufficiently represented in the very
limited number of weak learners, resulting in a biased feature
importance assessment. As we will see in the R example, by
constructing an ensemble tree model with 100 weak learners,
we will be able to conduct a more reliable feature importance
analysis.
Using R

a. Import the data from the HELOC_Data worksheet of the

HELOC data file into a data frame (table) and label it myData.
The following instructions and results are based on R version
3.5.3. To replicate the results with newer versions of R, enter:
>suppressWarnings(RNGversion(“3.5.3”))
b. Install and load the caret, gains, pROC, and page 460
randomForest packages. Enter:
>install.packages(“caret”, dependencies = c(“Depends”,
“Suggests”))
>install.packages(“gains”)
>install.packages(“pROC”)
>install.packages(“randomForest”)
>library(caret)
>library(gains)
>library(pROC)
>library(randomForest)
c. Because the randomForest package requires that all
categorical variables be declared as factor variables explicitly,
 we use the as.factor function to convert the categorical target
variable, HELOC, and categorical predictor variable, Sex, into
factor variables. Enter:
>myData$HELOC <- as.factor(myData$HELOC)
>myData$Sex <- as.factor(myData$Sex)
d. We partition the data into training (60%) and validation (40%)
data sets. By setting the random seed to 1, you will generate
the same partitions as shown in this example. Enter:
>set.seed(1)
>myIndex <- createDataPartition(myData$HELOC, p=0.6, list
= FALSE)
>trainSet <- myData[myIndex,]
>validationSet <- myData[-myIndex,]
e. We first use the randomForest function to construct the
ensemble tree model using the bagging strategy. We set the
ntree option equal to 100, which tells the function to build 100
single-tree models. As there are three predictor variables in
the data, we set the mtry option equal to 3, which uses all
three predictor variables in each single-tree model; this also
indicates that we are using the bagging strategy. We set the
importance option equal to TRUE to produce the feature
importance information. Again, by setting the random seed to
1, you will get the same results as in this example. Enter:
>set.seed(1)
>bagging_tree <- randomForest(HELOC ~ ., data=trainSet,
ntree= 100, mtry = 3, importance = TRUE)
f. To display the feature importance information, use the
varImpPlot function to display feature importance graphically.
We set the type option equal to 1 to show the feature
importance in terms of the average decrease (or mean
decrease) in overall accuracy. Alternately, if we set the type
option equal to 2, then R uses the average decrease in the
Gini impurity index to compare the feature importance. As
 Figure 10.28 shows, the most important predictor variable is
Sex as the model suffers the most decrease in classification
accuracy if the predictor variable is dropped. The second most
important predictor variable is Age.

FIGURE 10.28 R’s feature importance plot

>varImpPlot(bagging_tree, type=1)
g. The following commands create the confusion matrix by
comparing the predicted class memberships and actual class
memberships of the validation data set. Enter:
>predicted_class <- predict(bagging_tree, validationSet)
>confusionMatrix(predicted_class, validationSet$HELOC,
positive = “1”)
page 461
The confusion matrix is displayed in Figure 10.29.
The performance measures show that the bagging ensemble
tree model has an overall accuracy rate of 80% on the
validation data. As shown earlier, the model is much better at
 classifying Class 0 cases correctly (specificity = 0.8851) than
at classifying Class 1 cases correctly (sensitivity = 0.5577).

FIGURE 10.29 R’s summary measures

h. The following commands create the cumulative page 462

lift chart, the decile-wise lift chart, and the ROC
curve. As the syntax for creating these graphs has been
discussed in previous sections, we will not repeat it here.
Enter:
>predicted_prob<- predict(bagging_tree, validationSet,
type=“prob”)
>validationSet$HELOC <-
as.numeric(as.character(validationSet$ HELOC))
>gains_table <- gains(validationSet$HELOC,
predicted_prob[,2])
>gains_table
># cumulative lift chart
>plot(c(0,
gains_table$cume.pct.of.total*sum(validationSet$HELOC)) ~
c(0, gains_table$cume.obs),xlab = “# cases”,
ylab=“Cumulative”, type =“l”)
>lines(c(0, sum(validationSet$HELOC)) ~ c(0,
dim(validationSet)[1]), col=“red”, lty=2)
># decile-wise lift chart
>barplot(gains_table$mean.resp/mean(validationSet$HELOC)
, names.arg=gains_table$depth, xlab=“Percentile”, ylab=“Lift”,
ylim= c(0, 3), main = “Decile- Wise Lift Chart”)
># ROC curve
>roc_object<- roc(validationSet$HELOC, predicted_prob[,2])
>plot.roc(roc_object)
># compute auc
>auc(roc_object)
The cumulative lift chart, the decile-wise lift chart, and the
ROC curve are displayed in Figure 10.30. Because the
findings are comparable to those found using Analytic Solver,
we will not repeat the analysis here. Overall, the bagging
ensemble tree model provides a slight improvement in the
predictive performance as compared to the single-tree
classification model produced in R in Example 10.5.

FIGURE 10.30 R's performance charts

i. Finally, to score the 20 new cases, we import the data from
the HELOC_Score worksheet of the HELOC data file into a
data frame (table) called myScoreData and use the predict
function to produce the predicted class memberships and
probabilities for the new cases using the bagging ensemble
tree. We first convert the categorical predictor variable, Sex, to
a factor variable so that it is consistent with the Sex variable in
the bagging tree model created in step e. Enter:
>myScoreData$Sex <- as.factor(myScoreData$Sex)
>predicted_class_score <- predict(bagging_tree,
myScoreData, type=“class”)
>predicted_class_score
>predicted_class_prob <- predict(bagging_tree, myScoreData,
type=“prob”)
>predicted_class_prob
The scoring results are summarized in Table 10.15. As before,
the probabilities predicted by the bagging ensemble tree
constructed in R and Analytic Solver differ due to the different
built-in algorithms used in the two software products.
TABLE 10.15 R’s Scoring Results

page 463
To create an ensemble tree model using the
random forest strategy, simply replace the R commands in
step e with the following commands:
>set.seed(1)
>randomforest_tree <- randomForest(HELOC ~ .,
data=trainSet, ntree= 100, mtry = 2, importance = TRUE)
As you can see, the only difference between the commands
for bagging and random forest is the number you specify for
the mtry option in the randomForest function. Because
random forest selects a subset of the predictor variables for
building each single-tree model, we specify the number of
predictor variables to include in the subset. In this case, we ask
the function to randomly select two predictor variables for
building individual single-tree models.
R uses the boosting function of the adabag package to
create ensemble tree models that implement the boosting
 strategy. Due to the technical requirements of the adabag
package, we create the boosting tree model slightly differently
as shown in the example below:
a. Import the data from the HELOC_Data worksheet of the
HELOC data file into a data frame (table) and label it myData.
b. Install and load the caret, gains, pROC, and page 464
adabag packages. Enter:
>install.packages(“caret”, dependencies = c(“Depends”,
“Suggests”))
>install.packages(“gains”)
>install.packages(“pROC”)
>install.packages(“adabag”)
>library(caret)
>library(gains)
>library(pROC)
>library(adabag)
c. Because the adabag package requires the use of data frame
class objects, we convert myData to a data frame class object
using the data.frame function. We convert categorical
variables (i.e, Sex and HELOC) to factor variables as required
by the software. Enter:
>myData <- data.frame(myData)
>myData$HELOC <- as.factor(myData$HELOC)
>myData$Sex <- as.factor(myData$Sex)
d. We set the random seed to 1 and partition the data into
training (60%) and validation (40%) data sets. Enter:
>set.seed(1)
>myIndex <- createDataPartition(myData$HELOC, p=0.6, list
= FALSE)
>trainSet <- myData[myIndex,]
>validationSet <- myData[-myIndex,]
e. We use the boosting function to construct the ensemble tree
model using the boosting strategy. By setting the random seed
to 1, you will get the same results as in this example. The
mfinal option specifies the number of weak learners (single-
tree models) to create. Enter:
>set.seed(1)
>boosting_tree <- boosting(HELOC ~ ., data=trainSet, mfinal
= 100)
f. The following commands create the confusion matrix by
comparing the predicted class memberships and actual class
memberships of the validation data set. For boosting tree
models, the predict function produces a list (named prediction
here), which includes predicted class memberships and
probabilities. You can access the predicted class
memberships and predicted probabilities using
prediction$class and prediction$prob, respectively.
>prediction <- predict(boosting_tree, validationSet)
>confusionMatrix(as.factor(prediction$class),
validationSet$HELOC, positive = “1”)
Verify that the accuracy rate, sensitivity, and specificity are
0.8, 0.5577, and 0.8851, respectively. Using the default cutoff
value of 0.5, the boosting ensemble tree shows the same
predictive performance as the bagging ensemble tree does.
g. The following commands create the cumulative lift chart, the
decile-wise lift chart, and the ROC curve. As the syntax for
creating these graphs has been discussed in previous
sections, we will not repeat it here. Note that we access the
predicted probability of a validation case belonging the Class 1
(target class) using prediction$prob[, 2] because the Class 1
probabilities are listed in the second column of the object.
Enter:
>validationSet$HELOC <-
as.numeric(as.character(validationSet$HELOC))
 >gains_table <- gains(validationSet$HELOC,
prediction$prob[,2])
page 465
>gains_table
>#cumulative lift chart
>plot(c(0,
gains_table$cume.pct.of.total*sum(validationSet$HELOC)) ~
c(0, gains_table$cume.obs),xlab = “# cases”,
ylab=”Cumulative”, type =”l”)
>lines(c(0, sum(validationSet$HELOC)) ~ c(0,
dim(validationSet)[1]), col=”red”, lty=2)
># decile-wise lift chart
>barplot(gains_table$mean.resp/mean(validationSet$HELOC)
, names.arg=gains_table$depth, xlab=”Percentile”, ylab=”Lift”,
ylim= c(0, 3), main = “Decile-Wise Lift Chart”)
># ROC curve
>roc_object<- roc(validationSet$HELOC, prediction$prob[,2])
>plot.roc(roc_object)
># compute auc
>auc(roc_object)
Verify that the AUC value of the ROC curve is 0.8209, which is
slightly lower than the AUC value of the ROC curve derived
from the bagging ensemble tree.
h. Finally, to score the 20 new cases, we import the data from
the HELOC_Score worksheet of the HELOC data file into a
data frame (table) and label it myScoreData. Convert
myScoreData to a data frame class object and predictor
variable Sex to a factor variable as required by the software.
Enter:
>myScoreData <- data.frame(myScoreData)
>myScoreData$Sex <- as.factor(myScoreData$Sex)
 >predicted_class_score <- predict(boosting_tree,
myScoreData)
>predicted_class_score$class
>predicted_class_score$prob
The scoring results of the boosting ensemble model are
slightly different from those of the bagging ensemble model;
the 20th case is classified as a Class 0 case by the boosting
model, whereas it is classified as a Class 1 case by the
bagging model. As discussed before, unlike the bagging and
random forest models, the boosting model does not provide
variable importance information.

EXERCISES 10.4
Note: These exercises can be solved using Analytic Solver
and/or R. The answers, however, will depend on the software
package used. For Analytic Solver, partition data sets into 60%
training and 40% validation, and use 12345 as the default
random seed. Create 10 weak learners when constructing the
ensemble model. For R, partition data sets into 60% training
and 40% validation, and use the statement set.seed(1) to
specify the random seed for data partitioning and constructing
the ensemble tree models. Create 100 weak learners when
constructing the ensemble model. All answers in R are based
on version 3.5.3. To replicate the results with newer versions of
R, execute the following line of code at the beginning of the R
session: suppressWarnings(RNGversion(“3.5.3”)). If the
predictor variable values are in the character format, then treat
the predictor variable as a categorical variable. Otherwise, treat
the predictor variable as a numerical variable.

Mechanics
44. FILE Exercise_10.44. Create a bagging ensemble classification tree model
using the accompanying data set (predictor variables: x1 to x4; target: y).
a. What are the overall accuracy rate, sensitivity, and specificity of the model on
the validation data?
b. What is the AUC value of the model? page 466
c. Score a new record with the following values: x1 = 3.45, x2 =
1, x3 = 18, x4 = 5.80. How would the record be classified using the bagging
ensemble tree model? What is the probability of the record belonging to Class
1?
45. FILE Exercise_10.45. Create a boosting ensemble classification tree model
using the accompanying data set (predictor variables: x1 to x4; target: y).
a. What are the overall accuracy rate, sensitivity, and specificity of the model on
the validation data?
b. What is the AUC value of the model?
c. Score a new record with the following values: x1 = 3.45, x2 = 1, x3 = 18, x4 =
5.80. How would the record be classified using the boosting ensemble tree
model? What is the probability of the record belonging to Class 1?
46. FILE Exercise_10.46. Create a random forest ensemble classification tree
model using the accompanying data set (predictor variables: x1 to x4; target: y).
Select two predictor variables randomly to construct each weak learner.
a. What are the overall accuracy rate, sensitivity, and specificity of the model on
the validation data?
b. What is the AUC value of the model?
c. Which is the most important predictor variable?
d. Score a new record with the following values: x1 = 3.45, x2 = 1, x3 = 18, x4 =
5.80. How would the record be classified using the random forest ensemble
model? What is the probability of the record belonging to Class 1?
47. FILE Exercise_10.47. Create a bagging ensemble classification tree model
using the accompanying data set (predictor variables: x1 to x5; target: y).
a. What are the overall accuracy rate, sensitivity, and specificity of the model on
the validation data?
b. What is the lift value of the leftmost bar of the decile-wise lift chart?
c. Score a new record with the following values: x1 = 52.8, x2 = 230.50, x3 = 1, x4
= 144, x5 = 6.23. How would the record be classified using the bagging
ensemble tree model? What is the probability of the record belonging to Class
1?
48. FILE Exercise_10.48. Create a boosting ensemble classification tree model
using the accompanying data set (predictor variables: x1 to x5; target: y).
a. What are the overall accuracy rate, sensitivity, and specificity of the model on
the validation data?
 b. What is the lift value of the leftmost bar of the decile-wise lift chart?
c. Score a new record with the following values: x1 = 52.8, x2 = 230.50, x3 = 1, x4
= 144, x5 = 6.23. How would the record be classified using the boosting
ensemble tree model? What is the probability of the record belonging to Class
1?
49. FILE Exercise_10.49. Create a random forest ensemble classification tree
model using the accompanying data set (predictor variables: x1 to x5; target: y).
Select two predictor variables randomly to construct each weak learner.
a. What are the overall accuracy rate, sensitivity, and specificity of the model on
the validation data?
b. What is the lift value of the leftmost bar of the decile-wise lift chart?
c. Which is the most important predictor variable?
d. Score a new record with the following values: x1 = 52.8, x2 = 230.50, x3 = 1, x4
= 144, x5 = 6.23. How would the record be classified using the random forest
ensemble model? What is the probability of the record belonging to Class 1?

Applications
50. FILE Spam. Mateo Derby works as a cyber security analyst at a private equity
firm. His colleagues at the firm have been inundated by a large number of spam
e-mails. Mateo has been asked to implement a spam detection system on the
company’s e-mail server. He reviewed a sample of 500 spam and legitimate e-
mails with relevant variables: spam (1 if spam, 0 otherwise), the number of
recipients, the number of hyperlinks, and the number of characters in the
message. A portion of the Spam_Data worksheet is shown in the accompanying
table.

Spam Recipients Hyperlinks Characters

0 19 1 47

0 15 1 58

⋮ ⋮ ⋮ ⋮

1 13 2 32

page 467
a. Create a bagging ensemble classification tree model to
determine whether a future e-mail is spam. What are the overall accuracy rate,
sensitivity, and specificity of the model on the validation data? What is the AUC
value of the model?
 b. Create a random forest ensemble classification tree model. Select two predictor
variables randomly to construct each weak learner. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
What is the AUC value of the model? Which is the most important predictor
variable?
c. Score the new cases in the Spam_Score worksheet using the bagging
ensemble classification tree model. What percentage of the e-mails is spam?
51. FILE HR. Daniella Lara, a human resources manager at a large tech consulting
firm, has been reading about using analytics to predict the success of new
employees. With the fast-changing nature of the tech industry, some employees
have had difficulties staying current in their field and have
missed the opportunity to be promoted into a management
page 468
position. Daniella is particularly interested in whether or not a new employee is
likely to be promoted into a management role after 10 years with the company.
She gathers information on current employees who have worked for the firm for at
least 10 years. The information is based on the job application that the employees
provided when they originally applied for a job at the firm. For each employee, the
following variables are listed: Promoted (1 if promoted within 10 years, 0
otherwise), GPA (college GPA at graduation), Sports (number of athletic activities
during college), and Leadership (number of leadership roles in student
organizations). A portion of the HR_Data worksheet is shown in the
accompanying table.

Promoted GPA Sports Leadership

0 3.28 0 2

1 3.93 6 3

⋮ ⋮ ⋮ ⋮

0 3.54 5 0

a. Create a bagging ensemble classification tree model to determine whether a

new employee is likely to be promoted into a management role after 10 years
with the company. What are the overall accuracy rate, sensitivity, and specificity
of the model on the validation data? What is the AUC value of the model?
b. Create a random forest ensemble classification tree model. Select two predictor
variables randomly to construct each weak learner. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
What is the AUC value of the model? Which is the most important predictor
variable?
c. Score the new cases in the HR_Score worksheet using the random forest
ensemble classification tree model. What is the probability of the first new
employee being promoted within 10 years? How many new employees in the
 data set will likely be promoted within 10 years based on a cutoff probability
value of 0.5?
52. FILE Heart. In recent years, medical research has incorporated the use of data
analytics to find new ways to detect heart disease in its early stage. Medical
doctors are particularly interested in accurately identifying high-risk patients so
that preventive care and intervention can be administered in a timely manner. A
set of information is available regarding a patient’s age (Age), blood pressure (BP
Systolic and BP Diastolic), and BMI, along with an indicator of whether or not the
patient has heart disease (Disease = 1 if heart disease, 0 otherwise). A portion of
the data set is provided in the accompanying table.

a. Create a bagging ensemble classification tree model to determine whether a

patient has heart disease. What are the overall accuracy rate, sensitivity, and
specificity of the model on the validation data?
b. Create a boosting ensemble classification tree model. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
c. Compare the two ensemble models. Which model shows more robust
performance according to the AUC value?
53. FILE Solar. New Age Solar sells and installs solar panels for residential homes.
The company’s sales representatives contact and pay a personal visit to potential
customers to present the benefits of installing solar panels. This high-touch
approach works well as the customers feel that they receive personal services
that meet their individual needs, but it is more expensive than other, mass-
marketing approaches. The company wants to be very strategic about visiting
potential customers who are more likely to install solar panels. The company has
compiled a data set of past home visits by sales reps. The data include the age
(Age in years) and annual income (Income, in $1,000s) of the potential customer
and whether or not the customer purchases the solar panels (Install: Y/N). A
portion of the Solar_Data worksheet is shown in the accompanying table.

Income Age Install

115 45 N

68 31 Y

⋮ ⋮ ⋮
 Income Age Install

73 34 N

a. Create a bagging ensemble classification tree model to determine whether a

customer purchases the solar panels. What are the overall accuracy rate,
sensitivity, and specificity of the model on the validation data?
b. Create a boosting ensemble classification tree model. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
c. Compare the two ensemble models. Which model shows more robust
performance according to the lift value of the leftmost bar of the decile-wise lift
chart?
d. Score the new cases in the Solar_Score worksheet using the bagging
ensemble classification tree model. What percentage of the potential customers
are likely to purchase the solar panels based on a cutoff probability value of
0.5?
54. FILE MedSchool. Admission to medical school in the United States is highly
competitive. The acceptance rate to the top medical schools could be as low as
2% or 3%. With such a low acceptance rate, medical school
admissions consulting has become a growing business in many
page 469
cities. In order to better serve his clients, Paul Foster, a medical school
admissions consultant, wants to build a data-driven model to predict whether or
not a new applicant is likely to get accepted into one of the top 10 medical
schools. He collected a database of 1,992 past applicants to the top 10 medical
schools with the following information: Sex (F = female, M = male), CollegeParent
(1 if parents with college degrees, 0 otherwise), GPA (1 if undergraduate GPA of
3.50 or higher, 0 otherwise), and Med (1 if accepted to a top 10 medical school, 0
otherwise). A portion of the data set is shown in the accompanying table.

Sex CollegeParent GPA Med

F 1 1 1

M 1 0 1

⋮ ⋮ ⋮ ⋮

M 0 0 0

a. Create a bagging ensemble classification tree model to determine whether a

new applicant is likely to get accepted into one of the top 10 medical schools.
What are the overall accuracy rate, sensitivity, and specificity of the model on
the validation data? What is the AUC value of the model?
 b. Create a random forest ensemble classification tree model. Select two predictor
variables randomly to construct each weak learner. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
Which is the most important predictor variable?
c. Compare the two ensemble models. Which model shows more robust
performance according to the AUC value?
55. FILE Fraud. Credit card fraud is becoming a serious problem for the financial
industry and can pose a considerable cost to banks, credit card issuers, and
consumers. Fraud detection using data mining techniques has become an
indispensable tool for banks and credit card companies to combat fraudulent
transactions. A sample credit card data set contains the following variables: Fraud
(1 if fraudulent activities, 0 otherwise), Amount (1 if low, 2 if medium, 3 if high),
Online (1 if online transactions, 0 otherwise), and Prior (1 if products that the card
holder previously purchased, 0 otherwise). A portion of the data set is shown in
the accompanying table.

Fraud Amount Online Prior

0 2 0 1

0 3 0 0

⋮ ⋮ ⋮ ⋮

0 2 0 1

a. Create a bagging ensemble classification tree model to determine whether a

transaction is fraudulent. What are the overall accuracy rate, sensitivity, and
specificity of the model on the validation data? What is the AUC value of the
model?
b. Create a random forest ensemble classification tree model. Select two predictor
variables randomly to construct each weak learner. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
What is the AUC value of the model? Which is the most important predictor
variable?
c. Compare the two ensemble models. Which model shows more robust
performance according to the lift value of the leftmost bar of the decile-wise lift
chart?
56. FILE Travel_Plan. Refer to Exercise 11 for a description of the data set. Partition
the data into 60% training and 40% validation data. For Analytic Solver, use
12345 as the random seed and create 10 weak learners. For R, use one as the
random seed and create 100 weak learners.
a. Create a bagging ensemble classification tree model to determine whether a
customer will travel within the next year. What are the overall accuracy rate,
 sensitivity, and specificity of the model on the validation data? What is the AUC
value of the model?
b. Compare the performance of the bagging ensemble model to that of the single-
tree model created in Exercise 11 (for Analytic Solver) or Exercise 12 (for R).
Which model shows more robust performance? Explain.
c. Score the two new customers in the Travel_Plan_Score worksheet using the
bagging ensemble classification tree model. What is the probability of the first
customer having plans to travel within the next year? What is the probability for
the second customer?
57. FILE Continue_Edu. Refer to Exercise 13 for a description of the data set.
a. Create a boosting ensemble classification tree model. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
What is the AUC value of the model?
b. Compare the performance of the boosting ensemble model to that of the single-
tree model created in Exercise 13 (for R) or Exercise 14 (for Analytic Solver).
Which model shows more robust performance? Explain.
c. Score the two new cases in the Continue_Edu_Score worksheet using the
boosting ensemble classification tree model. What is the probability of the first
person enrolling in continuing education classes? What is the probability for the
second person?
58. FILE Church. Refer to Exercise 15 for a description of the data set.
a. Create a random forest ensemble classification tree model. Select two predictor
variables randomly to construct each weak learner. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
What is the AUC value of the model? Which is the most important predictor
variable?
b. Compare the performance of the random forest ensemble model to that of the
single-tree model created in Exercise 15 (for Analytic Solver) or Exercise 16 (for
R). Which model shows more robust performance? Explain.
c. Score the new cases in the Church_Score worksheet using the random forest
ensemble classification tree model. What percentage of the individuals in the
score data set are likely to go to church based on a cutoff probability value of
0.5?
59. FILE Graduate. Refer to Exercise 19 for a description of the data set.
a. Create a random forest ensemble classification tree model. Select three
predictor variables randomly to construct each weak learner. What are the
overall accuracy rate, sensitivity, and specificity of the model on the validation
data? What is the AUC value of the model? Which is the most important
predictor variable?
b. Compare the performance of the random forest ensemble model to that of the
single-tree model created in Exercise 19 (for R) or Exercise 20 (for Analytic
Solver). Which model shows more robust performance? Explain.
c. Score the new cases in the Graduate_Score worksheet using the random forest
ensemble classification tree model. What is the probability of the first student
 graduating within four years? What is the probability for the second and third
students?
60. FILE In_App_Pur. Refer to Exercise 21 for a description of the data set.
a. Create a boosting ensemble classification tree model. What are the overall
accuracy rate, sensitivity, and specificity of the model on the validation data?
What is the AUC value of the model?
b. Compare the performance of the boosting ensemble model to that of the single-
tree model created in Exercise 21 (for Analytic Solver) or Exercise 22 (for R).
Which model shows more robust performance? Explain.
c. Score the new cases in the In_App_Pur_Score worksheet using the boosting
ensemble classification tree model. What is the probability of the first gamer
making in-app purchases? What is the probability for the second gamer?

10.5 WRITING WITH BIG DATA

FILE
  NLS

Case Study
As millions of people in the U.S. are crippled by student loan debt and high
unemployment, policymakers are raising the question of whether college is even a
good investment. Richard Clancy, a sociology graduate student, is interested in
developing a model for predicting an individual’s income for his master’s thesis. He
found a rich data set maintained by the U.S. Bureau of Labor Statistics, called the
National Longitudinal Surveys (NLS), that follows over 12,000 individuals in the
United States over time. The data set focuses on labor force activities of these
individuals, but also includes information on a wide range of variables, such as
income, education, sex, race, personality trait, health, and marital history.

Sample Report—Predicting Personal Income

According to the U.S. Bureau of the Census, the median personal income in
2000 was $29,998. Because an individual’s personal income is determined by
a large number of factors, it would be useful to build a model that helps
predict whether an individual will earn an income that is at or above the
median personal income. Based on the literature concerning personal income
and data available from the U.S. Bureau of Labor Statistics, the following
predictor variables are used to explain an individual’s personal income:

Rawpixel.com/Shutterstock

page 470

Years of education
Mother’s years of education
Father’s years of education
Urban (equals 1 if the individual lived in an urban area
at the age of 14, 0 otherwise)
Black (equals 1 if the individual is black, 0 otherwise)
Hispanic (equals 1 if the individual is Hispanic, 0
otherwise)
White (equals 1 if the individual is white, 0 otherwise)
Male (equals 1 if the individual is male, 0 otherwise)
Self-Esteem (the individual’s self-esteem using the
Rosenberg Self-Esteem Scale; a higher score indicates
a higher self-esteem.)
Outgoing kid (equals 1 if the individual was outgoing at
the age of six, 0 otherwise)
 Outgoing adult (equals 1 if the individual is outgoing as
an adult, 0 otherwise)
As the objective is to determine whether an individual has an income that is at
or above the median personal income rather than an individual’s actual
income, the target variable income is converted into a categorical variable
that assumes the value one if income is greater than or equal to $29,998 and
0 otherwise. The final data set includes 5,821 observations, after
observations with missing values were removed from the analysis.
To build a predictive model and assess the model’s performance, the data
are partitioned into training (60%) and validation (40%) sets. As the predictor
variables include both numerical and categorical data types, the decision tree
methodology is a suitable technique to build a classification model for this
application. Figure 10.31 shows the best-pruned classification tree with eight
decision nodes and nine leaf nodes. A number of conclusions can be drawn
from the classification tree that shows that personal income level can be
predicted using an individual’s sex, education, race, and
mother’s education. For example, if an individual is male
page 471
and had more than 16 years of education, he has a high probability of earning
an income that is at or above the U.S. median income. If a person is female
and has less than 16 years of education, she has a high probability of earning
an income that is below the U.S. median income.

FIGURE 10.31 Best-pruned classification tree

 The performance of the classification tree is evaluated based on how
accurately the model classifies cases in the validation data. The results from
the confusion matrix, shown in Table 10.16, can be used to calculate an
overall accuracy rate of 69.5%, sensitivity of 0.5994, specificity of 0.7730, and
precision of 0.6830. Overall, the model demonstrates reasonably good
performance in predicting an individual’s income level.

TABLE 10.16 Confusion Matrix

Predicted Class

Actual Class Class 1 Class 0

 Predicted Class

Actual Class Class 1 Class 0

Class 1 627 291

Class 0 419 991

In an attempt to improve the predictive performance, a random forest

ensemble model is also developed. The ensemble model shows an accuracy
rate of 69.72%, sensitivity of 0.6138, specificity of 0.7652, and precision of
0.6808, which represents only a slight improvement over the single-tree
model. Because the single-tree classification model is easier to interpret than
the ensemble model, a single decision tree model is used as the final
predictive model.
The conclusions drawn from the classification tree in Figure 10.31 are not
surprising. For example, it has been well documented that higher education
leads to a higher income. Similarly, while the salary difference based on sex
and race has shrunk over the years, it still persists. According to a Pew
Research Center analysis, women earned 85% of what men earned in 2018
and households headed by a black person earned, on average, little more
than half of what the average white household earned.
What is surprising, however, is that an individual’s self-esteem level and
whether a person is outgoing as a kid or adult do not impact an individual’s
income level. Perhaps these psychological factors are related to multiple
indicators of later-year academic achievement. Therefore, an individual with
high self-esteem may go for higher education, which in turn impacts the
individual’s income.
The classification model developed in this report offers important and
actionable insights for policymakers. The model highlights that sex and racial
gaps continue to exist. Males and nonblack people are more likely to earn an
above-median income than females and black people do. Special attention
should be given to groups identified by the model as likely to earn below-
median income when developing equitable policies for improving economic
prospects.

Suggested Case Studies

Decision trees are considered to be one of the most transparent and easiest to
interpret supervised data mining techniques. They can be applied to many situations
and different data sets. Here are some suggestions using the big data that
accompany this text.

Report 10.1 FILE Car_Crash. Subset the data to include only the accidents that
occurred in one city or during one month. Develop a decision tree model that predicts
whether an automobile accident results in fatal or severe injuries using predictor
variables, such as traffic violation category, weather condition, type of collision,
location of the accident, and lighting condition. Note: Many of the predictor variables
are categorical; therefore, you will need to convert them to the appropriate data form
prior to the analysis.

Report 10.2 FILE Longitudinal_Survey. Subset the data to include only those
individuals who lived in an urban area. Develop a decision tree model to predict an
individual’s body mass index (BMI) or whether an individual is going to be overweight
or not using predictor variables, such as sex, number of years of education, parents’
years of education, self-esteem scale, and whether the person is
outgoing as a kid and/or adult. Note: The Centers for Disease
page 472
Control and Prevention (CDC) defines overweight as someone whose BMI is equal
to or greater than 25.

Report 10.3 FILE TechSales_Reps. Subset the data to include only college-
educated sales professionals in the software product group. Develop a decision tree
model to predict whether or not the sales professional will receive a high (9 or 10) net
promoter score (NPS), using predictor variables, such as age, sex, tenure with the
company, number of professional certificates acquired, annual evaluation score, and
personality type. Note: You may need to perform data transformation in order to meet
the requirements of the analytics technique.

Report 10.4 FILE College_Admissions. Subset the data to include only one of the
three colleges. Develop a series of decision tree models to predict which college is
most likely to accept a given university applicant based on the applicant’s sex, race,
high school GPA, SAT/ACT score, and parents’ years of education. Note: You will
construct a decision tree model for each college. The college whose model produces
the highest probability of acceptance is the one that is most likely to accept the
applicant.

Report 10.5 FILE House_Price. Develop a decision tree model to predict the sale
price of a house by using predictor variables, such as number of bedrooms, number
of bathrooms, home square footage, lot square footage, and age of the house. Note:
As real estate values can differ dramatically from region to region, you may want to
develop a model for each geographic location in the data set. Compare the
performance of the decision tree model with predictive models that use other
supervised learning techniques discussed in Chapter 9.

page 473
 page 474

11 Unsupervised Data
Mining

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 11.1  Conduct hierarchical cluster analysis.
LO 11.2  Conduct k-means cluster analysis.
LO 11.3  Conduct association rule analysis.

s discussed in earlier chapters, data mining is a process of

A applying analytical techniques to find hidden structures,
patterns, and relationships among data. It is categorized as
supervised (or predictive or directed) or unsupervised (or descriptive
or undirected) data mining. An overview of data mining concepts
including topics such as distance measures between observations
(or records) and principal component analysis for dimension
reduction was provided in Chapter 8. The focus of Chapters 9 and
10 was on supervised data mining. In this chapter, we turn our
attention to unsupervised data mining.
 We explore two popular unsupervised data mining techniques:
cluster analysis and association rule analysis. Cluster analysis finds
similarities among data and groups them into clusters of
observations that share similar characteristics. Clusters are formed
in such a way that objects are similar within a group but dissimilar
across groups. This analysis is valuable, for example, to marketers
who identify distinct segments in their customer bases and then use
this knowledge to develop targeted marketing programs. Association
rule analysis (or market basket analysis) identifies hidden
relationships and common occurrences among data. This analysis is
valuable, for example, to e-commerce companies that identify
associations to make product recommendations to customers based
on their previous purchases.

page 475
©Lissandra Melo/Shutterstock

FILE
Candy_Bars

INTRODUCTORY CASE
Nutritional Facts of Candy Bars
Aliyah Williams is an honors student at a prestigious business
school in Southern California. She is also a fledgling
entrepreneur and owns a vending machine business. Most of
her vending machines are located on campus and in the
 downtown area, and they are stocked with a variety of snacks,
including a large selection of candy bars. Aliyah is aware that
California consumers are becoming increasingly health
conscious when it comes to food purchase. She has learned
from her consumer research class that the U.S. Department of
Agriculture (USDA) maintains a website with a database of
nutritional facts about food products, including candy bars.
Aliyah wants to come up with a better selection of candy
bars and strategically group and display them in her vending
machines. She also wants to feature certain products more
prominently in different locations based on the type of
consumers that tend to frequent these locations. Table 11.1
shows a portion of the data that Aliyah has downloaded from
the USDA website. Listed for each candy bar are the calories
per serving (Calories), total fat (Fat in grams), protein (Protein
in grams), and carbohydrates (Carb in grams).

TABLE 11.1 Candy Bar Nutritional Facts

Aliyah wants to use the information in Table 11.1 to:

1. Analyze the nutritional fact data and group candy products
according to their nutritional content.
2. Select a variety of candy bars from each group to better meet
the taste of today’s consumers.
3. Select and display the candy bars in her vending machines
according to the grouping and the locations of the vending
machines.
 A synopsis of this introductory case is presented at the end of
Section 11.2.

page 476

11.1 HIERARCHICAL CLUSTER

ANALYSIS
Data mining uses many kinds of machine learning algorithms to
identify patterns in data. These algorithms are classified into
supervised (or predictive or directed) and unsupervised (or
descriptive or undirected) data mining, depending on the way they
“learn” about data to make predictions. Supervised data mining, also
called supervised learning, is the more commonly used methodology
for developing predictive models. The two primary types of
supervised data mining applications, classification and prediction,
were discussed in Chapters 6, 7, 9, and 10. These models are used
to predict the outcome of an output (response or target) variable
based on several input (predictor) variables. In some supervised
techniques such as regression, a mathematical model that relates
the outcome of the response variable (commonly referred to as y) to
one or more predictor variables (commonly referred to as x1, x2, . . . ,
xk) is estimated. A regression model, for example, is therefore
“trained” or “supervised” because the known values of the response
variable are used to build a model in the training set. In addition, the
performance of the model can be evaluated based on the extent to
which the predicted values deviate from the actual values of the
response variable in the validation set.
In this chapter, we focus on unsupervised data mining, also
referred to as unsupervised learning, that requires no knowledge of
the response variable. It is called unsupervised because, unlike
supervised learning, the algorithms allow the computer to identify
complex processes and patterns without any specific guidance from
the analyst. Unsupervised learning is an important part of exploratory
data analysis because it makes no distinction between the response
variable y and the predictor variable x; in this chapter, we represent
all k variables as x1, x2, . . . , xk. Several machine learning algorithms
are readily available to implement two core unsupervised data
mining techniques: cluster analysis and association rule analysis.
In Chapter 8, we explored several similarity measures to gauge
whether a group of observations (or records) are similar or dissimilar
to one another. In particular, we used Euclidian and Manhattan
distance measures for numerical variables, which were often first
standardized or normalized to make them independent of scale. In
addition, matching and Jaccard’s coefficients were used to measure
similarity for categorical variables. These measures are necessary
for both supervised and unsupervised data mining techniques. For
example, in cluster analysis, similarity measures are used to form
clusters in such a way that objects are similar within a group but
dissimilar across groups.

Hierarchical Cluster Analysis

LO 11.1
Conduct hierarchical cluster analysis.

Grouping objects into categories that share some similar

characteristic or trait is an important task that, as humans, we do
almost every day. We all have the ability to recognize patterns or
common characteristics among objects and classify them into groups
or clusters. Because grouping is such an important function in our
everyday lives, cluster analysis is one of the most widely used
unsupervised data mining techniques. It groups observations
(records) in a data set such that observations are similar within a
cluster and dissimilar across clusters, using similarity measures
discussed in Section 8.2.
 Cluster analysis is often considered part of descriptive analytics
because it performs useful exploratory analysis by summarizing a
large number of observations in a data set into a small number of
homogeneous clusters. The cluster characteristics or profiles help us
understand and describe the different groups of observations that
exist in the data.
While cluster analysis can be performed on its own, it is
sometimes performed prior to supervised learning techniques to help
separate observations that display very different patterns. For
example, pricing luxury or collectible vehicles may require a very
different model from the one that is used to price economy vehicles.
Therefore, it is reasonable to cluster different types of vehicles into
different groups and develop pricing models for each group
separately to achieve better predictive performance.

page 477

CLUSTER ANALYSIS
Cluster analysis is an unsupervised data mining technique that
groups data into categories that share some similar
characteristic or trait.

A popular application of cluster analysis is called customer or

market segmentation, where companies analyze a large amount of
customer-related demographic and behavioral data and group the
customers into different market segments. For example, a credit card
company might group customers into those who pay off their account
balance every month versus those who carry a monthly balance,
and, within these two customer segments, group them further
according to their spending habits. The company would likely target
each of the customer segments with different promotion and
advertising campaigns or design different financial products for each
group.
 Cluster analysis allows us to form groups of people, products,
business transactions, and many other types of observations into
internally homogenous groups where each group has unique
characteristics and can be treated differently from other groups.
Similarly, insurance companies might want to group motor insurance
policy holders based on their characteristics such as age, annual
miles driven, and average claim cost, thus allowing them to set
optimal insurance premiums.
Two common clustering techniques are hierarchical clustering
and k-means clustering. We discuss hierarchical clustering in this
section and k-means clustering in the next section. Hierarchical
clustering is a technique that uses an iterative process to group
data into a hierarchy of clusters. Common strategies of hierarchical
clustering usually follow one of two methods: agglomerative
clustering (or agglomerative nesting or AGNES) or divisive-
clustering (or divisive analysis or DIANA). AGNES starts with each
observation being its own cluster, with the algorithm iteratively
merging clusters that are similar to each other as one moves up the
hierarchy; it is loosely referred to as “the bottom-up” approach.
Conversely, DIANA starts by assigning all observations to one
cluster, with the algorithms iteratively separating the most dissimilar
observations as one moves down the hierarchy; it is loosely referred
to as a “top-down” approach.

HIERARCHICAL CLUSTERING
Hierarchical clustering is a technique that uses an iterative
process to group data into a hierarchy of clusters. Common
strategies of hierarchical clustering usually follow one of two
methods: agglomerative clustering or divisive clustering.
Agglomerative clustering is a “bottom-up” approach that
starts with each observation being its own cluster, with the
algorithm iteratively merging clusters that are similar to each
other as one moves up the hierarchy.
 Divisive clustering is a “top-down” approach that starts by
assigning all observations to one cluster, with the algorithms
iteratively separating the most dissimilar observations as one
moves down the hierarchy.

In this text, we focus on agglomerative clustering, which is the

more commonly used approach between the two methods for
hierarchical clustering; the computer applications for agglomerative
clustering can easily be modified to implement divisive clustering.
We first create agglomerative clusters with data on variables that are
either numerical or categorical. We then extend the analysis to
include mixed data, which combines data from both numerical and
categorical variables.

page 478

Agglomerative Clustering with Numerical or

Categorical Variables
As mentioned earlier, in AGNES, each observation in the data
initially forms its own cluster. The algorithm then successively
merges these clusters into larger clusters based on their similarity
until all observations are merged into one final cluster, referred to as
a root. We measure the (dis)similarity of observations using the
Euclidean distance or Manhattan distance measures for numerical
variables and matching or Jaccard’s coefficient for categorical
variables. Here, we use the z-score standardization to make the data
on numerical variables independent of scale. Distance measures
and z-score standardization were discussed in Section 8.2.
During the iterative process to merge clusters, the AGNES
algorithm uses one of several linkage methods to evaluate
(dis)similarity between clusters.
 The single linkage method uses the nearest distance between a
pair of observations that do not belong to the same cluster.
The complete linkage method uses the farthest distance
between a pair of observations that do not belong to the same
cluster.
The centroid method uses the centroid distance between
clusters, where the mean values for individual variables in the
cluster collectively represent the “center” or centroid of the cluster.
The average linkage method uses the average distance between
all pairs of observations that do not belong to the same cluster.
Ward’s method, developed by Joe H. Ward in 1963, uses a
slightly different algorithm to minimize the dissimilarity within
clusters. Specifically, it uses the “error sum of squares (ESS)” to
measure the loss of information that occurs when observations are
clustered, where ESS is defined as the squared difference
between individual observations and the cluster mean. Ward’s
method is widely implemented in many clustering analysis software
applications including Analytic Solver and R.
For illustration of the first three linkage methods, consider Figure
11.1 with two potential clusters. The blue line shows the single
linkage distance, and the green line represents the complete linkage
distance. The centroid distance is presented as a red dotted line.
The average linkage distance (not displayed in the figure) is the
average distance of all nine possible pairs between the two clusters.
Ward’s method (not displayed in the figure) involves computing the
error sum of squares to measure the loss of information when
observations are clustered.

FIGURE 11.1 Linkage distance for hierarchical clustering

 page 479

Dendrogram
Once the AGNES algorithm completes its clustering process, data
are usually represented in a treelike structure where each
observation can be thought of as a “leaf” on the tree. The treelike
structure is called a dendrogram. It allows users to visually inspect
the clustering result and determine the appropriate number of
clusters in the data. Determining the right number of clusters is
somewhat subjective, but we can visually inspect a dendrogram to
help guide our decision. Figure 11.2 shows an example of a
dendrogram.

FIGURE 11.2 Dendrogram

 Viewing from the top of the dendrogram, we can see two main
branches. Each branch represents a possible cluster of
observations. Collectively, the two branches consist of 10 sub-
branches or sub-clusters, labeled by the values below the horizontal
axis. The smaller branch on the right-hand side includes three sub-
branches or sub-clusters numbered 3, 8, and 10. The height of each
branch (cluster) or sub-branch (sub-cluster) indicates the distance,
or how dissimilar it is from the other branches or sub-branches with
which it is merged. The greater the height, the more distinctive the
cluster is from the other cluster. For example, sub-clusters 4 and 5
are very similar to each other and therefore connect at a lower
height. As a result, we can often determine an appropriate number of
clusters by visually inspecting the dendrogram. For the dendrogram
in Figure 11.2, we might decide that the observations should be
grouped into two clusters. As such, we may “cut the dendrogram
tree” horizontally to obtain the two main branches (two clusters), as
shown by the red line.
 Relying solely on the height of a dendrogram tree branch may
lead to statistically distinctive clusters that have little or no practical
meaning. Therefore, we often take into account both quantitative
measures (such as a dendrogram) and practical considerations to
determine an appropriate number of clusters. For example, suppose
the dendrogram in Figure 11.2 represents the process of clustering
customers into groups that are internally homogenous. If a firm
wants to target particular groups for one of the four different products
that it produces, then settling on four clusters seems appropriate in
this scenario; that is, in Figure 11.2 the tree would be cut at the
green line, rather than the red line.
We should also review the profile of each cluster using
descriptive statistics. For example, if age, sex, income, educational
attainment, and number of hours online per week are the variables
used in the cluster analysis to group customers, each resulting
cluster can be described with respect to its most defining
characteristics. One cluster may include page 480
predominantly high-income, highly educated female
customers between the ages of 35 and 45, and another cluster may
include predominantly customers between the ages of 18 and 25
who spend over 20 hours online weekly. We can label the first group
as “middle-age female professionals” and the second group as
“young Internet surfers.”
Another common approach to profile clusters is to incorporate
variables that were not used in clustering but of interest to the
decision maker. For example, we may find that a large percentage of
the customers in the first cluster have purchased our products at
least five times during the past year, while the majority of the
customers in the second cluster have not placed an order in the last
six months. Based on these insights gained from the cluster
analysis, the company may design different product offerings or
marketing messages to target different customer groups.

A DENDROGRAM
 A dendrogram is a graphical representation of data as a treelike
structure, where each branch can be considered a cluster of
observations.

A Note on the Choice of the Hierarchical Approach

The ability of a clustering method to discover useful hidden patterns
of the data depends on how it is implemented. The resulting
clustering structure depends on the data transformation (z-score
standardization or min-max normalization), the distance between
observations (Euclidian or Manhattan), the algorithm used (AGNES
or DIANA), and the linkage between clusters (single, complete,
average, centroid, or Ward’s method). There are no well-
documented guidelines for selecting a technique because the
discovery of patterns depends on the given data and the context.
Because clustering is essentially an unsupervised technique for data
exploration, the appropriate technique would be the one that makes
the most sense conceptually. It is always best to try several
techniques and select the cluster structure that allows us to generate
some insight.

Using Analytic Solver and R to Perform

Agglomerative Clustering
In Example 11.1 we demonstrate how to use Analytic Solver and R
to perform agglomerative hierarchical clustering. We should note that
cluster analysis is sensitive to the specific algorithms implemented in
different software packages. The algorithms used in Analytic Solver
differ from the algorithms used in R, and, thus, the clustering results
may differ.
FILE
Cities
 EXAMPLE 11.1
Table 11.2 shows a portion of data relating to the crime rate
(crimes per 100,000 residents), the poverty rate (in %), and the
median income (in $1,000s) for 41 cities in the United States.
Use Analytic Solver and R to perform agglomerative
hierarchical clustering. Interpret the results.

TABLE 11.2 Demographic Data for 41 Cities in the U.S.

City Crime Poverty Income

Barnstable 710.6 3.8 58.422

Boston 1317.7 16.7 48.729

⋮ ⋮ ⋮ ⋮

Warwick 139.7 3.9 59.445

page 481
SOLUTION:
Using Analytic Solver

a. Open the Cities data file.

b. From the menu, choose Data Mining > Cluster >
Hierarchical Clustering.

c. See Figure 11.3. Click on the ellipsis button next to the

Data range and select cells $A$1:$D$42. Make sure that the
box First Row Contains Headers is checked. The Variables in
Input Data box will populate. Select and move variables
 Crime, Poverty, and Income to the Selected Variables box.
Accept other defaults and click Next.

FIGURE 11.3 Hierarchical Clustering dialog boxi n Analytic

Solver

Source: Microsoft Excel

d. In the next dialog box, select Normalize input data. As

explained in Section 8.2, variables with a large scale (the
crime rate in this case) can dominate the results. Note that
Analytic Solver uses the input “Normalize” for a z-score
standardization. For Similarity Measure, we choose Euclidean
distance because all the input variables are numerical. For
Clustering Method, we choose to use Ward’s Method in this
example. Click Next.
e. In the final dialog box, under Output Options select Draw
Dendrogram and Show Cluster Membership. Set the
Maximum Number of Leaves at 10 and the Number of
Clusters at 3. Click Finish.
Next to the Cities worksheet, Analytic Solver will insert three
new worksheets: HC_Output, HC_Clusters, and
HC_Dendrogram. The HC_Output worksheet simply provides a
summary of the input parameters that we have entered. The
HC_Clusters worksheet lists each observation, its assigned
cluster, and its sub-cluster. What is of most interest to us at the
moment is the HC_Dendrogram worksheet. If you select this
worksheet, the dendrogram appears as a pop-up, and other
information appears in the accompanying worksheet.
The top half of Figure 11.4 shows the dendrogram and the
bottom half shows each observation’s cluster assignment.
Visual inspection of the dendrogram shows that the
observations in the data belong to three distinct clusters. We
can cut the dendrogram (e.g., where the y-axis value is 8) to
obtain these clusters where, for instance, the first cluster
consists of sub-clusters 1, 5, and 10. This cluster includes
observations 1 (from sub-cluster 1); 7, 9, 16, 20, 21, 29, 32,
and 41 (from sub-cluster 5); and 35 and 37 (from sub-cluster
10).

page 482

FIGURE 11.4 Analytic Solver’s HC_Dendrogram worksheet

 The height of each of the three branches is tall, indicating a
good clustering structure because the distances between the
clusters are large. This confirms that our decision to specify
three clusters in step e yields a reasonably good outcome. To
add the cluster membership to the original data in the Cities
worksheet, we can go to the HC_Clusters worksheet and copy
and paste the cluster numbers onto the Cities worksheet. Table
11.3 shows a portion of the updated Cities worksheet.

TABLE 11.3 Updated Cities Data using Analytic Solver for Example 11.1
 We can now compute summary statistics for each cluster to
review the cluster characteristics. Table 11.4 shows these
summary statistics, with the average crime rate, poverty rate,
and income for each cluster. We identify Cluster 1 as a group
of 11 cities with the lowest crime rate, the lowest poverty rate,
and the highest median income among the three clusters. On
the other hand, Cluster 2 represents a group of 15 cities with
the highest crime rate, highest poverty rate, and lowest median
income among the three clusters. Cluster 3 has a crime rate
closer to Cluster 1, a medium poverty rate, and a median
income closer to Cluster 2. Policymakers may find this type of
information useful when they make funding decisions or try to
understand the varying effect of an economic policy on different
cities.

TABLE 11.4 Hierarchical Clustering Results using Analytic Solver for Example
11.1

Cluster Crime Poverty Income

1 (N = 11) 235.90 5.51 61.39

2 (N = 15) 1,096.20 18.45 40.33

3 (N = 15) 362.03 11.61 44.61

 page 483

Compared to other clustering methods, the Ward’s method

in this case yields a clustering structure where each cluster has
a reasonable number of observations. Had we used the single
linkage or centroid clustering method, the analysis would have
resulted in several clusters that have only one observation,
which usually is not a desirable outcome. Students are
encouraged to experiment with different clustering methods
and examine the different outcomes.
Using R

a. Import the Cities data into a data frame (table) and label it
myData.
b. Install and load the cluster package. Enter:
> install.packages(“cluster”)
> library(cluster)
c. We exclude the City variable from the analysis; standardize
the Crime, Poverty, and Income variables; and store their
standardized values in a data frame called myData1. We use
the scale function to standardize the values of the three
variables. Enter:
> myData1 <- scale(myData[ , 2:4])
If no standardization is needed, we simply select the three
numerical variables and our R code would be > myData1 <-
myData[ , 2:4].
d. We use the dist function to determine similarity among
observations, and store the distance values in a new variable
called d. For options within the dist function, we use method
to specify the distance calculation. Options for the distance
calculation include “euclidean”, “manhattan”, “binary” (for
Jaccard’s coefficient), “maximum”, and “minkowski”. [The last
two options are beyond the scope of this text.] We specify
 “euclidean” to use the Euclidean distance as the similarity
measure. Enter:
> d <- dist(myData1, method = “euclidean”)
e. We use the agnes function to perform agglomerative
clustering and label the results as aResult. For options within
the agnes function, we use method to specify the clustering
method. Options for the clustering method include “single”
(single linkage), “complete” (complete linkage), “average”,
“ward” (Ward’s method), “weighted”, and “flexible”. We specify
“ward”. The diss option (meaning dissimilarity) is set equal to
TRUE to indicate that variable d contains a distance matrix
instead of the original variables. Enter “aResult” to obtain the
clustering results including an agglomerative coefficient, which
measures the strength of the clustering structure and whether
or not there is a natural clustering structure in the data.
Generally, a coefficient value of 0.75 or greater is indicative of
the existence of a good and natural clustering structure. Enter:
> aResult <- agnes(d, diss = TRUE, method = “ward”)
> aResult
R reports the following portion of results:
Call: agnes(x = myData, method = “ward”)
Agglomerative coefficient: 0.942097
The agglomerative coefficient (0.942097) suggests that a
good and natural clustering exists in the data.
f. We use the plot function to produce the dendrogram as well
as a banner plot. Enter:
> plot(aResult)
Hit <Return> to see next plot:
page 484
You will be prompted to press <Return> (the
“Enter” key) to display the plots. If you press <Return> once,
the banner plot appears.
 Figure 11.5(a) shows the banner plot. A banner plot is an
alternative to the dendrogram. The red bars in the banner plot
represent observations, and each empty gap between the red
bars represents potential clusters. As with the dendrogram, we
can cut the banner to obtain a desired number of clusters. In
this example, if we cut the banner where the Height value = 7
(shown as a blue line), we will obtain a result with three
clusters.

FIGURE 11.5 R’s banner plot and dendrogram plot

 If you press <Return> again, then the dendrogram
appears. Figure 11.5(b) shows the dendrogram; the results
are consistent with those obtained using Analytic Solver.
g. Based on inspection of the banner plot and the dendrogram,
we can determine the number of clusters. We use the cutree
function to create three distinct clusters. Enter:
> aClusters <- cutree(aResult, k = 3)
h. We can append the cluster membership to the original data
frame, myData, using the data.frame function. We then use
the View function to inspect the updated myData data frame.
Enter:
> myData <- data.frame(myData, aClusters)
> View(myData)
page 485
Table 11.5 shows a portion of the updated Cities
data frame.

TABLE 11.5 Updated Cities Data Using R for Example 11.1

i. We use the summary function to obtain summary statistics for

the three clusters. Enter:
> summary(subset(myData, aClusters==1))
> summary(subset(myData, aClusters==2))
> summary(subset(myData, aClusters==3))
j. To find out the number of observations in each cluster, we use
the as.factor function to convert aCluster to categorical data
 and then use the summary function to find out the number of
observations in each cluster. Enter:
> summary(as.factor(aClusters))
Table 11.6 shows average values for the crime rate, the
poverty rate, and median income for each of the three
clusters. The results do differ from those obtained by Analytic
Solver, but they are consistent. Cluster 1 is still a group of 11
cities with the lowest crime rate, the lowest poverty rate, and
the highest median income among the three clusters.
However, only nine cities fall in Cluster 3, with the highest
crime rate, the highest poverty rate, and the lowest median
income—Analytic Solver clustered 15 cities in this category. R
also groups 21 cities falling somewhere in the middle,
whereas Analytic Solver grouped 15 in this category.
TABLE 11.6 Hierarchical Clustering Results using R for
Example 11.1

Cluster Crime Poverty Income

1 (N = 11) 235.90 5.51 61.39

2 (N = 21) 544.2 12.28 45.64

3 (N = 9) 1,160.5 21.46 35.06

Agglomerative Clustering with Mixed Data

So far, we have discussed clustering with data on numerical
variables. We will now extend the analysis with data consisting of
both numerical and categorical variables, which is also referred to as
mixed data. Mixed data are quite often of interest in business
applications. For example, a firm may want to cluster the
observations based on numerical variables such as income and age
along with categorical variables such as sex and race.
In order to measure the distance between two observations with
mixed data, it is common to use the distance measure, referred to as
Gower’s coefficient, proposed by John C. Gower. Gower’s coefficient
computes the distance for each variable, converts it into a [0, 1]
scale, and calculates a weighted average of the scaled distances as
a measure of similarity between two observations. The conceptual
detail of the Gower’s similarity coefficient is beyond the scope of this
text.
The current version of Analytic Solver does not calculate Gower’s
similarity coefficient; therefore, we will use only R for agglomerative
clustering with mixed data. Consider the following example.
FILE
Subscribers

page 486

EXAMPLE 11.2
A national phone carrier conducted a socio-demographic study
of their current mobile phone subscribers. Subscribers were
asked to fill out survey questions about their current annual
salaries (Salary), whether or not they live in a city (City equals
1 if living in a city, 0 otherwise), and socio-demographic
information such as marital status (Married equals 1 if married,
0 otherwise), sex (Sex equals 1 if male, 0 otherwise), and
whether or not they have completed a college degree (College
equals 1 if college degree, 0 otherwise). Table 11.7 shows a
portion of the survey data collected from 196 subscribers.

TABLE 11.7 Mobile Phone Subscribers

 Perform agglomerative hierarchical clustering using
Gower’s similarity measure. Interpret the results.
SOLUTION:
We use R to perform agglomerative cluster analysis with the
Gower’s coefficients using many of the same steps that we
used in Example 11.1. For this reason, in some steps, we are
brief.
Using R

a. Import the Subscribers data into a data frame (table) and

label it myData. Install and load the cluster package, if you
have not already done so. Enter:
> install.packages(“cluster”)
> library(cluster)
b. The first variable in the data set contains the mobile
subscriber identification numbers and will not be used in the
cluster analysis. We use the daisy function to determine
similarity among observations and label the results as d. For
options within the daisy function, we use metric to specify the
distance calculation. Options for distance calculation include
“euclidean”, “manhattan”, and “gower”. We specify “gower”.
Enter:
> d <- daisy(myData[ , 2:6], metric = “gower”)
If you enter d again, R reports all “dissimilarity” measures for
all possible pairs of observations. Verify that the dissimilarity
 measure for observations 1 and 2 is 2.277405e-01. This is
simply 1 minus the Gower coefficient.
c. We use the agnes function with Ward’s method to perform
agglomerative clustering and label the results as mResult.
Enter:
> mResult <- agnes(d, method = “ward”)
> myResult
R reports an Agglomerative coefficient of 0.9972604.
d. We use the plot function to obtain the banner plot and the
dendrogram. Enter:
> plot(mResult)
page 487
Follow the prompts to view the banner plot and
the dendrogram. Figure 11.6 shows the dendrogram.

FIGURE 11.6 R’s dendrogram for Example 11.2

e. We use the cutree function to obtain four clusters, and then

add the cluster membership information to the myData data
frame. We use the summary function to obtain summary
statistics for each cluster. Table 11.8 provides mean values for
each cluster.

TABLE 11.8 Hierarchical Clustering Results for Example 11.2

 > mClusters <- cutree(mResult, k = 4)
> myData <- data.frame(myData, mClusters)
> summary(subset(myData, mClusters==1))
> summary(subset(myData, mClusters==2))
> summary(subset(myData, mClusters==3))
> summary(subset(myData, mClusters==4))
f. To find out the number of observations in each cluster, we use
the as.factor function to convert mCluster to categorical data
and then use the summary function to find out the number of
observations in each cluster. Enter:
> summary(as.factor(mClusters))
Based on these results, we can make the following
observations about each of the four clusters.
All subscribers in Cluster 1 are married males, most of whom
live in a city. Only about 40% of them completed a college
degree.
All subscribers in Cluster 2, the largest of the four clusters,
are not married and did not complete a college degree.
About 40% of these subscribers are male. Most subscribers
in this group live outside a city.
All subscribers in Cluster 3 are married females who did not
complete a college degree. Most of them live in a city.
All subscribers in Cluster 4 completed a college degree and,
on average, earn the highest salary among the four groups.
They tend to be unmarried, and most of them are female.
 page 488

The main appeal of hierarchical clustering is that it does not

require the pre-specification of the number of clusters and, therefore,
is purely data driven. It is also easy to understand and interpret. Note
that because hierarchical clustering evaluates individual
observations in order to merge or split them in each iteration, it is
inherently slow and does not scale efficiently to a large data set. This
poses a major challenge when dealing with Big Data. The method is
also not stable in that the interpretation may change with even a
small change in the sample data. To overcome these limitations,
practitioners often combine hierarchical clustering with other, more
efficient, clustering techniques. This is discussed in the next section.

EXERCISES 11.1
Mechanics

Notes: and indicate that the exercise must be solved

using Analytic Solver or R, respectively. If it is not specified,

then the exercise can be solved using Analytic Solver and/or R.
The answers, however, will depend on the software package
used. Please note that Analytics Solver uses the input
“Normalize” for z-score standardization.
1. FILE Exercise_11.1. Perform agglomerative clustering on the accompanying
data set.
a. Include all five variables, first standardized to z-scores, for the analysis. Choose
Euclidean for the distance between observations and single linkage for the
distance between clusters. Plot and inspect the dendrogram. How many
clusters are generated if the minimum distance between clusters is 5?
b. Repeat part a but choose the complete linkage clustering method.
c. Repeat part a but choose Ward’s clustering method.
2. FILE Exercise_11.2. Perform agglomerative clustering on the accompanying
data set. Include all seven variables, first standardized to z-scores, for the
analysis. Choose Euclidean for the distance between observations and Ward’s
method for the distance between clusters. Plot and inspect the dendrogram. How
many clusters are generated if the minimum distance between clusters is 5? How
many clusters are generated if the minimum distance between clusters is 3?
3. FILE Exercise_11.3. Perform agglomerative clustering on the accompanying
data set. Include all seven variables, first standardized to z-scores, for the
analysis. Choose Euclidean for the distance between observations and complete
linkage for the distance between clusters. Plot and inspect the dendrogram. How
many clusters are generated if the minimum distance between clusters is 5? How
many observations are in the largest cluster? What is the average value of x1 of
the largest cluster?
4. FILE Exercise_11.4. Perform agglomerative clustering on the accompanying
data set.
a. Include all five variables, first standardized to z-scores. Use the Euclidean
distance for similarity and single linkage for the clustering method to cluster the
data into three clusters. How many observations are in the largest cluster?
What is the average value of x4 of the largest cluster?
b. Repeat part a but choose the complete linkage clustering method.
c. Repeat part a but choose Ward’s clustering method.
5. FILE Exercise_11.5. Perform agglomerative clustering on the accompanying
data set. Use the first five variables, x1, x2, x3, x4, and x5, in the analysis. Do not
standardize the variables. Note: For R, use the method = “binary” option in the
dist function.
a. Use Jaccard’s coefficients for the similarity measure and average linkage for the
clustering method to cluster the data into four clusters. How many 1s are in x1
of the largest cluster?
b. Use Jaccard’s coefficients for the similarity measure and single linkage for the
clustering method to cluster the data into six clusters. How many 1s are in x1 of
the largest cluster?
6. FILE Exercise_11.6. Perform agglomerative clustering on the accompanying
data set, using all 11 binary variables. Use Jaccard’s coefficients for the similarity
measure and average linkage for the clustering method. Inspect the dendrogram.
How many clusters are generated if the minimum distance between clusters is
0.8? Note: For R, use the method = “binary” option in the dist function.
7. FILE Exercise_11.7. Perform agglomerative clustering on the accompanying
data set, using all 11 binary variables. Use Jaccard’s coefficients for the similarity
measure and complete linkage for the clustering method to cluster the data into 5
clusters. How many observations are in the largest cluster? How many 1s are in
x1 of the largest cluster? Note: For R, use the method = “binary” option in the dist
function.
8. FILE Exercise_11.8. Perform agglomerative clustering on the

accompanying data consisting of both numerical and categorical variables. Use

Gower’s coefficient for the distance between observations and Ward’s clustering
method. Plot and inspect the dendrogram. How many clusters are generated if the
minimum distance between clusters is 0.8? What is the average value of x5 of the
largest cluster?

page 489

Applications
9. FILE Colleges. Peter Lara, an aspiring college student, met with his high school
college advisor to discuss potential colleges to which he might apply. He was
advised to consult with the College Scorecard information on the Department of
Education website. After talking to his family, he downloaded a list of 116 colleges
and information on annual post-college earnings (Earnings in $), the average
annual cost (Cost in $), the graduation rate (Grad in %), the percentage of
students paying down debt (Debt in %), and whether or not a college is located in
a city (1 denotes a city location, 0 otherwise). Peter and his family want to group
these colleges based on the available information to help narrow down their
choices. The accompanying table shows a portion of the data.

a. Perform agglomerative clustering based on the earnings, cost, graduation rate,

and debt information, all standardized to z-scores. Use the Euclidean distance
and Ward clustering method to determine appropriate cluster membership for
the 116 colleges. How many clusters are generated if the minimum distance
between clusters is 15?
b. Separate the colleges based on whether or not they are located in the city into
two different data sets. Group the colleges that are located in the city into
clusters based on the earnings, cost, graduation rate, and debt information
using the Euclidean distance and Ward’s clustering method. Inspect the
dendrogram. How many clusters are generated if the minimum distance
between clusters is 10?
c. Repeat the analysis for the colleges not located in the city. Inspect the
dendrogram. How many clusters are generated if the minimum distance
between clusters is 10?
10. FILE Student_Body. Anne Cutberth has just started her new job as an
academic adviser at a small liberal arts college in Colorado. She is going through
a list of students in an academic department and wants to gain a better
understanding about the student body in the department. The accompanying table
shows a portion of the data that she has collected on 100 students. Note that all
variables are categorical with a value 1 implying that the given student is a
transfer student, was on the Dean’s list last semester, is currently receiving
financial aid, or is female, respectively. Note: For R, use the method = “binary”
option in the dist function.

Transfer Dean’s list Financial aid Female

0 1 1 1

1 0 1 0

⋮ ⋮ ⋮ ⋮

1 0 0 1

a. Perform agglomerative clustering to group the students in Anne’s data set. Use
Jaccard’s coefficients for the similarity measure and the complete linkage
clustering method. Inspect the dendrogram. How many clusters are generated if
the minimum distance between clusters is 0.8? How many transfer students are
in the largest cluster?
b. Repeat part a but choose average linkage clustering method. How many
students are on the Dean’s list in the largest cluster?
11. FILE Football_Players. Denise Lau is an avid football fan and religiously follows
every college football game. During the current season, she meticulously keeps a
record of how each quarterback has played throughout the season. Denise is
making a presentation at a local college football fan club about these
quarterbacks. The accompanying table shows a portion of the data that she has
recorded. Variables include the player number (Player), completed passes
(Comp), attempted passes (Att), completion percentage (Pct), total yards thrown
(Yds), average yards per attempt (Avg), yards thrown per game (Yds/G), number
of touchdowns (TD), and number of interceptions (Int).
 a. Perform agglomerative clustering to group the quarterbacks according to their
performance to help Denise prepare for this presentation. Standardize the
variables and use the Euclidean distance and the Ward clustering method.
Cluster the data into three clusters. How many players are in the largest
cluster? What is the average number of touchdowns of the largest cluster?
b. Select only the quarterbacks with at least 150 attempted passes and repeat the
cluster analysis performed in part a. Cluster the data into two clusters. How
many players are in the larger cluster? What is the average number of
touchdowns of the larger cluster? Note: For Analytic Solver, use the Filter option
in Excel to select the observations. Copy the observations to a new worksheet.
Do not alter the order of the observations.

12. FILE Baseball_Players. Ben Derby is a highly paid scout for a professional

baseball team. He attends at least five or six Major League Baseball games a
week and watches as many recorded games as he can in order to evaluate
potential players for his team. He also keeps detailed records about each
perspective player. His team is now seeking to add another hitter to its roster next
season. Luckily, Ben has information on 144 hitters in the league who have played
at least 100 games during the last season. The accompanying table shows a
portion of the data. Variables include the player number (Player), the number of
games played (G), at bats (AB), runs (R), hits (H), homeruns
(HR), runs batted in (RBI), batting average (AVG), on base
page 490
percentage (OBP), and slugging percentage (SLG).

Use agglomerative clustering to group the 144 players into three clusters using all
variables except Player. Standardize the variables and use the Euclidean distance
and single linkage clustering method. How many players are in each cluster?
Describe the characteristics of each cluster using the average values of the
variables.
13. FILE Internet_Addiction. Internet addiction has been found to be a widespread
problem among university students. A small liberal arts college in Colorado
conducted a survey of Internet addiction among its students using the Internet
Addiction Test (IAT) developed by Dr. Kimberly Young. The IAT contains 20
questions that measure three underlying psychometric factors. Questions 1
through 9 measure emotional/psychological conflicts, which refer to the degree to
which the individual uses the Internet as a means to avoid interactions with
friends and family. Questions 10 through 14 measure time management issues,
 which refer to the degree to which the individual chooses to spend time online at
the expense of other responsibilities. Finally, Questions 15 through 20 measure
mood modification, which refers to the degree to which the individual’s Internet
dependence is motivated by the need for mood improvement. The accompanying
table shows a portion of the responses from 350 students. Students respond on a
scale of 1 to 5, where higher values are indicative of possible problematic Internet
use.

a. The scores for the three psychometric factors (i.e., emotional/psychological

conflicts, time management issues, and mood modification) can be calculated
by averaging the scores for the questions that are designed to measure them.
First, find the average score for each psychometric factor for each student.
Then, perform agglomerative clustering to group the 350 university students
into five clusters based on the three underlying psychometric factors. Use
Euclidean distance and the single linkage clustering method. Do not
standardize the variables. How many students are in the largest cluster?
b. What are the average scores for emotional/psychological conflicts, time
management issues, and mood modification for students in the largest cluster?
14. FILE Health_Population2. The accompanying data set contains country-level
health and population measures for 38 countries from the World Bank’s 2000
Health Nutrition and Population Statistics database. For each country, the
measures include death rate per 1,000 people (Death Rate, in %), health
expenditure per capita (Health Expend, in US$), life expectancy at birth (Life Exp,
in years), male adult mortality rate per 1,000 male adults (Male Mortality), female
adult mortality rate per 1,000 female adults (Female Mortality), annual population
growth (Population Growth, in %), female population (Female Pop, in %), male
population (Male Pop, in %), total population (Total Pop), size of labor force
(Labor Force), births per woman (Fertility Rate), birth rate per 1,000 people (Birth
Rate), and gross national income per capita (GNI, in US$). The accompanying
table shows a portion of the data.
 a. Perform agglomerative clustering to group the 38 countries according to their
health measures (i.e., Death Rate, Health Expend, Life Exp, Male Mortality, and
Female Mortality) only. Use the Euclidean distance and the average linkage
clustering method to cluster the data into three clusters. Is data standardization
necessary in this case?
b. Describe the characteristics of each cluster. Compare the average GNI per
capita (US$) of each group and report your findings.
15. FILE Health_Population2. Refer to the previous exercise for a description of the
data set.
a. Perform agglomerative hierarchical clustering to group the 38 countries
according to their population measures (i.e., Population Growth, Female Pop,
Male Pop, Total Pop, Labor Force, Fertility Rate, and Birth Rate) only. Use the
Euclidean distance and the average linkage clustering method to cluster the
data into three clusters. Is data standardization necessary in this case?
b. Report the size and average GNI per capita (US$) of the largest cluster of
countries.
c. Compare your findings with those from the previous exercise. Discuss the
differences in cluster membership and characteristics.

16. FILE Pizza_Customers. A local pizza store wants to get a better sense of

who its customers are. The accompanying table shows a portion of data that it
collected on 30 randomly selected customers. Variables include
age, female (1 if female, 0 otherwise), annual income, married (1
page 491
if married, 0 otherwise), own (1 if own residence, 0 otherwise), college (1 if
completed college degree, 0 otherwise), household size (Size), and annual store
spending (Spending).

a. Perform agglomerative clustering to group the customers based on the

numerical variables only. Use the Euclidean distance and Ward’s clustering
method to cluster the customers into two clusters. Describe each cluster of
customers based on the cluster characteristics.
b. Use agglomerative clustering to group the customers based on the categorical
variables only. Use matching coefficients and the complete linkage clustering
method to cluster the customers into two clusters. Describe each cluster of
customers based on the cluster characteristics.
17. FILE Pizza_Customers. Refer to the previous exercise for a description of

the data set.

a. Perform agglomerative clustering on the accompanying data consisting of both
numerical and categorical variables. Use Gower’s coefficient for the distance
between observations and Ward’s clustering method. How many clusters can
be generated if the minimum distance between clusters is 1?
b. How many customers are in the largest cluster? What is the average income of
the customers in the largest cluster?
18. FILE Temperature. The accompanying table shows a portion of data consisting
of the January, April, July, and October average temperatures of 50 selected U.S.
cities.

a. Perform agglomerative hierarchical clustering to group the cities based on their

January, April, July, and October average temperatures. Standardize the
variables and use the Euclidean distance and the Ward’s clustering method to
cluster the cities into three clusters. How many cities are in the largest cluster?
b. What are the average January, April, July, and October average temperatures
for the largest cluster of cities?
19. FILE MLB. The accompanying table lists a portion of Major League Baseball’s
pitchers, their earned run average (ERA), and their salary (in millions of $).

Player ERA Salary

1 2.53 17

2 2.54 4

⋮ ⋮ ⋮

10 3.09 0.5

a. Perform agglomerative clustering to group the players based on ERA and

Salary. Standardize the variables and choose the Euclidean distance and the
single linkage clustering method to cluster the players into four clusters. How
many players are in the largest cluster?
 b. What are the average ERA and salary for the largest cluster of players?
20. FILE Chicago_CA. Sanjay Johnson is working on a research paper that studies
the relationship between the education level and the median income of a
community. The accompanying table shows a portion of the data that he has
collected on the educational attainment and the median income for 77 areas in
the city of Chicago. Sanjay plans to cluster the areas using the educational
attainment data and compare the average median incomes of the clusters. For
each community area, the measures include total number of residents 25 years
and over (25 or Over), number of residents with less than high school education
(Less than HS), number of residents with high school education (HS), number of
residents with some college (SC), number of residents with bachelor’s degree or
higher (Bachelor), and median household income (Income).

a. Should Sanjay standardize the data before cluster analysis? Explain.

b. Perform agglomerative clustering to group the community areas based on the
standardized variables related to the education attainment of the population
(i.e., Less than HS, HS, SC, and Bachelor). Use the Euclidean distance and the
Ward clustering method. Inspect the dendrogram. How many clusters are
generated if the minimum distance between clusters is 10?
c. What are the size and average median household income of the largest cluster
of community areas?

page 492
21. FILE Internet_Addiction2. Internet addiction has been

found to be a widespread problem among university students. A small liberal arts

college in Colorado conducted a survey of Internet addiction among its students
using the Internet Addiction Test (IAT) developed by Dr. Kimberly Young. The IAT
measures three underlying psychometric factors: emotional/psychological
conflicts (Factor1), time management issues (Factor2), and mood modification
(Factor3). Individuals who score high on the IAT are more likely to have
problematic Internet use. The accompanying table shows a portion of the scores
for the three psychometric factors for 350 students. The respondents also
provided a couple of pieces of demographic data including sex (equals 1 if
female, 0 otherwise) and level of study (equals 1 if graduate study, 0 otherwise).
 a. Perform hierarchical clustering to group the 350 university students using the
three psychometric factors as well as the other demographic variables. Use
Gower’s coefficient for the distance between observations and Ward’s
clustering method to cluster the students into three clusters. How many
students are in the largest cluster?
b. What are the average scores for emotional/psychological conflicts, time
management issues, and mood modification for students in the largest cluster?

22. FILE Longitudinal_Partial. The accompanying table contains a portion of

data from the National Longitudinal Survey (NLS), which follows over 12,000
individuals in the United States over time. Variables in this analysis include Urban
(1 if lives in urban area, 0 otherwise), Siblings (number of siblings), White (1 if
white, 0 otherwise), Christian (1 if Christian, 0 otherwise), FamilySize, Height,
Weight (in pounds), and Income (in $).

a. Use hierarchical clustering to group the respondents. Use Gower’s coefficient

for the distance between observations and Ward’s clustering method to cluster
the individuals into four clusters. How many individuals are in the largest
cluster?
b. Describe the characteristics of the clusters.

11.2 k-MEANS CLUSTER ANALYSIS

 LO 11.2
Conduct k-means cluster analysis.

Unlike hierarchical clustering, discussed in the previous section, in k-

means clustering, we need to specify the number of clusters, k,
prior to performing the analysis. The objective is to divide the sample
into a prespecified number k of nonoverlapping clusters so that each
of these k clusters is as homogenous as possible. In practice, we
may experiment with different values of k until we obtain a desired
result, or use hierarchical clustering methods to help determine the
appropriate k. In addition, we may have prior knowledge or theories
about the subjects under study and can determine the appropriate
number of clusters based on domain knowledge. For example, if we
were to sort through a stack of books, we might unconsciously group
them into fiction and nonfiction categories. Or while organizing a pile
of shoes, we might naturally place them into casual, athletic, or dress
shoe categories.
Once the number of clusters k is determined, the k-means
algorithm is based on the choice of the initial cluster centers. There
are several ways to find initial cluster centers. A popular approach is
to initiate the algorithm by randomly selecting k observations from
the data to serve as initial cluster centers. Each remaining
observation is then assigned to the nearest cluster center, forming k
clusters. Cluster centroids for the k clusters are calculated to be
used as the updated cluster centers. The algorithm then reassigns
each observation to its nearest cluster center and recalculates the
cluster centroids.

page 493

The reassignment of observations and recalculation of cluster

centroids are repeated, with the goal of minimizing the dispersion
within clusters, where dispersion is defined as the sum of Euclidean
distances of observations from their respective cluster centroid. The
cluster assignment and centroid update steps are iteratively
repeated until convergence is achieved and cluster dispersion can
no longer be improved. Because it uses only the Euclidean distance
as a similarity measure, the k-means algorithm can only handle data
with numerical variables. In other words, it cannot analyze
categorical variables unless they are transformed to numerical
variables prior to the cluster analysis.

THE k-MEANS CLUSTERING METHOD

The k-means clustering method assigns each observation to a
cluster, such that the observations assigned to the same cluster
are as similar as possible. The number of clusters k is
determined prior to estimation. The k-means clustering method
can only be applied to data with numerical variables. The
general process of the k-means clustering algorithm can be
summarized as follows:
1. Specify the k value.
2. Randomly assign k observations as cluster centers.
3. Assign each observation to its nearest cluster center.
4. Calculate cluster centroids.
5. Reassign each observation to a cluster with the nearest
centroid.
6. Recalculate the cluster centroids and repeat step 5.
7. Stop when reassigning observations can no longer improve
within-cluster dispersion.

Compared to the hierarchical clustering methods, the k-means

clustering method is more computationally efficient, especially when
dealing with large data sets. This is because hierarchical clustering
requires computation of distances between each possible pair of
observations while k-means clustering only requires computation of
distances between each observation and the cluster centroids during
each iteration. For example, with a relatively small data set that
contains 100 observations, hierarchical clustering would need to
compute the distances of 4,950 possible observation pairs in order to
form clusters. At the same time, for the same data set and a pre-
determined number of clusters (k) of 4, k-means clustering only
needs to compute 400 distance measures during each iteration.
Therefore, it is customary to cluster large data sets using the k-
means clustering method instead of the hierarchical clustering
methods.

Using Analytic Solver and R to Perform k-

Means Clustering
Results from k-means clustering are highly sensitive to the random
process for finding the initial cluster centers as well as implementing
specific algorithms. The initial cluster centers and the algorithms
used in Analytic Solver and R are not the same, and, therefore, their
outputs will differ. As mentioned earlier, because clustering is
essentially an unsupervised technique for data exploration, the
appropriate approach is the one that produces a clustering structure
that makes the most sense conceptually.
In the following example, we demonstrate how to perform k-
means clustering using Analytic Solver and R.

page 494
FILE

CandyBars

EXAMPLE 11.3
Recall that the objective outlined in the introductory case is to
help Aliyah Williams group candy bars into meaningful clusters
and improve product selection and display for her vending
 machines. Perform k-means clustering on the data assuming
four clusters. Interpret the results.
SOLUTION:
Using Analytic Solver

a. Open the CandyBars data file.

b. From the menu, choose Data Mining > Cluster > K-means
Clustering.

c. See Figure 11.7. Click on the ellipsis button next to the

Data range and select cells $A$1:$E$38. Make sure that the
box First Row Contains Headers is checked. The Variables in
Input Data box will populate. Select and move variables
Calories, Fat, Protein, and Carb to the Selected Variables box.
Click Next.

FIGURE 11.7 Analytic Solver’s k-Means Clustering dialog

box

Source: Microsoft Excel

d. In the next dialog box, select Normalize input data. Set #
Clusters to 4. Accept other defaults and click Next.
Note: Analytic Solver uses the term “Normalize” for z-
score standardization. Also, k-means clustering is a random
process for finding the initial cluster centers and implementing
specific algorithms. Analytic Solver allows you to control the k-
means algorithm by changing the random seed and the
maximum number of iterations. For simplicity, we will use the
defaults: 10 iterations, fixed start, and random seed of 12345.
e. In the final dialog box, select Show data summary and Show
distances from each cluster center. Click Finish.
Next to the CandyBars worksheet, Analytic Solver will insert
two new worksheets: KMC_Output and KMC_Clusters. The
KMC_Output worksheet contains three tables. See Figure 11.8.
The Cluster Centers table displays z-score values of calories,
fat, protein, and carbohydrate levels for each cluster centroid.
For example, Cluster 1 includes three candy bars with a
relatively high amount of carbohydrates (z-score of 2.5591) and
a low amount of fat (z-score of −1.8673), while Cluster 2
includes six candy bars that are relatively low in calories (z-
score of −1.4164) and in fat (z-score of −1.2567). The Inter-
Cluster Distances table shows the distance between the four
clusters. The Cluster Summary table shows the number of
candy bars in each cluster as well as the average distance
within each cluster, also referred to as intra-cluster distances.
We want the inter-cluster distance to be greater than the intra-
cluster distance.

page 495

FIGURE 11.8 Analytic Solver’s KMC_Output worksheet for Example 11.3

Source: Microsoft Excel

Figure 11.9 shows a portion of the results from the

KMC_Clusters worksheet. Analytic Solver displays each
observation, the cluster to which it is assigned, and distances
to the cluster centroids. For example, the first record is closest
to the centroid of Cluster 3 (a distance value of 2.0269,
compared to the distance from the other three clusters: 5.6182,
5.3371, and 3.3590), and, therefore, it is assigned to Cluster 3.
Similarly, records 2 through 5 are also assigned to Cluster 3
due to their proximity to the centroid of Cluster 3. Record 6 is
closest to the centroid of Cluster 1 (a distance value of 1.6274,
compared to the distance from the other three clusters) and is
assigned to Cluster 1.

FIGURE 11.9 A portion of the Analytic Solver’s KMC_Clusters worksheet

Source: Microsoft Excel

 Using R

a. Import the CandyBars data into a data frame (table) and label
it myData.
The following instructions and results are based on R version
3.5.3. To replicate the results with newer versions of R, enter:
> suppressWarnings(RNGversion(“3.5.3”))
b. Install and load the cluster package. Enter:
> install.packages(“cluster”)
> library(cluster)
c. We exclude the Brand variable from the analysis and
standardize the other variables using the scale function. The
new data frame is called myData1. Enter:
> myData1 <- scale(myData[ , 2:5])
d. We use the set.seed function to set the random seed and the
pam function to perform the k-means clustering. Within the
pam function, we set the option k to 4 because we have
preselected 4 clusters. We store the clustering results in a
variable called kResult. We use the summary function to view
the results. Enter:
> set.seed(1)
> kResult <- pam(myData1, k = 4)
> summary(kResult)
page 496
Table 11.9 shows a portion of the R results
that we have rearranged for presentation purposes. Despite
the fact that the results are quite different than the ones that
we obtained with Analytic Solver, there are some similarities.
For example, Cluster 2 in R, like Cluster 1 in Analytic Solver,
includes three candy bars with a relatively high amount of
carbohydrates (z-score of 2.0860) and a low amount of fat (z-
score of −1.9482). Similarly, Cluster 4 in R is like Cluster 2 in
 Analytic Solver and Cluster 1 in R is like Cluster 3 in Analytic
Solver.

TABLE 11.9 R’s clustering results with z-scores

e. In order to visually show the clusters and their members, we

use the plot function. Enter:
> plot(kResult)
Follow the prompts to view the cluster and silhouette plots
(Figure 11.10) that R produces. The cluster plot in Figure
11.10(a) incorporates principal component analysis (PCA) to
summarize information from the four variables (Calories, Fat,
Protein, and Carb). Because we cannot plot a page 497
four-dimensional graph based on the four
variables, we use PCA in order to plot the clusters onto a two-
dimensional graph with the first two principal components as
the x- and y-axes. Recall from Chapter 8 that these principal
components explain the highest degree of variability in the
data. In this case, the first two principal components explain
88.75% of the variability. Each blue oval represents a cluster,
and the observations in each cluster are displayed using a
different marker.

FIGURE 11.10 R’s cluster plot and silhouette plot for Example 11.3
 The silhouette plot in Figure 11.10(b) shows how close each
observation in one cluster is to observations in the other
clusters. It also allows us to visually determine the appropriate
number of clusters. The silhouette width for each observation,
si, ranges from −1 to +1, where a value closer to +1 indicates
that the observation is well matched to its own cluster and
poorly matched to neighboring clusters. If most observations
have values close to +1, then the clustering configuration is
appropriate. If many points have a low or negative value, then
the clustering configuration may have too many or too few
clusters. The results in Figure 11.10(b) seem to suggest that
the results are good, but not great. Clusters 2 and 4 seem well-
configured. Two of the observations in Cluster 1 and one
observation in Cluster 3 have silhouette widths that are
negative. The average silhouette width is 0.32 suggesting a
reasonable clustering configuration overall. It is not clear if a
different configuration would necessarily improve the results.

SYNOPSIS OF INTRODUCTORY CASE

Aliyah Williams is a fledgling entrepreneur who owns a vending machine business.
In order to strategically group and display candy bars in her vending machines, she
chooses four clusters to perform k-means clustering. Aliyah observes a certain
theme in the grouping of candy bars. For example, Three Musketeers, 100 Grand,
Raisinets, York Peppermint Pattie, After Eight Mints, and Twizzlers Cherry Nibs are
grouped in a cluster that is relatively low in calories, fat, and protein. In addition,
fruit-flavored candies like Skittles, Starburst, and Twizzlers, with a relatively low
amount of fat and protein but high in carbohydrates, are also assigned into a cluster.

Steve Degenhardt/Shutterstock

The defining characteristics of the two remaining groups are less obvious.
However, Aliyah observes that one of the remaining two clusters includes candy
bars, such as Reese’s Pieces, that have a high content of protein and calories.
Finally, the last cluster tends to include candy bars that contain a moderate amount
of the four nutrients.
Given the increase in health-conscious consumers in the market, Aliyah feels
that her vending machines would best meet customer needs by increasing the
proportion of candy bars that are low in calories and fat, found in the first two
 clusters. These candy bars would likely attract customers who crave for something
sweet but at the same time do not want to add too much calories and fat to their diet.
The fruit-flavored candies would likely cater to customers who are on a low-fat, high-
carb diet. She will also place the healthier candy bars in a more prominent place in
her vending machines located close to healthcare and fitness facilities and display a
sign listing the nutritional facts of each candy bar.
Aliyah also decides to increase the number of selected candy bars from the third
cluster for customers who are looking to consume a little more protein in their diet.
Finally, to maintain product variety, she will also choose an assortment of popular
candy bars from the remaining cluster to include in her vending machines.

page 498

EXERCISES 11.2

Notes: and indicate that the exercise must be solved

using Analytic Solver and R, respectively. If it is not specified,

then the exercise can be solved using Analytic Solver and/or R.
The answers, however, will depend on the software package
used. Use the following defaults in Analytic Solver: 10
iterations, fixed start, and random seed of 12345. For R, use
the set.seed(1) statement to set the random seed. All answers
in R are based on version 3.5.3. To replicate the results with
newer versions of R, execute the following line of code at the
beginning of the R session:
suppressWarnings(RNGversion(“3.5.3”)).

Mechanics
23. FILE Exercise_11.23. Perform k-means clustering on the accompanying data
set.
a. Use all variables in the analysis. Do not standardize the variables. Set the
number of clusters to 3. What are the size and average distance for the largest
cluster?
b. Specify the same settings as in part a, but standardize the variables. What are
the size and average distance for the largest cluster?
24. FILE Exercise_11.24. Perform k-means clustering on both variables in the
accompanying data set. Standardize the variables. Experiment with the k values
of 2, 3, and 4. Compare the number of observations and distance statistics of the
largest cluster for each k value.
25. FILE Exercise_11.25. Perform k-means clustering on the accompanying data
set.
a. Use variables x1, x2, and x3 in the analysis. Standardize the data. Specify the k
value as 2. What are the cluster center values for the larger cluster?
b. Specify the k value as 3. What are the cluster center values for the smallest
cluster?

26. FILE Exercise_11.26. Perform k-means clustering on the accompanying

data set. Use variables x4, x5, x6, and x7, standardized to z-scores, in the
analysis.
a. Specify the k value as 2 and plot the cluster membership using the cluster and
silhouette plots. What is the average silhouette width?
b. Specify the k value as 3 and plot the cluster membership using the cluster and
silhouette plots. What is the average silhouette width?
c. Specify the k value as 4 and plot the cluster membership using the cluster and
silhouette plots. What is the average silhouette width?
27. FILE Exercise_11.27. Perform k-means clustering on the accompanying data
set. Use variables x1, x3, and x5 in the analysis. Do not standardize the variables.
a. Set the number of clusters to 3. What are the size and cluster center values for
the largest cluster?
b. Perform the same analysis as in part a, but with variables standardized to z-
scores. What are the size and cluster center values for the largest cluster?

28. FILE Exercise_11.28. Perform k-means clustering on all the variables in the

accompanying data set. Do not standardize the variables.

a. Specify the k value as 2 and plot the cluster membership using the cluster and
silhouette plots.
b. Specify the k value as 3 and plot the cluster membership using the cluster and
silhouette plots.
c. Specify the k value as 4 and plot the cluster membership using the cluster and
silhouette plots.
29. FILE Exercise_11.29. Perform k-means clustering on the accompanying data
set.
a. Use variables x1, x3, and x5 in the analysis. Do not standardize the variables.
Set the number of clusters to 4. What are the size and cluster center values for
the largest cluster?
 b. Perform the same analysis as in part a, but with variables standardized to z-
scores. What are the size and cluster center values for the largest cluster?

30. FILE Exercise_11.30. Perform k-means clustering on all the variables in the

accompanying data set.

a. Standardize the data. Specify the k value as 2 and plot the cluster membership
using the cluster and silhouette plots. What is the average silhouette width?
b. Specify the k value as 3 and plot the cluster membership using the cluster and
silhouette plots. What is the average silhouette width?
c. Specify the k value as 4 and plot the cluster membership using the cluster and
silhouette plots. What is the average silhouette width?

Applications
31. FILE Iris. British biologist Ronald Fisher studied iris flowers and classified them
according to the width and length of the flower’s petals and sepals (a small, green
leafy part below the petal). The accompanying table shows a portion of the data
that Fisher used in his study.

Sepal length Sepal width Petal length Petal width

5.2 3.5 1.4 0.2

4.9 3.0 1.4 0.2

⋮ ⋮ ⋮ ⋮

5.9 3.0 5.1 1.8

a. Perform k-means clustering using k = 4. What are the size and cluster center
values of the largest cluster?
b. Experiment with k = 3 and k = 5. What are the size and cluster center values of
the largest cluster in each case?
32. FILE Football_Players. Denise Lau is an avid football fan and religiously follows
every college football game. During the current season, she
meticulously keeps a record of how each quarterback has played
page 499
throughout the season. Denise is making a presentation at the local college
football fan club about these quarterbacks. The accompanying table shows a
portion of the data that Denise has recorded, with the following variables: the
player number (Player), completed passes (Comp), attempted passes (Att),
completion percentage (Pct), total yards thrown (Yds), average yards per attempt
(Avg), yards thrown per game (Yds/G), number of touchdowns (TD), and number
of interceptions (Int).
 a. Perform k-means clustering using k = 3 on all variables except the player
number. Standardize the variables. What are the size and cluster center values
of the largest cluster?
b. Select only the quarterbacks with at least 150 attempted passes and repeat part
a. Note: For Analytic Solver, use the Filter option in Excel to select the
observations. Copy the observations to a new worksheet. Do not alter the order
of the observations.
c. Comment on the differences in the clustering results of parts a and b.
33. FILE Napa. Jennifer Gomez is moving to a small town in Napa Valley, California,
and has been house hunting for her new home. Her Realtor has given her a list of
35 homes with at least 2 bedrooms that were recently sold. Jennifer wants to see
if she can group them in some meaningful ways to help her narrow down her
options. A portion of the data is shown in the accompanying table with the
following variables: the sale price of a home (Price, in $), the number of bedrooms
(Beds), the number of bathrooms (Baths), and the square footage (Sqft).

Price Beds Baths Sqft

799,000 4 3 2,689

795,680 4 2.5 2,507

⋮ ⋮ ⋮ ⋮

327,900 3 1 1,459

a. Perform k-means clustering on the list of 35 homes using k = 3. Standardize the

variables. What are the size and cluster center values of the largest cluster?
b. Perform k-means clustering on the standardized data using k = 4. What are the
size and cluster center values of the largest cluster?
34. FILE Baseball_Players. Ben Derby is a highly paid scout for a professional
baseball team. He attends at least five or six Major League Baseball games a
week and watches as many recorded games as he can in order to evaluate
potential players for his team to recruit at the end of the season. He also keeps
detailed records about each perspective player. His team is now seeking to add
another hitter to its roster next season. Luckily Ben has information on 144 hitters
 in the league who have played at least 100 games during the last season. A
portion of the data is shown in the accompanying table with the following
variables: the player number (Player), the number of games played (G), at bats
(AB), runs (R), hits (H), homeruns (HR), runs batted in (RBI), batting average
(AVG), on base percentage (OBP), and slugging percentage (SLG).

a. Perform k-means clustering to group the 144 players into three clusters.
Standardize and include all the variables except the player number in the
analysis. What is the size of the largest cluster? Which cluster has the highest
average number of homeruns?
b. Select only the players with at least 500 at bats (AB ≥ 500). Perform k-means
clustering with k = 3. What is the size of the largest cluster? Which cluster has
the highest average number of homeruns? Note: For Analytic Solver, use the
Filter option in Excel to select the observations. Copy the observations to a new
worksheet. Do not alter the order of the observations.

35. FILE Health_Population2. The accompanying data set contains country-

level health and population measures for 38 countries from the World Bank’s 2000
Health Nutrition and Population Statistics database. For each country, the
measures include death rate per 1,000 people (Death Rate, in %), health
expenditure per capita (Health Expend, in US$), life expectancy at birth (Life Exp,
in years), male adult mortality rate per 1,000 male adults (Male Mortality), female
adult mortality rate per 1,000 female adults (Female Mortality), annual population
growth (Population Growth, in %), female population (Female Pop, in %), male
population (Male Pop, in %), total population (Total Pop), size of labor force
(Labor Force), births per woman (Fertility Rate), birth rate per 1,000 people (Birth
Rate), and gross national income per capita (GNI, in US$). The accompanying
table shows a portion of the data.
 page 500
a. Perform k-means clustering to group the 38 countries into four
clusters according to their health measures (i.e., Death Rate, Health Expend,
Life Exp, Male Mortality, and Female Mortality) only. Is data standardization
necessary in this case?
b. What are the size and the average GNI per capita for the largest cluster of
countries?

36. FILE Health_Population2. Refer to the previous exercise for a description

of the data set.

a. Perform k-means clustering to group the 38 countries into four clusters
according to their population measures (i.e., Population Growth, Female Pop,
Male Pop, Total Pop, Labor Force, Fertility Rate, and Birth Rate) only. Is data
standardization necessary in this case?
b. Describe the characteristics of each cluster. Compare the average GNI per
capita of each group and report your findings.

37. FILE Chicago_CA. Sanjay Johnson is working on a research paper that

studies the relationship between the education level and the median income of a
community. The accompanying table shows a portion of the data that he has
collected on the educational attainment and the median income for 77 areas in
the city of Chicago. Sanjay plans to cluster the areas using the educational
attainment data and compare the average median incomes of the clusters. For
each community area, the measures include total number of residents 25 years
and over (25 or Over), number of residents with less than a high school education
(Less than HS), number of residents with a high school education (HS), number of
residents with some college (SC), number of residents with a Bachelor’s degree
or higher (Bachelor), and median household income (Income, in $). The
accompanying table shows a portion of the data.

a. Does Sanjay need to standardize the data before performing cluster analysis?
Explain.
 b. Perform k-means clustering to group the community areas into three clusters
based on the variables related to educational attainment of the population (i.e.,
Less than HS, HS, SC, and Bachelor). Plot the three clusters using the cluster
and silhouette plots. What is the average silhouette width? What are the size
and cluster center values of the largest cluster? Which cluster of community
areas has the highest average median household income?

38. FILE Nutritional_Facts. The accompanying data set contains the nutrition

facts on 30 common food items; a portion of the data is shown. The values are
based on 100 grams of the food items. Perform k-means clustering using k = 3 on
the nutritional facts of the food items. Standardize the variables. Describe the
characteristics of each cluster.

39. FILE Websites. Jake Duffy is the e-commerce manager of a major

electronics retailer. He is researching his company’s e-commerce competitors and

wants to group the competitors based on their performance data. He has
compiled a data set that contains the following four performance measures of the
major competitors’ e-commerce operations: Annual Revenue (Revenue, in $),
Growth Rate (in %), Monthly Visits (Visits), and Conversion Rate (in %). In order
to avoid any personal bias in grouping these competitors, he decided to replace
the names of the competitors with a code. A portion of the data set is shown in the
accompanying table.

a. Perform k-means clustering to group the competitors into five clusters using all
four performance measures. Does the data set need to be standardized prior to
cluster analysis?
 b. Plot the cluster membership using the cluster and silhouette plots. What is the
average silhouette width? Describe the characteristics of each cluster. What are
the size and cluster center values of the largest cluster of competitors?

40. FILE Development. The accompanying data set contains economic

development indicators for 11 African countries collected by the World Bank in

2015. The economic development indicators include annual % growth in
agriculture (Agriculture), annual % growth in exports (Exports), annual % growth
in final consumption (Final consumption), annual % growth in GDP (GDP), and
annual % growth in GDP per capita (GDP Per Capita). A portion
of the data set is shown in the accompanying table.
page 501

a. Perform k-means clustering to group the countries into three clusters using all
the development indicators. Is data standardization necessary in this case?
b. What is the size of the largest cluster? Which cluster has the highest average
growth in GDP per capita?
c. Compare the cluster membership between parts a and b.

41. FILE SAT_NYC. The accompanying data set contains the school level

average SAT critical reading (CR), math (M), and writing (W) scores for the
graduating seniors from 100 high schools in New York City. The data set also
records the number of SAT test takers (Test Takers) from each school.
 a. Perform k-means clustering to group the 100 high schools into four clusters on
the critical reading, math, and writing scores. Is data standardization necessary
in this case? What are the size and cluster center values of the largest cluster?
Which cluster has the fewest average number of test takers?
b. Select only the schools with at least 200 SAT test takers and repeat the cluster
analysis performed in part a. What are the size and cluster center values of the
largest cluster? Note: Use the Filter option in Excel to select the observations.
Copy the observations to a new worksheet. Do not alter the order of the
observations.
42. FILE Telecom. A telecommunications company wants to identify customers who
are likely to unsubscribe to the telephone service. The company collects the
following information from 100 customers: customer ID (ID), age (Age), annual
income (Income), monthly usage (Usage, in minutes), tenure (Tenure, in months),
and whether the customer has unsubscribed to the telephone service
(Unsubscribe). A portion of the data set is shown in the accompanying table.

a. Perform k-means clustering to group the 100 customers into four clusters based
on age, income, monthly usage, and tenure. Describe the characteristics of
each cluster.
b. Compute the percent of customers that have unsubscribed to the telephone
service from each cluster. Which cluster has the highest percent of customers
who have unsubscribed to the telephone service?
43. FILE Information_Technology. A country’s information technology use has
been linked to economic and societal development. A non-profit organization
collects data that measure the use and impact of information technology in over
100 countries annually. The accompanying table shows a portion of the data
collected, with the following variables: country number (Country), hardware
industry (Hw), software industry (Sw), telecommunications industry (Tele),
individual usage (IU), business usage (BU), government usage (GU), diffusion of
e-business (EB), and diffusion of e-payment (EP). Each value in the table
represents the score a country has achieved for a given category. The range of
the score is from 1 (lowest) to 7 (highest).
 a. Perform k-means clustering to group the countries into three clusters using the
following variables: individual usage, business usage, and government usage.
Is data standardization necessary in this case?
b. What are the size and cluster center values of the largest cluster? Which cluster
has the highest average diffusion of e-business? Which cluster has the highest
average diffusion of e-payment?
c. Use k-means clustering to group the countries using all the variables except the
country number into the same number of clusters as in part a. What are the size
and cluster center values of the largest cluster?

page 502

11.3 ASSOCIATION RULE ANALYSIS

LO 11.3
Conduct association rule analysis.

Association rule analysis, also referred to as affinity analysis or

market basket analysis, is another widely used unsupervised data
mining technique. It is essentially a “what goes with what” study
designed to identify events that tend to occur together. In a classic
application of association rule analysis, retail companies seek to
identify products that consumers tend to purchase together. This
type of information is useful to retail store managers when they
display products next to each other on the shelf or when they
develop promotional campaigns to cross-sell (encourage a customer
to purchase complementary products or services) or up-sell
(encourage a customer to purchase higher-end products or
services).
In addition to the brick-and-mortar retailers, other companies also
use association rule analysis to improve sales and customer
services. For example, after purchasing a smartphone on a retail
website, a customer is likely to receive a recommendation to
purchase other related products, such as a smartphone case or a
screen protector. Or after subscribing to a movie streaming service,
a customer is likely to see a list of recommended movies or TV
shows based on the information obtained from other subscribers with
similar viewing habits. In health care, association rule analysis can
help doctors diagnose illnesses based on different symptoms that
tend to occur together.
The simplest way to understand an association rule is to think of
it as an If-Then statement. Association rules rely on If-Then logic to
discover hidden relationships among transactions (or observations or
records) in a data set.

ASSOCIATION RULES ANALYSIS

Association rule analysis is a series of If-Then logical
statements that represent relationships among different items
or item sets. The analysis is designed to identify hidden
patterns and co-occurring events in data.

Consider an example where an online retailer wants to know how

often customers simultaneously purchase the following three
seemingly unrelated products: smartwatches, LED lightbulbs, and
books on meditation. After extracting a data set with the 10,000 most
recent transactions, the retailer observes that 1,200 of the 10,000
transactions include smartwatches and LED lightbulbs, and out of
these 1,200 transactions, 750 of them include meditation books.
Based on this scenario, we might suggest an association rule that if
a transaction includes smartwatches and LED light bulbs, then a
meditation book is likely to be purchased as well. This association
rule can also be expressed as {smartwatch, LED lightbulb} =>
{meditation book}. With this information, the retailer might discover
later that these three products are often purchased together by
health-conscious consumers who tend to purchase eco-friendly
products (the LED lightbulbs) as well as products that help improve
their physical and mental health (the smartwatches and meditation
books).
The If portion of this statement is called the antecedent, and the
Then portion is called the consequent. Antecedents and
consequents can comprise a single product or a combination of
products. For the {smartwatch, LED lightbulb} => {meditation book}
example, the antecedent consists of two products, while the
consequent has only one product. In association rule analysis, we
refer to such product or a combination of products as an item or
item set.

ANTECEDENT AND CONSEQUENT

In an If-Then logical statement, the item or item set in the If
portion of the statement is called the antecedent, and the item
or item set in the Then portion of the statement is called the
consequent.

page 503

One inherent problem with searching for hidden relationships

between items or item sets is dealing with the extremely large
number of possible combinations. For example, the three items
mentioned in the previous example would result in 12 possible
combinations that need to be examined. These combinations include
{smartwatch} => {LED lightbulb}, {smartwatch} => {meditation book},
{smartwatch} => {LED lightbulb, meditation book}, and so on. Four
items would result in 50 possible combinations. As you can see, the
number of possible combinations is exponential to the number of
items, and it can be estimated using the following formula:

where n is the number of items. Even a small convenience store with

a modest number of items, say 100, would lead to 5.15378E+47
possible combinations. With the number of items often in thousands
or even tens of thousands in many cases, the search problem
becomes extremely computationally intensive and time-consuming.
There are several algorithms that can be used to perform
association rule analysis in a more efficient manner. They all focus
on the frequency of item sets. One of the most widely used methods
is called the Apriori algorithm. The Apriori algorithm is designed to
recursively generate item sets that exceed a predetermined
frequency threshold; the term support of the item or item set is
often used synonymously with the frequency of the item or item set.
The Apriori method allows us to set a minimum support value, below
which an item or item set is excluded, thus making the analysis more
computationally feasible. For example, if it is found that
smartwatches are included in 2,000 of the 10,000 transactions, then
the support of smartwatches is calculated as 2,000/10,000 = 0.20.
The Apriori method eliminates infrequent items that are below a
predetermined frequency threshold, and makes it easier to analyze
relevant information in a large data set with thousands of items and
millions of transactions.

THE APRIORI ALGORITHM

The Apriori algorithm eliminates items or item sets that do not
meet a predetermined frequency threshold in order to reduce
the computational intensity of the association rule analysis.
 With enough data, we can propose many of these If-Then
association rules. However, we need a way to evaluate the
effectiveness of these rules; this is the strength of association among
products. Only the strong associations that occur frequently have the
potential to reappear consistently in the future and lead to financially
rewarding strategies. The support of the association rule is a ratio
of the number of transactions that include all the items mentioned in
the If-Then statement to the total number of transactions. In other
words, we examine the probability of the If-Then statement in the
data set. The support of the association rule is calculated as:

page 504

In the {smartwatch, LED lightbulb} => {meditation book} example,

750 of the 10,000 transactions included smartwatches, LED
lightbulbs, and meditation books, and, therefore, the support of the
{smartwatch, LED lightbulb} => {meditation book} association rule is
750/10,000 = 0.075.
The confidence of the association rule is another measure that
is interpreted as a conditional probability of the consequent
occurring, given that the antecedent occurs. The confidence of the
association rule is calculated as:

In the {smartwatch, LED lightbulb} => {meditation book} example,

smartwatches and LED lightbulbs are the products in the
antecedent, and 1,200 of the 10,000 transactions included both
smartwatches and LED lightbulbs. Therefore, the {smartwatch, LED
lightbulb} => {meditation book} association rule has a confidence of
750/1,200 = 0.625. This is the conditional probability that a customer
would purchase a meditation book, given that a smartwatch and an
LED lightbulb are included in the transaction. Naturally, a retailer
would look for an association rule that has high support and
confidence values as they suggest strong associations that occur
frequently.
However, sometimes, simply looking at support and confidence to
determine the strength of association can be misleading. For
example, suppose a store in town offers the lowest prices on broccoli
and milk, which attracts many customers to the store to purchase
these two items. As a result, you will get high support and
confidence values for the {broccoli} => {milk} association simply
because a large percent of customers purchase these two items. In
this case, there is not necessarily a relationship between the two
items. In order to avoid this common trap in association rules, we
often turn to the lift ratio to evaluate the strength of the association.
The lift ratio explicitly compares the confidence with a benchmark
value called “expected confidence.” It is calculated as:

where

In the {smartwatch, LED lightbulb} => {meditation book} example,

the expected confidence is simply an overall, unconditional
probability of a meditation book purchase, regardless of other
products. The lift ratio is a useful measure that compares the
confidence of the association rule with the overall unconditional
probability. A lift ratio of 1 suggests that the level of association
between the antecedent and the consequent is just as likely as if
there was no rule at all.
A strong and positive association rule should demonstrate a
better outcome (i.e., confidence) than the overall probability (i.e.,
expected confidence), and, therefore, we should expect a lift ratio to
be greater than 1 if a strong and positive association exists. The
higher the lift ratio, the stronger the association between the
antecedent and consequent. Conversely, a lift ratio between 0 and 1
indicates a negative association. The closer the lift ratio is to 0, the
stronger the negative association between the antecedent and
consequent is. If we assume that the number of transactions that
include a meditation book is 1,000, the expected confidence would
be 1,000/10,000 = 0.10. The lift ratio would be 0.625/0.10 = 6.25,
suggesting a strong positive association between {smartwatch, LED
lightbulb} and {meditation book} item sets. In practice, a common
approach used in association rule analysis is to sort all the rules by
lift ratio in descending order and focus on strong associations only.

ASSESSING AN ASSOCIATION RULE

An association rule can be evaluated using quantitative
measures such as the support of the association rule, the
confidence of the association rule, and the lift ratio.

The support of an association rule is calculated as:

Support values that are closer to one are preferred to values

closer to 0 because they suggest high occurrence frequency of
the association rule.

page 505

The confidence of an association rule is calculated as:

Confidence values that are closer to one are preferred to

values closer to 0 because they suggest strong associations.

The lift ratio compares the confidence with a benchmark value

called expected confidence. It is calculated as:

where

A lift ratio greater than 1 implies a strong and positive

association. A lift ratio between 0 and 1 indicates a negative
relationship.

EXAMPLE 11.4
 Consider the 10 retail transactions of cosmetics in Table 11.10.
For example, the first transaction includes a lipstick, mascara,
and an eye liner, while the second transaction includes an eye
shadow and a mascara.

TABLE 11.10 Retail Transactions of Cosmetics

Transaction Items

001 lipstick, mascara, eye liner

002 eye shadow, mascara

003 primer, concealer, lip gloss

004 face brush, concealer, primer

005 mascara, lipstick, eye liner

006 mascara, eye liner, eye shadow

007 lip gloss, eye liner, mascara

008 mascara, eye shadow, lip gloss

009 face brush, mascara, primer, eye liner

010 eye liner, lipstick, primer

a. Construct a frequency distribution that summarizes the

number of purchases for each cosmetic product. Based on the
frequency distribution, develop an association rule.
b. Compute and interpret the support, the confidence, and the lift
ratio of the association rule developed in part a.
SOLUTION:

a. Table 11.11 shows the frequency distribution. This helps us

gauge the support for each product. Seven of the 10
transactions include mascara (support = 7/10 = 0.7). The next
most frequently purchased product is eye liner (support = 6/10
= 0.6). Thus, it seems reasonable to learn how often
customers purchase mascara and eye liner together. The
corresponding association rule can be written as page 506
{mascara} => {eye liner}. Both antecedent
(mascara) and consequent (eye liner) of the association rule
have only one item.

TABLE 11.11 Frequency Distribution for Cosmetics

b. To assess the quality of this association rule, we first calculate

the support of the association rule. We find the proportion of
the transactions that include both mascara and eyeliner. Of
the 10 transactions, mascara and eye liner are purchased in
five transactions.

We then calculate the confidence of the association rule.

Here, we find the probability that the transaction includes
eyeliner, given that the transaction includes mascara. Of the
seven transactions that include mascara, five of them include
both mascara and eyeliner.

Finally, we calculate the lift ratio, which entails first

calculating the expected confidence, or, equivalently, the
probability that a transaction includes eye liner. Of the 10
transactions, there are six transactions with eye liner. So, the
expected confidence is:

The lift ratio is calculated as:

Thus, with a lift ratio greater than one, a strong and positive
association between the purchase of mascara and eye liner
appears to exist, relative to having no rule at all. The lift ratio of
1.19 implies that identifying a customer who purchased a
mascara as one who also purchased an eyeliner is 19% better
than just guessing that a random customer purchased an
eyeliner.
Using Analytic Solver and R to Perform
Association Rule Analysis
With a small data set, it is relatively easy to manually describe and
assess association rules with one item in the antecedent and
consequent. Discovering more complex associations with multiple
items in the antecedents and/or consequents in a larger data set will
require computer applications. The following example demonstrates
how to use association rule analysis in Analytic Solver and R.

page 507

Note: Association rule algorithms in Analytic Solver and R use

comma-separated data as an input.
FILE
Transaction

EXAMPLE 11.5
The store manager at an electronics store collects data on the
last 100 transactions. Five possible products were purchased:
a keyboard, an SD card, a mouse, a USB drive, and/or a
headphone. Table 11.12 shows a portion of the data. For
example, transaction 1 shows that the customer purchased a
keyboard, a mouse, and a headphone. This type of data is
called a “basket” of products, which is often used in a market
basket analysis.

TABLE 11.12 Sample of Transactions at Electronics Store

Transaction 1 keyboard, mouse, headphone

Transaction 2 keyboard, SD card, mouse

 ⋮ ⋮

Transaction 100 mouse, headphone

Conduct association rule analysis to help the store manager

understand the relationships among these products. How might
the store manager use this information?
SOLUTION:
Using Analytic Solver

a. Open the Transaction data file. [Note: The data are already in
a comma separated format.]
b. From the menu, choose Data Mining > Associate >
Association Rules.

c. See Figure 11.11. Click on the ellipsis button next to the

Data range option and highlight cells $A$1:$E$100. Make

sure that the box First Row Contains Headers is unchecked.
Accept other defaults and click OK.

FIGURE 11.11 Association Rules dialog box in Analytic

Solver
 Source: Microsoft Excel

d. Next to the Transaction worksheet, Analytic Solver will insert

two new worksheets: AR_Output and AR_PMMLModel. The
two worksheets contain the same output information, but in a
different format. We will focus on the Rules table from the
AR_Output worksheet that is shown in Figure 11.12. The
Rules table lists all the rules that meet the 10 page 508
minimum transactions and 50% minimum
confidence specified in the default option in step c. Figure
11.12 shows the rules sorted by the lift ratio. Of the nine rules
in the output, rules 2, 3, 6, 7, 8, and 9 have lift ratios greater
than 1 and relatively high confidence values.

FIGURE 11.12 Association rules results from Analytic Solver

 Source: Microsoft Excel

We find that rule 2, {keyboard} => {SDcard}, and rule 3,

{SDcard} => {keyboard}, have the highest lift ratios, suggesting
the strongest associations. One recommendation for the store
manager might be to place these two products next to each
other.
Now let’s assess the effectiveness of rule 2: {keyboard} =>
{SDcard}. The support frequency of 55 implies that 55 of the
100 transactions included both a keyboard and an SD card.
The confidence for rule 2 implies that there is an 82.09%
chance that a transaction includes an SD card, given that the
transaction includes a keyboard. Interestingly, rule 3, {SDcard}
=> {keyboard}, has higher confidence because SDcard has a
lower support value than keyboard has. And, finally, the lift ratio
of 1.32 means that identifying a customer who purchased a
keyboard as one who also purchased an SD card is 32% better
than just guessing that a random customer purchased an SD
card.
Using R

a. Install and load the arules and the arulesViz packages. Enter:
> install.packages(“arules”)
> install.packages(“arulesViz”)
> library(arules)
> library(arulesViz)
b. We use the read.transactions function to read the data and
store the data in a new variable called myData. This example
assumes that the Transaction data file is saved in the C:
folder (or the root directory of the C: drive). You will need to
make changes according to the location of the file on your
computer. Make sure to specify the format option as “basket”
and the sep option as “,”. Enter:
> myData <- read.transactions(“C:/Transaction.csv”, format =
“basket”, sep = “,”)
c. We use the inspect function to display the first five
transactions. Enter:
> inspect(myData[1:5])
Figure 11.13 shows the results.

FIGURE 11.13 A sample of the transaction data

page 509
d. We use the itemFrequency and the
itemFrequencyPlot functions to check the frequency of the
items. Enter:
> itemFrequency(myData)
> itemFrequencyPlot(myData)
Figure 11.14 shows the item frequency table and plot. The
results indicate that headphones appear in 32% of the
transactions (support of {headphone} = 0.32), while USB
drives appear in only 19% of the transactions (support of {USB
drive} = 0.19).
 FIGURE 11.14 R’s item frequency table and item frequency
plot

e. We use the apriori function to conduct association rule

analysis. We set minimum values of supp (support) and conf
(confidence) at 0.1 and 0.5, respectively. Practitioners usually
experiment with different support and confidence thresholds to
see whether or not interesting results might emerge.
Inspecting the lift ratios (see step f below) can also help
determine these thresholds. [If you do not specify any
parameter values, the default minimum support and
confidence values will be set at 0.1 and 0.8, respectively.] We
also set the minlen option to 2 to prevent R from displaying
association rules without an antecedent. Enter:
> rules <- apriori(myData, parameter = list(minlen = 2,
supp=0.1, conf=0.5))
f. To view the rules sorted by their lift ratios, we use the sort
function to sort the rules. We use the by and decreasing
options to sort by the lift ratio in descending order. Enter:
> srules <- sort(rules, by = ‘lift’, decreasing = TRUE)
> inspect(srules)
 Figure 11.15 shows the results. As expected, the association
rules generated by R are identical to those generated by
Analytic Solver, and, thus, the conclusions are the same as
those outlined in the Analytic Solver section.

FIGURE 11.15 Association rules generated by R

g. If we have a larger data set and the apriori function produces

many association rules, we can visually inspect the rules using
the plot function. The scatterplot makes it easy to identify the
rules with high lift ratios, confidence values, and support
values. Enter:
> plot(rules)
page 510
Figure 11.16 shows the results. The rules with
higher lift ratios are displayed in a more intense color. As the
plot shows, two rules have high lift ratios, confidence values,
and support values.

FIGURE 11.16 R’s scatterplot for the rules

EXERCISES 11.3
Notes: All data for these exercises are already in a comma-

separated (csv) format. and indicate that the exercises

must be solved using Analytic Solver and R, respectively. When

asked to identify the top association rule, select the rule with
the highest lift ratio. If multiple rules have the same lift ratio,
select the rule with the highest confidence value as the top rule.

Mechanics

44. FILE Exercise_11.44.csv Consider the following portion of data consisting

of 25 transactions.

Transaction 1 a, i, j, k

Transaction 2 c, g, i, k

⋮ ⋮

Transaction 25 a, e, f, j
 a. Generate association rules with a minimum support of 10 transactions and
minimum confidence of 75%. Sort the rules by lift ratio. What is the lift ratio for
the top rule?
b. Interpret the support count of the top rule.
c. Generate association rules with a minimum support of 5 transactions and
minimum confidence of 50%. What is the lift ratio for the top rule?
d. Interpret the confidence of the top rule.

45. FILE Exercise_11.45.csv The data are the same as in the preceding

exercise. Read the data file into R using the readtransactions function and
perform the following tasks.
a. Produce an item frequency plot and a frequency table. Which item is the most
frequent item?
b. Generate association rules with a minimum support of 0.25 and minimum
confidence of 0.75. Sort the rules by lift ratio. What is the lift ratio for the top
rule?

page 511
c. Interpret the support proportion of the rule with the largest lift
ratio in part b.
d. Generate association rules with a minimum support of 0.15 and minimum
confidence of 0.60. Sort the rules by lift ratio. What is the lift ratio for the top
rule?
e. Interpret the confidence of the rule with the largest lift ratio in part d.
f. Generate and interpret the scatterplot for the rules generated in part d.

46. FILE Exercise_11.46.csv Consider the following portion of data consisting

of 40 transactions. Read the data file into R using the readtransactions function
and perform the following tasks.

Transaction 1 a, d

Transaction 2 a, b

⋮ ⋮

Transaction 40 b, d, f

a. Produce an item frequency plot and frequency table. Which item is the least
frequent item?
 b. Generate association rules with a minimum support of 0.25 and minimum
confidence of 0.50. Sort the rules by lift ratio. What is the lift ratio for the top
rule?
c. Generate association rules with a minimum support of 0.10 and minimum
confidence of 0.50. Sort the rules by lift ratio. What is the lift ratio for the top
rule?

47. FILE Exercise_11.47.csv The data are the same as in the previous

exercise. Perform association rule analysis using the following settings.

a. Generate association rules with a minimum support of 10 transactions and
minimum confidence of 75%. How many rules are generated?
b. Generate association rules with a minimum support of 5 transactions and
minimum confidence of 50%. How many rules are generated? Sort the rules by
lift ratio. What is the lift ratio for the top rule?

48. FILE Exercise_11.48.csv Consider the following portion of data consisting

of 100 transactions.

Transaction 1 a, c, d

Transaction 2 c, e, g

⋮ ⋮

Transaction 100 c, d, e, g

a. Generate association rules with minimum support of 10 transactions and

minimum confidence of 50%. How many rules are generated? Sort the rules by
lift ratio. Which is the top rule? What is the lift ratio for the top rule?
b. Generate association rules with minimum support of 20 transactions and
minimum confidence of 80%. How many rules are generated? Sort the rules by
lift ratio. Which is the top rule? What is the lift ratio for the top rule?

49. FILE Exercise_11.49.csv The data are the same as in the previous

exercise. Read the data file using the readtransactions function and perform the
following tasks.
a. Produce an item frequency plot and frequency table. Which item is the most
frequent item?
b. Generate association rules with a minimum support of 0.1 and minimum
confidence of 0.75. Sort the rules by the lift ratio. Report and interpret the lift
ratio for the top rule.
 c. Generate association rules with a minimum support of 0.25 and minimum
confidence of 0.75. Sort the rules by the lift ratio. Report and interpret the lift
ratio for the top rule.
d. Generate a scatterplot to display the rules obtained in part c.

50. FILE Exercise_11.50.csv Consider the following portion of data consisting

of 41 transactions. Read the data file using the readtransactions function and
perform the following tasks.

Transaction 1 a, b, c, e, f

Transaction 2 a, b, c, d, e

⋮ ⋮

Transaction 41 a, b, c, e, f, g

a. Produce an item frequency plot and frequency table. Which item is the least
frequent item?
b. Generate association rules with a minimum support of 0.25 and minimum
confidence of 0.7. Sort the rules by lift ratio. Report and interpret the lift ratio for
the top rule.
c. Generate a scatterplot comparing the rules obtained in part b.

51. FILE Exercise_11.51.csv The data are the same as in the previous

exercise. Perform association rule analysis using the following settings.

a. Generate association rules with a minimum support of 10 transactions and
minimum confidence of 50%. Sort the rules by lift ratio. Report and interpret the
lift ratio for the top rule.
b. Generate association rules with a minimum of 10 transactions and minimum
confidence of 70%. How many rules are generated? Sort the rules by lift ratio.
Report and interpret the lift ratio and confidence for the top rule.

Applications

52. FILE Movies.csv An online movie streaming company conducts a

consumer study to find out the movie genres that its customers watch. Eighty-
eight households volunteered to participate in the study and
allow the company to track the genres of movies they watch over
page 512
a one-week period. A portion of the data is shown in the accompanying table.
 Record 1 shows that members of that household watched action, romance, and
drama movies during the week, while the second household watched only action
movies during the same week.

Record 1 action, romance, drama

Record 2 action

⋮ ⋮

Record 88 romance, drama

Perform association rule analysis using the following settings.

a. Specify a minimum support of 9 and a minimum confidence of 50%. Sort the
rules by lift ratio. What is the top rule? Report and interpret the lift ratio for the
top rule.
b. Specify a minimum support of 5 and a minimum confidence of 25%. Sort the
rules by lift ratio. What is the top rule? Report and interpret the lift ratio for the
top rule.

53. FILE Movies.csv Use the movie data set from the previous exercise and R

to perform association rule analysis. Make sure to read the data file using the
readtransactions function first.
a. Explore the data using an item frequency plot and a frequency table. Which
genre of movie is the most frequently watched?
b. Generate association rules with a minimum support of 0.1 and minimum
confidence of 0.5. How many rules are generated?
c. Sort the rules by lift ratio. What’s the lift ratio of the top rule? What does it
mean?

54. FILE Fruits.csv A local grocery store keeps track of individual products that

customers purchase. Natalie Jackson, the manager in charge of the fresh fruits
and produce section, wants to learn more about the customer purchasing patterns
of apples, bananas, cherries, oranges, and watermelons, the five most frequently
purchased fruit items in the store. She gathers data on the last 50 transactions, a
portion of which is shown in the accompanying table. Transaction 1 shows that
the customer purchased apples, bananas, and watermelon, while the second
customer purchased bananas, cherries, and oranges. Find association rules for
this consumer study. Use 10 as the minimum support and 50% as the confidence
values. Report and interpret the lift ratio of the top three rules.

Transaction 1 apple, banana, watermelon

 Transaction 2 banana, cherry, orange

⋮ ⋮

Transaction 50 apple, banana, watermelon, orange

55. FILE Beatles_Songs.csv. An online music streaming service wants to find

out which popular Beatles songs are frequently downloaded together by its users.
The service collects the download logs for 100 users over the past month, where
the download logs show which songs were downloaded during the same session.
A portion of the data is shown in the accompanying table.

User 1 Yesterday, All You Need Is Love . . . Here Comes the Sun

User 2 Hey Jude, Come Together . . . Here Comes the Sun

⋮ ⋮

User 100 Hey Jude, Come Together . . . Here Comes the Sun

Perform association rule analysis using the following settings.

a. Use a minimum support of 20 and a minimum confidence value of 60%. Sort
the rules by lift ratio. Report and interpret the lift ratio of the top rule.
b. Use a minimum support of 15 and a minimum confidence value of 50%. Sort
the rules by lift ratio. Report and interpret the lift ratio of the top rule.

56. FILE Beatles_Songs.csv. Use the online music data set from the previous

exercise to perform association rule analysis. Make sure to read the data file into
R using the readtransactions function first.
a. Explore the data using an item frequency plot and a frequency table. What is
the most frequently downloaded song?
b. Generate association rules with a minimum support of 0.1 and minimum
confidence of 0.5. Sort the rules by lift ratio. Report and interpret the lift ratio of
the top rule.
c. Display a scatterplot comparing the rules obtained in part b. How many rules
have at least 0.75 confidence, 0.15 support, and 1.2 lift ratio?

57. FILE Condition_Symptoms.csv Allen Chu is a first-year medical student.

He is interested in finding out which symptoms of medical conditions tend to

 appear together in his research. He compiles a data set for 36 patients where the
first entry lists the medical condition and the remaining entries are common
symptoms associated with the medical condition. A portion of the data is shown in
the accompanying table.

Patient Acute Sinusitis, Fever . . . Decreased smell

1

Patient Alcohol Intoxication, Abnormal gait (walking) . . . Stool leaking

2 (incontinence)

⋮ ⋮

Patient West Nile Virus, Chills . . . Confusion, Fever

36

page 513
Make sure to read the data file into R using the
readtransactions function first.
a. Generate association rules with a minimum support of 0.1 and minimum
confidence of 0.5. Sort the rules by lift ratio. Report and interpret the lift ratio of
the top three association rules.
b. Display a scatterplot comparing the rules obtained in part a. How many rules
have at least 0.75 confidence, 0.15 support, and 1.2 lift ratio?

58. FILE Crime_Analysis.csv Assistant police chief Todd Beck wants to

analyze the city’s historic crime data in order to better allocate police resources in
the future. He collects data over the past two years. Each record in the data
shows the type of crime reported and the location of the crime. Todd is interested
in answering the following question: Which types of crimes are often associated
with which locations?

Record 1 NARCOTICS, STREET

Record 2 ASSAULT, RESIDENCE

⋮ ⋮

Record 2,500 NARCOTICS, SIDEWALK

a. Generate association rules with a minimum support of 50 records and minimum

confidence of 30%. How many rules are generated?
 b. Based on the rules, which type of crime is most likely to be committed in a
department store? Which type of crime is most likely to be committed on the
sidewalk? Which type of crime is most likely to be committed in apartments?

59. FILE Crime_Analysis.csv Use the crime data from the previous exercise to

perform association rule analysis. Make sure to read the data file using the
readtransactions function first.
a. Explore the data using an item frequency plot and a frequency table. What type
of crime occurred most frequently? What location is associated with crimes
most frequently?
b. Generate association rules with a minimum support of 0.02 and minimum
confidence of 0.30. How many rules are generated?
c. Based on the rules, which type of crime is most likely to be committed in a
department store? Which type of crime is most likely to be committed on the
sidewalk? Which type of crime is most likely to be committed in apartments?

60. FILE Social_Media.csv Adrian Brown is a researcher studying social

media usage patterns. In his research, he noticed that people tend to use multiple
social media applications, and he wants to find out which popular social media
applications are often used together by the same user. He surveyed 100 users
about which social media applications they use on a regular basis. A sample of
the data set is shown in the accompanying table.

User 1 Facebook, Snapchat, LinkedIn

User 2 Facebook, Snapchat, Instagram, Pinterest, LinkedIn, Twitter

⋮ ⋮

User 100 Instagram, Tumbler, LinkedIn

a. Generate association rules with a minimum support of 20 and minimum

confidence of 60%. How many rules are generated?
b. Sort the rules by lift ratio. What is the top rule? Report and interpret the lift ratio
of the top rule.

61. FILE Business_Licenses.csv Claire Williams is an intern working for the

city government of New York City (NYC). She is given an assignment to find out
which types of business are most likely to open in which area of NYC. She
located a data file that contains the business license types and location of over
70,000 NYC businesses. A portion of the data is shown in the accompanying
 table. Make sure to read the data file into R using the readtransactions function
first.

License 1 Electronics Store,BROOKLYN

License 2 Home Improvement Salesperson,CORONA

⋮ ⋮

License 76,705 Secondhand Dealer - General,NEW YORK

Source: New York City Open Data, https://opendata.cityofnewyork.us/.

a. Generate association rules with a minimum support of 0.01 and minimum
confidence of 0.25. How many rules are generated?
b. Sort the rules by lift ratio. What is the top rule? Report and interpret the lift ratio
of the top rule.

62. FILE Bank_Accounts.csv Beau Crew is the branch manager of a major

bank. He is interested in understanding which types of accounts a customer tends

to have simultaneously so that he can offer additional financial services to his
clients. He has compiled a list of customers and the accounts they have. A portion
of the data is shown in the accompanying table. Find association rules for Beau’s
study. Use 10 as the minimum support and 50% as the minimum confidence
values. How many rules are generated. Report and interpret the lift ratio of the top
rule.

Customer 1 HELOC,IRA,Checking

Customer 2 Mortgage,CD,Checking,CD

⋮ ⋮

Customer 81 Checking,Savings

page 514
63. FILE Bank_Accounts.csv Use the bank account data set

from the previous exercise to perform association rule analysis. Make sure to read
the data file into R using the readtransactions function first.
a. Explore the data using an item frequency plot and a frequency table. Which
account type is the most frequent item?
 b. Generate association rules with a minimum support of 0.1 and minimum
confidence of 50%. Sort the rules by lift ratio. How many rules are generated?
Report and interpret the lift ratio of the top rule.

11.4 WRITING WITH BIG DATA

As explained earlier in this chapter, sometimes we need to convert the data into a
format that is appropriate for the analysis. With big data, it is impractical to transform
the data manually. In order to replicate the results for the big data case in this section
and complete the accompanying exercises, you will need to consult Chapter 2 on the
topic related to the transformation of categorical variables into binary variables.

Case Study
FILE Car_Crash. Ramona Kim is a California Highway Patrol (CHP) officer who
works in the city of San Diego. Having lost her own uncle in a car accident, she is
particularly interested in educating local drivers about driver safety. After discussing
this idea with her commanding officer, she learns that since 2005 the CHP
headquarters receives traffic-related collision information from local and state
agencies and makes them publicly available on its website. Her commanding officer
recommends that she focus on car accidents that result in deaths and severe injuries
and encourages Officer Kim to share her findings with her colleagues as well as
make presentations at local community meetings to raise awareness about car
accidents among local drivers.

Sample Report—Severe Car Crashes in San

Diego, CA
Many factors contribute to car accidents. Bad weather, reckless driving,
distracted driving, lack of visibility at night, and failure to observe traffic laws
can all lead to car accidents. In 2013, there were 3,000 deaths caused by
motor-vehicle-related accidents in California. That is almost 10 motor-vehicle-
related deaths per day. Conventional wisdom tells us that to stay safe, one
should be careful in severe weather or while driving at night, refrain from
driving under the influence, and always observe traffic laws. Beyond these
conventional safety rules, it is critical to gain a better understanding of the
different circumstances under which car accidents result in fatalities and
severe injuries.

John Panella/Alamy Stock Photo

Relevant San Diego traffic-accident data are extracted from the Statewide
Integrated Traffic Records System database from January 1, 2013, through
February 28, 2013. It is the rainy season in California during this time period;
thus, drivers often have to travel in heavy rain and other bad weather
conditions. Table 11.13 shows a portion of the data from the 785 accidents
that occurred in the city of San Diego over this time
period. The five variables of interest for the analysis
page 515
include the day of the week (WEEKEND =1 if weekend, 0 otherwise), crash
severity (CRASHSEV =1 if severe injury or fatality, 0 otherwise), whether or
not there was inclement weather (WEATHER = 1 if inclement, 0 otherwise),
whether or not the accident occurred on a highway (HIGHWAY = 1 if highway,
0 otherwise), and whether or not there was daylight (LIGHTING = 1 if
daylight, 0 otherwise). For example, the first observation represents an
accident that occurred on a weekday that resulted in a severe injury or fatality.
At the time of the accident, the weather was not inclement, but there was no
daylight. And finally, the accident did not occur on a highway.

TABLE 11.13 San Diego Traffic-Accident Data, January–February 2013

 Agglomerative clustering analysis is performed using the matching
coefficients as the distance measure and average linkage for the clustering
method. Because the values of 0 and 1 (e.g., for the WEEKEND variable, 0 =
weekdays, and 1 = weekends) are equally important in this analysis, the
matching coefficients distance measure is appropriate. Figure 11.17 shows
the resulting dendrogram plot.

FIGURE 11.17 Dendrogram for traffic-accident data

A careful inspection of the dendrogram suggests that the 785

observations should be grouped into four clusters, as shown in Figure 11.17.
Each cluster forms a meaningful group of observations that are distinctive
from one another. The cluster membership is assigned to each observation.
In order to gain a better understanding of these clusters, the average value of
each variable is calculated for each cluster. These averages are shown in
Table 11.14. The last column shows the number of cases in each cluster.

TABLE 11.14 Clustering Results for Traffic Accidents in San Diego

 page 516
Based on these results, the following observations can be made.

1. All of the accidents that resulted in fatalities or severe

injuries are grouped into Cluster 1 (53 cases) and
Cluster 2 (8 cases).
2. An overwhelming majority of fatal and severe-injury
cases in Cluster 1 occurred when the weather conditions
were clear.
3. Surprisingly, most fatal and severe-injury cases occurred
on a weekday; only 32.1% of cases in Cluster 1 and
25% of cases in Cluster 2 occurred over the weekend. A
possible explanation is that there are more weekdays
than weekends in a week and, therefore, weekdays
would naturally have a larger portion of the total fatal or
severe-injury cases.
4. Cluster 2 consists of eight cases that involved fatalities
or severe injuries. All eight cases happened during bad
weather conditions and after dark, and the majority of
them occurred on the highway.
5. The vast majority of the accidents are grouped into
Cluster 3. These accidents tended to happen during the
weekdays on nonhighway roads when the weather
condition was not inclement. None of them resulted in
fatalities or severe injuries.
6. All the accidents in Cluster 4 happened on the highway.
However, none of them resulted in fatality or severe
 injuries. Comparing Cluster 4 to Cluster 2, where the
majority of the accidents also happened on the highway,
but all cases resulted in fatality or severe injuries, we
notice that the key differences between the two clusters
are that all the accidents in Cluster 2 happened on
weekends during the day when the weather was not
inclement. While none of the variables may be a strong
indicator for fatal or severe-injury cases on its own, a
combination of traveling on the highway after dark on
weekdays during bad weather seems to increase the
chance of an accident resulting in fatality or severe
injuries.
The key finding from the cluster analysis is that many fatal and severe-
injury accidents occur on weekdays when the weather condition is clear.
Perhaps, when the weather condition is favorable, drivers tend to speed,
causing severe injuries or fatalities in an accident. During the weekdays,
drivers might also be in a rush to get home or to get to work, adding to the
severity of a car accident. This study can be used to design an educational
campaign to raise awareness among local drivers in San Diego.

Suggested Case Studies

Cluster analysis can be applied to many situations with the big data sets that
accompany this text. Here are some suggestions.

Report 11.1 FILE College_Admissions. Subset the data to include only one of the
three colleges. Group college applicants into clusters based on categorical variables
of your choice (e.g., sex, parent’s education). Determine the appropriate number of
clusters. Report the characteristics of each cluster and explain how they are different
from each other. Note: You will need to transform the categorical values into binary
variables in order to perform cluster analysis. See Chapter 2 for procedures used to
transform categorical variables.

Report 11.2 FILE College_Admissions. Subset the data to include only one of the
three colleges. Find a combination of numerical and categorical variables to include
in cluster analysis. Determine the appropriate number of clusters and write a report
to explain your decision and describe how each cluster is different
from other clusters. Note: Cluster analysis with mixed data can only
page 517
be performed in R.

Report 11.3 FILE TechSales_Reps. Subset the data set to include only sales
professionals in one of the two product groups. Cluster the sales professionals based
on numerical variables of your choice (e.g., age, salary). Determine the appropriate
number of clusters. Describe each cluster and compare the average net promoter
scores of the clusters.

Report 11.4 FILE NBA. Group NBA players into clusters based on either their
career performance or performance statistics from a particular season (e.g., 2013–
2014). Determine an appropriate number of clusters, and write a report based on the
clustering results.

Report 11.5 FILE Longitudinal_Survey. Subset the data set to include only those
individuals who lived in an urban area. Cluster the individuals using a combination of
numerical and categorical variables. Determine the appropriate number of clusters
and write a report to describe the differences between clusters. Note: Cluster
analysis with mixed data can only be performed in R.
 page 518

12 Forecasting with Time

Series Data

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 12.1  Describe the time series forecasting
process.
LO 12.2  Use smoothing techniques to make
forecasts.
LO 12.3  Use linear regression models to make
forecasts.
LO 12.4  Use nonlinear regression models to make
forecasts.
LO 12.5  Apply cross-validation techniques for
model selection.
LO 12.6  Use advanced smoothing methods to
make forecasts.
bservations of any variable recorded over time in sequential order
are considered a time series. Forecasting with time series is an
O important aspect of analytics, providing guidance for decisions
in all areas of business. In fact, the success of any business
depends on the ability to accurately forecast vital variables. Sound
forecasts not only improve the quality of business plans, but also
help identify and evaluate potential risks. Examples include
forecasting product sales, product defects, the inflation rate, cyber
attacks, or a company’s cash flows.
In this chapter, we focus on the trend, the seasonal, and the
random components of a time series. Several models are introduced
that capture one or more of these components. In particular, we use
simple smoothing techniques for making forecasts when short-term
fluctuations in the data represent random departures from the overall
pattern with no discernible trend or seasonal fluctuations.
Forecasting models based on regression and advanced smoothing
are introduced when trend and seasonal fluctuations are present in
the time series. Because it is highly unlikely to know a priori which of
the competing models is likely to provide the best forecast, we apply
in-sample and out-of-sample criteria to select the preferred model for
forecasting.

page 519
©icon Stocker/Shutterstock

INTRODUCTORY CASE
Apple Revenue Forecast
On August 2, 2018, Apple Inc. reported its fourth consecutive
quarter of record revenue and became the first publicly traded
American company to surpass $1 trillion in market value. Its
explosive growth has played a big role in the technology
industry’s ascent to the forefront of the global market economy
(The Wall Street Journal, Aug. 2, 2018). Although the company
designs, develops, and sells consumer electronics, computer
 software, and online services, the iPhones segment continues
to be the company’s core source of revenue.
Cadence Johnson, a research analyst at a small investment
firm, is evaluating Apple’s performance by analyzing the firm’s
revenue. She is aware that Apple could be seeing some
resistance to its newly revamped and high-priced line of
iPhones, stoking fears among investors that demand for
iPhones is waning. Cadence hopes that Apple’s past
performance will aid in predicting its future performance. She
collects quarterly data on Apple’s revenue for the fiscal years
2010 through 2018, with the fiscal year concluding at the end of
September. A portion of the data is shown in Table 12.1.
FILE
Revenue_Apple

TABLE 12.1 Quarterly Revenue for Apple Inc. (in $ millions)

Year Quarter Revenue

2010 1 15,683

2010 2 13,499

⋮ ⋮ ⋮

2018 4 62,900

Cadence would like to use the information in Table 12.1 to

1. Explore models that capture the trend and seasonal
components of Apple’s revenue.
2. Forecast Apple’s revenue for fiscal year 2019.
A synopsis of this case is provided at the end of Section 12.4.
 page 520

12.1 THE FORECASTING PROCESS

FOR TIME SERIES
LO 12.1
Describe the time series forecasting process.

In this chapter, we focus our attention on time series data.

Observations of any variable recorded over time in sequential order
are considered a time series. The time period can be expressed in
terms of a year, a quarter, a month, a week, a day, an hour, or even
a second. Examples of time series include the hourly volume of
stocks traded on the New York Stock Exchange (NYSE) on five
consecutive trading days; the daily number of loan applications over
the months of June and July; the monthly sales for a retailer over a
five-year period; and the annual growth rate of a country over the
past 30 years.
Let y1, y2, . . . , yT represent a sample of T observations of a
variable y with yt denoting the value of y at time t. With time series
data, it is customary to use the notation T, instead of n, to represent
the number of sample observations and to use a subscript t to
identify time. For instance, if the number of daily loan applications
over five days are 100, 94, 98, 110, 102, then y1 = 100, y2 = 94, . . . ,
y5 = 102.
Time series consist of the trend, the seasonal, the cyclical, and
the random components. The trend component represents long-
term upward or downward movements of the series. For example,
product sales or a firm’s stock price may go up (or go down) over a
certain time period. The seasonal component typically represents
repetitions over a one-year period. For example, every year, sales of
retail goods increase during the holiday season, and the number of
vacation packages sold goes up during the summer. The cyclical
component represents wavelike fluctuations or business cycles,
often caused by expansion and contraction of the economy. The
main distinction between seasonal and cyclical patterns is that
seasonal patterns tend to repeat within periods of one year or less,
whereas cyclical patterns last for one to several years—plus, the
duration of a cycle differs from one cycle to the next. In addition, the
magnitude of the up-and-down swings of the time series are more
predictable with seasonal patterns as opposed to cyclical patterns.
The random component is difficult to identify as it captures the
unexplained movements of the time series. For example, there may
be an increase or decrease in customers at a retail store for no
apparent reason. In this text, we will focus on the trend, the
seasonal, and the random components of a time series when making
forecasts.

TIME SERIES
A time series is a set of sequential observations of a variable
over time. It is generally characterized by the trend, the
seasonal, the cyclical, and the random components.

FILE
Revenue_Apple

In the introductory case, we considered Apple’s quarterly revenue

from 2010 through 2018, with the fiscal year concluding at the end of
September. Figure 12.1 is a scatterplot of the series, with the dots
connected, where we have relabeled the nine years of quarterly
observations from 1 to 36.

FIGURE 12.1 Scatterplot of Apple’s quarterly revenue (in $ millions)

 page 521

The graph highlights some important characteristics of Apple’s

revenue. First, there is a persistent upward movement with the
series plateauing near the end of the observation period. Second, a
seasonal pattern repeats itself year after year. For instance, revenue
is consistently higher in the first quarter as compared to the other
quarters. Note that given Apple’s fiscal calendar, the first quarter,
ending in December, encompasses the holiday period with usual
strong sales.

Forecasting Methods
Forecasting methods are broadly classified as quantitative or
qualitative. Qualitative forecasting methods are based on the
judgment of the forecaster, who uses prior experience and expertise
to make forecasts. On the other hand, quantitative forecasting
methods use a formal model along with historical data for the
variable of interest.
Qualitative forecasting is especially attractive when historical data
are not available. For instance, a manager may use qualitative
forecasts when she attempts to project sales for a new product.
Similarly, we rely on qualitative forecasts when future results are
suspected to depart markedly from results in prior periods, and,
therefore, cannot be based on historical data. For example, major
changes in market conditions or government policies will render the
analysis from historical data misleading.
Although attractive in certain scenarios, qualitative forecasts are
often criticized on the grounds that they are prone to some well-
documented biases such as optimism and overconfidence.
Decisions based on the judgment of an overly optimistic manager
may prove costly to the business. Furthermore, qualitative
forecasting is difficult to document, and its quality is totally
dependent on the judgment and skill of the forecaster. Two people
with access to similar information may offer different qualitative
forecasts.
Formal quantitative models have been used extensively to
forecast variables, such as product sales, product defects, house
prices, inflation, stock prices, and cash flows.

FORECASTING METHODS
Forecasting methods are broadly classified as quantitative or
qualitative. Qualitative methods are based on the judgment of
the forecaster, whereas quantitative methods use a formal
model to project historical data.

In this chapter, we focus on quantitative models to project

historical data, where each model is specially designed to capture
one or more components of a time series.

Model Selection Criteria

Numerous models can be used to make a forecast, with each model
well-suited to capture a particular feature of the time series. It would
be easy to choose the right model if we knew for certain which
feature describes the given series. Unfortunately, such certainty
rarely exists in the business world. Because we do not know which
of the competing models is likely to provide the best forecast, it is
common to consider several models. Model selection is one of the
most important steps in forecasting. Therefore, it is important to
understand model selection criteria before we even introduce any of
the formal models.
We can assess competing forecasting models, using what is
referred to as in-sample criteria, where the same sample period is
used for both the estimation of the model and its assessment. In-
sample criteria do not help us assess how well an estimated model
will predict in an unseen sample. In Chapters 7–10, we used cross-
validation techniques to evaluate competing models by partitioning
the sample into a training set to build (train) the model and a
validation set to assess (validate) it. Because it is possible for a
model to perform well in the training set but poorly in the validation
set, an independent assessment of the model performance is
achieved in the validation set, using what is often referred to as out-
of-sample criteria.

page 522

Both in-sample and out-of-sample criteria are based on the

forecast error where yt denotes the value of the series at

time t and denotes its forecast. Commonly used performance

measures for the comparison of competing forecasting models are

the familiar mean square error (MSE), the mean absolute deviation
(MAD), and the mean absolute percentage error (MAPE). These
performance measures were also discussed in Chapter 8; recall that
MAD is also referred to as the mean absolute error (MAE). The
preferred model will have the lowest MSE, MAD, or MAPE values.
Because there is no ‘primary’ model selection criterion, multiple
performance measures are used for model selection. The formulas
for MSE, MAD, and MAPE are shown in the following definition box.
 PERFORMANCE MEASURES
In-sample and out-of-sample criteria are based on the forecast
error where yt denotes the value of the series at time t

and denotes its forecast. We use et to compute the mean

square error (MSE), the mean absolute deviation (MAD), and

the mean absolute percentage error (MAPE) as

where n is the number of observations used in the computation.

We choose the model with the lowest MSE, MAD, or MAPE
values. Because there is no ‘primary’ performance measure,
multiple measures are used for model selection.

Recall from Chapter 8 that for any given model, there is no

universal “good” value for the above performance measures;
therefore, these are best used as model selection criteria. MSE
heavily penalizes models with large forecast errors and, therefore, is
preferred if relatively large forecast errors are particularly
undesirable. Another popular measure is the root mean square error
(RMSE), which is simply the square root of MSE. A large RMSE
relative to MAD is indicative of relatively large errors in the forecast.
The main attraction of using MAPE is that it shows the error as a
percentage of the actual value, giving a sense of the magnitude of
the errors.
 As mentioned earlier, in-sample and out-of-sample criteria give
rise to two important questions: How well does a model explain in-
sample forecasts in the training set? And how well does a model
make out-of-sample forecasts in the validation set? Ideally, the
chosen model is best in terms of in-sample and out-of-sample
criteria. We focus on in-sample criteria in Sections 12.2 through
12.4; we extend the discussion to include out-of-sample criteria in
Sections 12.5 and 12.6.

12.2 SIMPLE SMOOTHING

TECHNIQUES
LO 12.2
Use smoothing techniques to make forecasts.

As mentioned earlier, a time series is a sequence of observations

that are ordered in time. Inherently, any data collected over time is
likely to exhibit some form of random variation. For instance, the
checkout time at a campus bookstore or weekly sales at a
convenience store encounter random variations for no apparent
reason. In this section we focus on applications where the time
series is described primarily by random variations around an
unknown level. In other words, there are no variations due to trend
and/or seasonality.
A simple plot of the time series provides insights into its
components. A jagged appearance, caused by abrupt changes in the
series, indicates random variations. Smoothing page 523
techniques are employed to reduce the effect of
random fluctuations. These techniques can also be used to provide
forecasts if short-term fluctuations represent random departures from
the structure, with no discernible patterns. These techniques are
especially attractive when forecasts of multiple variables need to be
updated frequently. For example, consider a manager of a
convenience store who has to update the inventories of numerous
items on a weekly basis. It is not practical in such situations to
develop complex forecasting models for each item. We discuss two
distinct smoothing techniques: the moving average technique and
the simple exponential smoothing technique.

The Moving Average Technique

Due to its simplicity, the moving average technique ranks among
the most popular techniques for processing time series. The method
is based on computing the average from a fixed number m of the
most recent observations. For instance, a 3-period moving average
is formed by averaging the three most recent observations. The term
“moving” is used because as a new observation becomes available,
the average is updated by including the newest observation and
dropping the oldest observation.

CALCULATING A MOVING AVERAGE

An m-period moving average is computed as

EXAMPLE 12.1
In preparation for staffing during the upcoming summer
months, an online retailer reviews the number of customer
service calls received over the past three weeks (21 days).
Table 12.2 shows a portion of the time series.
a. Construct a 3-period moving average series for the data.
b. Plot the time series and its corresponding 3-period moving
average, and comment on any differences.
c. Using the 3-period moving average series, forecast the
number of customer service calls for the 22nd day.
d. Calculate MSE, MAD, and MAPE.
FILE
Service_Calls

TABLE 12.2 Daily Customer Service Calls

Day Calls

1 309

2 292

3 284

4 294

5 292

⋮ ⋮

19 326

20 327

21 309

page 524
 SOLUTION:

a. We would like to point out that the calculations are based on

unrounded values even though we show rounded values in
the text. For notational simplicity, let Calls be denoted by yt
and the corresponding moving average be denoted by We

form a 3-period moving average series by averaging all sets of

three consecutive values of the original series. The first value
of a 3-period moving average is calculated as

We designate this value because it represents the average

in days 1 through 3. The next moving average, representing

the average in days 2 through 4, is

Other values of are calculated similarly and are presented in

column 3 of Table 12.3. Note that we lose one observation at

the beginning and one observation at the end of the 3-period
moving average series (If it were a 5-period moving

average, we would lose two observations at the beginning and

two at the end.)

TABLE 12.3 3-Period Moving Averages, Forecasts, and

Errors
b. In Figure 12.2, we plot the time series and its corresponding 3-
period moving average against days. Note that the original
time series has a jagged appearance, suggesting the
presence of an important random component of the series.
The series of moving averages, on the other hand, presents a
much smoother picture.

FIGURE 12.2 Number of customer service calls and 3-period

moving average

page 525
c. As mentioned earlier, if the series exhibits
primarily random variations, we can use moving averages to
 generate forecasts. Because represents the average for

days 1 through 3, it is the most updated estimate of the series

prior to period 4. Therefore, we use where is the in-

sample forecast for period 4. Similarly, is the forecast for

period 5, where is the average for days 2 through 4, and so

on. These forecasts, derived as are shown in

column 4 of Table 12.3. Following this simple process, we

compute the out-of-sample forecast for day 22 as

Therefore, the forecast for the number of customer service

calls for the 22nd day is 321 calls. One potential weakness
when using the moving average technique is that all future
forecasts take on the same value as the first out-of-sample
forecast; that is, the forecast for the 23rd day is also 321 calls.
d. To calculate MSE, MAD, and MAPE, we first compute the
forecast error, as shown in column 5 of Table 12.3.
 These performance measures will prove useful when
comparing alternative models.

The Simple Exponential Smoothing

Technique
Although the moving average approach is popular, it has some
shortcomings. First, the choice of the order m is arbitrary, although
we can use trial and error to choose the value of m that results in the
smallest MSE, MAD, or MAPE. Second, it may not be appropriate to
give equal weight to all recent m observations. Whereas the moving
average technique weighs all recent observations equally, the
method called simple exponential smoothing assigns
exponentially decreasing weights as the observations get older. As in
the case of moving averages, exponential smoothing is a procedure
for continually revising a forecast in light of more recent
observations.
Let Lt denote the estimated level of the series at time t, where Lt
is defined as
That is, Lt is simply a weighted average of exponentially declining
weights, with α dictating the speed of decline. For example, with α =
0.8,

Similarly, with α = 0.2,

Note that the speed of decline is higher when α = 0.8 as compared

to when α = 0.2.

page 526

Using algebra, it can be shown that the initial equation simplifies

to

We use this representation to define the formula for exponential

smoothing. Because Lt represents the most updated level at time t,

we can use it to make a one-period-ahead forecast as This

equation shows that the forecast depends on the current

value, yt, and its earlier forecast, In other words, we

continually revise the forecast in the light of more recent
observations.

CALCULATING A SIMPLE EXPONENTIALLY SMOOTHED

SERIES
The simple exponential smoothing technique continually
updates the level of the series as

where α represents the speed of decline. Forecasts are made

as

In order to implement this method, we need to determine α and

the initial value of the series, L1. Typically, the initial value is set
equal to the first value of the time series, that is, L1 = y1; the choice
of the initial value is less important if the number of observations is
large. The optimal value for α is determined by a trial-and-error
method. We evaluate various values of α and choose the one that
results in the smallest MSE, MAD, MAPE, or some other selection
criteria. In Section 12.6, we show how to use Analytic Solver and R
for simple as well as advanced exponential smoothing methods.
Both software packages allow user-supplied and computer-
generated values for the smoothing parameters, including α.

EXAMPLE 12.2
Revisit the Service_Calls data from Example 12.1.
a. Construct the simple exponentially smoothed series with α =
0.20 and L1 = y1.
b. Plot time series and its corresponding exponentially smoothed
series against days. Comment on any differences.
c. Using the exponentially smoothed series, forecast the number
of customer service calls for the 22nd day.
d. Calculate MSE, MAD, and MAPE. Compare these values with
those obtained using the 3-period moving average technique
in Example 12.1.
SOLUTION: Again, the calculations are based on unrounded
values even though we show rounded values in the text.
a. In Column 3 of Table 12.4, we present sequential estimates of
Lt with the initial value L1 = y1 = 309. We use Lt = αyt + (1 −
α)Lt−1 to continuously update the level with α = 0.2. For
instance, for periods 2 and 3 we calculate

All other estimates of Lt are found in a like manner.

TABLE 12.4 Exponentially Smoothed Series with α = 0.20,

Forecasts, and Errors
b. In Figure 12.3, we plot the original time series and its
corresponding exponentially smoothed series against days. As
mentioned earlier, while the original time series has the jagged
appearance, the exponentially smoothed series removes most
of the sharp points and, like the moving average series,
presents a much smoother picture.
page 527
FIGURE 12.3 Number of customer service calls and
exponentially smoothed series

c. Forecasts, given by are presented in column 4 of Table

12.4. For instance, for period 2, Similarly, L2 =

305.60 is the forecast for Therefore, the forecast for the

22nd day is computed as

or 311 customer service calls. As with the moving average

technique, any further out-of-sample forecasts also assume
this same value; for instance, customer service calls.
To calculate MSE, MAD, and MAPE, we first compute the
forecast error, as shown in column 5 of Table 12.4.

There is nothing special about α = 0.2; we used this value

primarily to illustrate the exponential smoothing technique. As
we noted earlier, it is common to evaluate various values of α
and choose the one that produces the smallest MSE, MAD, or
MAPE values. In order to illustrate how α is chosen, we
generate MSE, MAD, and MAPE with α values ranging from
0.1 to 0.9 with increments of 0.1. The results are summarized
in Table 12.5.

TABLE 12.5 Various Values of α and the Resulting MSE,

MAD, and MAPE

Here, the choice of α depends on whether we employ

MSE, MAD, or MAPE for model comparison. In this example,
it may be appropriate to select α = 0.5 because it leads to the
smallest value for two out of three performance measures
(MAD and MAPE). This model (with α = 0.5) page 528
 also outperforms the moving average model as measured by
lower MSE, MAD, and MAPE values.

Using Excel for Moving Averages and

Exponential Smoothing
Obtaining the Moving Averages in Table 12.3
FILE
Service_Calls

. Open the Service_Calls data file.

. From the menu, choose Data > Data Analysis > Moving Average.
Click OK.
. Click on the box next to Input Range, select the Calls data
(including the heading), and then check the box in front of Labels in
First Row. Next to Interval, enter 3 because we want to generate a
3-period moving average. Finally, indicate an Output Range; we
enter D3, thus replicating the forecasts shown in column 4 of Table
12.3. Click OK.
Obtaining the Exponentially Smoothed Series in Table 12.4
. Open the Service_Calls data file.
. From the menu, choose Data > Data Analysis > Exponential
Smoothing. Click OK.
. Click on the box next to Input Range, select the Calls data
(including the heading), and then check the box in front of Labels.
Select the box next to Damping Factor. If we want to construct an
exponentially smoothed series with α = 0.2, then for Damping
Factor we enter 1 − α = 0.8. Finally, indicate an Output Range; we
enter D2, thus replicating the forecasts shown in column 4 of Table
12.4. Click OK.
 Note: In Section 12.6, we will show how to use Analytic Solver
and R for simple as well as advanced exponential smoothing
techniques. Both software packages allow user-supplied and
computer-generated values for the smoothing parameters, including
α.

EXERCISES 12.2
Applications
1. FILE Convenience_Store. The owner of a convenience store near Salt Lake
City in Utah has been tabulating weekly sales at the store, excluding gas. The
accompanying table shows a portion of the sales for 30 weeks.

Week Sales

1 5387

2 5522

⋮ ⋮

30 5206

a. Use the 3-period moving average to forecast sales for the 31st week.
b. Use simple exponential smoothing with α = 0.3 to forecast sales for the 31st
week.
c. Which is the preferred technique for making the forecast based on MSE, MAD,
and MAPE?
2. FILE Spotify. Spotify is a music streaming platform that gives access to songs
from artists all over the world. On February 28, 2018, Spotify filed for an initial
public offering (IPO) on the New York Stock Exchange. The accompanying table
shows a portion of the adjusted monthly stock price of Spotify from April 1, 2018,
to February 1, 2019.

Date Stock Price

Apr-18 161.67

May-18 157.71
 Date Stock Price

⋮ ⋮

Feb-19 134.71

page 529
a. Use the 3-period moving average to forecast Spotify’s stock
price for March 2019.
b. Use simple exponential smoothing with α = 0.2 to forecast Spotify’s stock price
for March 2019.
c. Which is the preferred technique for making the forecast based on MSE, MAD,
and MAPE?
3. FILE FoodTruck. Food trucks have become a common sight on American
campuses. They serve scores of hungry students strolling through campus and
looking for trendy food served fast. The owner of a food truck collects data on the
number of students he serves on weekdays on a small campus in California. A
portion of the data is shown in the accompanying table.

Weekday Students

1 84

2 66

⋮ ⋮

40 166

a. Use the 3-period moving average to make a forecast for Weekday 41.
b. Use the 5-period moving average to make a forecast for Weekday 41.
c. Which is the preferred technique for making the forecast based on MSE, MAD,
and MAPE?
4. Exchange_Rate. Consider the exchange rate of the $ (USD) with € (Euro) and $
(USD) with £ (Pound). The accompanying table shows a portion of the exchange
rates from January 2017 to January 2019.

Date Euro Pound

Jan-17 1.0635 1.2367

 Date Euro Pound

Feb-17 1.0650 1.2495

⋮ ⋮ ⋮

Jan-19 1.1414 1.2845

a. Find the 3-period and the 5-period moving averages for Euro. Based on MSE,
MAD, and MAPE, use the preferred model to forecast Euro for February 2019.
b. Find the simple exponential smoothing series for Pound with possible α values
of 0.2, 0.4, 0.6. Based on MSE, MAD, and MAPE, use the preferred model to
forecast Pound for February 2019.
5. FILE Downtown_Cafe. The manager of a trendy downtown café in Columbus,
Ohio, collects weekly data on the number of customers it serves. A portion of the
data is shown in the accompanying table.

Week Customers

1 944

2 997

⋮ ⋮

52 1365

a. Use the simple exponential smoothing technique with α = 0.2 to make a

forecast for Week 53.
b. Use the simple exponential smoothing technique with α = 0.4 to make a
forecast for Week 53.
c. Which is the preferred technique for making the forecast based on MSE, MAD,
and MAPE?
6. FILE Gas_Prices. It is difficult to predict gas prices given a multitude of factors
affecting them. Consider 22 weeks of the average weekly regular gasoline price ($
per gallon) in New England and the West Coast.

Date New England West Coast

9/3/2018 2.855 3.329

 Date New England West Coast

9/10/2018 2.864 3.336

⋮ ⋮ ⋮

1/28/2019 2.346 2.928

c. Find the 3-period and the 5-period moving averages for gas prices in New
England. Based on MSE, MAD, and MAPE, use the preferred model to forecast
gas prices for the first week of February 2019.
d. Find the simple exponential smoothing series with possible α values of 0.2, 0.4,
0.6 for gas prices on the West Coast. Based on MSE, MAD, and MAPE, use the
preferred model to forecast gas prices for the first week of February 2019.

12.3 LINEAR REGRESSION MODELS

FOR TREND AND SEASONALITY
LO 12.3
Use linear regression models to make forecasts.

The smoothing techniques discussed in Section 12.2 are used when

the time series represent random fluctuations with no discernible
trend or seasonal fluctuations. When trend and seasonal variations
are present, we need to use special models for the analysis. In this
section, we first focus on trend analysis, which extracts long-term
upward or downward movements of the series. We then incorporate
seasonal dummy variables that extract the repetitive movement of
the series within a one-year period.

page 530

The Linear Trend Model

We estimate a linear trend model using the regression techniques
described in earlier chapters. Let yt be the value of the response
variable at time t. Here we use t as the predictor variable
corresponding to consecutive time periods, such as 1, 2, 3, and so
on.

THE LINEAR TREND MODEL

The linear trend model is used for a time series that is expected
to grow by a fixed amount each time period. It is specified as yt
= β0 + β1t + εt, where yt is the value of the time series at time t.

The estimated model is used to make forecasts as

where b0 and b1 are the coefficient estimates.

Example 12.3 provides an application of the linear trend model

for making forecasts.

EXAMPLE 12.3
A local organic food store carries several food products for
health-conscious consumers. The store has witnessed a
steady growth in the sale of chef-designed meals, which are
especially popular with college-educated millennials. For
planning purposes, the manager of the store would like to
extract useful information from the weekly sales of chef-
designed meals for the past year, a portion of which is shown in
Table 12.6.
FILE
Organic

Table 12.6 Weekly Sales (in $) at Organic Food Store

 Week Sales

1 1925

2 2978

⋮ ⋮

52 6281

a. Visually inspect the time series to confirm the existence of a

trend.
b. Estimate and interpret the linear trend model for the sale of
chef-designed meals.
c. Forecast the sale of chef-designed meals for the next four
weeks.
SOLUTION:

a. As a first step, it is advisable to visually inspect the time

series. Figure 12.4 is a scatterplot of the weekly sales of chef-
designed meals for the past year with a superimposed linear
trend line. We see an upward movement of the series over
this time period.

FIGURE 12.4 Scatterplot of weekly sales of chef-designed

meals
 page 531
b. The linear trend model is specified as Sales = β 0
+ β 1 Week + ε. The estimated linear trend equation is

Week, implying that every week, sales

increase by about $96.84. The estimated R2 = 0.7538

suggests that about 75.38% of the sample variations in sales
are explained by the sample trend line.
Note: Please refer to Chapter 6 for Excel and R instructions
on how to estimate the linear model.
c. We use Week = 53 to forecast sales for the next week as
Similarly, we use Week =

54, 55, and 56 to forecast sales for the subsequent three

weeks as $7,227.49, $7,324.33, $7,421.17, respectively.

Trend forecasting models, like the model used in Example 12.3,

extract long-term upward or downward movements of a time series.
These models are appropriate when the time series does not exhibit
seasonal variations or has been stripped of its seasonal variation;
that is, it has been deseasonalized. We now move our attention to
making forecasts that extract both trend and seasonal variations.

The Linear Trend Model with Seasonality

With seasonal data, we estimate a linear trend model that also
includes dummy variables to capture the seasonal variations. Recall
that a dummy variable is commonly used to describe a categorical
variable with two categories. Here, we use dummy variables to
describe seasons. For quarterly data, we need to define only three
dummy variables representing three quarters, using the fourth
quarter as reference.

LINEAR TREND MODEL WITH SEASONAL DUMMY

VARIABLES
With quarterly data, a linear trend model with seasonal dummy
variables can be specified as

where d1, d2, and d3 are the dummy variables representing the
first three quarters. Forecasts based on the estimated model
are made as follows:

Here, b0, b1, . . . b4 are the coefficient estimates.

 Note that the above forecasting equations can easily be modified
if a different quarter is used as reference. Also, the model for the
quarterly data can be modified to make forecasts with monthly or
other forms of seasonal data. Example 12.4 provides an application
with quarterly data.

EXAMPLE 12.4
With Amazon.com at the lead, e-commerce retail sales have
increased substantially over the last decade. Consider
quarterly data on e-commerce retail sales in the U.S. from the
first quarter of 2010 to the first quarter of 2019, a portion of
which is shown in Table 12.7.

page 532
FILE
ECommerce

Table 12.7 Quarterly e-Commerce Retail Sales (in $ millions)

Year Quarter Sales

2010 1 37059

2010 2 38467

⋮ ⋮ ⋮

2019 1 127265

a. Visually inspect the data to confirm the existence of trend and

seasonality.
b. Estimate and interpret the linear trend model with seasonal
dummy variables for e-commerce retail sales.
c. Forecast e-commerce retail sales for the last three quarters of
2019.
SOLUTION: Given quarterly data, we have to first construct
relevant variables for the linear trend models with seasonal
dummy variables. Table 12.8 presents a portion of the
constructed data for sales, seasonal dummy variables d1, d2,
and d3 representing the first three quarters (using the fourth
quarter as reference), and the trend variable t representing the
37 quarters of data. (Shortly, we will provide R instructions for
replicating the results.)

Table 12.8 Constructing Variables for Example 12.4

a. Figure 12.5 is a scatterplot of quarterly sales, with the dots

connected, and the quarterly data relabeled from 1 to 37. The
graph highlights some important characteristics of e-
commerce retail sales. First, there is a persistent upward
movement in sales. Second, a seasonal pattern repeats itself
year after year. For instance, sales are consistently higher in
the fourth quarter as compared to the other quarters. The
graph makes a strong case for a model that captures both
seasonality and trend.
 FIGURE 12.5 Scatterplot of e-commerce retail sales (in $
millions)

page 533
b. The estimated linear trend model with seasonal
dummy variables is:

The coefficients for the seasonal dummy variables indicate

that, relative to the 4th quarter, e-commerce sales are about
$22,000 lower in the other three quarters. The estimated
coefficient for the trend variable suggests that the predicted
quarterly sales increase by about $2,650 million every quarter,
in addition to the seasonal variations.
c. For the 2nd quarter of 2019, we use d 1 = 0, d 2 = 1, d 3 = 0,
and t = 38 to forecast

Similarly, we use d 1 = 0, d 2 = 0, d 3 = 1, and t = 39 for the 3rd

quarter and d 1 = 0, d 2 = 0, d 3 = 0, t = 40 for the 4th quarter to
forecast sales as $127,434 and $152,678 million, respectively.
Estimating a Linear Trend Model with Seasonality with R
FILE
ECommerce

In order to replicate the results in Example 12.4, we follow these

steps.
. Import the ECommerce data into a data frame (table) and label it
myData.
. Install and load the forecast package. Enter:

> install.packages(“forecast”)
> library(forecast)

. We use the ts function to create a time series object and call it

newData. Within ts, we specify the start and end periods as well as
frequency, denoting the number of seasons in a year. Enter:

> newData <- ts(myData$Sales, start = c(2010,1), end = c(2019,1),

frequency=4)
. We use the tslm function to estimate the model and the summary
function to view the regression output. Enter:

> TSReg <- tslm(newData ~ trend + season)

> summary(TSReg)

Note that by default R uses the first quarter as the reference

season, which explains why the intercept and the coefficient
estimates differ from those reported in Example 12.4. The forecasts,
however, are not influenced by the choice of the reference dummy
variable.
. Use the forecast function where h denotes the number of forecasts.
Enter:
> forecast(TSReg, h=3)

A Note on Causal Models for Forecasting

So far, we have discussed noncausal, or purely time series, models
that capture trend and seasonality for making forecasts. These
models do not offer any explanation of the mechanism generating
the target variable and simply provide a method for projecting
historical data. Causal models, on the other hand, are standard
regression models that exploit the relationship between the target
and the predictor variables for making forecasts. For example, we
can use causal models to forecast product sales using predictor
variables such as the firm’s advertising budget and its pricing
strategy. The regression framework also allows us the flexibility to
combine causal with time series effects. In other words, the list of
predictor variables may include k causal variables x 1, x 2, . . . x k
along with the trend variable t and seasonal dummy variables d 1, d
2, . . . d p−1 for p seasons.

page 534

For example, consider quarterly data for developing a forecasting

model for product sales y. Let the predictor variables include the
firm’s advertising budget x, the trend variable t, and seasonal dummy
variables, d1, d2, d3, representing the first three quarters. We can
easily estimate the regression model to make forecasts as:

Note that in addition to the known future values of t, d1, d2, and d3,
this approach will work only if we also know, or can predict, the
future value of the variable x. In other words, we cannot forecast
product sales if the advertising budget in the future is not known.
Sometimes, we can justify the use of lagged values of causal
variables for making forecasts. In the product sales example, the
relationship between advertising budget and sales may not be
contemporaneous. Therefore, we can specify a model where product
sales is related to the trend term, seasonal dummy variables, and
the value of the advertising budget in the previous period. Further
discussion of lagged regression models is beyond the scope of this
text.

EXERCISES 12.3
Applications
7. FILE Inquiries Morgan Bank has been encouraging its customers to use its new
mobile banking app. While this may be good for business, the bank has to deal
with a number of inquiries it receives about the new app. The following table
contains a portion of weekly inquiries the bank has received over the past 30
weeks.

Week Inquiries

1 286

2 331

⋮ ⋮

30 219

Estimate the linear trend model to forecast the number of inquiries over the next
two weeks.
8. FILE Apple_Price. Apple Inc. has performed extremely well in the last decade.
After its stock price dropped to below 90 in May 2016, it made a tremendous
comeback to reach about 146 by May 2017 (SeekingAlpha.com, May 1, 2017). An
investor seeking to gain from the positive momentum of Apple’s stock price
analyzes 53 weeks of stock price data from 5/30/16 to 5/26/17. A portion of the
data is shown in the accompanying table.

Date Price
 Date Price

5/30/2016  97.92

6/6/2016  98.83

⋮ ⋮

5/26/2017 153.57

a. Estimate and interpret the linear trend model (no seasonality).

b. Make a forecast for the next week (54th week).
9. FILE Tax_Revenue. In Colorado, sales of medical marijuana began in
November 2012; however, the Department of Revenue did not report tax
collection data until February of 2014. The accompanying table shows a portion of
the monthly revenue from medical and retail marijuana tax and fee collections as
posted in the Colorado state accounting system.

Date Revenue

Feb-14 3,519,756

Mar-14 4,092,575

⋮ ⋮

Oct-18 22,589,679

Use the linear trend model (no seasonality) to forecast the tax revenue for
November and December of 2018.
10. FILE Revenue_Lowes. Lowe’s Companies, Inc., is a home improvement
company offering a range of products for maintenance, repair, remodeling, and
decorating. During the recovery phase since the financial crisis of 2008, Lowe’s
has enjoyed a steady growth in revenue. The following table contains a portion of
quarterly data on Lowe’s revenue (in $ millions) with its fiscal year concluding at
the end of January.

Year Quarter Revenue

2010 1 12,388

2010 2 14,361
 Year Quarter Revenue

⋮ ⋮ ⋮

2018 3 17,415

a. Estimate and interpret the linear trend model with seasonal dummy variables.
b. Use the estimated model to forecast Lowe’s revenue for the fourth quarter of
2018.

page 535
11. FILE Vacation Vacation destinations often run on a seasonal
basis, depending on the primary activities in that location. Amanda Wang is the
owner of a travel agency in Cincinnati, Ohio. She has built a database of the
number of vacation packages (Vacation) that she has sold over the last twelve
years. The following table contains a portion of quarterly data on the number of
vacation packages sold.

Year Quarter Vacation

2008 1 500

2008 2 147

⋮ ⋮ ⋮

2019 4 923

a. Estimate the linear regression models using seasonal dummy variables with
and without the trend term.
b. Determine the preferred model and use it to forecast the quarterly number of
vacation packages sold in the first two quarters of 2020.
12. FILE Consumer_Sentiment. The following table lists a portion of the University
of Michigan’s Consumer Sentiment index. This index is normalized to have a
value of 100 in 1966 and is used to record changes in consumer morale.

Date Consumer Sentiment

Jan-10 74.4

Feb-10 73.6
 Date Consumer Sentiment

⋮ ⋮

Nov-18 97.5

a. Estimate and interpret the linear trend model with seasonal dummy variables.
b. Use the estimated model to make a consumer sentiment index forecast for
December 2018.
13. FILE UsedCars Used car dealerships generally have sales quotas that they
strive to hit each month, quarter, and calendar year. Consequently, buying a used
car toward the end of those periods presents a great opportunity to get a good
deal on the car. A local dealership has compiled monthly sales data for used cars
(Cars) from 2014-2019, a portion of which is shown in the accompanying table.

Date Cars

Jan-2014 138

Feb-2014 179

⋮ ⋮

Dec-2019 195

a. Estimate the linear regression models using seasonal dummy variables with
and without the trend term.
b. Determine the preferred model and use it to forecast used cars sales in the first
two months of 2020.

12.4 NONLINEAR REGRESSION

MODELS FOR TREND AND
SEASONALITY
LO 12.4
Use nonlinear regression models to make forecasts.
Although the linear relationship assumed in Section 12.3 can be
adequate, there are many cases in which a nonlinear functional form
is more suitable. In this section, we discuss the exponential, the
quadratic, and the cubic trend models with and without seasonal
dummy variables.

The Exponential Trend Model

A linear trend model uses a straight line to capture the trend, thus
implying that for each period, the value of the series is expected to
change by a fixed amount, given by the estimated coefficient b1. The
exponential trend model is attractive when the expected increase
in the series gets larger over time. It is not uncommon for some
variables to exhibit exponential growth over time. For example, in
recent years, technology firms such as Amazon, Netflix, Spotify,
Airbnb, and Paypal have exhibited exponential growth for which the
linear trend model would clearly be inadequate.

page 536

Figure 12.6 presents a scatterplot of a time series with

superimposed linear and exponential trend lines. While both trend
lines capture positive growth, the exponential trend line (green)
correctly allows the values to grow by an increasing amount over
time. Here, the linear trend line (blue) would under-forecast future
values.

FIGURE 12.6 Scatterplot with superimposed linear and exponential trend lines
 Recall from Chapter 7 that we specify an exponential model as
ln(yt) = β0 + β1t + εt. In order to estimate this model, we first generate
the series in natural logs, ln(yt), and then run a regression of ln(yt) on
t. Because, in the exponential model, the response variable is
measured in logs, we make forecasts in regular units as
where se is the standard error of the estimate.

THE EXPONENTIAL TREND MODEL

The exponential trend model is appropriate for a time series
that is expected to grow by an increasing amount each time
period. It is specified as ln(yt) = β0 + β1t + εt, where ln(yt) is the
natural log of yt. The estimated model is used to make

forecasts as where b0 and b1 are the

coefficient estimates and se is the standard error of the

estimate.
 Note: It is advisable to use unrounded values for making
forecasts in an exponential model because even a small difference,
when exponentiated, can make a big difference in the forecast.
Example 12.5 provides an application of the linear and the
exponential trend models for making forecasts.

EXAMPLE 12.5
According to data compiled by the World Bank, the world
population has increased from 3.03 billion in 1960 to 7.53
billion in 2017. This rapid increase concerns environmentalists,
who believe that our natural resources may not be able to
support the ever-increasing population. Additionally, most of
the rapid population growth has been in 34 low-income
countries, many of which are located in Africa. Consider the
population data, in millions, for low-income countries from 1960
to 2017, a portion of which is shown in Table 12.9.
FILE
Population_LowInc

TABLE 12.9 Population in Low-Income Countries

Year Population

1960 166.5028

1961 170.2108

⋮ ⋮

2017 732.4486
 page 537

a. Estimate and interpret the linear and the exponential trend

models.
b. Use MSE, MAD, and MAPE to select the appropriate model.
c. Forecast the population in low-income countries for 2018.
SOLUTION: As a first step, it is advisable to inspect the data
visually. Figure 12.7 is a scatterplot of the population in low-
income countries from 1960 through 2017. We relabel the 58
years of annual observations from 1 to 58 and superimpose the
linear and exponential trends to the data.

FIGURE 12.7 Population in low-income countries with superimposed trends

The scatterplot makes a strong case for an exponential growth

in the population in low-income countries.
a. We estimate the linear and the exponential trend models,
where the linear model is used for comparison with the
visually preferred exponential model. To estimate the
exponential trend model, we first transform the population
series to natural logs. Table 12.10 shows a portion of the log-
transformed population along with the predictor variable t,
relabeled from 1 to 58.
TABLE 12.10 Constructing Variables for Example 12.5
 Year Population t ln(Population)

1960 166.5028 1 5.1150

1961 170.2108 2 5.1370

⋮ ⋮ ⋮ ⋮

2017 732.4486 58 6.5964

The estimated trend models, where y denotes population, are:

Note: Please refer to Chapter 7 for the Excel and R

instructions for estimating the exponential model.
The slope coefficient of the linear trend model implies that
the population in low-income countries has grown, on
average, by approximately 9.6799 million each year. The
slope coefficient of the exponential trend model implies that
the population in low-income countries has grown, on
average, by approximately 2.64% (= 0.0264 × 100). The exact
growth rate can be calculated as 2.68% (= (exp(0.0264) − 1) ×
100).
b. In Table 12.11, we present a portion of the series yt along with

for both models using unrounded values for the estimates.

As discussed in Section 12.2, we use the forecast error,

to compute MSE, MAD, and MAPE. These
 performance measures are used to find the preferred model
for forecasting. The MSE, MAD, and MAPE page 538
values for the linear and the exponential models
are shown in the last two rows of Table 12.11.
TABLE 12.11 Analysis of Linear and Exponential Trend
Models for Example 12.5

t y

1 166.5028 102.6011 162.0398

2 170.2108 112.2810 166.3719

⋮ ⋮ ⋮ ⋮

58 732.4486 654.3529 729.0290

MSE 1,176.74 12.28

MAD    29.58  2.87

MAPE     9.36  0.82

Consistent with the scatterplot, the exponential trend is clearly

better suited to describe the population in low-income
countries because it has lower MSE, MAD, and MAPE values
than the linear trend model.
c. For 2018 (t = 59), we use unrounded values of the estimates
to forecast the population in low-income countries as
The Polynomial Trend Model
Sometimes a time series reverses direction, due to any number of
circumstances. A common polynomial function that allows for
curvature in the series is the quadratic trend model. The quadratic
regression model was first introduced in Chapter 7. This quadratic
trend model allows one change in the direction of a series and is
estimated as

The coefficient β2 determines whether the trend is U-shaped or

inverted U-shaped.

FIGURE 12.8 Scatterplots with superimposed linear and quadratic trend lines

Figure 12.8 presents a scatterplot of a time series with

superimposed linear (blue) and quadratic (green) trend lines. Here,
the linear trend line would under-forecast future values when β 2 > 0
and over-forecast future values when β 2 < 0.
 In order to estimate the quadratic trend model, we generate t2,
which is simply the square of t. Then we run a multiple regression
model that uses y as the response variable and both t and t2 as the
predictor variables. The estimated model is used to make forecasts
as

Higher-order polynomial functions can be estimated similarly. For

instance, the cubic trend model is specified as

page 539

The cubic trend model allows for two changes in the direction of a
series. In the cubic trend model, we basically generate two additional
variables, t2 and t3, for the regression. A multiple regression model is
run that uses y as the response variable and t, t2, and t3 as the
predictor variables. The estimated model is used to make forecasts
as

We cannot use MSE, MAD, and MAPE to compare polynomial

trend models because these values decrease as the order of the
polynomial increases. The problem is similar to that of the coefficient
of determination R2 discussed in earlier chapters where R2 increases
as the number of predictor variables increases. When comparing
polynomial trend models, we use adjusted R2, which imposes a
penalty for over-parameterization.
 THE POLYNOMIAL TREND MODEL
The polynomial trend model of order q is estimated as

This model specializes to a linear trend model, quadratic trend

model, and cubic trend model for q = 1, 2, and 3, respectively.
The estimated model is used to make forecasts as
where b0, b1, . . . , bq are the

coefficient estimates. We use adjusted R2 to compare

polynomial trend models with different orders.

It is important to be mindful of overfitting when using higher-level

polynomials. Recall that overfitting occurs when an estimated model
begins to describe the quirks of the data rather than the real
relationships between variables. By making the model conform too
closely to the given data, the model’s predictive power is
compromised. Although adjusted R2 imposes a penalty for additional
predictor variables, the measure still suffers from the fact that the
same data are used for both estimation and assessment. In Section
12.5, we discuss cross-validation techniques to assess forecasting
models.

The Nonlinear Trend Models with Seasonality

In Section 12.3, we developed linear forecasting models for trend
and seasonality; here we extend the analysis to include nonlinear
models. With quarterly data, the exponential, and the quadratic,
trend models with seasonal dummy variables are summarized in the
following definition boxes. These models can be easily modified to
forecast with monthly or other forms of seasonal data.

EXPONENTIAL TREND MODEL WITH SEASONAL DUMMY

VARIABLES
With quarterly data, an exponential trend model with seasonal
dummy variables is specified as

Forecasts based on the estimated model are as follows:

Here, b0, b1, . . . b4 are the coefficient estimates and se is the

standard error of the estimate.

page 540

Note that for the exponential model, we compute in regular

units and not in natural logs. The resulting also enables us to

compare the linear and the exponential models in terms of their

MSE, MAD, and MAPE.
 QUADRATIC TREND MODEL WITH SEASONAL DUMMY
VARIABLES
With quarterly data, a quadratic trend model with seasonal
dummy variables is specified as

Forecasts based on the estimated model are as follows:

Here, b0, b1, . . . b4 are the coefficient estimates.

We simply use adjusted R2 to compare the linear and the

quadratic models. Example 12.6 provides an application of the
quadratic trend model with seasonal dummy variables.

EXAMPLE 12.6
FILE
Revenue_Apple

The objective outlined in the introductory case is to forecast

Apple’s quarterly revenue, in $ millions, from 2010 to 2018.
Use the data in Table 12.1 to
a. Estimate the linear and the quadratic trend models with
seasonal dummy variables for Apple’s revenue.
b. Determine the preferred model and use it to forecast Apple’s
revenue for fiscal year 2019.
SOLUTION: In Section 12.1, we used Figure 12.1 to highlight
important characteristics of Apple’s revenue. First, there is a
persistent upward movement with the series plateauing near
the end of the observation period, which is suggestive of a
quadratic trend model. Second, a seasonal pattern repeats
itself year after year. For instance, revenue is consistently
higher in the first quarter (September–December), as
compared to the other quarters.
a. Given quarterly data, we first construct relevant variables for
the linear and the quadratic trend models with seasonal
dummy variables. Table 12.12 presents a portion of the
constructed data representing the revenue variable y; three
seasonal dummy variables d1, d2, and d3 representing the first
three quarters (using the fourth quarter as reference); and the
trend variable t and its square t2.

TABLE 12.12 Constructing Variables for Example 12.6

page 541
The estimated trend models are:
 The coefficients for the seasonal dummy variables indicate
that the revenue is about $19,757 million, or $19.76 billion,
higher in the first quarter as compared to the fourth quarter.
The results also suggest that compared to the fourth quarter,
the revenue is somewhat higher in the second quarter and
lower in the third quarter. The positive coefficient for the trend
variable t in the linear model indicates an upward movement
of the revenue. The positive coefficient for t along with a
negative coefficient for t2 in the quadratic model captures the
inverted U-shape of the series. Given the coefficients, holding
seasonality constant, the revenue reaches its maximum at
which suggests that Apple’s revenues

reached their maximum in the fourth quarter of 2018.

b. The quadratic trend model with seasonal dummy variables is
preferred for making forecasts because of a higher adjusted
R2 value (0.8824 > 0.8324). Therefore, the revenue forecasts
for fiscal year 2019 are:

The quarterly forecasts result in a sum of $246,230 million in

revenue for fiscal year 2019.
SYNOPSIS OF INTRODUCTORY CASE
Apple Inc. is an American multinational technology company headquartered in
Cupertino, California. It designs, manufactures, and markets mobile communication
and media devices, personal computers, and portable digital music players. It also
sells a range of related software, streaming services, accessories, networking
solutions, and third-party digital content and applications.

Pavel L Photo and Video/Shutterstock

For several years, Apple’s smartphone segment has been the company’s core
source of revenue, resulting in record revenue. A scatterplot of Apple’s quarterly
revenue for the fiscal years 2010 through 2018 highlights some important
characteristics. First, there is a persistent upward movement with the revenue
plateauing near the end of the observation period. Second, a seasonal pattern
repeats itself. For each year, the revenue is the highest in the first quarter (October–
December) followed by the second (January–March), fourth (July–September), and
third (April–June) quarters.
The coefficients of the estimated quadratic trend model with seasonal dummy
variables suggest that the revenue is about $20 billion higher in the first quarter as
compared to the fourth quarter. This is not surprising because given Apple’s fiscal
calendar, the first quarter encompasses the holiday period with usual strong sales.
The positive coefficient for the trend variable t along with a negative coefficient for t2
captures the plateauing of the series. In fact, given the coefficients, holding
seasonality constant, the revenue reaches its maximum in the fourth quarter of
2018. This finding is consistent with the concern that while Apple is clearly doing
well for now, its future growth may be murky partly because, in terms of the
smartphone market, it only sells on the somewhat saturated mid-to-high-end range
market. The quarterly revenue forecasts for 2019 are $76.28, $60.38, $53.62, and
$55.95 billion, respectively, resulting in a whopping $246 billion in revenue for fiscal
year 2019.
 page 542

EXERCISES 12.4
Applications
14. FILE Whites. In 2016, demographers reported that deaths outnumbered births
among white Americans in more than half the states in the U.S. (The New York
Times, June 20, 2018). Consider the white American population, in millions, from
2005 through 2017; a portion of the data is shown in accompanying table.

Year Whites

2005 215.33

2006 221.33

⋮ ⋮

2017 235.51

a. Use the scatterplot to explore linear and quadratic trends; the cubic trend is not
considered. Which trend model do you think describes the data better?
b. Validate your intuition by comparing adjusted R2 of the two models. Use the
preferred model to forecast the white population in 2018 and 2019.
15. FILE TrueCar. Investors are always reviewing past pricing history and using it to
influence their future investment decisions. On May 16, 2014, online car buying
system TrueCar launched its initial public offering (IPO), raising $70 million in the
stock offering. An investor, looking for a promising return, analyzes the monthly
stock price data of TrueCar from June 2014 to May 2017. A portion of the data is
shown in the accompanying table.

Date Price

Jun-14 14.78

Jul-14 13.57
 Date Price

⋮ ⋮

May-17 17.51

a. Estimate the linear, the quadratic, and the cubic trend models.
b. Determine the preferred model and use it to make a forecast for June 2017.
16. FILE Miles_Traveled. The number of cars sold in the United States in 2016
reached a record high for the seventh year in a row (CNNMoney, January 4,
2017). Consider monthly total miles traveled (in billions) in the United States from
January 2010 to December 2016. A portion of the data is shown in the
accompanying table.

Date Miles

Jan-10 2953.305

Feb-10 2946.689

⋮ ⋮

Dec-16 3169.501

a. Estimate the quadratic and the cubic trend models.

b. Determine the preferred model and use it to make a forecast for Miles in
January 2017.
17. FILE Café_Sales. With a new chef and a creative menu, Café Venetian has
witnessed a huge surge in sales. The following data show a portion of daily sales
(in $) at Café Venetian in the first 100 days after the changes.

Day Sales

1 263

2 215

⋮ ⋮

100 2020

Estimate the exponential trend model to forecast sales for the 101st day.
18. FILE Population_Japan. For several years, Japan’s declining population has led
experts and lawmakers to consider its economic and social repercussions (NPR,
December 21, 2018). Consider the population data, in millions, for Japan from
1960 to 2017; a portion of the data is shown in the accompanying table.

Year Population

1960  92.50

1961  94.94

⋮ ⋮

2017 126.79

a. Use a scatterplot of the series to suggest the appropriate polynomial trend

model and follow it up with a formal model selection measure.
b. Use the preferred model to forecast the population in Japan for 2018 and 2019.
19. FILE Case_Shiller. The S&P’s Case-Shiller home price index measures repeat-
sales house price indices for the United States. The index, normalized to have a
value of 100 in January 2000, captures price movements of the same homes
relative to January 2000. Consider the seasonally adjusted monthly series from
January 2016 to November 2018; a portion of the data is shown in the
accompanying table.

Date Case Shiller

Jan-16 177.412

Feb-16 177.828

⋮ ⋮

Nov-18 206.263

page 543
a. Estimate the linear and the exponential trend models and
calculate their MSE, MAD, and MAPE.
b. Use the preferred model to forecast the Case-Shiller index for December 2018.
20. FILE Expenses. The controller of a small construction company is attempting to
forecast expenses for the next year. He collects quarterly data on expenses (in
$1,000s) over the past five years, a portion of which is shown in the
accompanying table.
 Year Quarter Expenses

2008 1 96.50

2008 2 54.00

⋮ ⋮ ⋮

2017 4 22335.30

a. Estimate and interpret the exponential trend model with seasonal dummy
variables.
b. Use the estimated model to forecast expenses for the first two quarters of 2018.
21. FILE Treasury_Securities. Treasury securities are bonds issued by the U.S.
government. Consider a portion of quarterly data on treasury securities, measured
in millions of U.S. dollars.

Year Quarter Securities

2010 1  927527

2010 2 1038881

⋮ ⋮ ⋮

2018 3 2284572

Estimate the exponential trend model with seasonal dummy variables to make a
forecast for the fourth quarter of 2018.
22. FILE House_Price. The West Census region for the U.S. includes Montana,
Wyoming, Colorado, New Mexico, Idaho, Utah, Arizona, Nevada, California,
Oregon, and Washington. Consider the median house prices in the West Census
region from 2010:01 through 2018:03, a portion of which is shown in the
accompanying table.

Year Quarter Price

2010 1 263600

2010 2 264100

⋮ ⋮ ⋮
 Year Quarter Price

2018 3 404300

a. Estimate and interpret the quadratic trend model with seasonal dummy
variables.
b. Use the estimated model to forecast the median house price in the West
Census region for the fourth quarter of 2018.
23. FILE Vehicle_Miles. The United States economy picked up speed in 2012 as
businesses substantially built up their inventories and consumers increased their
spending. This also led to an increase in domestic travel. Consider the vehicle
miles traveled in the U.S. (in millions) from January 2012 through September
2018, a portion of which is shown in the accompanying table.

Date Vehicle Miles

Jan-12 227527

Feb-12 218196

⋮ ⋮

Sep-18 260555

a. Estimate the linear and the exponential trend models with seasonal dummy
variables for vehicle miles and calculate their MSE, MAD, and MAPE.
b. Use the preferred model to forecast vehicle miles for the last three months of
2018.
24. FILE Housing_Starts. Housing starts are the number of new residential
construction projects that have begun during any given month. It is considered to
be a leading indicator of economic strength. The following table contains a portion
of monthly data on housing starts (in 1,000s) in the U.S. from Jan-11 to Nov-18.

Date Housing Starts

Jan-11 40.2

Feb-11 35.4

⋮ ⋮

Nov-18 95.9
 a. Estimate and interpret the exponential seasonal trend model.
b. Use the estimated model to forecast housing starts for December 2018.
25. FILE Weekly_Earnings. Data on weekly earnings are collected as part of the
Current Population Survey, a nationwide sample survey of households in which
respondents are asked how much each worker usually earns. The following table
contains a portion of quarterly data on weekly earnings (Earnings, adjusted for
inflation) in the U.S. from 2010–2017.

page 544
Year Quarter Earnings

2010 1 347

2010 2 340

⋮ ⋮ ⋮

2017 4 347

a. Estimate the linear and the quadratic trend models with seasonal dummy
variables.
b. Determine the preferred model and use it to forecast earnings for the first two
quarters of 2018.

12.5 DATA PARTITIONING AND

MODEL SELECTION
LO 12.5
Apply cross-validation techniques for model selection.

So far, we have assessed competing forecasting models on the

basis of in-sample criteria where the predictability of a model is
assessed in the sample period that was also used to build the model.
Unfortunately, these measures do not help us gauge how well an
estimated model will predict in an unseen sample period.
 As in Chapters 7–10, we will apply cross-validation techniques to
assess forecasting models. Using the holdout method, we partition
the series into a training set to build (train) the model, and a
validation set to assess (validate) it. As discussed in Section 12.1,
out-of-sample criteria are based on the forecast error where

yt and denote the value of the series and its forecast, respectively,

at time t in the validation period. We use performance measures

MSE, MAD, and MAPE in the validation period to determine the
preferred model for forecasting.
It is important to note key differences between cross-validation
techniques applied to cross-sectional data and time series data.
First, unlike random partitioning used for cross-sectional data, we
use sequential partitioning for time series data. Here, the series is
split into early and later periods, representing the training and the
validation sets, respectively. Because all forecasting models exploit
intrinsic patterns in time series, random partitioning would create
unnecessary holes in the data. It is common to split the series so
that the validation period consists of about 20% of the total sample
data; however, this value depends on the series length. Sometimes,
the series is not sufficiently long for meaningful partitioning and,
therefore, for accurate estimation and assessment. For example, if
we have only 20 observations, an 80/20% split leaves only 16
observations for training and 4 for validation. Here, it might be better
to assess competing forecasting models in terms of the in-sample
criteria discussed earlier.
We always use the entire data, that combines training and
validation sets, for estimating the preferred model for making
forecasts. There are several reasons for doing so. First, by
combining the training and validation sets, we create a bigger
sample needed for accurate estimation. Second, and more
importantly, the validation set contains the most recent information
on the series, which is useful for projecting into the future; otherwise
you will be introducing unnecessary noise by using the training
period to project past the validation period.

CROSS-VALIDATION WITH TIME SERIES

Cross-validation with time series involves the following steps.
A. Split the series into early and later periods, representing the
training and the validation sets, respectively.
B. Explore suitable forecasting models for the training set and
use the forecast errors in the validation set to compute MSE,
MAD, and MAPE. Choose the model with the lowest MSE,
MAD, or MAPE.
C. Use the entire data set, one that combines the training and the
validation sets, to reestimate the preferred model for making
forecasts.

page 545

In Examples 12.7 and 12.8, we use cross-validation to find the

appropriate models for trend, and trend with seasonality,
respectively.

EXAMPLE 12.7
FILE
Population_LowInc

Revisit the population data (in millions) for low-income

countries from 1960 to 2017. These data were earlier used in
Example 12.5. Let the training and the validation sets comprise
the periods from 1960 to 2000 and 2001 to 2017, respectively.
a. Use the training set to estimate the linear and the exponential
trend models and compute the resulting MSE, MAD, and
MAPE for the validation set.
b. Determine the preferred model and reestimate it with the
entire data set to forecast the population in low-income
countries for 2018.

SOLUTION: As in Example 12.5, the trend variable t is relabeled

from 1 to 58 denoting the period between 1960 and 2017. For
the exponential trend model, we first transform the population
series into natural logs.
a. The estimated trend models with the training set, where y
denotes population, are:

The derivations for the resulting MSE, MAD, and MAPE in the
validation set are:

TABLE 12.13 Cross-Validation for Example 12.7

b. Consistent with the in-sample criteria used in Example 12.5,

the exponential model is preferred because it has lower MSE,
MAD, and MAPE values. We reestimate the exponential
 model with the entire data set from 1960 through 2017, as in
Example 12.5. For 2018 (t = 59), the population forecast is

EXAMPLE 12.8
FILE
Revenue_Apple

Revisit Apple’s quarterly revenue data (in $ millions) from 2010

to 2018. This data set was outlined in the introductory case and
used in Example 12.6. Let the training and the validation sets
comprise the periods from 2010 to 2016 and 2017 to 2018,
respectively.
a. Use the training set to estimate the linear and the quadratic
trend models with seasonal dummy variables and compute the
resulting MSE, MAD, and MAPE for the validation set.
b. Determine the preferred model and reestimate it with the
entire data set to forecast Apple’s revenue for fiscal year
2019.

page 546

SOLUTION: As in Example 12.6, the trend variable t is relabeled

from 1 to 36 denoting the quarterly period between 2000 and
2018. In addition to t and its square t2, we also compute three
seasonal dummy variables d1, d2, and d3 representing the first
three quarters, using the fourth quarter as reference. (Shortly,
we will provide R instructions for replicating the results.)
a. The estimated trend models with the training set, where y
denotes revenue, are:
 The derivations for the resulting MSE, MAD, and MAPE in the
validation set are:

TABLE 12.14 Cross-Validation for Example 12.8

b. The linear model is preferred over the quadratic model

because it has markedly lower MSE and MAD and only a
slightly higher MAPE. We reestimate the linear model with the
entire data set from 2010–2017 as

to make

forecasts for fiscal year 2019.

 The quarterly forecasts result in a sum of $292,999.61 million
in revenue for fiscal year 2019.
Note that the results based on out-of-sample criteria are
not consistent with those based on in-sample criteria used in
Example 12.6, where the quadratic model was preferred. So,
which model should we use for making forecasts?
Unfortunately, there is no definitive answer for this question.
Arguably, out-of-sample criteria are more important because
forecasting is essentially an out-of-sample exercise and it also
avoids overfitting. In this example, however, we have a small
sample of 36 observations that is further split into training and
validation sets. Furthermore, the plateauing of the revenue is
most evident in 2017 and 2018, a period not included in the
training set. Therefore, the estimates in the training set, and
the resulting MSE, MAD, and MAPE in the validation set may
not have been accurate. In such conflicting situations, we
sometimes rely on the experience of the forecaster for making
the right decision.

page 547

Cross-validation of Regression Models with

R
FILE
Revenue_Apple

In order to replicate the results in Example 12.8, we follow these

steps.
. Import the Revenue _Apple data into a data frame (table) and label
it myData.
. Install and load the forecast package. Enter:
 > install.packages(“forecast”)
> library(forecast)

. We use the ts function to create a time series object and call it

newData. Within ts, we specify the start and end periods as well as
frequency, denoting the number of seasons in a year. Enter:

> newData <- ts(myData$Revenue, start = c(2010,1), end =

c(2018,4), frequency=4)
. We use the window function to partition the series into training and
validation sets, labeled TData and VData, respectively. Enter:

> TData <- window(newData, end = c(2016, 4))

> VData <- window(newData, start = c(2017, 1))

. We use the tslm function to estimate seasonal linear and quadratic

models. Enter:

> Reg1 <- tslm(TData ~ trend + season)

> Reg2 <- tslm(TData ~ trend + I(trend^2) + season)

. We use the length function to find the number of observations in

the validation set, the forecast function to make h number of
forecasts for the validation set, and the accuracy function to view
the resulting performance measures. Note that R denotes validation
set as test set and MAD as MAE and MSE is found by squaring the
reported RMSE. Enter:

> nV <- length(VData)

> fReg1 <-forecast(Reg1, h=nV)
> fReg2 <-forecast(Reg2, h=nV)
> accuracy(fReg1,VData)
 > accuracy(fReg2,VData)

. We use the entire data, that combines the training and the
validation sets, to re-estimate the preferred linear model for
forecasting Apple’s revenue for fiscal year 2019. Enter:

> RegFin <- tslm(newData ~ trend + season)

> forecast(RegFin, h=4)

EXERCISES 12.5
Applications
26. FILE Population_Japan. The accompanying data file contains annual
population data (in millions) for Japan from 1960 to 2017. For cross-validation, let
the training and the validation sets comprise the periods from 1960 to 2005 and
2006 to 2017, respectively.
a. Use the training set to estimate the linear, the quadratic, and the cubic trend
models and compute the resulting MSE, MAD, and MAPE for the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast the population in Japan for 2018.
27. FILE Tax_Revenue. The accompanying data file contains 57 months of tax
revenue from medical and retail marijuana tax and fee collections. For cross-
validation, let the training and the validation sets comprise the first 45 months and
the last 12 months, respectively.

page 548
a. Use the training set to estimate the linear, the quadratic, and
the cubic trend models and compute the resulting MSE, MAD, and MAPE for
the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast tax revenue for the 58th month.
28. FILE Cafe_Sales. The accompanying data file contains daily sales (in $) at Café
Venetian for 100 days. For cross-validation, let the training and the validation sets
comprise the first 80 days and the last 20 days, respectively.
a. Use the training set to estimate the linear and the exponential trend models and
compute the resulting MSE, MAD, and MAPE for the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast sale for the 101st day.
29. FILE Apple_Price. The accompanying data file contains 53 weeks of Apple’s
stock price data. For cross-validation, let the training and the validation sets
comprise the first 40 weeks and the last 13 weeks, respectively.
a. Use the training set to estimate the linear and the exponential trend models and
compute the resulting MSE, MAD, and MAPE for the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast Apple’s stock price for the 54th week.
30. FILE Expenses. The accompanying data file contains quarterly data on
expenses (in $1,000s) over five years. For cross-validation, let the training and
the validation sets comprise the periods from 2008:01 to 2015:04 and 2016:01 to
2017:04, respectively.
a. Use the training set to estimate the linear and the exponential trend models with
seasonal dummy variables and compute the resulting MSE, MAD, and MAPE
for the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast expenses for the first quarter of 2018.
31. FILE House_Price. The accompanying data file lists quarterly data on median
house prices in the West Census region. For cross-validation, let the training and
the validation sets comprise the periods from 2010:01 to 2016:04 and 2017:01 to
2018:03, respectively.
a. Use the training set to estimate the linear and the quadratic trend models with
seasonal dummy variables and compute the resulting MSE, MAD, and MAPE
for the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast house price for the fourth quarter of 2018.
32. FILE Vehicle_Miles. The accompanying data file lists monthly data on vehicle
miles traveled in the U.S. (in millions). For cross-validation, let the training and the
validation sets comprise the periods from Jan-12 to Dec-16 and Jan-17 to Sep-18,
respectively.
a. Use the training set to estimate the linear and the exponential trend models with
seasonal dummy variables and compute the resulting MSE, MAD, and MAPE
for the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast vehicle miles for October 2018.
33. FILE Weekly_Earnings. The accompanying data file contains quarterly data on
weekly earnings (Earnings, adjusted for inflation) in the U.S. For cross-validation,
let the training and the validation sets comprise the periods from 2010:01 to
2015:04 and 2016:01 to 2017:04, respectively.
a. Use the training set to estimate the linear and the quadratic trend models with
seasonal dummy variables and compute the resulting MSE, MAD, and MAPE
for the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast earnings for the first quarter of 2018.
 34. FILE Housing_Starts. The accompanying data file contains monthly data on
housing starts (in 1,000s) in the U.S. For cross-validation, let the training and the
validation sets comprise the period from Jan-11 to Dec-16 and Jan-17 to Nov-18,
respectively.
a. Use the training set to estimate the linear and the exponential trend models with
seasonal dummy variables and compute the resulting MSE, MAD, and MAPE
for the validation set.
b. Determine the preferred model and reestimate it with the entire data set to
forecast housing starts for December of 2018.

12.6 ADVANCED EXPONENTIAL

SMOOTHING METHODS
LO 12.6
Use advanced smoothing methods to make forecasts.

In Section 12.2, we used the simple exponential smoothing method

to forecast a series y1, y2, …, yT described by random variations
around an unknown level. Given no trend and/or seasonality, we
employed the recursive equation Lt = αyt + (1 − α)Lt−1 to continually
update the level. We then used the most current level of the series at
time T to forecast for the next period as Because there are no

trend and/or seasonal variations, the level cannot be updated and,

therefore, the s-step ahead forecasts are also made as

As discussed earlier, smoothing is a procedure for continually

revising the forecast in the light of more recent observations. In the
case of simple exponential smoothing, forecasts are weighted
averages of past observations, with the weights decaying
exponentially as the observations get older. Recall that the
smoothing parameter α, representing the speed of decline, is the key
to the analysis. With larger values of α, we pay attention mainly to
the most recent observations, whereas with smaller values, greater
emphasis is placed on past observations.
In this section, we extend the analysis to include variations due to
trend and/or seasonality and implement it with Analytic Solver and R.
These software packages allow implementation with user-supplied
as well as computer-generated smoothing parameters. It is important
to note that the results may differ not only between Analytic Solver
and R but also between the different versions of the software
packages used. This is due to the choice of initial values used in the
recursive equations as well as the performance measures used to
obtain optimal smoothing parameters. The differences are negligible
when the sample size is large.

The Holt Exponential Smoothing Method

The Holt exponential smoothing method, also referred to as the
double exponential smoothing method, incorporates long-term
upward or downward movements of a time series. It is appropriate
when the time series does not exhibit seasonal variations or the
series has been deseasonalized.
The Holt exponential smoothing method uses two recursive
equations to smooth the series for level, L, as well as trend, T. (A
formal discussion of these equations is beyond the scope of this
text.) The recursive equations depend on initial values L1 and T1 as
well as the values of the smoothing parameters α and β, for level
and trend, respectively. Recall that the smoothing parameters
indicate the speed of decline. The parameter values can be user-
supplied or computer-generated by minimizing MSE, or some other
performance measure. Computer-generated values are prone to
over-fitting where the selected model does well in the sample period
but not so well in the future. Intuition and prior experience, as well as
cross-validation, dictate the approach used. Common user-supplied
values are α = 0.20 and β = 0.15.
 THE HOLT EXPONENTIAL SMOOTHING METHOD
The Holt exponential smoothing method is an extension of the
simple exponential smoothing method in that the level Lt as well
as the trend Tt adapt over time. It is appropriate when the time
series exhibits trend but no seasonality (or has been
deseasonalized).

In Example 12.9, we will use Analytic Solver and R to implement

the Holt exponential smoothing method. As noted earlier, results may
differ between the software packages and their different versions.

EXAMPLE 12.9
FILE
Population_LowInc

Revisit the population data (in millions) for low-income

countries from 1960 to 2017. As in Example 12.7, let the
training and the validation sets comprise the periods from 1960
to 2000 and 2001 to 2017, respectively. Use the training set to
implement the Holt exponential smoothing model with user-
supplied parameters (α = 0.20 and β = 0.15), as well as
computer-generated smoothing parameters, and page 549
compute the resulting MSE, MAD, and MAPE for
the validation set. Determine the preferred model and
reimplement it with the entire data set to forecast the
population in low-income countries for 2018.
SOLUTION:
Using Analytic Solver

a. Open the Population_LowInc data file.

b. From the menu, choose Data Mining > Time Series >
Partition. Specify the data range by highlighting cells A1:B59
 and check First Row Contains Headers. Select and move Year
to Time Variable and Population to Variables in the Partition
Data. For both Specify Partitioning Options and Specify #
Records for Partitioning, choose Specify # records and input
41 in Training Set, which automatically creates 17
observations in the validation set. Click OK.
c. From the TSPartition worksheet, choose Time Series >
Smoothing > Double Exponential. Select and move
Population to Selected Variable. Check Optimize for
computer-generated parameters or input user-supplied
parameters, α = 0.20 and β = 0.15. We will do both, one at a
time. Check Produce forecast on validation. Click OK.
As expected, the forecasts are a lot closer to the actual
values for the training set than for the validation set. The
DoubleExpo worksheet contains a lot of other useful
information. In Table 12.15, we present the smoothing
parameters and the MSE, MAD, and MAPE values found
under the heading ‘Error Measures: Validation’.
TABLE 12.15 Example 12.9 Results with Analytic Solver

Values User-supplied Computer-generated

α     0.20   0.9137

β     0.15   0.4158

MSE 2,563.58 1,086.32

MAD    44.07    26.44

MAPE     6.94     4.06

 Here, computer-generated smoothing parameters are
preferred as they result in lower MSE, MAD, and MAPE
values. (We would like to point out that for simple exponential
smoothing, discussed in Section 12.2, we choose Time
Series > Smoothing > Exponential.)
d. We now use the entire data, which combine the training and
the validation sets to reimplement the preferred model for
making forecasts. From the original data worksheet, choose
Time Series > Smoothing > Double Exponential. Specify
the Data range by highlighting cells A1:B59. Select and move
Year to Time variable and Population to Selected Variable.
Check Optimize and Produce forecast and input 1 in the #
Forecasts box. This generates a one-year-ahead forecast of
750.08 million for 2018.
Using R

a. Import the Population_LowInc data into a data frame (table)

and label it myData.
b. Install and load the forecast package. Enter:

> install.packages(“forecast”)
> library(forecast)

page 550
c. We use the ts function to create a time series
object and call it newData. Within ts, we specify the start and
end periods as well as frequency, denoting the number of
seasons in a year. Enter:

> newData <- ts(myData$Population, start = c(1960), end =

c(2017), frequency=1)
d. We use the window function to partition the series into
training and validation sets, labeled TData and VData,
respectively. Enter:
 > TData <- window(newData, end = c(2000))
> VData <- window(newData, start = c(2001))

e. We use the ets function, denoting error, trend, and

seasonality, respectively. Within ets, we use a three-letter
string in model, where the first letter represents the error type
(A, M, or Z), the second letter represents the trend type (N, A,
M, or Z), and the third letter represents the season type (N, A,
M, or Z); N, A, M, and Z stand for none, additive,
multiplicative, and automatically selected, respectively. In this
text, we will always use the additive level for the error type as
well as trend. For example, for the Holt method, we specify A
for error, A for trend, and N for seasonality. The ets function
allows both user-supplied and computer-generated smoothing
parameters. Finally, we use the summary function to view the
results. Enter:

> HUser <- ets(TData, model = “AAN”, alpha=0.2, beta=0.15)

> HCmp <- ets(TData, model = “AAN”)
> summary(HUser)
> summary(HCmp)

Here, HUser and HCmp implement the Holt exponential

smoothing model with user-supplied and computer-generated
smoothing parameters, respectively. (We would like to point
out that for the simple exponential smoothing model,
discussed in Section 12.2, we would set model = “ANN”.)
f. We use the length function to find the number of observations
in the validation set, the forecast function to make h number
of forecasts for the validation set, and the accuracy function
to view the resulting performance measures. Enter:

> nV <- length(VData)

 > fUser <-forecast(HUser, h=nV)
> fCmp <-forecast(HCmp, h=nV)
> accuracy(fUser,VData)
> accuracy(fCmp,VData)

In Table 12.16, we present the smoothing parameters and the

resulting MSE, MAD, and MAPE values found by the
summary and accuracy functions. Recall that R denotes
validation set as test set and MAD as MAE. MSE is found by
squaring the reported RMSE.
TABLE 12.16 Example 12.9 Results with R

Values User-supplied Computer-generated

α   0.20 0.9999

β   0.15 0.9999

MSE 988.34 801.62

MAD   25.12  22.27

MAPE    3.86   3.40

page 551
Again, computer-generated smoothing
parameters are preferred as they result in lower MSE, MAD,
and MAPE values.
g. We use the entire data, which combine the training and the
validation sets, to reimplement the preferred model for making
forecasts. Enter:
 > HFinal <- ets(newData, model = “AAN”)
> forecast(HFinal, h=1)

This generates a one-year-ahead forecast of 750.87 million for

2018.
Summary:
The results reported in Tables 12.15 and 12.16 differ because
of the algorithms used. Here, the algorithm used in R produces
lower MSE, MAD, and MAPE values. Fortunately, the one-
year-ahead forecasts for 2018 are almost the same (750.87
versus 750.08 million), suggesting that the differences between
software packages are negligible when the sample size is
large. Furthermore, with both software packages, the Holt
exponential smoothing method is inferior to the exponential
trend model estimated in Example 12.7 with MSE, MAD, and
MAPE values of 276.15, 16.00, and 2.62, respectively (see
Table 12.13). This result is not surprising as the scatterplot in
Figure 12.6 suggested an exponential trend. The Holt
exponential smoothing method with a multiplicative trend might
have produced better results; however, we do not discuss it in
this text.

The Holt-Winters Exponential Smoothing

Method
We will now extend the Holt exponential smoothing method to
include seasonality. This method is often referred to as the Holt-
Winters exponential smoothing method or the triple exponential
smoothing method. It incorporates the long-term upward or
downward movements of the time series as well as seasonality.
The Holt-Winters exponential smoothing method uses three
recursive equations to smooth the series for level, L; trend, T; and
seasonality, S. The recursive equations depend on initial values L1,
T1, and S1 as well as the values of three smoothing parameters α, β,
and γ. Recall that the values of the smoothing parameters indicate
the speed of decline.
The Holt-Winters exponential smoothing method is further divided
into additive and multiplicative structures depending on the type of
seasonality exhibited by the series. The additive method is preferred
when the seasonal variations are roughly constant through the
series, while the multiplicative method is preferred when the
seasonal variations are proportional to the level of the series.

THE HOLT-WINTERS EXPONENTIAL SMOOTHING METHOD

The Holt-Winters exponential smoothing method is an
extension of the Holt exponential smoothing method in that the
seasonality St along with the level Lt and the trend Tt adapt
over time. It is appropriate when the time series exhibits trend
and seasonality.

In Example 12.10, we will use Analytic Solver and R for the Holt-
Winters exponential smoothing method using computer-generated
smoothing parameters. As noted earlier, results may differ between
the software packages and their different versions.

page 552

EXAMPLE 12.10
FILE
Revenue_Apple

Revisit Apple’s quarterly revenue data (in $ millions) from 2010

to 2018. As in Example 12.8, let the training and the validation
sets comprise the periods from 2010 to 2016 and 2017 to
 2018, respectively. Use the training set to implement the Holt-
Winters smoothing method with additive and multiplicative
seasonality and compute the resulting MSE, MAD, and MAPE
for the validation set. Determine the preferred model and
reimplement it with the entire data set to forecast Apple’s
revenue for fiscal year 2019.

SOLUTION: Here, the instructions for Analytic Solver and R are

brief because the general implementation is similar to the Holt
exponential smoothing method that was outlined in Example
12.9.

Using Analytic Solver

a. Open the Revenue_Apple data file.

b. Prior to using Analytic Solver, enter the column heading Time
in cell D1. Populate cells D2:D37 with values 1, 2, . . . 36 that
correspond to the nine years of quarterly observations. These
non-repetitive values in column D will be used for partitioning
the series.
c. From the menu, choose Data Mining > Time Series >
Partition. Specify the Data range by highlighting cells A1:D37
and check First Row Contains Headers. Select and move
Time to Time Variable and Revenue to Variables in the
Partition Data. For both Specify Partitioning Options and
Specify # Records for Partitioning, choose Specify # records
and input 28 in Training Set, creating 8 observations in the
validation set. Click OK.
d. From the TSPartition worksheet, choose Time Series >
Smoothing > Holt-Winters > Additive. Select and move
Revenue to Selected Variable. Input 4 in the Period box,
denoting the number of seasons. Check Optimize and
Produce forecast on validation. Click OK. Repeat the process,
replacing Additive with Multiplicative. In Table 12.17, we
present the relevant information for additive and multiplicative
seasonality.
 TABLE 12.17 Example 12.10 Results with Analytic Solver

Values Additive Multiplicative

α 0.4202 0.6746

β 0.0323 0.0321

γ 0.9155 0.9156

MSE 10,837,599 15,529,454

MAD 2,583.34 3,017.45

MAPE 3.88 4.36

The additive method with lower MSE, MAD, and MAPE values
is preferred.
e. From the original data worksheet, choose Time Series >
Smoothing > Holt-Winters > Additive. Select and move
Revenue to Selected Variable. Input 4 in the Period box and
check Optimize. Also, check Produce forecast and input 4 in
the # Forecasts box. The revenue forecasts (in $ millions) are
$95,206.29 for quarter 1, $67,774.63 for quarter 2, $60,275.86
for quarter 3, and $69,155,45 for quarter 4. The quarterly
forecasts result in a sum of $292,412.23 million in Apple’s
revenue for fiscal year 2019.
Using R

a. Import the Revenue_Apple data into a data frame (table) and

label it myData.
b. Install and load the forecast package if you have not done so
already.
 page 553
c. We use the ts function to create a time series
object and call it newData. Within ts, we specify the start and
end periods as well as frequency, denoting the number of
seasons in a year. Enter:

> newData <- ts(myData$Revenue, start = c(2010,1), end =

c(2018,4), frequency=4)
d. We use the window function to partition the series into
training and validation sets, labeled TData and VData,
respectively. Enter:

> TData <- window(newData, end = c(2016, 4))

> VData <- window(newData, start = c(2017, 1))

e. We use the ets function, with model inputs “AAA” for additive
seasonality and “AAM” for multiplicative seasonality. Also, we
enter FALSE in restrict; sometimes the default setting of TRUE
results in an error. Finally, we use the summary function to
view the results. Enter:

> WAdd <- ets(TData, model = “AAA”, restrict = FALSE)

> WMlt <- ets(TData, model = “AAM”, restrict = FALSE)
> summary(WAdd)
> summary(WMlt)

f. We use the length function to find the number of observations

in the validation set, the forecast function to make h number
of forecasts for the validation set, and the accuracy function
to view the resulting performance measures. Enter:

> nV <- length(VData)

> fAdd <-forecast(WAdd, h=nV)
 > fMlt <-forecast(WMlt, h=nV)
> accuracy(fAdd,VData)
> accuracy(fMlt,VData)

In Table 12.18, we present the relevant information for additive

and multiplicative seasonality; note that R denotes validation
set as test set and MAD as MAE. MSE is found by squaring
the reported RMSE.
TABLE 12.18 Example 12.10 Results with R

Values Additive Multiplicative

α 0.3967 0.9995

β 0.0001 0.0857

γ 0.6033 0.0005

MSE 60,955,870 32,459,808

MAD 6,560 4,589

MAPE 10.02 6.81

Here, the multiplicative method with lower MSE, MAD, and

MAPE values is preferred.
g. We use the entire data, which combine the training and the
validation sets, to reimplement the preferred model for
forecasting Apple’s revenue for fiscal year 2019. Enter:

> WFinal <- ets(newData, model = “AAM”, restrict = FALSE)

> forecast(WFinal, h=4)
 The revenue forecasts, in millions, are $98,643.87 for quarter
1, $70,144.73 for quarter 2, $61,848.01 for quarter 3, and
$69,723.77 for quarter 4. The quarterly forecasts page 554
result in a sum of about $300,360 million in
Apple’s revenue for fiscal year 2019.
Summary:
Again, the Analytic Solver and R results in Tables 12.17 and
12.18 differ because of the algorithms used. The resulting
MSE, MAD, and MAPE suggest that the Holt-Winters model
outperforms the trend model with seasonal dummy variables
that was estimated in Example 12.8.

EXERCISES 12.6
Applications
Note: These exercises can be solved using Analytic Solver and/or R. The answers,
however, will depend on the software package used. In R, use AAN for the Holt
method, and use AAA and AAM for the Holt-Winters method with additive and
multiplicative seasonality, respectively.
35. FILE Population_Japan. The accompanying data file contains annual
population data (in millions) for Japan from 1960 to 2017. For cross-validation, let
the training and the validation sets comprise the periods from 1960 to 2005 and
2006 to 2017, respectively.
a. Use the training set to implement the Holt exponential smoothing method with
user-supplied (α = 0.20 and β = 0.15), as well as the computer-generated,
smoothing parameters and compute the resulting MSE, MAD, and MAPE for
the validation set.
b. Determine the preferred model and reimplement it with the entire data set to
forecast the population in Japan for 2018.
36. FILE Tax_Revenue. The accompanying data file contains 57 months of tax
revenue from medical and retail marijuana tax and fee collections. For cross-
validation, let the training and the validation sets comprise the first 45 months and
the last 12 months, respectively.
a. Use the training set to implement the Holt exponential smoothing method with
user-supplied (α = 0.30 and β = 0.20), as well as the computer-generated,
 smoothing parameters and compute the resulting MSE, MAD, and MAPE for
the validation set.
b. Determine the preferred model and reimplement it with the entire data set to
forecast tax revenue for the 58th month.
37. FILE Cafe_Sales. The accompanying data file contains daily sales (in $) at Café
Venetian for 100 days. For cross-validation, let the training and the validation sets
comprise the first 80 days and the last 20 days, respectively.
a. Use the training set to implement the Holt exponential smoothing method with
two sets of user-supplied smoothing parameters, α = 0.20 and β = 0.10 and α =
0.30 and β = 0.20. Compute the resulting MSE, MAD, and MAPE for the
validation set.
b. Determine the preferred model and reimplement it with the entire data set to
forecast sales for the 101st day.
38. FILE Apple_Price. The accompanying data file contains 53 weeks of Apple’s
stock price data. For cross-validation, let the training and the validation sets
comprise the first 40 weeks and the last 13 weeks, respectively.
a. Use the training set to implement the Holt exponential smoothing method with
user-supplied (α = 0.20 and β = 0.10), as well as the computer-generated,
smoothing parameters and compute the resulting MSE, MAD, and MAPE for
the validation set.
b. Determine the preferred model and reimplement it with the entire data set to
forecast Apple’s stock price for the 54th week.
39. FILE Expenses. The accompanying data file contains quarterly data on
expenses (in $1,000s) over five years. For cross-validation, let the training and
the validation sets comprise the periods from 2008:01 to 2015:04 and 2016:01 to
2017:04, respectively.
a. Use the training set to implement the Holt-Winters exponential smoothing
method with additive and multiplicative seasonality and compute the resulting
MSE, MAD, and MAPE for the validation set.
b. Determine the preferred model and reimplement it with the entire data set to
forecast expenses for the first quarter of 2018.
40. FILE House_Price. The accompanying data file lists quarterly data on median
house prices in the West Census region. For cross-validation, let the training and
the validation sets comprise the periods from 2010:01 to 2016:04 and 2017:01 to
2018:03, respectively.
a. Use the training set to implement the Holt-Winters exponential smoothing
method with additive and multiplicative seasonality and compute the resulting
MSE, MAD, and MAPE for the validation set.
b. Determine the preferred model and reimplement it with the entire data set to
forecast house price for the fourth quarter of 2018.

page 555
41. FILE Vehicle_Miles. The accompanying data file lists monthly
data on vehicle miles traveled in the U.S. (in millions). For cross-validation, let the
 training and the validation sets comprise the periods from Jan-12 to Dec-16 and
Jan-17 to Sep-18, respectively.
a. Use the training set to implement the Holt-Winters exponential smoothing
method with additive and multiplicative seasonality and compute the resulting
MSE, MAD, and MAPE for the validation set.
b. Determine the preferred model and reimplement it with the entire data set to
forecast vehicle miles for the last three months of 2018.
42. FILE Housing_Starts. The accompanying data file contains monthly data on
housing starts (in 1,000s) in the U.S. For cross-validation, let the training and the
validation sets comprise the period from Jan-11 to Dec-16 and Jan-17 to Nov-18,
respectively.
a. Use the training set to implement the Holt-Winters exponential smoothing
method with additive and multiplicative seasonality and compute the resulting
MSE, MAD, and MAPE for the validation set.
b. Determine the preferred model and reimplement it with the entire data set to
forecast housing starts for December of 2018.

page 556

12.7 WRITING WITH DATA

Case Study
Leading economic indicators, such as the stock market or the housing market, often
change prior to large economic adjustments. For example, a rise in stock prices often
means that investors are more confident of future growth in the economy. Or, a fall in
building permits is likely a signal that the housing market is weakening – which is
often a sign that other sectors of the economy are on the downturn.
feverpitched/123RF

Consider what happened prior to the 2008 recession. As early as October 2006,
building permits for new homes were down 28% from October 2005. Analysts use
economic indicators to predict future trends and gauge where the economy is
heading. The information provided by economic indicators helps firms implement or
alter business strategies.
Pooja Nanda is an analyst for a large investment firm in Chicago. She covers the
construction industry and has been given the challenging task of forecasting housing
starts for June 2019. She has access to seasonally adjusted monthly housing starts
in the United States from January 2016 to May 2019. A portion of the data is shown
in Table 12.19.

FILE
Starts

TABLE 12.19 Monthly Housing Starts (in 1,000s)

Date Housing Starts

Jan-16 1114

Feb-16 1208

⋮ ⋮

May-19 1269
Pooja would like to use the sample information to identify the best-fitting model to
forecast housing starts for June 2019.

page 557

Sample Report— Forecasting Monthly

Housing Starts
Leading economic indicators are often used to gauge where the economy is
heading. The housing market is one of the most important indicators because
it is a significant component of the economy. When this sector weakens, just
about everyone and everything feels it – from homeowners and construction
workers to government municipalities that rely on property taxes to operate.
Given the importance of the housing market, this report will forecast next
month’s housing starts using historical data.
A scatterplot of housing starts from January 2016 to May 2019 is shown in
Figure 12.9. A casual observation of the scatterplot suggests quite a bit of
random variation and possibly a slight upward trend. There is no concern for
seasonality as the housing starts data represent seasonally adjusted annual
rates.

FIGURE 12.9 Scatterplot of housing starts (in 1,000s)

Given the findings from Figure 12.9, three trend models are estimated.
1. The three-period moving average model.
2. The simple exponential smoothed model with various values for the speed
of decline α.
3. The simple linear trend model, y t = β 0 + β 1Time t + ε t, where yt
represents housing starts.
Three performance measures are used for model selection: mean square
error (MSE), mean absolute deviation (MAD), and the mean absolute
percentage error (MAPE). The preferred model will have the lowest MSE,
MAD, or MAPE. Table 12.20 shows the values of these three performance
measures for the models.

Table 12.20 Performance Measures of Competing Models

3-period Exponential Linear

Moving Smoothed (α = Regression
Average 0.20)* Model

MSE 5,069.50 4,798.44 4,091.72

MAD 60.55    57.13    54.24   

MAPE 4.96   4.65   4.46  

*For the exponential smoothing model, α = 0.20 provides the lowest values
for MSE, MAD, and MAPE.

The linear trend model provides the best sample fit, as it has the lowest
values for MSE, MAD, and MAPE. Therefore, the estimated linear trend
model is used to derive the forecast for June 2019 as

Housing starts plays a key role in determining the health of the economy and
is, therefore, always under scrutiny. The U.S. housing market seems to be on
solid ground even though there has been slowdown from its peak in early
2018. In this report, we employ simple time series models to project historical
data on housing starts.

page 558
Suggested Case Studies
Report 12.1 FILE Fried_Dough. Fried dough is a popular North American food
associated with outdoor food stands at carnivals, amusement parks, fairs, festivals,
and so on. Usually dusted with powdered sugar and drenched in oil, it is not
particularly healthy, but it sure is tasty! Jose Sanchez owns a small stall at Boston
Commons in Boston, Massachusetts, where he sells fried dough and soft drinks.
Although business is good, he is apprehensive about the variation in sales for no
apparent reason. The accompanying data file contains information on the number of
plates of fried dough and soft drinks that he sold over the last 30 days. In a report,
use the sample information to
Explore forecasting models, including moving averages and the simple exponential
method, to smooth the time series for fried dough and soft drinks.
Use the preferred method to forecast sales of fried dough and soft drinks for the
31st day.

Report 12.2 FILE India_China. According to United Nations estimates, more than
half of the world population live in just seven countries, with China, closely followed
by India, leading the pack. The other five countries on the list include the United
States, Indonesia, Brazil, Pakistan, and Nigeria. It is believed that India will overtake
China to become the world’s most populous nation much sooner than previously
thought (CNN, June 2019). The accompanying data file, compiled by the World
Bank, contains the population data, in millions, for India and China from 1960 to
2017. In a report, use the sample information to
Explore forecasting models, including trend regression models and the Holt
exponential smoothing model, to capture the population trend for both China and
India.
Use the preferred model to forecast the population of China and India from 2018–
2020.

Report 12.3 FILE Revenue_Amazon. Amazon.com, Inc., was a money-losing

company when it went public on May 15, 1997, with an IPO valued at a modest $438
million. Amazon has since had an epic 20-year run as a public company, worth about
$460 billion in 2017. Many analysts attribute the success of the company to its
dynamic leader, Jeff Bezos, who, according to the Bloomberg Billionaires Index, is
the richest man in modern history (CNBC, July 2018). An important question for
investors and other stakeholders is if Amazon’s growth is sustainable. The
accompanying data file contains quarterly data on Amazon’s revenue for the fiscal
years 2010 through 2018, with the fiscal year concluding at the end of December. In
a report, use the sample information to
Explore forecasting models, including trend regression models with seasonal
dummy variables and the Holt-Winters exponential smoothing model, to capture
the trend and seasonality of Amazon’s revenue.
Use the preferred model to forecast Amazon’s revenue for fiscal year 2019.
page 559
 page 560

13 Introduction to
Prescriptive Analytics

page 561

LEARNING OBJECTIVES
After reading this chapter, you should be able to:
LO 13.1  Generate values for random variables.
LO 13.2  Develop and apply Monte Carlo
simulation models.
LO 13.3  Formulate a linear programming model.
LO 13.4  Solve and interpret a linear programming
model.
LO 13.5  Formulate and solve a linear integer
programming model.

n the previous chapters, we focused on descriptive and predictive

I analytics to extract value from data. Descriptive techniques were
used to explore the data, and predictive methods were
implemented to classify or predict new records. In this chapter, we
introduce prescriptive analytics, which is the third and final stage of
business analytics where analytical models are developed, and the
results are analyzed to improve decision making. In particular,
prescriptive analytics uses simulation and optimization algorithms to
quantify the effect of different possible actions of a decision maker to
make a more informed decision.
For simulation, we use random variables to represent the risks
and uncertainties that are encountered in many day-to-day, real-life
events. We learn how to generate values for various random
variables. We then focus on a technique called the Monte Carlo
simulation method, which is widely used by many practitioners and
researchers. Finally, we demonstrate how to use the linear and
integer programming techniques to develop optimization models
under constraints.
INTRODUCTORY CASE
FashionTech: Operation and Staffing
Strategy
FashionTech is an online apparel retailer that specializes in
activewear for both men and women with an active and
outdoors lifestyle. FashionTech’s best-selling item is a red
pullover jacket for women. Abigail Kwan, the company’s
operations manager, is looking for a better, more data-driven
strategy for production and staffing for this product.
 The selling price of each pullover jacket is $50. Abigail also
obtains the weekly demand and production data from the
corporate data warehouse. She believes that the weekly
demand for the pullover follows a normal distribution with an
average of 139 units and a standard deviation of 14.61 units.
The material-per-unit cost of a pullover is also normally
distributed with a mean of $12 and a standard deviation of
$1.37. The weekly production rate for each tailor varies
between 31 and 41 units and follows a discrete uniform
distribution. FashionTech usually pays each tailor $700 per
week. The overhead cost allocated to the pullover jacket
production is approximately $800 per week.
Although the production process for the pullover is partly
automated, many important steps are done manually by skilled
tailors. Abigail wants to get a better estimate of the demand of
the pullover and determine how many tailors she needs to hire.
In addition, she is considering increasing the level of
automation in the manufacturing process. While the added
automation will incur a one-time installation and training cost as
well as raise the overhead cost to $900 per week, it will allow
FashionTech to use less-skilled tailors at a lower weekly wage
of $550, and the weekly production rate will likely have less
variability. Based on the data from other manufacturers, Abigail
believes that with increased automation, the weekly production
rate will be uniformly distributed with lower and upper limits of
36 and 41 units, respectively. Abigail wants to analyze her
data, and based on the results, confirm her plan to send out a
request for proposals from automation vendors for the possible
installation and training cost.
Abigail would like to use the information described above to
accomplish the following tasks.
1. Develop a simulation to examine the profit at each staffing
level as well as evaluate the plan to increase the automation
level in the production process.
2. Examine the profit variability given the uncertainty of the
demand and the production rate.
A synopsis of this case is provided in Section 13.2.

page 562

13.1 OVERVIEW OF PRESCRIPTIVE

ANALYTICS
Prescriptive analytics is a process of using analysis tools to improve
decision making. In many cases, prescriptive analytics can help a
decision maker determine the best course of action among different
alternatives. Some people regard prescriptive analytics as a very
broad term that encompasses all analytics techniques with an overall
goal of improving business decision making. However, most people
prefer making a distinction between descriptive, predictive, and
prescriptive analytics, where prescriptive analytics uses simulation
and optimization algorithms to quantify the effect of different possible
actions by a decision maker to help make a more informed decision.
Recall from Chapter 1 that these three types of analytics
techniques—descriptive, predictive, and prescriptive—are used to
extract value from data and make better business decisions.
Descriptive analytics refers to gathering, organizing, tabulating, and
visualizing data to summarize “what has happened.” Predictive
analytics refers to using historical data to determine “what could
happen in the future.” Prescriptive analytics, our focus here, refers to
using simulation and optimization algorithms to advise on “what
businesses should do.” Simulation is an attempt to imitate a real-
world process that produces several business scenarios, whereas
optimization is an attempt to find an optimal way to achieve a
business objective under constraints, such as limited capacities,
financial resources, and competing priorities.
 PRESCRIPTIVE ANALYTICS
An analytics process of using decision analysis tools such as
simulation and optimization to improve decision making.
Simulation is an attempt to imitate a real-world process to
demonstrate the outcome from different scenarios. identified by
decision makers, whereas optimization is an attempt to find an
optimal way to achieve a business objective under given
constraints.

Section 13.2 discusses a specific type of simulation—Monte

Carlo simulation. Sections 13.3 and 13.4 explore two optimization
techniques: linear and integer programming.

13.2 MONTE CARLO SIMULATION

To imitate real-world business scenarios, simulations are often
performed using computer software with relevant variables and
relationships among those variables. The usefulness of a computer
simulation largely depends on how we select variables and construct
the simulation to adequately and accurately reflect the real world.
Once a computer simulation is developed, we can study how
changes in certain variables might impact the overall result.
Simulation has been extensively used in a wide variety of
business settings to explore alternative solutions or improve
business processes. Companies may use it in engineering and
manufacturing to experiment with different product designs, changes
in production lines, and streamlining a supply chain. Financial
institutions often use simulation to explore different investment
decisions in various market scenarios. Retail stores, restaurants,
commercial airlines, hospitals, and many other service providers use
simulation to study the fluctuation of consumer demand to design a
better work schedule of their employees as well as manage their
inventory in order to meet the demand. A computer simulation also
allows us to quickly experiment with different values and parameters
of the variables without disrupting the actual business process and
operation.

page 563

In business analytics, computer simulation is often synonymous

with the Monte Carlo method, which is a family of computer
algorithms used to model the risk or uncertainty of a real-world
process or system. Stanislaw Ulam and John von Neumann are
often credited for inventing the Monte Carlo simulation, named after
the Monte Carlo Casino in Monaco. The technique was used to
study nuclear fission by Ulam and von Neumann, who worked on the
Manhattan Project at Los Alamos National Laboratory during World
War II. Because the technique relies on random sampling to mimic
the odds of all possible outcomes, Monte Carlo simulation is also
called probabilistic or stochastic simulation, as opposed to a
deterministic process that can be accurately described by a formula
and whose outcome can be precisely estimated.

MONTE CARLO SIMULATION

A statistical technique used to model probabilistic (stochastic)
processes. By re-creating a real-world process, Monte Carlo
simulation is used to understand the impact of the risk and
uncertainty in a wide variety of business settings.

To understand the probabilistic nature of Monte Carlo simulation,

let’s first consider a deterministic scenario where a simulation may
not be needed. For example, if we deposit $1,000 in a savings
account that yields a 2% annual return, we would certainly end up
with $1,020 in our bank account at the end of the year. In this
deterministic case, repeating the same scenario (i.e., investing
$1,000 in the same savings account) would yield the exact same
outcome (i.e., $1,020 in our bank account) every time, and no
simulation is needed to predict the outcome. However, most real-life
situations are much more complicated. An investment of the same
$1,000 in the stock market would involve a much higher risk and
uncertainty. In a good year, the return may be in double digits,
whereas in a bad year the return may be negative. Here, a stock
investor needs to make an investment decision when the actual
return on the investment is not known with certainty. Similarly, a
retailer has to determine the inventory level in the store when the
precise demand for each product is unknown. In all these cases, we
can use a Monte Carlo simulation with relevant variables to capture
the inherently probabilistic nature of the real-world financial and
economic markets to make better-informed decisions.
Here, one might ask why we cannot simply consider the average
values such as the average return on the stock investment or the
average customer demand for a product to help make a decision for
the stock investor or the retailer. This is because the actual outcome
(i.e., actual return on the investment or the actual customer demand)
can wildly fluctuate and deviate significantly from the averages.
Many real-life events serve as a cautionary tale that relying solely on
average values could lead to a negative, and sometimes disastrous,
consequence. In 1994, the government of Orange County in
California made an investment decision based on an average value
of the then-very-low interest rate, instead of considering a range of
possible outcomes and the uncertainty of the future interest rate. In
the end, the investment decision resulted in a loss of almost $2
billion, and Orange County became the largest municipality in the
U.S. forced to declare bankruptcy (Source: The Flaw of Averages by
Sam. L. Savage).
Even outside of the financial market, the pitfalls of relying on
averages have long been well known and documented. Back in the
1950s, the U.S. Air Force set out to design a better cockpit in a
fighter jet by measuring more than 100 body parts of over 4,000
pilots, including the size of their chest, neck, torso, hip, and thighs as
well as the distance between their eyes and ears. The main objective
of the project was to build a better-fitting cockpit to reduce the
number of accidents and plane crashes. However, an early attempt
to build a cockpit to fit an “average” pilot turned out to be a mistake,
as one researcher on the project quickly realized that none of the
thousands of pilots would comfortably fit within the overall average
specifications. A pilot with an average height might have a longer-
than-average arm length or a shorter-than-average torso. A pilot with
some average measurements often turned out to be page 564
an outlier on many other body dimensions. Had the
Air Force built a cockpit for an “average” pilot, the cockpit would
have fit no one! (Source: The “Average Man” by Gilbert S. Daniels.)
Instead of relying solely on the average value, Monte Carlo
simulation is useful because it considers all possible values using a
random variable to capture the risk and uncertainty. From Chapter 4,
recall that a random variable is a function that assigns numerical
values to the outcomes of an experiment. A discrete random variable
assumes a countable number of distinct values. A continuous
random variable, on the other hand, is characterized by uncountable
values in an interval. Examples of discrete random variables include
the number of students receiving the grade A in a class or the
number of firms filing for bankruptcy in a given month. Similarly, the
return on an investment or the height of an individual are examples
of continuous random variables because they are both associated
with uncountable outcomes. The following subsections use Monte
Carlo methods to model the risk and uncertainty using random
variables based on discrete and continuous probability distributions.

LO 13.1
Generate values for random variables.

Modeling Risk and Uncertainty

We generally rely on random variables to capture the risk and
uncertainty in the real world. Monte Carlo simulations are based on
both discrete and continuous probability distributions of random
variables. Students are advised to review the discussion of these
probability distributions in Chapter 4.
Two of the most relevant discrete probability distributions for
Monte Carlo simulation are binomial and Poisson distributions.
Recall from Chapter 4 that a binomial random variable must first
satisfy the conditions of a Bernoulli process. A binomial random
variable is defined as the number of successes achieved in the n
trials of a Bernoulli process. Examples include the number of
customers, in a sample of 100, whose loan application will be
approved by a bank and the number of customers, in a sample of 20,
who will return a purchased product.
The Poisson distribution is particularly useful for modeling when
we are interested in finding the number of occurrences (successes)
of a certain event over a given interval of time or space. Examples
include the number of customers arriving at a retail store between 10
am and noon on a weekday, the number of bankruptcies that are
filed in a month, and the number of defects in a 50-yard roll of fabric.
The discrete uniform distribution, which was not discussed in
Chapter 4, is another important probability distribution for conducting
Monte Carlo simulation. A discrete uniform random variable has a
finite number of specified values, with each value being equally
likely. When rolling a fair six-sided die, the number rolled is an
example of a discrete uniform random variable. The distribution is
symmetric. To generate random observations from these probability
distributions, we can use computer software such as Excel and R.

Using Excel and R to Generate Random

Observations from a Discrete Probability
Distribution
In Example 13.1, we show how to use Excel and R to generate
random observations based on the binomial probability distribution.
We also modify the instructions to accommodate Poisson and
discrete uniform distributions.
EXAMPLE 13.1
In the United States, about 30% of adults have four-year
college degrees (U.S. Census, July 31, 2018). In Example
4.10, we demonstrated how to calculate probabilities based on
this application of the binomial distribution. Here, we use a
similar scenario to demonstrate how to generate random
observations that represent the number of four-year college
degree holders in a randomly selected group of 10 individuals
in the U.S. Use Excel and R to simulate random observations
based on 100 selections of 10 individuals.

page 565

SOLUTION: Recall from Chapter 4 that this example satisfies the

conditions for a Bernoulli process and the binomial probability
distribution. Moreover, the mean and the standard deviation of
a binomial probability distribution are μ = np and ,

respectively, where n represents the number of trials and p

represents the probability of success on each trial. In this
example, we randomly select 10 individuals where each
individual has a 30% chance of holding a college degree, so n
= 10 and p = 0.30. Thus, the mean and the standard deviation
of this binomial probability distribution are 3 (= 10 × 0.30) and
1.4491 , respectively. The process of

selecting 10 individuals will be repeated 100 times where, for

each random selection, the number of college-degree holders
will vary between 0 to 10. This variation represents the
uncertainty in the process.
Using Excel
a. To model the uncertainty, we use the RAND and the
BINOM.INV functions. The RAND function generates a
random value between 0 and 1. The BINOM.INV(n, p, α)
function returns the smallest value for which the cumulative
binomial distribution is greater than or equal to a criterion
value denoted as α. Here, we open a new Excel workbook
and enter =BINOM.INV(10, 0.30, RAND()) in cell A1.
Whenever the F9 key is pressed, Excel redraws a new
observation, and the output changes. The output can
potentially vary from 0 to 10 but most values will be clustered
around the mean of 3.
b. To repeat the random selection 100 times, we copy and paste
the formula in step a to cells A2 through A100. We can
calculate the average number of college degree holders from
the 100 random observations using the =AVERAGE(A1:A100)
function in cell B1. As expected, the average value is
approximately 3. We calculate the sample standard deviation
from the 100 random observations using the
=STDEV.S(A1:A100) function in cell B2. The answer should
be close to 1.4491.

Excel can also generate random observations based on

other discrete probability distributions. For the discrete uniform
distribution, we use =RANDBETWEEN(a, b) to generate
random values between a and b, inclusive. For example, if we
enter =RANDBETWEEN(2, 5) in a cell, then Excel will generate
2, 3, 4, or 5. There is no built-in Excel function to generate
random observations for the Poisson probability distribution.
When using built-in functions in Excel to generate random
observations, we cannot set a random seed, and, therefore, the
random observations we obtain will vary each time. To store
random numbers obtained from the built-in function, we can
use the Paste Values feature in Excel to copy and paste only
the values into other blank cells (e.g., C1 through C100).
Excel’s Analysis ToolPak, discussed in earlier chapters, allows
 us to set a random seed so we can obtain consistent outputs
each time. Excel’s Analysis ToolPak also allows us to generate
random observations for a Poisson random variable. We use
Excel for Microsoft Windows for examples in this Section. Excel
for macOS genrates different random numbers and produces
slightly different results, as indicated in each example in this
Section. The instructions for these examples are applicable for
both Windows and macOS.
Using Excel’s Analysis ToolPak

a. Choose Data > Data Analysis > Random Number

Generation from the menu. (Note: If you do not see the Data
Analysis option under Data, you must add in the Analysis
Toolpak option. From the menu, choose File > Options >
Add-Ins and choose Go at the bottom of the dialog box.
Select Analysis Toolpak and then click OK. If page 566
you have installed this option properly, you
should now see Data Analysis under Data.)
b. See Figure 13.1. In this example, we set Number of Variables
to 1 and Number of Random Numbers to 100. For Distribution
we select Binomial, and for Parameters we set p Value
(probability) equal to 0.3 and the Number of trials equal to 10.
For consistency of results, we set Random Seed equal to 1.
We choose to place the output starting in cell A1. Click OK.

FIGURE 13.1 Excel’s Random Number Generation dialog

box
 Source: Microsoft Excel

c. Excel will generate 100 random observations in cells A1

through A100. For the 100 random observations that we
generated, we use the AVERAGE and STDEV.S functions and
find that the mean and the standard deviation are 2.95 and
1.3734, respectively. (For Excel on macOs, the mean and
standard deviation are 2.93 and 1.3652, respectively.) As
expected, the values are close to the population mean of 3
and the population standard deviation of 1.4491 of the
binomial probability distribution.
To generate random observations from other probability
distributions, we can choose other options in the Distribution
box in Figure 13.1 and specify appropriate parameter values.
For example, in order to generate random observations from
the Poisson distribution, we select Poisson in the Distribution
box and specify the value of lambda (λ), where λ is the mean
and variance for the Poisson distribution. In Chapter 4, we
used the Greek letter μ in place of λ.
 Using R

a. In order to ensure consistency of the output, we use the

set.seed function. Enter:
> set.seed(1)
b. We use the rbinom function to generate random observations
for a binomial random variable. We enter rbinom(numobs, n,
p) where numobs denotes the number of observations that R
will generate for a binomial process with n trials and a
probability p of success on each trial. We store the random
observations in an object called output and retype output to
display the random observations. Enter:
> output <- rbinom(100,10,0.3)
> output
page 567
c. In order to calculate the mean and standard
deviation of the 100 random observations, we use the mean
and the sd functions. Enter:
> mean(output)
> sd(output)
The mean and the standard deviation for the 100 random
observations are 3.08 and 1.3156, respectively. As expected,
the values are close to the population mean of 3 and the
population standard deviation of 1.4491 of the binomial
probability distribution.
R can also generate random observations based on other
probability distributions. For the discrete uniform and the
Poisson distributions, instead of the rbinom function, we use
the sample and rpois functions, respectively. For example, if
we enter sample(2:5, 100, replace = TRUE), then R generates
100 random observations between the values of 2 and 5
(inclusive) using a sampling method with replacement. If we
enter rpois(100, 3), then R generates 100 random observations
from a Poisson distribution with a mean of 3. Make sure to use
 the set.seed function before generating each set of random
observations to ensure consistency of the output.

So far, we have generated random observations from discrete

probability distributions. We will now turn our attention to simulating
random observations for continuous random variables. The normal,
the continuous uniform, and the exponential distributions are among
the most widely used probability distributions for the Monte Carlo
method.
As discussed extensively in Chapters 3 and 4, the normal
distribution is bell-shaped and symmetric around its mean. It is the
most extensively used probability distribution for statistical work and
analytical modeling and accurately describes numerous random
variables of interest. Examples include the amount of rainfall during
a planting season, students’ exam scores, or monthly usage of cell
phones among consumers.
Other important distributions, not discussed in Chapter 4, include
the continuous uniform and the exponential distributions. The
continuous uniform distribution is also referred to as the rectangular
distribution as it represents a constant probability within a specified
range. Examples include the delivery time of an appliance, the arrival
time of an elevator in a building, or the scheduled flight time between
destinations. Unlike the normal and the continuous uniform
distributions, the exponential distribution is appropriate when the
data are believed to be generated from a nonsymmetric distribution.
The exponential distribution is also related to the Poisson
distribution. While a Poisson random variable counts the number of
occurrences of an event over a given interval of time or space, an
exponential random variable captures the time that elapses between
such occurrences. Examples include the time between text
messages or between customer purchases. The exponential
distribution is also used in modeling failure times such as the life of
an appliance or a person. It has a noted “memoryless” feature that
implies a constant failure rate. It assumes, for example, that the
probability of an electric bulb burning out on a given day is
independent of whether the bulb has lasted for 10, 100, or 1,000
hours. It is common to define the exponential probability distribution
in terms of its rate parameter λ, which is the inverse of its mean.

Using Excel and R to Generate Random

Observations from a Continuous Probability
Distribution
In Example 13.2, we show how to use Excel and R to generate
random observations based on the normal distribution. We also
modify the instructions to accommodate the continuous uniform and
the exponential distributions.

page 568

EXAMPLE 13.2
From the introductory case, the material cost of a red pullover
jacket sold by FashionTech is normally distributed with a mean
of $12 per unit and a standard deviation of $1.37. Generate the
material cost for 100 randomly selected jackets.

SOLUTION: Similar to Example 13.1, we can use Excel or R to

generate the material cost of 100 randomly selected jackets.
Based on the 100 observations, we should expect the average
value to be approximately $12 with a standard deviation of
approximately $1.37; however, the exact cost of individual
jackets may vary.
Using Excel

a. We use the RAND and the NORM.INV functions. As in

Example 13.1, the RAND function generates random
 observations between 0 and 1. The NORM.INV(α, μ, σ)
function returns the smallest value for which the cumulative
normal distribution is greater than or equal to a criterion value
denoted as α. Here, we open a new Excel workbook and enter
=NORM.INV(RAND(), 12, 1.37) in cell A1. Whenever the F9
key is pressed, Excel redraws a new observation, and the
output changes.
b. To repeat the random selection, we copy and paste the
formula in cell A1 to cells A2 through A100. We can use the
AVERAGE and STDEV.S functions to find the mean and the
standard deviation of the 100 random observations.
We can use Excel to generate random observations from
other continuous probability distributions. For the continuous
uniform distribution with a lower limit of a and an upper limit of
b, we enter =a + (b − a) * RAND () to generate random
observations between a and b, inclusive. In order to generate
random observations from the exponential distribution with a
rate parameter of λ, we use (−1/λ) * ln(1 − RAND()). Recall that
1/λ is the mean of the exponential distribution.
Using Excel’s Analysis ToolPak

a. Similar to Example 13.1, from the menu we choose Data >

Data Analysis > Random Number Generation.
b. In this example, we set Number of Variables to 1 and Number
of Random Numbers to 100. For Distribution we select
Normal, and for Parameters, we set Mean equal to 12 and
Standard Deviation equal to 1.37. For consistency of the
output, we set Random Seed equal to 1. We choose to place
the output starting in cell A1. Click OK.
c. We can use the AVERAGE and STDEV.S functions to find that
the mean and the standard deviation of the 100 observations
as 11.9628 and 1.5055, respectively. (For Excel on macOs,
the mean and standard deviation are 12.2010 and 1.5423,
respectively.) To generate random observations from the
 continuous uniform distribution, for Distribution we select
Uniform and specify the values for the lower and upper limits
of the distribution. The exponential distribution is not one of
the options in Excel’s Analysis ToolPak, but we can simulate
an exponential random variable, y, from a continuous uniform
variable, x, using the following transformation, y = (−1/λ) * ln(1
− x), where 1/λ is the mean of the exponential distribution. For
example, to generate an exponential random variable whose λ
is 3, start by selecting the uniform distribution, specify [0, 1] for
the lower and upper limits, and place the random numbers in
cells A1:A100. In cell B1, enter =(−1/3) * ln(1 − A1) and copy
and paste the formula to cells B2:B100.
Using R

a. Again, we use the set.seed(1) statement to ensure consistent

results. We use the rnorm function to generate random
observations for a normal random variable. We enter
rnorm(numobs, μ, σ), where numobs denotes the number of
observations that R will generate for a normal page 569
process with mean μ and standard deviation σ.
We store the random observations in an object called output
and retype output to display the random observations. Enter:
> set.seed(1)
> output <- rnorm(100, 12, 1.37)
> output
b. We use the mean and sd functions to find the mean and the
standard deviation of the 100 random observations. Enter:
> mean(output)
> sd(output)
For the 100 observations generated in step b, the mean and
standard deviation are 12.1492 and 1.2305, respectively.

For the continuous uniform distribution, we use runif(100,

20, 21) to generate 100 random observations with values
 between 20 and 21, inclusive. To generate 100 observations
from the exponential distribution with λ = 3, we use rexp(100,
3).

So far, we have demonstrated how to model uncertainty based

on theoretical probability distributions. In many real-world settings,
an appropriate distribution is not known in advance. Sometimes, we
may be able to plot the data and inspect the histogram for distinctive
shapes corresponding to these theoretical distributions. For instance,
if the histogram displays data that are evenly distributed, we might
assume a continuous uniform distribution. If the histogram displays
data that are symmetrically clustered around the middle with
decreasing probabilities in the tails, we might assume a normal
distribution. Formal methods, including the chi-square goodness of fit
test, are used to determine which theoretical probability distribution
is appropriate. Distribution fitting functions can also be found in many
software applications. In this chapter, we focus on developing
simulation models based on known theoretical distributions.

LO 13.2
Develop and apply Monte Carlo simulation models.

Formulating and Developing a Monte Carlo

Simulation
In any business setting, when the probability distributions are known
for all relevant random variables, we can easily formulate and
develop a Monte Carlo simulation. The next crucial step is to
construct a quantitative model that represents the relationship
among relevant variables. In the introductory case, Abigail wants to
examine the profit at each staffing level for the production of a red
pullover jacket for women. A quantitative formulation that relates the
random variables, constants, and the profit can be constructed
based on basic accounting principles, as shown in the following
equation:

where R = the selling price per jacket; Qs = the number of jackets

sold; F = the fixed overhead cost; V = the variable cost per jacket
consisting of the material cost and weekly wages of the tailors; and
Qp = the number of jackets produced.
Of course, for other scenarios, the quantitative formulation would
be different. Regardless of the context or underlying situations, the
Monte Carlo formulation needs to capture the uncertainty with
random variables, in relation to the outcome. Based on the
quantitative formulation, we can repeat the simulation and generate
possible outcomes of interest.
If the simulation process is repeated many times, relevant
information such as average values, best and worst cases, and a
range of outcome values can be gathered to inform decision makers
about the appropriate course of actions. For example, in the
introductory case, Abigail wants to examine the page 570
demand and production rate of the pullover jacket to
see how they affect the profit. A Monte Carlo simulation can be used
to estimate the average profit or a range of best- and worst-case
scenarios at different staffing levels. More importantly, we can easily
change the parameter values in the simulation to help FashionTech
make a decision regarding the appropriate staffing level and whether
or not to automate the manufacturing process. This is shown in
Example 13.3.

EXAMPLE 13.3
From the introductory case, Abigail wants to develop an
analytical model to analyze the demand and production
information for a red pullover jacket for women, the top-selling
product for FashionTech. She obtains relevant demand and
cost information from her corporate database, which is
summarized in Table 13.1. Create a Monte Carlo simulation to
help Abigail examine the profit at each staffing level. Also, use
the simulation results to determine whether or not FashionTech
should consider increasing the automation in its production
process.

TABLE 13.1 Summary of FashionTech’s Demand and Cost Data

Description Data

Weekly demand Normal distribution with μ = 139 units

and σ = 14.61 units

Weekly production Uniform distribution between 31 and

rate per tailor 41 units

Selling price per $50

jacket

Material cost per Normal distribution with μ = $12 and

jacket σ = $1.37

Weekly wage per $700

tailor

Overhead cost per $800

week
 SOLUTION: We develop a Monte Carlo simulation based on the
profit equation presented earlier. In Example 13.2, we
generated random observations of the material cost for the red
pullover jacket. We use the same technique to generate
random observations for the weekly demand and production
rate. Because the number of jackets should be an integer, we
will generate random numbers and convert them into whole
numbers. For the material cost, we will use two decimal places.
FILE
Jacket

Using Excel
There are many ways to develop a Monte Carl simulation in
Excel; however, for consistency, we recommend that students
use a template provided in the Jacket worksheet for this
example. Also, to obtain consistent results, we use the random
number generation feature in Excel’s Analysis ToolPak and set
the random seed equal to 1. We will repeat the simulation for
100 iterations and summarize the results on the Jacket
worksheet.
a. Open the Jacket worksheet. Use Data > Data Analysis >
Random Number Generation to generate 100 random
numbers for the weekly demand in cells L11:L110, the
production rate in cells M11:M110, and the material cost in
cells N11: N110. Figure 13.2 shows the Random Number
Generation options for the three random variables; note that
the random seed is set equal to 1 in all simulations.
page 571
FIGURE 13.2 Random number generations for Demand,
Production Rate, and Material Cost
 Source: Microsoft Excel

b. We convert the random numbers for the weekly demand and

production rate into integers and use two decimal places for
the material cost using Excel’s ROUND function. In cells B11,
C11, and D11, enter the formulas =ROUND(L11, 0),
=ROUND(M11, 0), and =ROUND(N11, 2), respectively. Copy
and paste the formulas to cells B12:D110. Figure 13.3 shows
a portion of the random numbers. (Note that Excel on macOs
will generate different random numbers and the final results
will be slightly different, as shown in step e)

FIGURE 13.3 A portion of the Jacket worksheet for

FashionTech

Source: Microsoft Excel

c. Enter the selling price, $50, in cell B2; weekly overhead cost,
$800, in cell B3; and the tailor’s weekly wage, $700, in cell B4.
Label these input parameters, as shown in Figure 13.3. Recall
that Abigail wants to examine the possibility of increasing the
automation in her production process. Storing the input
 parameters in their own cells will allow us to later change
these values and obtain results from different scenarios.
d. Recall that the profit is calculated as revenue minus cost. To
compute the revenue, we multiply the selling price by the
number of jackets sold. Note that the number of jackets sold is
the lower value between the demand and the number of
jackets produced, and we use the MIN function in Excel to
determine the number of jackets sold. The total cost is an
aggregate of the overhead cost, the tailors’ wages, and the
material cost. To calculate the weekly profit for hiring one tailor
in cell E11, we enter the following formula.

To calculate the profit for other cells and different staffing

levels, copy and paste the formula to the remaining cells in the
E11:J110 range. Figure 13.4 shows a portion of the profit
values.

FIGURE 13.4 Simulation results for FashionTech

 page 572
It is important to pay close attention to the
dollar signs ($) in the above profit formula. A column name or
row number with a preceding dollar sign will not change even
if the formula is copied elsewhere on the worksheet. For
example, $B11 in the formula will always refer to column B
and E$10 will consistently refer to row 10, even if the formula
is copied to other cells. Similarly, $B$2 ensures that the
formula will always refer to cell B2 after being copied to other
cells.
e. We now use the Simulation results section on the Jacket
worksheet to calculate relevant statistics for each staffing
level. Enter the following formula in cells E3 through E6,
respectively: =AVERAGE(E11:E110), =STDEV.S(E11:E110),
=MAX(E11:E110), and =MIN(E11:E110). Copy and paste the
formula from cells E3:E6 to cells F3:J6. Additional statistical
information can also be obtained by using other Excel
functions; see Chapter 3 for a review. The Jacket worksheet
and the simulation results should now look similar to Figure
13.4. (For Excel on macOS, verify that the mean and standard
deviation for the three-tailor scenario in cells G3 and G4 are
$1,213.46 and $333.93, respectively. For the four-tailor
scenario, the mean and standard deviation in cells H3 and H4
are $1,615.11 and $596.03, respectively.) Figure 13.5 displays
a scatterplot of the profits from the 100 simulations for the
scenarios with three and four tailors (cells G11:G110 and
H11:H110). Figures 13.4 and 13.5 show that hiring four tailors
generates the highest profit, but variability of the profit levels is
also much greater than in other scenarios (i.e., a higher risk as
indicated by a larger standard deviation). Alternately, hiring
three tailors generates the second-highest profit but presents
a much lower risk than hiring four tailors.

FIGURE 13.5 Profit distribution for FashionTech

f. To evaluate the possibility of increasing automation in the
production process, change the weekly overhead cost and the
tailor’s wage in cells B3 and B4 to 900 and 550, respectively.
Also, generate new random numbers in column M by
specifying a uniform distribution with values between 36 and
41. The Jacket worksheet will repeat the 100 simulations for
the new scenario and re-calculate the simulation results as
shown in Figure 13.6. (For Excel on macOS, verify that the
mean and standard deviation for the three-tailor scenario in
cells G3 and G4 are $1,836.86 and $224.84, respectively. For
the four-tailor scenario, the mean and standard deviation in
cells H3 and H4 are $1,909.15 and $473.31, respectively.) In
this new scenario, hiring 4 tailors remains the most profitable
option and poses the highest risk. However, the expected
profit from hiring 3 tailors is now much closer to the 4-tailor
scenario, but at a much lower risk. Further discussion of the
results is included in the synopsis of the introductory case.
page 573
FIGURE 13.6 Simulation results with updated input
parameters
 As mentioned earlier, alternative spreadsheet designs can be
used to achieve similar Monte Carlo simulation results, as well
as to add more complexity and gain additional insights. For
example, one might consider treating the production rate for
each tailor as a separate random variable (e.g., six random
variables for the six tailors), or separating the demand for
various jacket sizes (i.e., S, M, and L) into different random
variables or simulation templates.
Using R

a. To ensure that we obtain consistent results, use the set.seed

function. Enter:
> set.seed(1)
b. Generate 100 random observations for the demand, the
production rate, and the material cost variables using the
designated distributions. Then use the round function to
convert the random numbers for the weekly demand and
production rate into integers and use two decimal places for
the material cost. Enter:
> demand <- rnorm(100, 139, 14.61)
> demand <- round(demand, digits = 0)
> productionRate <- sample(31:41, 100, replace = TRUE)
> productionRate<-round(productionRate, digits = 0)
> materialCost <- rnorm(100, 12, 1.37)
> materialCost<-round(materialCost, digits = 2)
c. Assign the per-unit price, the overhead cost, and the tailor’s
weekly wage to the corresponding variables. Enter:
> price <- 50
> overheadCost <- 800
> wageCost <- 700
d. Create a function called profitCalc that includes the formula for
computing the profit value. Enter:
> profitCalc <- function(staffLvl, priceLvl, overhead, wage,
demandLvl, productionLvl, material)
{priceLvl * pmin(demandLvl, productionLvl * staffLvl) -
overhead - (wage * staffLvl) - (productionLvl * staffLvl *
material)}
In the function statement, we list all the necessary arguments
(staffLvl for staffing level, priceLvL for the selling price, and so
forth) within the parentheses and define the profit formula
within the curly brackets ({ }). Recall that the profit is
calculated as revenue minus cost. To compute the revenue,
we multiply the selling price by the number of jackets sold.
Note that the number of jackets sold is the lower value
between the demand and the number of jackets produced,
and we use the pmin function to determine the number of
jackets sold. The total cost is an aggregate of the overhead
cost, the tailors’ wages, and the material cost.
page 574
e. The sapply function applies the profitCalc
formula defined in step d for each of the staffing levels (i.e.,
hiring 1–6 tailors) using each of the random observations for
the demand, production rate, and material cost. The output is
stored in the profitResults data frame. Enter:
> profitResults <- sapply(1:6, profitCalc, price, overheadCost,
wageCost, demand, productionRate, materialCost)
The first argument (1:6) in the sapply function represents the
six scenarios for hiring one to six tailors. The second
 argument in the sapply function recalls the profitCalc
function defined in step d. The remaining arguments apply the
selling price, overhead and wage costs, product demand,
production rate, and material cost to the profitCalc function
based on the corresponding variables defined in steps b and
c. Note that we list these arguments in the same order as the
corresponding variables defined in step d. If we were to
evaluate staffing levels between three and eight tailors, we
would change the R statement to profitResults <-
sapply(3:8, profitCalc, price, overheadCost, wageCost,
demand, productionRate, materialCost)
f. To calculate and view the average profit and identify the
minimum and maximum profits for each staffing level, use the
summary function. Enter:
> summary(profitResults)
As shown in Figure 13.7, hiring 3 and 4 tailors are likely to
generate the highest profit, with an average of $1,110.50 and
$1,396.50, respectively. However, the range of profit values is
larger in the four-tailor scenario.

FIGURE 13.7 Summary statistics

g. To calculate the standard deviation for the profits at each

staffing level, we use the sd function. The following
commands compute the standard deviation for the scenarios
with 3 and 4 tailors; the standard deviations for the other
staffing levels can be computed similarly. Enter:
> sd(profitResults[,3])
> sd(profitResults[,4])
 The standard deviation for the three-tailor scenario is
380.7242, and the standard deviation for four tailors is
507.4703. This is consistent with our observation in step f that
hiring four tailors yields a higher average profit but is also
riskier than hiring three tailors.
h. To evaluate the possibility of increasing automation, we start
by changing the overhead cost to $900 and the weekly wage
of each tailor to $550. We also redraw random observations
for the production rate variable using a uniform distribution
between 36 and 41. Enter:
> set.seed(1)
> overheadCost <- 900
> wageCost <- 550
> productionRate <- sample(36:41, 100, replace = TRUE)
> productionRate<-round(productionRate, digits = 0)
page 575
i. Recalculate the profits using the sapply function
and store the new output in the profitResultsAutomation data
frame. Also, calculate the summary statistics by using the
summary and sd functions. Enter:
> profitResultsAutomation <- sapply(1:6, profitCalc, price,
overheadCost, wageCost, demand, productionRate,
materialCost)
> summary(profitResultsAutomation)
> sd(profitResultsAutomation[,3])
> sd(profitResultsAutomation[,4])
Figure 13.8 shows the new output from the summary function.
In addition, the standard deviations for hiring three and four
tailors are 255.2004 and 646.9132, respectively.

FIGURE 13.8 Summary statistics for increasing automation

 Additional discussion of the results is presented in the synopsis
of the introductory case.

SYNOPSIS OF INTRODUCTORY CASE

Abigail Kwan develops an analytical model to analyze the demand and production
information for a red pullover jacket for women, the top-selling product for
FashionTech. She obtains relevant demand and cost information from her corporate
database. After reviewing the Monte Carlo simulation results, Abigail determines that
hiring three or four tailors per week would likely be the best scenarios for
FashionTech to generate a higher average weekly profit for the pullover jacket than
at other staffing levels. On average, hiring four tailors would generate the highest
profit, but it also poses a higher risk than hiring fewer tailors. With hiring three tailors
per week, the average profit is slightly lower, but it is less risky with a narrower range
of possible profits and a smaller standard deviation.
Abigail also evaluates the simulation results where additional automation is
implemented. Being able to hire less-skilled tailors would reduce the labor cost, and
the added automation would help decrease the variability in the production rate.
With these potential benefits, the three-tailor scenario becomes much more viable
as the average profit is now much closer to the average profit in the four-tailor
scenario, but with much less variability. However, this plan would also incur a one-
time installation and training cost for the added automation, which will be determined
through a request for proposals (RFP) from potential vendors. At her next meeting
with the company owner, Abigail plans to recommend that FashionTech move
forward with the RFP process.
lzf/Shutterstock

EXERCISES 13.2
For exercises in this section, set the random seed at 1 for the
Random Number Generation feature in Excel’s Analysis
ToolPak. In R, use the statement set.seed(1). This ensures
consistency of the results.

Mechanics
1. Use Excel’s Analysis ToolPak or R, both with a seed of 1, to simulate 25 random
observations based on a binomial distribution with five trials and p = 0.2. What are
the mean and standard deviation of the 25 observations?
page 576
2. Use Excel’s Analysis ToolPak or R, both with a seed of 1, to
simulate 50 random observations based on a continuous uniform distribution over
the interval [23, 37]. What is the range of simulated observations?
3. Use Excel’s Analysis ToolPak or R, both with a seed of 1, to simulate 200 random
observations based on a Poisson distribution with λ = 2. What are the mean and
standard deviation of the 200 observations?
4. Use Excel’s Analysis ToolPak or R, both with a seed of 1, to simulate 120 random
observations of a continuous uniform random variable over the interval [10, 75].
What are the mean, the standard deviation, and the range of the 120
observations? How many observations are greater than 65?
5. Use Excel’s Analysis ToolPak or R, both with a seed of 1, to simulate 1,000
random observations of a normally distributed random variable with μ = 9.23 and
 σ = 0.87. Report the mean and standard deviation of the 1,000 observations. How
many of the 1,000 observations have a value less than 8?

Applications
6. A manager at a local grocery store learns that 75% of her customers will use a
credit card when making a purchase. Use Excel’s Analysis ToolPak or R, both
with a seed of 1, to simulate a sample of 12 customers waiting in line at the
grocery store to make a purchase and repeat the simulation 500 times. Based on
the 500 simulations, report the average number of customers who use a credit
card to make a purchase. What is the range of the number of customers using a
credit card?
7. On a given day, about 85% of Internet users in the U.S. visit a social media site.
Use Analysis ToolPak or R, both with a seed of 1, to generate 100 simulations of
a sample of 10 Internet users and report the mean and the standard deviation of
the number of users in the sample who visit a social media site.
8. At a highly selective 4-year college, only 25% of transfer students go on to
graduate on time. Use Analysis ToolPak or R, both with a seed of 1, to generate
400 simulations where each simulation contains a group of 10 transfer students.
Report the mean, the standard deviation, and the range of the number of transfer
students with on-time graduation.
9. A local grocery store observes that on average 4 customers enter the store every
5 minutes during the hour between 5:30 pm and 6:30 pm each day. Use Analysis
ToolPak or R, both with a seed of 1, to generate a simulation for a period of 50
days. Report the mean and the standard deviation from the 50 simulations.
10. Monthly demand for a 60" TV at a local appliance store is normally distributed with
a mean of 11 units and standard deviation of 4.17 units. Use Analysis ToolPak or
R, both with a seed of 1, to develop a simulation for 300 months and report the
mean and the range of the demand for the TV.
11. Peter plans to invest $25,000 in a mutual fund whose annual returns are normally
distributed with a mean of 7.4% and standard deviation of 2.65%. Use Analysis
ToolPak or R, both with a seed of 1, to generate 100 trials to estimate the return
on Peter’s investment after one year. What are the mean and the range of the
investment returns?
12. Every Saturday morning between 9:00 am and 10:00 am, customers, on average,
arrive at a local coffee shop every 1.2 minutes, and the customer arrival follows
an exponential distribution. Use Analysis ToolPak or R, both with a seed of 1, to
generate 500 trials to simulate the time between customer arrivals and report the
sample mean and the standard deviation. Repeat the simulation with 1,000 trials
and compare the new sample mean and the standard deviation with the
theoretical values.
13. On average, a smartphone battery lasts about 6.5 hours with normal usage. The
smartphone battery life follows an exponential distribution. Use Analysis ToolPak
or R, both with a seed of 1, to generate 50 smartphone battery life simulations,
 and report the sample mean and the standard deviation. If the number of
simulations is increased to 500, compare the sample mean and the standard
deviation to the theoretical values.
14. A manufacturer of a smartphone battery estimates that monthly demand follows a
normal distribution with a mean of 400 units and standard deviation of 26. Material
cost is uniformly distributed between $7.00 and $8.50. Fixed costs are $2,700 per
month, regardless of the production rate. The selling price is $15 per unit.
a. Use Analysis ToolPak or R, both with a seed of 1, to simulate 1,000 trials to
estimate the expected monthly profit and standard deviation. Demand values
need to be rounded to integers, and use two decimal places for the material
cost.
b. What are the best and worst profit scenarios for the company?
15. Peter has $30,000 to invest in a mutual fund whose annual returns are normally
distributed with a mean of 5% and standard deviation of 4.2%.
a. Use Analysis ToolPak or R, both with a seed of 1, to simulate 5,000 trials to
estimate the mean balance after one year.
b. What is the probability of a balance of $32,000 or more?
c. Compare your results to another investment option at a fixed annual return of
3% per year.
16. Each week, a grocery store purchases eggs from a local ranch for $1.99 for each
12-egg carton and sells it for $3.89. Any cartons not sold within a week will be on
“manager’s special” sales or sold to a low-cost outlet for $1.25. If the eggs sell out
before the end of the week, an estimated opportunity cost of not meeting demand
is $1.75 per carton. The demand distribution is normally distributed with a mean of
75 cartons and a standard deviation of 12.5 cartons. Use Analysis ToolPak or R,
both with a seed of 1, to develop a Monte Carlo simulation for 500 weeks to
answer parts a and b.

page 577
a. If the store has been ordering exactly 75 cartons per week
from the rancher, what are the likelihood and opportunity cost of not meeting the
weekly demand?
b. If the store increases the weekly order to 85 cartons, what is the estimated cost
of having too many eggs in store?
17. A local coffee shop observes that, on average, four customers enter the store
every 5 minutes during the rush hour between 6:30 am and 7:30 am each day.
The number of customers arriving at the coffee shop follows a Poisson
distribution. Each barista can serve 2 or 3 customers every 8 minutes, a pattern
that follows a uniform distribution. The shop owner staffs her coffee shop with two
baristas. During the rush hour in the morning, customers are in a hurry to get to
work or school and will balk when there is a line. The opportunity cost when a
customer balks is $4.25 per customer. The profit generated from each customer is
normally distributed with a mean of $6.50 and standard deviation of $2.37. Each
barista is paid $20 per hour.
 a. Use Analysis ToolPak or R, both with a seed of 1, to develop a Monte Carlo
simulation with 500 trials to examine the current staffing level and report the
average profit or loss during the rush hour.
b. If the owner hires a third barista, what is the impact of the new hire on the
profit?
18. Hoping to increase its sales, a pizzeria wants to start a new marketing campaign
promising its customers that if their order does not get delivered within an hour,
the pizzas are free. Historically, the probability of on-time pizza delivery follows a
binomial distribution with n = 50 and p = 0.88. The order amount follows a normal
distribution with a mean of $35 and a standard deviation of $11.
a. Use Analysis ToolPak or R, both with a seed of 1, to simulate 1,000 pizza
orders. What is the average loss of revenue based on the 1,000 simulations?
b. In order to break even, how many new orders does the marketing campaign
need to generate?
19. An investor wants to invest $300,000 in a portfolio of three mutual funds. The
annual fund returns are normally distributed with a mean of 2.00% and standard
deviation of 0.30% for the short-term investment fund, a mean of 5.00% and
standard deviation of 2.50% for the intermediate-term fund, and a mean of 6.25%
and standard deviation of 5.50% for the long-term fund. An initial plan for the
investment allocation is 45% in the short-term fund, 35% in the intermediate-term
fund, and 20% in the long-term fund.
a. Use Analysis ToolPak or R, both with a seed of 1, to simulate 100 trials to
estimate the mean ending balance after the first year and assess the risk of this
investment.
b. If the allocation is changed to 30% short-term, 55% intermediate-term, and 15%
long-term, estimate the ending balance after the first year and the risk of the
investment.
20. At a local appliance store, weekly demand for a large screen TV during the week
prior to the Super Bowl is normally distributed with a mean of 35 units and
standard deviation of 12 units. The store usually keeps 40 units in its inventory.
Use Analysis ToolPak or R, both with a seed of 1, to simulate 500 trials to
estimate the likelihood of overstocking and understocking.
21. Following an exponential distribution, the average lifespan of a smartphone
battery is 2.3 years. The battery manufacturer wants to offer a warranty for its
customers to receive a free replacement if the battery fails during the first year.
Each battery generates a profit of $10.85, and the replacement cost is $6.35.
Develop a Monte Carlo simulation using Analysis ToolPak or R, both with a seed
of 1, for 100 battery units sold.
a. What is the expected total cost of this warranty program?
b. In order to cover the cost of the warranty program, how many additional battery
units does the company need to sell?
22. An amusement park has a roller coaster that can accommodate up to 45 park
goers per ride, and each ride lasts about 20 minutes. For each hour, the park
gives out 200 tickets for the ride, and on average, only 60% of the ticket holders
 come to ride on the roller coaster. The number of roller coaster riders follows a
normal distribution with a mean of 120 riders per hour and standard deviation of
35 riders. Develop a Monte Carlo simulation using Analysis ToolPak or R, both
with a seed of 1, of 400 trials. What is the average number of riders unable to get
on the roller coaster?

13.3 OPTIMIZATION WITH LINEAR

PROGRAMMING
LO 13.3
Formulate a linear programming model.

Optimization is a family of quantitative techniques where an objective

function is maximized or minimized for obtaining optimal solutions to
complex business problems. We usually start by clearly defining a
business objective. For example, the objective may be to maximize
profits or to minimize costs subject to resource constraints, such as a
fixed amount of available materials, limited financial resources, and a
finite amount of time.
Linear programming (LP), a subset of the optimization
techniques, is one of the simplest ways to perform optimization. By
making a few simplifying assumptions, such as linear relationships
among relevant variables, it helps in solving some very complex
optimization problems. The LP technique is widely page 578
used, and common LP applications include
developing a mix of financial investment options, choosing a
combination of products to manufacture, and minimizing the delivery
cost for shipping products and packages.
This section discusses a class of LP problems where optimal
solutions can take on fractional values. Section 13.4 discusses the
linear integer programming technique where optimal solutions are
integers. Other optimization techniques where nonlinear
relationships exist among relevant variables are beyond the scope of
this text.

OPTIMIZATION AND LINEAR PROGRAMMING

Optimization: A family of quantitative techniques for obtaining
optimal solutions to complex problems with resource
constraints.

Linear programming (LP): A simple way to perform optimization

by assuming linear relationships among relevant variables.

The first step in performing LP is to formulate a problem into a

series of mathematical expressions. Naturally, different applications
require different formulations, but regardless of the context, the
mathematical representation of most LP problems has four essential
components: an objective function, decision variables, constraints,
and parameters.
In any optimization problem, a single objective with a
mathematical representation is referred to as the objective
function. In LP, the objective function is assumed to be linear. Two
general types of an objective function are a maximization function
(e.g., for profit, revenue, or the return on investment) or a
minimization function (e.g., for manufacturing cost, the amount of
time to complete a project, or the distance traveled to complete a
delivery).
Decision variables refer to the different choices or alternatives
from which a decision maker has to choose, in order to maximize or
minimize the value of the objective function. For example, an
investor might choose from several stocks or bonds to maximize the
potential return on investment. Similarly, a manager might choose
from several production inputs to minimize manufacturing cost.
An LP optimization problem includes a number of constraints or
limitations within which a decision maker needs to operate. These
constraints represent the limited resources available to the decision
maker. Examples of constraints include a fixed amount of money an
investor contributes toward retirement, a limited amount of raw
materials a manufacturer can procure each month, or a certain
number of hours a delivery truck driver is available to work.
Sometimes, a constraint also reflects the knowledge gained from
historical data or human expertise. For example, a retail store owner
might learn from experience that there is a limit to the demand for
certain products in the local market and maintains the quantity of
products in the retail store accordingly. Other constraints are based
on intuitive assumptions about the nature of the LP problem. When a
manufacturer decides how many units of products to produce or
when an investor decides how much money to invest in each
financial asset, the units to produce and the amount of investment
are assumed to be non-negative.
Finally, parameters (or input parameters) are numerical values
associated with the objective function, decision variables, and
constraints. For example, a delivery truck driver in the United States
can only work up to 14 consecutive hours within a 24-hour period,
and, therefore, the value 14 is the parameter value for the working
hour constraint. For other LP problems, parameter values might be
derived from historical data (e.g., the amount of time and raw
materials needed to manufacture a product might be obtained from
the manufacturing data from the previous month). The input
parameters usually remain constant and do not change when the
solution to the LP problem is implemented, unless there are
significant changes to the external factors surrounding the LP
problem (e.g., changes in the manufacturing process).

page 579

FOUR COMPONENTS OF LINEAR PROGRAMMING

Objective function: A mathematical representation of an
objective, usually expressed as a maximization or minimization
 function of a single variable such as profit or cost.

Decision variables: Choices or alternatives available for a

decision maker to choose, in order to maximize or minimize the
value of the objective function.

Constraints: Limitations within which a decision maker needs to

operate that usually represent the limited resources available to
the decision makers.

Parameters: Numerical values in the LP mathematical

expression associated with the objective function, the decision
variables, and the constraints.

Formulating a Linear Programming Model

In Example 13.4 and Example 13.5 we show how to specify a
maximization function and a minimization function, respectively,
given various constraints.

EXAMPLE 13.4
After obtaining useful results and actionable insights from the
Monte Carlo simulation, Abigail Kwan wants to continue using
prescriptive analytics to solve other business problems for
FashionTech. One of the difficult decisions that she has to
make every winter season is deciding how many parkas and
winter jackets for women to produce. Both products use the
same materials and require very similar sewing and stitching
skills and manufacturing steps, and, therefore, compete for the
same resources. The main difference between the two
products is the length—jackets are generally shorter and end at
the waist or just below, while parkas are longer fitting, and
therefore offer more warmth. Due to their length, parkas have a
higher manufacturing cost, but also have a higher selling price
and generate a greater profit per unit.

Gordana Sermek/Shutterstock

Jackets and parkas come in different sizes, but the most

popular ones are the medium size. Abigail will start by focusing
on analyzing data for the medium-sized jackets and parkas.
Based on her records, the per-unit profits for a winter jacket
and parka are $9 and $12.50, respectively. Also, a medium-
sized jacket requires approximately 8.5 feet of fabric, while a
medium-sized parka requires about 12.5 feet of the same
fabric. The amount of machine time needed to produce a jacket
and a parka are 1.5 hours and 2 hours, respectively. A tailor
usually spends about 2 hours on sewing and stitching for a
winter jacket and about 3 hours for a parka.
FashionTech has a monthly contract with its supplier to
procure 4,000 feet of the required fabric. The amount of
machine time allocated to producing the two products is 650
hours per month, and each month a number of skilled tailors
are assigned to these products for approximately 900 hours.
Historically, FashionTech has never sold more than 150 parkas
or 400 jackets in a month, and, therefore, Abigail wants to
make sure that the monthly production rates of parkas and
winter jackets are no more than these historical limits. Use the
information provided to help Abigail formulate this problem into
an LP model.

page 580

SOLUTION: The objective of this analysis is to find a solution that

maximizes the profit contribution from producing and selling the
medium-sized jackets and parkas. The two decision variables
are the number of winter jackets and parkas to produce. The
objective function for this LP model can be formulated as
follows.

where x1 and x2 are the decision variables that represent the

number of jackets and parkas produced, respectively. For
tractability, we assume that the quantity produced and the
quantity sold are the same. Note that the per-unit profits, $9
and $12.50, are the parameter values. The parameter values in
the objective function are also referred to as the objective
function coefficients.
Furthermore, FashionTech needs to operate within the five
constraints mentioned in the problem description: the amount
of fabric available from the supplier, the limited machine time,
the number of hours the skilled tailors are assigned to the two
products, and the historical demand for parkas and winter
jackets. The five constraints and their corresponding
parameters can be formulated as follows.
 A maximization LP model usually has at least one
constraint with the ≤ sign, which reflects the fact that a decision
maker needs to operate within these limitations. In addition to
the five constraints, the number of winter jackets and parkas
are non-negative. Consequently, we add the last constraint to
the LP formulation.

As mentioned earlier, another type of linear programming

problem is to minimize the value of the objective function. In these
cases, the objective function would be formulated in a similar fashion
as in Example 13.4, but we would instead aim to minimize its value.
In addition, the constraints in the minimization problem are usually
formulated to describe minimum requirements that need to be met.
For example, a shipping company may want to minimize the total
delivery cost of packages (i.e., the objective function) while making
sure that at least a certain number of customers will receive their
packages on a given day (e.g., each delivery truck driver must
deliver at least 35 packages a day). As such, a minimization LP
model usually has at least one constraint with the ≥ sign. Example
13.5 demonstrates how to formulate a minimization LP problem.
EXAMPLE 13.5
Yerba Buena Tea is a manufacturer and wholesaler of high-
quality tea products. Spiced chai is becoming increasingly
popular among younger consumers, and Yerba Buena offers
two chai products: spiced chai powder mix and spiced chai tea
concentrate. An older production facility costs $190 per hour to
operate and, for each hour of operation, can produce 295
ounces of the powder mix and 260 ounces of the tea
concentrate.

JulijaDmitrijeva/Shutterstock

page 581

Because of its age, the old facility cannot operate longer

than 8 consecutive hours at a time. A newer facility costs $260
per hour to operate and produces 385 ounces of the powder
mix and 375 ounces of the tea concentrate per hour. Yerba
Buena also owns a small, single-purpose facility that costs
$150 per hour to operate and produces 350 ounces of the
powder mix per hour. Several coffee shops have placed orders
for a total of 5,500 ounces of the chai powder mix and 4,000
ounces of the tea concentrate.
 Anika Patel, the general manager at Yerba Buena, has
none of the two products left in inventory and needs to decide
how many hours to operate each facility to fulfill the orders
while minimizing the production cost. Formulate this problem
into an LP model for Yerba Buena.

SOLUTION: The objective for Yerba Buena is to minimize the

production cost while being able to fulfill the orders. The
objective function represents the total cost of operating the
three production facilities and can be formulated as follows.

Subject to:

where x1, x2 and x3 represent the number of hours to operate

the three facilities, which are the decision variables. The first
two constraints in the LP formulation represent the order
amounts for the chai powder mix and tea concentrate with the
corresponding parameters. The third constraint indicates that
the old facility cannot operate longer than 8 hours. Following a
similar assumption in Example 13.4, we include the last three
non-negativity constraints in the LP formulation.

In the next subsection, we demonstrate how to solve the LP

model and obtain the optimal solution to help a decision maker, such
as Abigail Kwan at FashionTech or Anika Patel at Yerba Buena Tea,
make an informed business decision.

LO 13.4
Solve and interpret a linear programming model.

Solving a Linear Programming Problem

In this chapter, we solve linear programming problems using Excel
and R. However, to gain an understanding of a linear programming
technique, we first demonstrate how to solve an LP problem with two
decision variables through graphical representation. Given the
linearity of the objective function as well as the constraints, we can
plot the mathematical expressions in a line chart and identify
possible solutions that satisfy the constraints. Consider the
FashionTech example where Abigail Kwan tries to decide how many
winter jackets and parkas to produce. The mathematical expression
for the objective function and constraints can all be presented
graphically in a line chart. To find an optimal solution to the LP
model, we identify the values of the decision variables, in this case
x1 and x2, that maximize the profit while satisfying all of the
constraints. In the Yerba Buena case, we find the values of the
decision variables, x1, x2, and x3, that minimize the production cost
subject to the constraints.
We will illustrate the graphical approach with the FashionTech
example. We start by plotting the machine time and the two demand
constraints. Recall that x1 and x2 denote the number of jackets and
parkas produced, respectively. The orange line in Figure 13.9
represents the machine time constraint, 1.5x1 + 2x2 ≤ 650. Note that
x1 reaches a maximum value of 433.33 (= 650/1.5) page 582
when x2 = 0, and x2 reaches a maximum value of
325.00 (= 650/2) when x1 = 0. These values provide the coordinates
for the line representing the machine time constraint. Furthermore,
because the machine time constraint is an inequality with the ≤ sign,
the area under the orange line represents the values of x1 and x2
that satisfy this constraint. We also plot the demand limits for x1 and
x2 as two dotted purple lines. Similar to the machine time constraint,
the two demand constraints, x1 ≤ 400 and x2 ≤ 150 have the ≤ sign,
and, therefore, the area to the left of the vertical dotted line and the
area under the horizontal dotted line represent the values of x1 and
x2 that satisfy the constraints. Recall also that x1 and x2 must be
non-negative, and, therefore, we only consider the first quadrant of
x1 and x2. The shaded area in Figure 13.9 represents the values of
x1 and x2 that satisfy the machine time constraint and the two
demand constraints.

FIGURE 13.9 Graphical representation of the machine time constraint for

FashionTech
 The other two constraints for fabric and labor as well as the
objective function can also be plotted in a similar fashion. Figure
13.10 shows a line chart, and the resulting shaded area represents
the values of x1 and x2 that satisfy all five constraints. This shaded
area is referred to as the feasibility region of the LP model. Because
we do not yet know the maximum profit we can achieve, we start by
superimposing three black solid lines representing the objective
function at a profit value of $500, $1,500, and $2,500. As we
increase the profit, the line shifts further to the right-hand side of the
chart. Note that at a profit value of $2,500, part of the profit line
already falls outside of the feasibility region. The values of x1 and x2
outside of the feasibility region cannot be considered in the LP
solution as they violate at least one of the constraints (e.g., the
number of parkas cannot exceed 150). Conversely, page 583
any part of the profit line that stays within the
feasibility region satisfies all of the constraints, and the
corresponding values of x1 and x2 are possible candidates for a
solution to the LP problem.

FIGURE 13.10 Graphical representation of the LP model for FashionTech

 In order to maximize the profit while satisfying all the constraints,
we continue to increase the profit as long as part of the profit line
remains within the feasibility region. Normally, an optimal solution to
an LP model can be found at the boundary or at a corner of the
feasibility region. In this example, if we continue to shift the profit line
further to the right-hand side of the chart, we will find that an optimal
solution is where x1 = 300 and x2 = 100 (see the dotted black lines in
Figure 13.10). At these optimal values, we calculate the maximum
profit as $3,950 (= 9 × 300 + 12.5 × 100), as shown in black dotted
lines.
Note that at the optimal solution of 300 winter jackets and 100
parkas, FashionTech would have used only 3,800 (= 8.5 × 300 +
12.5 × 100) of the 4,000 feet of fabric available from the supplier.
The remaining 200 feet of fabric is called “slack,” indicating an
excess supply that can be used for other products; FashionTech
could also renegotiate with its supplier to order less fabric per month.
For a minimization problem, this would be called “surplus,” instead of
slack.
Constraints with slack or surplus in the LP solution are called
nonbinding constraints. Adding or reducing the nonbinding
constraint value by one unit (i.e., increasing the amount of available
fabric to 4,001 feet) will not change the LP solution. At the optimized
values, however, FashionTech will have used all of the allotted
machine time (650 hours) and labor (900 hours), so there is no slack
in these two resources. As a result, the machine page 584
time and labor are called binding constraints.
Increasing or decreasing the available resource of a binding
constraint will impact the LP solution and, in most cases, the
optimized value of the objective function.
Another important concept related to LP is that of a shadow
price, also referred to as a dual price, of a constraint. It indicates
the change in the optimized value of an objective function due to a 1-
unit change in a binding constraint, holding the other parameters
constant. In the FashionTech example, it can be shown that reducing
the available machine time by one hour (i.e., from 650 to 649 hours)
would cause the optimized profit to decrease from $3,950 to $3,946,
making the shadow price of the machine time $4. Note that the
optimized value of an objective function will not change due to a 1-
unit change in a nonbinding constraint.

CONSTRAINTS IN A LINEAR PROGRAMMING PROBLEM

A constraint in an LP problem is considered to be binding if
changing it also changes the optimized value of the objective
function. Nonbinding constraints are those that do not affect the
optimized value.
The shadow price of a constraint measures the change in
the optimized value of the objective function when a binding
constraint is changed by one unit.

With only two decision variables, we are able to plot the LP

formulation onto a two-dimensional line chart and find an optimal
solution. This approach, however, is not suitable for LP problems
with more than two decision variables, such as the Yerba Buena Tea
example. Conceptually, we could also use an algebraic procedure
called the Simplex method to manually solve the LP problem. In
most situations, even for a problem with two decision variables, it is
much more convenient to solve the LP formulation using computer
software. In Examples 13.6 and 13.7, we use Excel’s Solver and R
to solve a maximization problem and a minimization problem,
respectively.
FILE
Parkas

EXAMPLE 13.6
Recall from Example 13.4 that Abigail Kwan, the operations
manager at FashionTech, needs to decide how many winter
 jackets and parkas to produce in order to maximize profits. Use
Excel’s Solver and R to maximize the objective function, given
the five constraints. Summarize the results.
SOLUTION:
Using Excel
Excel has an add-in feature called Solver that implements the
Simplex method and performs necessary computations to
solve LP problems. There are many ways to structure an Excel
spreadsheet for an LP problem. For consistency, we will use
the Parkas worksheet as a template to demonstrate how to
use Solver. Similar to the Monte Carlo simulation example,
creating an Excel worksheet with input parameters in separate
cells, as in the Parkas worksheet, will allow us to easily change
the parameter values in the LP model and evaluate alternative
scenarios.
a. Open the Parkas worksheet and make sure Solver is
activated by going to File > Options > Add-ins > Go (next to
Manage Excel Add-ins). Check the Solver checkbox and click
OK. Verify that a button labeled Solver is in the Analyze
group under the Data tab.
b. Navigate to the Parameters section of the Parkas worksheet
and, in cells B5 through B8, enter the values 8.5, 1.5, 2, and 9,
which correspond to the amount of fabric, the page 585
machine time, the amount of labor, and the per-
unit profit that are associated with the production of one winter
jacket. In cells C5 through C8, enter the corresponding values
for the production of one parka.
c. Enter the amount of fabric available to FashionTech each
month (4,000 feet) in cell D14 and enter the remaining
applicable information in cells D15 through D19. Your Parkas
worksheet should look similar to Figure 13.11.

FIGURE 13.11 Excel template for Example 13.6

 Source: Microsoft Excel

d. We use cell E4 to represent the objective function and cells

B11 and C11 to represent the two decision variables: the
number of winter jackets and parkas that FashionTech should
produce. To formulate the objective function from Example
13.4, enter =SUMPRODUCT(B8:C8, B11:C11) in cell E4. Note
that the SUMPRODUCT formula is equivalent to inputting
=B8*B11+C8*C11. Verify that the profit calculation is correct
by entering 1 in cells B11 and C11. Cell E4 should display a
profit of $21.50 (= 9 × 1 + 12.50 × 1).
e. Enter =SUMPRODUCT(B5:C5, B11:C11) in cell B14 to
calculate the amount of fabric used; =SUMPRODUCT(B6:C6,
B11:C11) in cell B15 for the total machine time; and
=SUMPRODUCT(B7:C7, B11:C11) in cell B16 for the total
number of hours the tailors will spend making the jackets and
parkas each month.
f. Enter =B11 and =C11 in cells B18 and B19, respectively, to
replicate the numbers of winter jackets and parkas to be
produced. Your Parkas worksheet should now look similar to
Figure 13.12.

FIGURE 13.12 Completed Excel template for Example 13.6

 Source: Microsoft Excel

page 586
g. Launch Solver by navigating to the Data tab and
click the Solver button in the Analyze group.
h. In the Solver Parameters dialog box, enter E4 for the Set
Objective box. Make sure to choose the Max option to
maximize the profit value in cell E4.
i. Enter B11:C11 in the By Changing Variable Cells box. Recall
that these two cells represent the decision variables.
j. Click the Add button to add the constraints for the amount of
fabric, total machine time, and labor available to FashionTech.
In the Change Constraint dialog box, enter B14:B16 for Cell
Reference, choose <= from the drop-down list of operators,
and enter D14:D16 for the Constraint option; see Figure
13.13. Click Add.

FIGURE 13.13 Adding the constraints in Excel’s Solver

 Source: Microsoft Excel

k. Click the Add button again, and in the next Change Constraint
dialog box, enter B18:B19 in the Cell Reference box for the
demand constraints, choose the <= option, and enter
“D18:D19” for the Constraint option. Click OK.
l. Make sure the option Make Unconstrained Variables Non-
Negative is checked to enforce the non-negativity constraints.
For the Select a Solving Method option, choose Simplex LP.
Your Solver dialog box should look similar to Figure 13.14.
Click Solve.

FIGURE 13.14 Solver Parameters

 Source: Microsoft Excel

m. In the Solver Results dialog box, click the Answer and

Sensitivity options under the Reports section and then click
OK. The solution to the LP model is shown in page 587
Figure 13.15. Solver will also create two new
worksheets labeled Answer Report 1 and Sensitivity Report 1.
[The actual worksheet number might be different.]

FIGURE 13.15 Solver optimal LP solution

Source: Microsoft Excel

As shown in Figure 13.15, Solver indicates in cells B11 and

C11 on the Parkas worksheet that FashionTech should
produce 300 winter jackets and 100 parkas, well within the two
demand constraints of 400 winter jackets and 150 parkas.
Recall that this is the same solution found on the line chart in
Figure 13.10. The maximum profit from this optimal solution is
$3,950 as calculated in cell E4. It is important to note that for
other LP problems, there may be multiple optimal solutions that
maximize the objective function value.
Figure 13.16 shows a relevant portion of the Answer Report
worksheet produced by Solver. As noted earlier, the amount of
fabric is a nonbinding constraint because only 3,800 of the
4,000 feet of fabric are used, resulting in a slack of 200 feet.
The machine time and labor are binding constraints because all
of the allotted machine time (650 hours) and labor (900 hours)
are used at the optimal values.

FIGURE 13.16 Excel’s Solver Answer Report

Source: Microsoft Excel

Relevant information from the Sensitivity Report worksheet

is shown in Figure 13.17. The Variable Cells section describes
the optimal LP solution where 300 jackets and 100 parkas are
produced and sold at the per-unit profit of $9 and $12.50.
Recall that the per-unit profits are called objective function
coefficients. A change in these coefficients beyond the
allowable increase or decrease (the last two columns in Figure
13.17) will cause the optimal LP solution to change. For
example, if the per-unit profit of the winter jacket decreases by
more than $0.6667, we would need to solve the LP model
again to obtain a new solution; that is, producing 300 jackets
and 100 parkas would no longer generate the maximum profit.
Likewise, if the per-unit profit of a jacket increases by more
than $0.3750, we will need to find a new LP solution. In such
scenarios, we will need to replace the per-unit profits with the
new values on the Parkas worksheet and rerun Solver to
obtain a new solution. Any other changes smaller than the
allowable amount would not alter the optimal page 588
solution, even though the actual profit amount
will change. In other words, as long as the per-unit profit of the
jackets falls between $8.3333 and $9.375 (or $9 − $0.6667 and
$9 + $0.375), it will be optimal for FashionTech to continue to
produce 300 jackets and 100 parkas per month, but the total
profit will change depending on the per-unit value. This range
of values between the allowable increase and decrease is
referred to as the range of optimality; the optimal solution will
remain the same as long as the objective function coefficient
stays within this range. The range of optimality for the per-unit
profit of the parkas is between $12 and $13.5 (or $12.5 − $0.5
and $12.5 + $1). The range of optimality is useful for business
managers like Abigail Kwan at FashionTech who regularly
make decisions about product discounts, sales, and marketing
campaigns in relation to their manufacturing plan.

FIGURE 13.17 Solver’s Sensitivity Report

Source: Microsoft Excel

The Constraints section of Figure 13.17 also shows the

allowable increase and decrease for each constraint. Because
the Fabric constraint is nonbinding, increasing the amount of
fabric in the production process would not have any impact on
the LP solution (the 1E+30 value implies that even an infinite
increase in the amount of fabric would not alter the LP
solution). The allowable decrease is simply the slack value, or
the difference between the amount of Fabric available (4,000
feet) and the amount used (3,800 feet). Decreasing the amount
of fabric below 3,800 feet per month will alter the LP solution.
This is also the case for the two demand constraints at the
bottom of Figure 13.17.
The shadow price column in Figure 13.17 is the change in
the profit for each unitary change in the binding constraint
within the allowable increase and decrease. For example, the
shadow price of the machine time is $4. This means that
reducing the available machine time by one hour (from 650 to
649 hours) would cause the total profit to decrease from $3,950
to $3,946. In other words, a $4 decrease in profit is the “price”
the company will have to pay for not allocating the one hour of
machine time to the production of jackets and parkas. We can
verify this by entering 649 for the available machine time in cell
D15 in the Parkas worksheet. Rerunning Solver would yield a
new optimal solution of 294 winter jackets and 104 parkas in
the decision variables (cells B11 and C11), and the new profit
of $3,946 in cell E4. Likewise, allocating one additional hour of
machine time (i.e., from 650 to 651 hours) would increase the
profit from $3,950 to $3,954 by producing 306 jackets and 96
parkas.
The shadow price remains constant only within the
allowable increase and decrease. For the machine time, the
shadow price stays at $4 as long as the available machine time
is between 637.5 (i.e., 650 − 12.5) and 666.67 (i.e., 650 +
16.67) hours. This range is called the range of feasibility, over
which the shadow price remains unchanged. The labor
constraint has a shadow price of $1.50 with a range of
feasibility between 875 and 916.67 hours.

page 589

In certain situations, the shadow price of a binding

constraint could also be zero. The zero value of the shadow
price of a binding constraint simply means that changing the
available amount of resources within the allowable range will
cause the LP solution to change, but the optimized objective
 function value (e.g., profit) will remain the same. The shadow
price information is particularly useful for business managers
like Abigail Kwan at FashionTech who have to regularly decide
where to allocate additional resources. Intuitively, allocating
additional resources to the constraint with the highest shadow
price would yield the greatest financial return.
Using R
We use the lpSolve package in R to find an optimal solution for
the FashionTech example. Like Excel’s Solver and many other
linear programming software, the lpSolve package also
implements the Simplex method. The optimal solution found by
Excel’s Solver and R for the LP examples in this chapter are
identical. The answers may differ when an LP model has
multiple optimal solutions, which we will not pursue in this
chapter.
a. Install and load the lpSolve package. Enter:
> install.packages(“lpSolve”)
> library(lpSolve)
b. Specify the two objective function coefficients for the per-unit
profits of winter jackets ($9) and parkas ($12.5). Enter:
> lp.objective <- c(9, 12.5)
c. Store the parameters from the left-hand side of the five
constraints in a matrix using the matrix command. Each
constraint has two left-hand side parameters, which are
specified in pairs. Enter:
> lp.constraints <- matrix(c(8.5, 12.5, 1.5, 2, 2, 3, 1, 0, 0, 1),
nrow=5, byrow=TRUE)
Recall that the first pair of parameters, 8.5 and 12.5, are the
amount of fabric required to produce one winter jacket and
one parka; the second pair, 1.5 and 2, are the amounts of
machine time needed for one winter jacket and one parka, and
so forth. It is important to note that the lpSolve package
requires that we specify the parameter values for all of the
 decision variables (x1 and x2 in this case), even for decision
variables that may not appear in a constraint. For example, for
the demand constraint for jackets (x1 ≤ 400), even though x2
does not appear in this constraint, we need to specify that its
parameter is zero. As a result, the parameter values of x1 and
x2 for the jacket demand constraint are 1 and 0, respectively.
Likewise, for the demand constraint for parkas (x2 ≤ 100), the
parameter values for x1 and x2 are 0 and 1. This is why the
last four parameters for the two demand constraints appear as
1, 0, 0, 1.
d. Specify the ≤ sign for each of the five constraints. Enter:
> lp.directions <- c(“<=”, “<=”, “<=”, “<=”, “<=”)
e. Specify the available amount of fabric, machine time, and
labor as well as the two demand limitations (i.e., the values on
the right-hand side of the five constraints). Enter:
> lp.rhs <- c(4000, 650, 900, 400, 150)
f. By default, the lpSolve package imposes non-negativity
constraints on all decision variables, and, therefore, we do not
need to explicitly specify these constraints. We use the lp
function to maximize the objective function given the
constraints and store the results in an object called lp.output.
To display the maximum profit and LP solution, page 590
we retype lp.output as well as lp.output$solution.
Enter:
> lp.output <- lp(“max”, lp.objective, lp.constraints,
lp.directions, lp.rhs, compute.sens = TRUE)
> lp.output
> lp.output$solution
R reports that the maximum profit is $3,950, and the LP
solution is for FashionTech to produce 300 winter jackets and
100 parkas. These are the same results that we obtained
graphically as well as with Excel Solver.
g. To display the range of optimality for the two objective function
coefficients (the per-unit profit of the jackets and parkas),
enter:
> lp.output$sens.coef.from
> lp.output$sens.coef.to
The range of optimality for the per-unit profit of a winter jacket
is between $8.333 and $9.375, while the range for the per-unit
profit of parkas is between $12.000 and $13.500. The optimal
number of jackets and parkas to be produced will remain at
300 and 100 units as long as the objective function
coefficients stay within these ranges.
h. To display the shadow price for the constraints (called the dual
price in the lpSolve package) and the range of feasibility,
enter:
> lp.output$duals
> lp.output$duals.from
> lp.output$duals.to
For the two binding constraints, machine time and labor, the
shadow prices are $4 and $1.5, respectively. The ranges of
feasibility for the two resources are between 637.50 and
666.67 hours and between 875 and 916.67 hours,
respectively. As mentioned earlier, the optimal solution found
by Excel’s Solver and R for the LP examples in this chapter
are identical. For further discussion of the range of optimality,
the shadow price, and the range of feasibility, refer to the
discussion of the Excel results.

FILE
Chai
 EXAMPLE 13.7
Recall from Example 13.5 that Anika Patel, the general
manager at Yerba Buena Tea, needs to decide how many
hours to operate each of the three production facilities. This
decision must be made to ensure that all orders for the spiced
chai powder mix and the chai tea concentrate are filled, while
the production cost is kept at minimum. Use Excel’s Solver and
R to minimize the objective function given the four constraints.
Summarize the results.
SOLUTION:
Using Excel
In this example, we use a template in the Chai worksheet for
consistency. If you have not already done so, activate the
Solver software by launching Excel and go to File > Options >
Add-ins > Go. Check the Solver checkbox and click OK.
Because many of the steps are analogous to those that we
outlined in Example 13.6, we are brief. Because Excel and R
produce identical results, we summarize the results after the R
instructions.

page 591

a. Open the Chai worksheet and enter the parameter values

from Example 13.5 in cells B5:D7 and cells D13:D14 and D16.
See Figure 13.18.

FIGURE 13.18 Excel template for minimization problems

 Source: Microsoft Excel

b. Enter =SUMPRODUCT(B5:D5, B10:D10) in cell B13,

=SUMPRODUCT(B6:D6, B10:D10) in cell B14, and =B10 in
cell B16.
c. Enter =SUMPRODUCT(B7:D7, B10:D10) in cell F4. Cell F4
represents the value of the objective function (i.e., the total
production cost). Remember that our objective is to minimize
the production cost.
d. Verify the Excel formula by entering 1 in cells B10, C10, and
D10. We use these three cells to store the number of hours to
operate the three production facilities (i.e., the decision
variables). Your worksheet should look similar to Figure 13.18.
e. Go to Data > Solver. Enter the options and constraints as
shown in Figure 13.19 and click Solve. Note that we choose
the Min option for the minimization problem and use the ≥ sign
in the first two constraints.

FIGURE 13.19 Excel’s Solver Parameters dialog box for

Example 13.7
 Source: Microsoft Excel

f. In the Solver Results dialog box, click the Answer and

Sensitivity reports. Then click OK.

page 592
Using R
Because many of the steps are analogous to those that we
outlined in Example 13.6, we are brief.
a. If you have not already done so, install and load the lpSolve
package.
b. We specify the three objective function coefficients, the
parameters in the three constraints, the ≥ and ≤ signs, and the
amounts on the right-hand side of the three constraints. Note
that the constraint for the order amount for the tea concentrate
 (260x1 + 375x2 ≥ 4,000) only has two decision variables, so
the parameter value for x3 is specified as 0. Similarly, the limit
on the operating hours for the old facility (x1 ≤ 8) has only one
decision variable, so the parameter values for x2 and x3 are
both 0. As a result, the last four parameters in the matrix
function are 0, 1, 0, 0. Enter:
> lp.objective <- c(190, 260, 150)
> lp.constraints <- matrix(c(295, 385, 350, 260, 375, 0, 1, 0,
0), nrow=3, byrow=TRUE)
> lp.directions <- c(“>=”, “>=”, “<=”)
> lp.rhs <- c(5500, 4000, 8)
c. In order to solve the LP minimization model, we use the lp
function with the min option. Enter:
> lp.output <- lp (“min”, lp.objective, lp.constraints,
lp.directions, lp.rhs, compute.sens = TRUE)
d. In order to display the LP solution and relevant output, enter:
> lp.output
> lp.output$solution
> lp.output$sens.coef.from
> lp.output$sens.coef.to
> lp.output$duals
> lp.output$duals.from
> lp.output$duals.to
Summary of the Results
Based on the optimal solution given by Excel’s Solver and R,
Anika should operate the older facility for 8 hours (x1 = 8), the
newer facility for 5.12 hours (x2 = 5.12), and the single-purpose
facility for approximately 3.34 hours (x3 = 3.34). This plan
would produce 5,500 ounces of the spiced chai powder mix
and 4,000 ounces of the spiced chai tea concentrate, which are
precisely the total amounts ordered. This optimal solution will
incur the total production cost of $3,352.11, which is the lowest
cost for the current parameters and constraints. All three
constraints in the LP model are binding, with no surplus. Based
on the range of optimality, the older facility should be used for 8
hours unless its operating cost increases beyond $192.30 per
hour. Reducing the operating cost of this facility does not have
any impact on the operating hours. For the newer facility, if its
operating cost can be reduced below $256.69 per hour, Anika
should consider using this facility for longer hours. For the
operating cost of the single-purpose facility, the range of
optimality is between $121.38 and $236.36 per hour.
The shadow price for the chai powder mix constraint is
about $0.43, which means that if the order amount for the
powder mix increases by one ounce, the total production cost
will increase by $0.43. This shadow price remains the same as
long as the order amount for the chai powder mix is not less
than 4,331.20 ounces. The shadow price for the tea
concentrate is about $0.25, which indicates that if the order
amount for the tea concentrate increases by one ounce, the
total production cost will increase by $0.25. The range of
feasibility for this shadow price is between 2,080.00 ounces
and 5,138.44 ounces.

page 593

The shadow price for the constraint on the operating hours

of the older facility is about −$2.30. Recall that unlike the other
constraints in the LP problem, this constraint (x1 ≤ 8) has a ≤
sign. The negative shadow price means that if Anika is able to
use the older facility for one additional hour of operation, the
total production cost will decrease by $2.30. This shadow price
remains unchanged as long as the facility does not operate
longer than 15.38 hours. To decide whether or not to operate
the older facility longer than normal, Anika would need to take
 into account the potential cost saving along with the risk of
manufacturing equipment breaking down from extended
operation.

EXERCISES 13.3
Mechanics
23. Consider the following LP problem.

Subject to:

where x1 and x2 represent the decision variables. Solve the LP problem to

answer the following questions.
a. What are the values of x1 and x2 at the optimal solution? What is the maximum
value of z?
b. Identify the binding and nonbinding constraints and report the slack value, as
applicable.
c. Report the shadow price and range of feasibility of each binding constraint.
d. What is the range of optimality for the two objective function coefficients?
24. Consider the following LP problem.

Subject to:
 where x1 and x2 represent the decision variables. Solve the LP problem to
answer the following questions.
a. What are the values of x1 and x2 at the optimal solution? What is the maximum
value of z?
b. Identify the binding and nonbinding constraints and report the slack value, as
applicable.
c. Report the shadow price and the range of feasibility of each binding constraint.
Interpret the results.
d. What is the range of optimality for the two objective function coefficients?
Interpret the results.
25. Consider the following LP problem.

Subject to:

where x1 and x2 represent the decision variables. Solve the LP problem to

answer the following questions.
a. What are the values of x1 and x2 at the optimal solution? What is the minimum
value of z?
b. Identify the binding and nonbinding constraints and report the surplus value, as
applicable.
c. Report the shadow price and range of feasibility of each binding constraint.
d. What is the range of optimality for the two objective function coefficients?
26. Consider the following LP problem.

Subject to:

where x1 and x2 represents the decision variables. Solve the LP problem to

answer the following questions.
 a. What are the values of x1 and x2 at the optimal solution? What is the minimum
value of z?
b. Identify the binding and nonbinding constraints and report the surplus value, as
applicable.
c. Report the shadow price and the range of feasibility of each binding
constraint(s). Interpret the results.
d. What is the range of optimality for the two objective function coefficients?
Interpret the results.

page 594
27. Consider the following LP problem.

Subject to:

where x1, x2, and x3 represent the decision variables. Solve the LP problem to
answer the following questions.
a. What are the values of x1, x2, and x3 at the optimal solution? What is the
maximum value of z?
b. Identify the binding and nonbinding constraints and report the slack value, as
appropriate.
c. Report the shadow price and range of feasibility of each binding constraint.
Interpret the results.
d. What is the range of optimality for the three objective function coefficients?
Interpret the results.
28. Consider the following LP problem.

Subject to:
 where x1, x2, and x3 represent the decision variables. Solve the LP problem to
answer the following questions.
a. What are the values of x1, x2, and x3 at the optimal solution? What is the
minimum value of z?
b. Identify the binding and nonbinding constraints and report the surplus value, as
appropriate.
c. Report the values and ranges of feasibility of the shadow price of each binding
constraint. Interpret the results.
d. What is the range of optimality for the three objective function coefficients?
Interpret the results.

Applications
29. Big Sur Taffy Company makes two types of candies: salt water taffy and special
home-recipe taffy. Big Sur wants to use a more quantitative approach to decide
how much salt water and special taffy to make each day. Molasses, honey, and
butter are the main ingredients that Big Sur uses to make taffy candies. For a
pound of salt water taffy, Big Sur uses 8 cups of molasses, 4 cups of honey, and
0.7 cup of butter, and the selling price is $7.50/lb. For a pound of special taffy, Big
Sur uses 6 cups of molasses, 6 cups of honey, and 0.3 cup of butter, and the
selling price is $9.25/lb. Taffy candies are made fresh at dawn each morning, and
Big Sur uses ingredients from a very exclusive supplier who delivers 400 cups of
molasses, 300 cups of honey, and 32 cups of butter once a day before sunrise.
a. Formulate and solve the LP model that maximizes revenue given the
constraints. What is the maximum revenue that Big Sur can generate? How
much salt water and special taffy does Big Sur make each day?
b. Identify the binding and nonbinding constraints and report the slack value, as
appropriate.
c. Report the shadow price and the range of feasibility of each binding constraint.
d. What is the range of optimality for the objective function coefficients?
30. A French vineyard in the Chablis region uses Chardonnay grapes to make
Chardonnay wine and Blanc de Blancs blended champagne. To produce 1 liter of
wine, the vineyard uses about 8 kilograms of Chardonnay grapes, and the
winemakers usually spend about 2.5 hours in the winemaking process, including
pressing, blending, and processing. To produce 1 liter of Blanc de Blancs
champagne, about 6 kilograms of Chardonnay grapes are used, with the
pressing, blending, and processing time of 3 hours. A liter of Chardonnay from the
vineyard is usually sold at $55, and each liter of champagne is sold at $45. Each
week, 400 kilograms of Chardonnay grapes and 150 hours for pressing, blending,
and processing are available.
a. Formulate and solve the LP model that maximizes revenue given the
constraints. What is the maximum revenue that the vineyard can make from
selling Chardonnay wine and champagne? How many liters of Chardonnay
wine and champagne should the vineyard make each week?
 b. If the amount of Chardonnay grapes available to the vineyard increases to 500
kilograms per week, how much wine and champagne should be made? Explain.
c. If the price of champagne drops to $30 per liter, how much wine and
champagne should the vineyard make? Explain and discuss your answer.
31. A consumer product company makes two types of dishwasher detergents: regular
and concentrate. The company has two manufacturing facilities for making
detergent products. The first facility has an operating cost of $120 an hour and
can produce 300 ounces of regular detergent products per hour and 220 ounces
of concentrate detergent per hour. The second facility has an hourly operating
cost of $220 and produces 350 ounces of regular detergent and 450 ounces of
concentrate detergent per hour. The company received wholesale orders totaling
4,500 ounces of regular detergent and 5,200 ounces of concentrate detergent.
a. Formulate and solve the LP model that minimizes costs given the constraints.
How many hours should the company operate each facility in order to fulfill the
orders and minimize the operating cost?
b. Identify each binding constraint and report the shadow price and the range of
feasibility.

page 595
c. If the hourly operating cost of the second facility is reduced to
$200, how many hours should the company operate each facility? What is the
minimum operating cost in this scenario?
32. Calcium and vitamin D are some of the most essential nutrients for bone health.
According to the Institute of Medicine, an average adult should have in his or her
daily diet 1,000 mg of calcium and 600 IU of vitamin D. These two nutrients are
found in milk and cold cereal, which many people eat for breakfast. One cup of
whole milk has 270 mg of calcium and 124 IU of vitamin D. A cup of a popular
whole grain cereal brand contains 150 mg of calcium and 120 IU of vitamin D.
One gallon of whole milk contains about 16 cups and costs $4.89. A box of the
popular cereal also contains about 16 cups and costs $3.19. Suppose an adult
relies on getting these two nutrients only from breakfast. Formulate and solve the
LP model that minimizes the cost of milk and cereal but satisfies the daily
requirement of calcium and vitamin D. How much milk and cereal should an adult
consume? What is the minimum cost?
33. A renowned chocolatier, Francesco Schröeder, makes three kinds of chocolate
confectionery: artisanal truffles, handcrafted chocolate nuggets, and premium
gourmet chocolate bars. He uses the highest quality of cacao butter, dairy cream,
and honey as the main ingredients. Francesco makes his chocolates each
morning, and they are usually sold out by the early afternoon. For a pound of
artisanal truffles, Francesco uses 1 cup of cacao butter, 1 cup of honey, and ½
cup of cream. The handcrafted nuggets are milk chocolate and take ½ cup of
cacao, ⅔ cup of honey, and 2/3 cup of cream for each pound. Each pound of the
chocolate bars uses 1 cup of cacao butter, ½ cup of honey, and ½ cup of cream.
One pound of truffles, nuggets, and chocolate bars can be purchased for $35,
$25, and $20, respectively. A local store places a daily order of 10 pounds of
chocolate nuggets, which means that Francesco needs to make at least 10
 pounds of the chocolate nuggets each day. Before sunrise each day, Francesco
receives a delivery of 50 cups of cocao butter, 50 cups of honey, and 30 cups of
dairy cream.
a. Formulate and solve the LP model that maximizes revenue given the
constraints. How much of each chocolate product should Francesco make each
morning? What is the maximum daily revenue that he can make?
b. Report the shadow price and the range of feasibility of each binding constraint.
c. If the local store increases the daily order to 25 pounds of chocolate nuggets,
how much of each product should Francesco make?
34. CaseInPoint is a start-up company that makes ultra-slim, protective cases in
stylish bold colors for a highly popular smartphone with two models: a large MÒR
model and a compact BEAG model. CaseInPoint uses thermoplastic
polyurethane, or TPU, to make all of its phone cases. The MÒR case is larger but
is mostly solid color in design. To produce a MÒR case, 30 grams of TPU are
used with 35 minutes of manufacturing time. The BEAG model is smaller but
more elaborate in design and requires 22 grams of TPU and 53 minutes of
manufacturing time. For each production period, 700 grams of TPU and 1,500
minutes of machine time are available. The profits from the MÒR and BEAG
cases are $9 and $7.50, respectively.
a. Formulate and solve the LP model that maximizes revenue given the
constraints. How many MÒR and BEAG cases should CaseInPoint make for
each production period?
b. If the amount of TPU available increases to 1,000 grams for each production
period, report the new LP solution and discuss your answer.
35. Many vitamin C products are made from ascorbic acid, which is derived from corn.
A small dietary supplement company makes two types of vitamin C supplements:
capsules and chewable tablets. To produce 1 kilogram of vitamin C capsules, 30
kilograms of corn, 2.3 hours of manufacturing time, and 2 hours of packaging time
are required. To produce 1 kilogram of vitamin C tablets, 40 kilograms of corn, 4.2
hours of manufacturing time, and 1 hour of packaging time are required. Each
week, the company has 2,000 kilograms of corn from its suppliers and allocates
180 hours of equipment time for manufacturing and 110 hours of packaging
equipment time for vitamin C supplements. One kilogram of vitamin C capsules
generates a profit of $9.50, and one kilogram of chewable tablets generates a
profit of $12.
a. Formulate and solve an LP model to maximize the profit contribution from the
two vitamin C products. How many kilograms of vitamin C capsules and tablets
should the company produce each week? What is the maximum profit the
company can make?
b. If the per-kilogram profits of vitamin C capsules and tablets increase to $11.50
and $16, respectively, how much of each product should the company make?
What is the corresponding profit amount?
36. A rancher has a six-year old pony that weighs about 180 pounds. A pony of this
age and size needs about 6.2 Mcal of digestible energy, 260 grams of protein, and
9,700 IU of vitamin A in her daily diet. EquineHealth and PonyEssentials are two
 popular brands of horse feed. One serving of the EquineHealth feed costs $10.50
and provides 1.5 Mcal, 52 grams of protein, and 1,800 IU of vitamin A. One
serving of the PonyEssentials feed costs $12 and provides 1.8 Mcal, 58 grams of
protein, and 2,200 IU of vitamin A. If the rancher were to blend the two feed
products to make horse feed for his pony, how much of each brand should he use
each week to minimize the cost? Explain and discuss your answer.

page 596

13.4 OPTIMIZATION WITH INTEGER

PROGRAMMING
LO 13.5
Formulate and solve a linear integer programming model.

Linear programming (LP) discussed in Section 13.3 finds an optimal

solution where the values of the decision variables are fractional. In
many cases, the assumption of fractional decision variables is
realistic. In Example 3.5, it was realistic for Yerba Buena Tea to
operate the older facility for 8 hours, the newer facility for 5.12 hours,
and the single-purpose facility for 3.34 hours. Similarly, it is realistic
to produce products in fractional quantities such as 20.5 gallons of
wine, 15.8 pounds of chocolate, etc.
In some other situations, it is critical that decision variables are
integers. Consider a capital budgeting problem where a
pharmaceutical company has to decide which of the several new
drug development projects it should fund. Here, each decision
variable represents whether or not the company should fund a
certain project; it takes a value of 1 if the company decides to fund
the project and 0 otherwise. The LP framework might find an
unrealistic value, such as 0.3, for a decision variable. In such
situations, we use linear integer programming, or simply integer
programming (IP), to provide integer solutions for the decision
variables.
We would like to point out that in certain situations we could
possibly use LP where we round optimal fractional values to the
nearest integer. In general, rounding optimal values to the nearest
integer leads to a suboptimal solution. However, if the objective
function coefficients are relatively small, and the values of the
decision variables are relatively large, rounding the decision
variables to the nearest integer would not have a significant impact
on the underlying business decision.
In this section, we discuss two common types of IP problems: (1)
a capital-budgeting problem for selection of investment projects and
(2) a transportation problem for selection of supply locations for
product delivery.

Capital Budgeting
In a typical capital-budgeting problem, a decision maker tries to
choose from a number of potential projects. Examples include
choosing among possible manufacturing plant locations or among
new drug development projects. In these situations, it does not make
sense to partially fund individual projects. There are no real benefits
from building just 45% of a manufacturing plant or partially
developing a new drug. In Example 13.8, we demonstrate how to
mathematically formulate a capital budgeting problem as an IP
model and use Excel’s Solver and R to find an optimal solution.

EXAMPLE 13.8
North Star Biotech is exploring six new drug development
projects. Each project has a five-year cash investment estimate
and an expected return on investment as shown in Table 13.2.
For example, to develop Drug 1, North Star has to make a cash
investment of $2 million in Year 1 and $1 million in Years 2, 3,
and 4, and no cash investment is needed in Year 5. Once
completed, the company expects a return of $12.5 million. To
develop Drug 2, the company needs to make a $2 million cash
investment in Years 1 and 2, and $1 million in Years 3, 4, and
5, and expects a return of $13.5 million.

TABLE 13.2 North Star Biotech’s New Drug Development Projects

Even though each drug development project has a positive

return, after accounting for the cash investment, North Star
Biotech cannot possibly fund all six projects because of
financial constraints. The amount of cash available to North
Star Biotech is $10 million in Year 1, $8 million in Year 2, $7
million in Year 3, $5 million in Year 4, and $3 million in Year 5.
Develop an integer programming model to help North Star
Biotech decide which of the six projects to fund in order to
maximize its total expected return.

page 597
SOLUTION:
Integer Programming Formulation
North Star needs to decide whether or not to fund each project;
therefore, each decision variable xi, where i = 1, 2, . . . , 6, has
binary values. For example, x1 = 1 if North Star invests in
developing Drug 1, and 0 otherwise. We also define ci as the
expected return on investment for Drug i. For example, the
expected return for Drug 1 is $12.5 million. For brevity, we
show the financial amounts in millions of dollars, and, hence, c1
= 12.5. We formulate the objective function as follows.

In addition, each project requires an annual cash

investment over a five-year period. We define aij as the cash
investment required for Drug i in Year j, where j = 1, 2, . . . , 5.
For example, the annual cash investments, in $ millions, for
Drug 1 are a11 = 2 for Year 1, a12 = 1 for Year 2, a13 = 1 for
Year 3, a14 = 1 for Year 4, and a15 = 0 for Year 5. The cash
investments for other projects are defined similarly. The
constraints for the integer programming formulation can be
written as follows.

The right-hand side of these constraints represent the amount

of cash North Star Biotech has available in each of the five
years. Finally, we specify that the decision variables must be
binary.
 xi = 0 or 1, where i = 1, 2, . . . , 6

We now show how to use Excel Solver and R to solve the

integer programming problem.
FILE
Biotech

Using Excel Solver

a. Open the Biotech worksheet and make sure the Solver add-in
is activated. Note that cells B10:G10 represent the decision
variables that will have a value of either 0 or 1 (i.e., whether or
not each drug development project is funded). To help verify
our Excel formula, we initially have 1 in cell B10 and 0
elsewhere.
page 598
b. Cells B2:G6 store the annual cash investments
required for each project. To calculate the total cash
investment spent in Year 1, enter =SUMPRODUCT(B2:G2,
B$10:G$10) in cell H2. Verify that the value in cell H2 is 2.
Copy and paste the formula to cells H3:H6.
c. Cells B8:G8 contain the expected returns from the six
projects. To calculate the total return, enter
=SUMPRODUCT(B8:G8, B10:G10) in cell I8 (the objective
function for the IP problem). Verify that the result in cell I8 is
12.5.
d. Navigate to the Data tab and click on Solver. Enter cell I8 in
the Set Objective box. Make sure to choose the Max option to
maximize the total return value in cell I8. See Figure 13.20 for
all solver parameters.

FIGURE 13.20 Solver parameters for Example 13.8

 Source: Microsoft Excel

e. Enter B10:G10 in the By Changing Variable Cells box. Recall

that these cells represent the binary decision variables.
f. Click the Add button to add the constraints for the amount of
cash available. In the Add Constraint dialog box, enter H2:H6
for Cell Reference, choose <= from the drop-down list of
operators, and enter I2:I6 for the Constraint option. Click Add.
g. In the next Add Constraint dialog box, enter B10:G10 in the
Cell Reference box and choose the bin option from the drop-
down list of operators to indicate that cells B10:G10 must have
binary values. Click OK.
h. Choose Simplex LP in the Select a Solving Method box. Click
Solve. Verify that your results are similar to Figure 13.21.

FIGURE 13.21 Results for the North Star Biotech problem

 Source: Microsoft Excel

page 599
Using R

a. If you have not already done so, install and load the lpSolve
package.
b. As in the previous examples, we specify the five objective
function coefficients, the parameters in the constraints, the ≤
signs, and the available annual cash amounts on the right-
hand side of the constraints. Enter:

c. We use the lp function with the max option. We also add the
all.bin = TRUE parameter to specify that the decision variables
are binary. (For other types of problems where the decision
variables can take on any integer values, we would use the
all.int = TRUE parameter instead.) Enter:
 > lp.output <- lp(“max”, lp.objective, lp.constraints,
lp.directions, lp.rhs, all.bin = TRUE)
d. Display the integer programming solution. Enter:
> lp.output
> lp.output$solution
Summary of the Results
Both the Excel Solver and R outputs indicate that North Star
Biotech should invest in Projects 2, 3, 5, and 6, and not in
Projects 1 and 4. Therefore, the company would make the
annual cash investments of $9, $5, $7, $5, and $2 million over
the five-year period. This investment plan results in $1 million
of excess cash in Year 1, $3 million in Year 2, and $1 million in
Year 5. Recall that for a maximization problem, these are called
slack in the linear and integer programing constraints. Based
on this recommendation, North Star Biotech would maximize its
expected total return at $57 million.

Transportation Problem
A classic example of a transportation problem involves a manager
who needs to decide which of the warehouses to use in order to
deliver products to retail stores or customers. The manager’s goal
(or the objective function) is usually to minimize the overall shipment
cost while meeting the demand at each retail store or from individual
customers. Intuitively, a transportation problem is feasible only when
the total supply (i.e., the combined capacity of all warehouses) is at
least equal to the total demand (i.e., the total amount being ordered).
In many situations, the transportation model deals with a large
bundle of goods, such as a pallet of 1,600 bottles of water or a
bundle of 300 2 × 4 wood planks. As such, it is impractical to
transport these bundles of goods in fractional units, and we use the
IP technique to solve a transportation model to obtain integer
solutions, instead of using the LP technique and rounding the
solutions to the nearest integers. Example 13.9 demonstrates how to
formulate an IP model for a simple transportation problem and use
Excel’s Solver and R to find an optimal solution.

page 600

EXAMPLE 13.9
Rainier Spring Water has two warehouses that service three
retail stores. Per week, the first warehouse can supply up to
160 pallets of bottled water and the second warehouse can
supply up to 155 pallets. The manager at Rainier Spring Water
receives a weekly order from the three retail stores for 85, 125,
and 100 pallets, respectively. Note that the combined capacity
of the two warehouses is larger than the total number of pallets
of bottled water ordered by the three stores, a required
condition for a feasible solution. The shipping cost for each
pallet varies among the warehouses and retail stores,
depending on the distance, traffic, and road conditions. As
shown in Table 13.3, to ship one pallet of bottled water from
Warehouse 1 to the three stores, it would cost $4.15, $5.95,
and $6.25, respectively. Likewise, to ship one pallet from
Warehouse 2 to the three stores, it would cost $3.75, $4.25,
and $8.25, respectively.

TABLE 13.3 Shipping Cost per Pallet of Bottled Water

Store 1 Store 2 Store 3

Warehouse 1 $4.15 $5.95 $6.25

Warehouse 2 $3.75 $4.25 $8.25

Develop an integer programming model to help the manager
decide how many pallets of bottled water to ship from each
warehouse to individual stores in order to fulfill the orders while
minimizing the total shipping cost.
SOLUTION:
Integer Programming Formulation
We require that the number of pallets that Rainier Spring Water
ships from its warehouse be an integer value. We define our
decision variables xij as the number of pallets to be shipped
from warehouse i to retail store j, where i = 1 and 2, and j = 1,
2, and 3. We assume xij to be non-negative integers. The per-
pallet shipping cost cij is defined as the cost of shipping one
pallet of bottled water from warehouse i to retail store j. For
example, as shown in Table 13.3, the values of c11, c12, and c13
are $4.15, $5.95, and $6.25, respectively. We formulate the
objective function as follows.

The constraints in the transportation problem reflect the

demand for bottled water from each store and the capacity of
each of the two warehouses. The first three constraints below
specify that the total number of pallets shipped to each retail
store must at least meet the order amount. Alternatively, we
can use an equal sign (=) to specify that the number of pallets
shipped to the store must be exactly the same as the amount
ordered. The next two constraints indicate that the number of
pallets shipped from each warehouse cannot exceed its
capacity.
 We now show how to use Excel Solver and R to solve the
transportation problem.
FILE
Rainier

page 601
Using Excel Solver

a. Open the Rainier worksheet and make sure the Solver add-in
is activated. Note that cells B5:D6 represent the decision
variables and will have only integer values (i.e., the number of
pallets of bottled water to ship from a warehouse to a retail
store). Initially these cells are blank.
b. We use cells E5 and E6 to show the total number of pallets
shipped from the two warehouses, respectively. Enter
=SUM(B5:D5) in cell E5. Copy and paste the formula to cell
E6.
c. We use cells B7:D7 to show the total number of pallets
shipped to the three stores, respectively. Enter =SUM(B5:B6)
in cell B7, and copy and paste it to cells C7 and D7.
d. We use cell F13 to calculate the total shipping cost (the
objective function), based on the per-unit cost in cells B13:D14
and the number of units shipped in cells B5:D6. Enter
=SUMPRODUCT(B13:D14, B5:D6) in cell F13.
e. Navigate to the Data tab and click on Solver. Enter cell F13 in
the Set Objective box. Make sure to choose the Min option to
minimize the total shipping cost. See Figure 13.22 for all
solver parameters.

FIGURE 13.22 Solver parameters for Example 13.9

Source: Microsoft Excel

f. Enter B5:D6 in the By Changing Variable Cells box. Recall

that these cells represent the integer decision variables.
g. Click the Add button to add the integer constraints for the
decision variables. In the Add Constraint dialog box, enter
B5:D6 in the Cell Reference box and choose the int option
from the drop-down list of operators to indicate that cells
B5:D6 must have integer values. Click Add.
h. In the next Add Constraint dialog box, enter E5:E6 for Cell
Reference, choose <= from the drop-down list of operators,
and enter F5:F6 for the Constraint option. This is to ensure
 that the number of pallets shipped from each warehouse does
not exceed the capacity of each warehouse. Click Add.
i. In the next Add Constraint dialog box, enter B7:D7 for Cell
Reference, choose >= from the drop-down list of operators,
and enter B8:D8 for the Constraint option. This is to ensure
that each retail store receives at least the ordered amount.
Click OK.
page 602
j. Choose Simplex LP in the Select a Solving
Method box. Click the Option button and make sure that the
Ignore Integer Constraints box is unchecked. Click OK.
k. Click Solve and verify that your results are similar to Figure
13.23.

FIGURE 13.23 Results for the Rainier Spring Water problem

Source: Microsoft Excel

Using R

a. If you have not already done so, install and load the lpSolve
package. The lpSolve package includes the lp.transport
function developed especially for solving transportation
problems.
b. We store the per-pallet shipping costs (shown in Table 13.3) in
an object called unit.costs, using the matrix function. Enter:
c. We specify the signs and order amount from the retail stores
for the three demand constraints. Enter:
> order.signs <- c(“>=”, “>=”, “>=”)
> order.amount <- c(85, 125, 100)
d. We specify the signs and the capacity of the two warehouses
for the two capacity constraints. Enter:
> capacity.signs <- c(“<=”, “<=”)
> capacity.limits <- c(160, 155)
e. As mentioned above, we use the lp.transport function to
obtain the optimal solution for transportation problems. By
default, the lp.transport function assumes that all decision
variables are integers. We use the min option to minimize the
total shipping cost and include the shipping cost matrix and all
the constraints. Enter:
> lp.output <- lp.transport(unit.costs, “min”, capacity.signs,
capacity.limits, order.signs, order.amount)
f. To display the optimal shipping cost and solution to the
transportation model, we use the following statements. Enter:
> lp.output
> lp.output$solution
Summary of the Results
The Excel Solver and R outputs indicate that from Warehouse
1, the manager at Rainier Spring Water should ship 55 pallets
of bottled water to Retail Store 1 and 100 pallets to Retail Store
3. From Warehouse 2, the manager should ship 30 pallets to
Retail Store 1 and 125 pallets to Retail Store 2. page 603
This shipment schedule would fulfill the orders
from all three stores. Warehouse 1 would deliver 155 out of the
 160 pallets available in its weekly inventory, whereas
Warehouse 2 would use up all of its weekly bottled water
inventory. The five remaining pallets in Warehouse 1 are
available for shipping to other stores. Recall that in a
minimization problem, the excess quantity at Warehouse 1 is
called a surplus. With this shipment plan, Rainier Spring Water
will incur the minimum overall shipment cost of $1,497 per
week.

EXERCISES 13.4
Mechanics
37. Consider the following IP problem.

Subject to:

where x1, x2, and x3 represent the decision variables. Solve the IP problem to
answer the following questions.
a. What are the values of x1, x2, and x3 at the optimal solution?
b. What is the maximum value of z?
38. Consider the following IP problem.

Subject to:
 where xij represent the decision variables. Solve the IP problem to answer the
following questions.
a. What are the values of xij at the optimal solution?
b. What is the minimum value of z?

Applications
39. An online apparel retailer runs regular marketing campaigns on social media
channels. The retailer is considering four social media marketing campaigns on
Facebook, Instagram, Pinterest, and Twitter for the four weeks prior to the
December holiday season. However, due to its limited marketing budget, the
retailer cannot run all four proposed campaigns. The following table describes the
weekly cost (in $1,000s) and the expected numbers of consumers each campaign
will reach (in 1,000s).

a. Formulate and solve the IP model to determine which of the four marketing
campaigns the retailer should run in order to maximize the number of
consumers reached, without going over its budget. What is the maximum
number of consumers that the retailer can reach?
b. Which social media channels should the retailer use in order to maximize the
number of consumers reached?
c. Will the retailer use up all of its marketing budget each week?
40. Akkadian Capital, a real estate investment firm, owns five apartment buildings
around a four-year college campus. The five buildings need significant repairs and
renovation, and each renovation project will take about three months to complete.
 The company expects each renovated apartment building to generate a
substantial rental income from college students. However, Akkadian Capital has a
limited budget during the next three months and cannot repair and remodel all five
buildings. The remodeling cost and the expected annual income after expenses
from each apartment are shown in the following table. Also included in the table is
Akkadian’s budget for the next three months.

a. Formulate and solve an IP model to help Akkadian determine which apartment

buildings to renovate in order to maximize its rental income,
while staying within its budget. What is the maximum amount
page 604
of rental income that Akkadian can generate?
b. Which apartment buildings should Akkadian remodel? Why?
c. Will Akkadian use up all of its budget in each month?
41. A microbrewery company in Portland, Oregon, operates two breweries that supply
its popular craft beer to five local restaurants. The first brewery can produce 56
barrels of craft beer per week, and the second brewery can produce 72 barrels
per week. The delivery cost per barrel and the average weekly order from each of
the five restaurants are summarized in the following table.

a. Formulate and solve an IP model to determine the number of barrels of craft

beer to be shipped from each brewery to the restaurants. What is the lowest
delivery cost that the microbrewery would incur?
b. How many barrels of craft beer should the microbrewery ship from each
brewery to each of the five restaurants in order to minimize the total delivery
cost?
42. An electric car company owns two facilities that manufacture and supply high-
capacity lithium-ion batteries to three vehicle assembly plants. The first battery
facility can produce and ship up to 10,000 lithium-ion batteries per month, and the
second facility can produce and ship up to 8,000 batteries per month. The per-unit
shipping cost and the battery demand from the three assembly plants are
provided in the following table.

Shipment Cost per Battery Plant 1 Plant 2 Plant 3

Manufacturer 1 $15.67 $12.30 $13.55

Manufacturer 2 $13.75 $13.90 $14.00

Monthly Demand 5,000 6,600 6,400

a. Formulate and solve an IP model to determine the number of batteries to be

shipped from each manufacturer to the three assembly plants. What is the
lowest shipping cost that the electric car company would incur?
b. How many batteries should the company ship from each manufacturer to each
of the three assembly plants in order to minimize the shipping cost?

13.5 WRITING WITH DATA

Case Study
FamVeld/Shutterstock

Demand for organic food has grown steadily over the years. Organic strawberry
farming is one of the segments in the food and agriculture industry that is projected
to continue expanding in the foreseeable future. Estuary Organic, a fruit farming
company on the coast of California, struggles to reduce its production cost while
staying true to its mission of offering fully certified organic produce to its customers.
To address the cost issue, Estuary Organic hires Professor Tom Richards as a
consultant to review different types of organic fertilizers that the company has used.
His task is to find a combination of organic fertilizers that can meet the three primary
nutrient requirements in the soil, nitrogen (N), phosphate (P2O5), and potassium
(K2O). At the same time, Professor Richards needs to take into consideration the
cost of certified organic fertilizers in his analysis.
For each strawberry planting season, growers need to have the right combination
of nitrogen, phosphate, and potassium in the soil. Adding too much or too little of
these nutrients is detrimental to the plant growth and its ability to bear fruit. This
combination of N, P2O5, and K2O nutrients is often called the NPK value. Generally,
the minimum amount of NPK per acre is 100 lbs, 50 lbs, and 50 lbs, respectively.
Based on the current soil condition and soil testing results, Professor Richards
also determines that the maximum amount of NPK per acre is 125 lbs, 55 lbs, and 55
lbs, respectively. Some of the fertilizers commonly used in the local area and by
Estuary Organic include alfalfa meal (typically made from fermented alfalfa), soybean
meal, fish meal (typically pollock, mackerel, and/or anchovies), and compost made
from animal manure. The cost and NPK composition of each fertilizer is listed in
Table 13.4.
TABLE 13.4 Organic Fertilizers, Their Cost, and Nutrient Contents

page 605
The goal of this consulting engagement is to provide Estuary Organic a
recommended combination of fertilizers that meet the NPK requirements at a minimal
cost. After analyzing relevant information, Professor Richards submits the following
report to the Estuary Organic owner.

Sample Report—Organic Strawberry Farm

In this report, we consider the cost and NPK information provided by Estuary
Organic Farm and develop an optimization model that represents the
relationship among the cost and NPK values of each type of organic fertilizer
as well as the NPK requirements for each acre of land. We use a linear
programming technique and develop the following formulation.

Subject to:
 The four decision variables, x1, x2, x3, and x4, represent the amount of
alfalfa meal, soybean meal, fish meal, and compost (in pounds) to be mixed
into the fertilizer compound. We also convert the cost of each fertilizer into a
per-pound basis (e.g., $0.28 per pound for the alfalfa meal, $0.46 per pound
for the soybean meal, and so on). The cost-per-pound values are used as the
objective function coefficients.
Based on the results obtained from the linear programming formulation,
we recommend the mix of organic fertilizers for each acre of land as shown in
Table 13.5.

TABLE 13.5 Recommended Mix of Fertilizers per Acre

Organic Recommended
Cost
Fertilizers Amount (lbs)

Alfalfa meal  750 $210.00

Soybean meal  875 $402.50

Fish meal   0 $ 0.00

Animal manure 2,250 $292.50

compost

Total cost: $905.00

Total NPK amount (in lbs): 100-55-50

The optimal combination provides 100 pounds of nitrogen, 55 pounds of

phosphate, and 50 pounds of potassium per acre. The base for the fertilizer
compound is the animal manure compost, which is the most inexpensive
fertilizer but has the least NPK values among the four fertilizer choices. We
then supplement the compost with alfalfa and soybean meals to reach the
required amount of nitrogen and potassium. Because we do not use any fish
meal, the most expensive of the four fertilizers, in the recommended fertilizer
mix, we are able to keep the cost per acre to a minimal amount of $905.
We also examine the impact of possible changes in the cost of fertilizers.
For example, if the cost of animal compost increases from $6.50 to $7.00 per
 50 pounds, we will have to reformulate our fertilizer mix.
In such a scenario, one possible option is to use only
page 606
alfalfa and soybean meals to achieve the NPK requirements while keeping
the cost reasonably low. Based on our initial analysis, we would recommend
using 1,578.95 pounds of alfalfa meal and 1,052.63 pounds of soybean meal,
which will result in a total cost of $926.31 per acre.
Another scenario that we consider in our analysis is if the cost of alfalfa
meal drops from $14 to $13 per 50 pounds (while the cost of compost
remains at $6.50 per 50 pounds). With its well-balanced NPK value, we could
use a much higher amount of alfalfa meal while keeping the total cost low.
Specifically, one possible fertilizer mix in this scenario would be to use 1,750
pounds of alfalfa meal, 875 pounds of soybean meal, and only 250 pounds of
animal manure compost. This would result in a total cost of $890 per acre.

Suggested Case Studies

Prescriptive analytics can be applied to many situations. Here are some suggestions
and examples of Monte Carlo simulations and linear programming problems.

Report 13.1. A grocery store manager wants to analyze customer spending data by
product categories: fresh baked goods, meat and dairy, produce, and frozen food.
For each shopping trip, about 30% of shoppers purchase baked good items, and the
spending in this category tends to follow a continuous uniform distribution between
$3 and $19. For the meat and dairy products, 70% of shoppers make regular
purchases from this category, and their spending is normally distributed with a mean
of $21 and standard deviation of $5.27. Eighty percent of the shoppers spend an
average of $15 on produce, and their spending follows a normal distribution with a
standard deviation of $2.31. Sixty-five percent of shoppers purchase at least an item
from the frozen food aisles; the spending amount in this category follows a uniform
distribution between $7.25 and $28.50. On average, approximately 220 customers
make a trip to the grocery store each day. Develop a Monte Carlo simulation to
analyze the customer spending and the fluctuation in daily revenue.

Report 13.2. An oil and gas company has two refineries that produce light and heavy
crude oil. The first refinery can produce 500 barrels of light crude oil and 300 barrels
of heavy crude oil per day, and the second refinery can produce 600 barrels of light
crude oil and 450 barrels of heavy crude oil per day. The daily operating cost of each
refinery is $15,000 and $20,000, respectively. The company has to fulfill production
orders totaling 3,200 barrels of light crude oil and 2,100 barrels of heavy crude oil.
Analyze the information and recommend an appropriate production plan.

Report 13.3. South Bay Candles makes two types of scented candles: 10-inch pillars
and 2-inch decorative gel. The production requirements for each product and
available resources are shown in Table 13.6. The sales price for each pillar and gel
candle is $2.15 and $3.55, respectively. Analyze the data provided and create a
report to recommend how many pillar and gel candles South Bay should make.
TABLE 13.6 Product Information for South Bay Candles

Required Materials for Each Candle

Type
Materials/Resources
Amount
Pillar Gel
Available

Cotton wick (ft.)  12,000 0.580 0.800

Beeswax (lb.)  7,000 0.500 0.325

Essential oils  3,200 0.240 0.130

Manufacturing time 120,000 5.320 9.460

(min.)

page 607
 page 608

APPENDIX A

Big Data Sets: Variable

Description and Data Dictionary
A distinctive feature of the text is the access to select big data sets
with relevance to numerous applications to which students can
relate. Although the data sets represent a highly simplified (modified)
portion of the actual data, they still retain several features of big
data. Throughout the text, we use these big data sets to help
introduce problems, formulate possible solutions, and communicate
the findings, based on the concepts introduced in the chapters. Data
dictionaries for the six big data sets used in the text are described
below.

Data 1: Car Crash Data

FILE
Car_Crash

The Car_Crash data are extracted from the Statewide Integrated

Traffic Records System database that serves as a means to collect
and process data gathered from a collision scene in California
(http://www.chp.ca.gov). The data are based on the information
provided to the California Highway Patrol (CHP) from both local and
government agencies. CHP collected data regarding the actual
collision, the parties involved, as well as information on the victims.
The data set represents a highly simplified/modified sample of about
113,000 collisions within the span of a single calendar year.

Table A.1 Data Dictionary for Car Crash Data

Variable
Description or Possible Values
Name

ID A modified ID for each motor vehicle accident

County Actual county name

City Actual city name

Weekday 1 – Monday
2 – Tuesday
3 – Wednesday
4 – Thursday
5 – Friday
6 – Saturday
7 – Sunday

Severity 1 – Fatal or severe injury

0 – Others
Variable
Description or Possible Values
Name

ViolCat 01 – Driving or Bicycling Under the Influence of

Alcohol or Drug
02 – Unsafe Speed
03 – Following Too Closely
04 – Improper Passing
05 – Unsafe Lane Change
06 – Improper Turning
07 – Automobile Right of Way
08 – Pedestrian Right of Way
09 – Pedestrian Violation
10 – Traffic Signals and Signs
11 – Fell Asleep

ClearWeather 1 – Clear weather

0 – Not clear weather
Variable
Description or Possible Values
Name

Month 1 – January
2 – February
3 – March
4 – April
5 – May
6 – June
7 – July
8 – August
9 – September
10 – October
11 – November
12 – December

CrashType A – Head-On
B – Sideswipe
C – Rear End
D – Broadside
E – Hit Object
F – Overturned
G – Vehicle/Pedestrian

Highway 1 – Highway
0 – Not highway

Daylight 1 – Daylight
0 – Not daylight
 page 609

Data 2: College Admissions Data

FILE
College_Admissions

The College_Admissions data contain important college admission

and enrollment information related to a selective four-year university
in North America. The data shed light on admission and enrollment
decisions made by the admissions office and the applicants. Based
on the admissions dashboard, there are about 18,000 applicants for
fall admissions in three colleges within the university. Also included
in the data is the college GPA of enrolled applicants. The data
represent a highly simplified/modified sample of anonymized
applicants.

Table A.2 Data Dictionary for College Admissions Data

Variable
Description or Possible Values
Name

Applicant A modified ID for each applicant

Edu_Parent1 1 – No High School

2 – Some High School
3 – High School Graduate
4 – Some College
Edu_Parent2
6 – 4-Year College Graduate
7 – Postgraduate

Gender M – Male
F – Female
Variable
Description or Possible Values
Name

White 1 – White
0 – Not White

Asian 1 – Asian
0 – Not Asian

HSGPA High school weighted GPA, ranging from 0 to 5.

SAT/ACT The higher of the SAT/ACT score, where the ACT

score is first converted into the equivalent SAT
score for English and Math

College Arts & Letters

Business & Economics
Math & Science

Admitted 1 – Admitted by College

0 – Not Admitted by College

Enrolled 1 – Applicant enrolled

0 – Applicant not enrolled

College_GPA College GPA, ranging from 0 to 4, four years after

enrollment; blanks for those who did not enroll

page 610

Data 3: House Price Data

FILE
House_Price

Investment in college town real estate is attractive because students

offer a steady stream of rental demand as cash-strapped public
universities are unable to house their students beyond freshman
year. Data on about 11,000 homes sold in a single year are extracted
from Zillow for 50 campus towns in the U.S. (https://www.zillow.com).
The House_Price data represent a highly simplified/modified sample
of anonymized homes, excluding homes with observations that seem
erroneous and/or not pertaining to sales.

Table A.3 Data Dictionary for House Price Data

Variable Name Description or Possible Values

Record A modified ID for each house

Sale_amount Sale price of the house in U.S. dollars

Sale_date Sale date of the house

Beds Number of bedrooms in the house

Baths Number of bathrooms in the house

Sqft_home Square footage of the house

Sqft_lot Square footage of the lot

Type Multiple Family

Multiple Occupancy
Single Family
 Variable Name Description or Possible Values

Build_year Year the house was built

Town Name of the campus town

University Name of the university

Data 4: Longitudinal Survey Data

FILE
Longitudinal_Survey

The National Longitudinal Survey (NLS) follows over 12,000

individuals in the United States. Starting in 1979, individuals were
surveyed annually up until 1994; thereafter, surveys were conducted
every other year. When first surveyed in 1979, individuals ranged in
age from 14 to 21 years old. Survey questions focused on “labor
force activity;” however, the survey also covered a large variety of
subjects including, “educational attainment, training investments,
income and assets, health conditions, workplace injuries, insurance
coverage, alcohol and substance abuse, sexual activity, and marital
and fertility histories.” The Longitudinal_Survey data represent a
highly simplified/modified sample of anonymized respondents.

page 611

Table A.4 Data Dictionary for Longitudinal Survey Data

Variable Name Description or Possible Values

ID A modified ID for each respondent

Variable Name Description or Possible Values

Age Age, as of 1979

Urban At the age of 14, respondent lived in:

1 – Urban area
0 – Non-urban area

Mother_ Edu Number of years of mother’s education

Father_ Edu Number of years of father’s education

Siblings Number of siblings in 1979

Black 1 – Black
0 – Non-Black

Hispanic 1 – Hispanic
0 – Non-Hispanic

White 1 – White
0 – Non-White

Christian As of 1979,
1 – Christian
0 – Non-Christian

WomenPlace Opinion on: “A woman’s place is in the home,

not in the office or shop.”
1 – Agree
0 – Disagree
Variable Name Description or Possible Values

Male 1 – Male
0 – Female

FamilySize Family size in 1979

Self_Esteem The Rosenberg Self-Esteem Scale measures

individuals’ self-evaluation. The overall score
can range between zero and 30 points, where
higher scores indicate higher self-esteem.
Actual scores in 1980.

Height Height in inches in 1981

Weight Weight in pounds in 1981

Outgoing_Kid At the age of 6, the respondent was:

1 – Outgoing
0 – Shy

Outgoing_Adult In adulthood, the respondent was:

1 – Outgoing
0 – Shy

HealthPlan Covered by health insurance/health care plan

in 2000:
1 – Yes
0 – No

Income Income in 2000

 Variable Name Description or Possible Values

Marital_Status Marital status in 2000:

0 – Never Married
1 – Married
2 – Separated
3 – Divorced
4 – Remarried
5 – Widowed

Education Number of years of education as of 2000

WeeksEmployed Number of weeks employed in past calendar

year in 2000

NumberSpouses Number of spouses/partners reported in 2000

page 612

Data 5: NBA Data

FILE
NBA

In this text, we use a highly simplified/modified sample of the

National Basketball Association (NBA) data collected for 2016. This
information is valuable to sports enthusiasts who use data on
players’ statistics and other sports-related information to inform
decisions. The data represent a highly simplified/modified sample of
30 competing teams and 457 players. The player statistics are for
several seasons as well as for their career. The NBA data were
collected from the following sources:
 http://stats.nba.com/help/glossary/
http://www.basketball-tips-and-training.com/basketball-
statistics.html
http://www.sportingcharts.com/NBA/dictionary/
http://espn.go.com/nba/dailyleaders

Table A.5 Data Dictionary for NBA Data

Variable Name Description or Possible Values

Player A unique ID for each NBA player

Name Player’s name

Position C – Center
PF – Power Forward
SF – Small Forward
PG – Point Guard
SG – Shooting Guard

Age Player’s age in 2015–2016

Height Player’s height in inches as of 2015–2016

Weight Player’s weight in pounds as of 2015–2016

Salary Player’s annual salary in 2015–2016

Season 1999–2000 to 2015–2016

Variable Name Description or Possible Values

Postseason TRUE – Playoff

FALSE – Regular season

Team Player’s team in 2015–2016

Games_played Total number of games played

Games_started Number of games played as a starter

Minutes Average minutes per game

FG_made Total number of baskets made

FG_attempted Total number of baskets attempted

FG_percent Field goal percentage

3P_made Total number of 3 pointers made

3P_attempted Total number of 3 pointers attempted

3P_percent Three pointer percentage

FT_made Free throws made

FT_attempted Free throws attempted

FT_percent Free throw percentage

 Variable Name Description or Possible Values

Rebounds_off Total times player caught basketball after

missed shot when playing offense

Rebounds_def Total times player caught basketball after

missed shot when playing defense

Assists Total times player passed to another player who

then scored a basket

Blocks Total times player stopped the opponent from

taking a shot (the ball must “be on its way up” to
count)

Steals Total times player gained legal control of the

basketball from opponent

Fouls Total times player made illegal physical contact

with an opponent

Turnovers Total times player lost possession of basketball

to opponent

Points Average points per game

page 613

Data 6: Tech Sales Reps Data

FILE
TechSales_Reps
Analytics can help managers and organizations make well-informed
decisions about their workforce. Instead of making personnel
decisions based on perception and instinct, managers can use data
to help identify, attract, and retain employees. The TechSales_Reps
data contain records on 21,990 sales representatives from the
hardware and software product groups of a high-tech company. For
each employee, the data include socio-demographic and education
information, salary, sales performance, and a personality indicator.
Also included in the data is the net promoter score, which is an
indicator of customer satisfaction with each sales rep. In this text, the
TechSales_Reps data represent a highly simplified/modified
sample.

Table A.6 Data Dictionary for Tech Sales Reps Data

Variable
Description or Possible Values
Name

Rep A unique ID for each sales representative

Business One of the two product groups: Hardware and

Software

Age Employee’s actual age

Female 1 – female
0 – otherwise

Years The number of years the employee has been

employed at the company

College Whether or not the employee has a four-year

college degree (Yes/No)
Variable
Description or Possible Values
Name

Personality Analyst: This personality type exemplifies rationality.

Analysts tend to be open-minded and strong-willed.
They like to work independently and usually
approach things from a very practical perspective.
Diplomat: Diplomats care about people and tend to
have a lot of empathy toward others. They exemplify
cooperation and diplomacy.
Explorer: Explorers are highly practical and can
think on their feet. They tend to be very good at
making quick, rational decisions in difficult
situations.
Sentinel: Sentinels are cooperative and practical.
They like stability, order, and security. People with
this personality type tend to be hardworking and
meticulous.

Certificates The number of relevant professional certifications

each employee has earned

Feedback The average feedback score that each employee

receives from his or her peers and supervisor on the
360-degree annual evaluation. The possible scores
range from 0 (lowest) to 4 (highest).

Salary Annual base salary of each employee

NPS The net promoter score (NPS) is a key indicator of

customer satisfaction and loyalty.
 page 614

APPENDIX B

Getting Started with Excel and

Excel Add-Ins
Microsoft Excel
Microsoft Excel is arguably the most widely used computer
application among business professionals. Accountants, economists,
financial analysts, marketers, HR managers, and many others use
Excel spreadsheets for everyday business tasks. Oftentimes, these
tasks involve entering, editing, and formatting data as well as
performing data analysis. We rely heavily on Excel and Excel add-ins
in virtually every chapter of this text. Analytic Solver is the most
relevant Excel add-in that offers a powerful tool for business analytics.
In this appendix, we summarize Excel formulas, references, and
functions that are commonly used for calculations and data analysis.
We also show how to build a simple spreadsheet model for solving a
business problem. Finally, we provide a brief description of Analytic
Solver and other add-ins at the end of the appendix; detailed
instructions on the use of these add-ins are included in the relevant
chapters of the text.

Formulas
In Excel, we use formulas to perform basic calculations. When we
enter a formula in a cell, Excel carries out the specified calculation
and returns the result in the same cell. We also use formulas to
manipulate the cell content such as rounding a number. A formula in
Excel always starts with an equal sign (=) and usually includes cell
addresses. A cell address or cell reference consists of a column
name and a row number. For example, cell reference A1 refers to the
top and leftmost cell in column A and row 1. Basic calculations can be
performed using arithmetic operations such as addition ( + ),
subtraction ( − ), multiplication ( * ), division ( / ), and exponentiation (
^ ). For example, we select an empty cell and use the formula
=A1+B1+C1 to add values from cells A1, B1, and C1, and the formula
=A1^2 to square the value in cell A1.

Relative, Absolute, and Mixed References

There are three types of cell references: relative, absolute, and
mixed. The three cell reference types behave differently when copied
elsewhere on the worksheet. By default, cell addresses in a formula,
such as =B1+C1, are relative references and will change when a
formula is copied to another cell. For example, if we enter the formula
=B1+C1 in cell A1 and copy it to cell D4, the formula in cell D4 will
appear as =E4+F4. In other words, references to cells B1 and C1 are
relative to where the formula is placed on the worksheet. We use
relative references when we want to repeat similar calculations while
allowing column names and row numbers to change.
Absolute references allow us to maintain the original cell
references when a formula is copied elsewhere. We specify absolute
references by adding a dollar sign ($) in front of the column name and
row number (e.g., $B$1). From the previous example, if we enter the
formula =$B$1+$C$1 in cell A1 and copy it to cell D4, the formula in
cell D4 will remain =$B$1+$C$1. Unlike relative references, absolute
references remain unchanged when copied elsewhere on the
worksheet.
We use mixed references by adding a dollar sign ($) in front of
either the column name or the row number (e.g., $B1 or B$1), but not
both. This will keep the reference to the specific column name or the
row number constant. From the previous example, if we enter the
formula =$B1+$C1 in cell A1 and copy it to cell D4, the formula in cell
D4 will become =$B4+$C4. Similarly, if we enter the formula
=B$1+C$1 in cell A1 and copy it to cell D4, the formula in cell D4 will
become =E$1+F$1.

page 615

In Table B.1, we summarize the different results when using

relative, absolute, and mixed references with the same formula.

Table B.1 Relative, Absolute, and Mixed References

Formula in Cell Formula in Cell A1 Copied to

Reference
A1 Cell D4

Relative =B1+C1 =E4+F4

Absolute =$B$1+$C$1 =$B$1+$C$1

Mixed =$B1+$C1 =$B4+$C4

Mixed =B$1+C$1 =E$1+F$1

Functions
Functions in Excel are predefined formulas. Like a formula, a function
always begins with an equal sign (=) and must be written with the
correct syntax enclosed within parentheses. Most functions require at
least one argument. Arguments are the values that Excel uses to
perform calculations. For example, the COUNT function is used to
count the number of cells that contain numerical values and has the
syntax =COUNT(A1:A10), where A1:A10 is the argument indicating
the array of cells to be counted. Table B.2 provides a summary of
some of the most basic descriptive functions in Excel. The notation
array in the function’s argument in the table specifies the range of cell
addresses to be included in the calculation.

Important: Due to different fonts and type settings, copying and

pasting formulas and functions from this text directly into Excel may
cause errors. When such errors occur, you may need to replace
special characters such as quotation marks and parentheses or
delete extra spaces in the functions.

Table B.2 Basic Descriptive Functions in Excel

Function and Syntax Description Example

=COUNT(array) Returns the =COUNT(A1:A10)

number of
cells in the
array with
numerical
values.

=COUNTA(array) Returns the =COUNTA(A1:A10)

number of
cells in the
array that are
not blank.

=COUNTBLANK(array) Returns the =COUNTBLANK(A1:A10)

number of
cells in the
array that are
blank.
Function and Syntax Description Example

=COUNTIF(array, Returns the =COUNTIF(A1:A10,

criteria) number of “>10”)
cells in the
array that
meet a
specific
selection
criterion.

=IF(logical statement, Returns a =IF(A1=“Yes”, 1, 0). If A1

result if the statement result based = “Yes”, returns a 1. If
is true, result if the on the not, returns a 0. See also
statement is false) outcome of Example B.1
the logical
statement.

=SUM(array) Adds and =SUM(A1:A10)

returns the
sum of the
numbers in
the array.

=VLOOKUP(lookup Searches See Example B.1

value, reference table, and retrieves
column number in the information
reference table from a
containing results) specified
column in a
reference
table.
Table B.3 provides a summary of some basic date and time functions
in Excel.

page 616

Table B.3 Basic Date and Time Functions in Excel

Function and
Description Example
Syntax

=DAY(date Returns the day of =DAY(“3/15/2019”)

value) the month as a Excel returns 15.
number between 1
and 31 given the
date value in the
argument.

=MONTH(date Returns the month =MONTH(“3/15/2019”)

value) of the year as a Excel returns 3.
number between 1
and 12 given the
date value in the
argument.

=NOW( ) Returns the current =NOW( )

date and time. The
NOW function
does not require
any argument.
 Function and
Description Example
Syntax

=TODAY( ) Returns the current =TODAY( )

date. The TODAY
function does not
require any
argument.

=WEEKDAY(date Returns the day of =WEEKDAY(“3/15/2019”)

value) the week as a Excel returns 6, meaning
number between 1 that March 15, 2019, was
(Sunday) and 7 a Friday.
(Saturday) given
the date value in
the argument.

=YEAR(date Returns the year =YEAR(“3/15/2019”)

value) portion of the data Excel returns 2019.
value in the
argument.

Table B.4 provides a summary of some basic statistical functions

in Excel. Other functions are introduced and discussed in the relevant
chapters of the text.

Table B.4 Basic Statistical functions

Function and Syntax Description Example

Function and Syntax Description Example

=AVERAGE(array) Returns the =AVERAGE(A1:A10)

arithmetic
mean of the
array.

=MAX(array) Returns the =MAX(A1:A10)

largest
number in
the array.

=MEDIAN(array) Returns the =MEDIAN(A1:A10)

median of
the array.

=MIN(array) Returns the =MIN(A1:A10)

smallest
number in
the array.

=RAND( ) Returns a =RAND( )

uniformly
distributed
random
number that
is greater
than or
equal to 0
and less
than 1.
 Function and Syntax Description Example

=RANDBETWEEN(lowest Returns a =RANDBETWEEN(1,10)

value, highest value) uniformly Excel returns a random
distributed number between 1 and
random 10.
number
between the
specified
lowest and
highest
values,
inclusive.

=ROUND(value, decimal Rounds the =ROUND(2.6589,2)

digit) numerical Excel returns 2.66.
value to the
specified
decimal
digit.

=STDEV.S(array) Estimates =STDEV.S(A1:A10)

and returns
the sample
standard
deviation of
the array.

page 617

Developing a Spreadsheet Model

While using formulas, references, and functions involves an
understanding of Excel syntax, crafting a well-structured spreadsheet
requires critical thinking and creativity. We usually build a spreadsheet
model to represent simplified reality with the goals of solving a
business problem, performing data analysis, and gaining insightful
information. Ideally, a spreadsheet model should allow users to easily
experiment with different assumptions, evaluate and compare
different scenarios, and, more importantly, improve decision making.
Spreadsheet modeling is more an art than a science, and it takes
experience to hone the craft. A good spreadsheet model empowers
both model developers and model users with the knowledge to
improve its effectiveness.
In this appendix, we offer a few general guidelines for developing
a spreadsheet model. Students are encouraged to find additional
resources on spreadsheet modeling and to design their own
spreadsheet whenever possible. When constructing a spreadsheet
model, we consider the following general principles.
The spreadsheet model should be well-organized. Similar
information should be grouped together, be clearly labeled, and use
the same format.
The spreadsheet model should allow users to experiment with
different scenarios based on input parameters, which are stored in
individual cells and separately from the data.
When entering a formula, careful consideration should be given to
the choice of absolute, relative, and mixed references so that the
formula can be copied to other cells that require a similar
calculation.
Output, including charts, should be clearly labeled and appropriately
formatted. Users should be able to easily identify and understand
the results.
Example B.1 demonstrates how to construct a simple spreadsheet
model that follows these guidelines. In Chapter 13, we follow these
guidelines to implement Monte Carlo simulations and linear
programming analysis in Excel.

EXAMPLE B.1
FILE
Consulting

Sarah Washington, a manager at Iniesta Consulting Group, oversees

a team of business consultants. Each of her team members has a
highly specialized expertise that is crucial to the success of her
consulting group. Sarah wants to examine the after-tax earnings of
her team members to ensure that their salary remains competitive
compared to the salaries offered by competing firms. Iniesta offers
each employee medical benefits, but each employee is required to
pay a co-premium (a portion of the medical insurance cost). The
annual co-premium amount differs between a management employee
($1,800) and a non-management employee ($1,250). Even though
Iniesta’s medical benefits are much better than those offered by other
employers, the co-premium reduces an employee’s net earnings.
Sarah collects the salary information of her team members and
income tax rates in an Excel spreadsheet, shown in Figure B.1. She
includes her own information in the table in order to verify the
accuracy of her results. Develop a spreadsheet model to help Sarah
analyze the salary, income tax, co-premium, and net earnings of each
employee.

Developing a Spreadsheet Model in Excel

Sarah’s spreadsheet organizes the five employees into a table with
each column containing similar information. Each row is a simplified
model representing an employee. All columns are clearly labeled with
column headings and appropriately formatted (e.g., a dollar sign and
no decimal places for the salary column). In addition, the input
parameters, such as the medical co-premium and income tax rates,
are stored in separate cells below the data. This way, if Sarah wants
to experiment with different scenarios (e.g., reducing the medical co-
premium cost for non-management employees), she can simply
change the value in these individual cells.

page 618
FIGURE B.1 Salary and Tax Information of Employees

Source: Microsoft Excel

In order to perform relevant calculations for the spreadsheet

model, we follow these steps.
. Enter column headings Co-premium in cell F3, Tax amount in cell
G3, Net earnings in cell H3, and Net earnings as percent of salary in
cell I3. Format the column headings and cells in these columns
similar to the first five columns. We will use columns F, G, H, and I to
perform calculations and display analysis results.
. Enter the formula =IF(D4=”Management”, B$12, C$12) in cell F4.
We use the IF statement to evaluate whether the employee has a
management-level position. If true, the co-premium is $1,800 (from
cell B12), otherwise the co-premium is $1,250 (from cell C12). Verify
that the output in cell F4 is $1,800.00.
Copy and paste the formula to cells F5 through F8. In the formula
entered in cell F4, we used a relative reference (without dollar signs)
for cell D4 and mixed references for cell B12 and C12 with a dollar
sign in front of the row numbers. The job classification varies from
one employee to another, and therefore we use the relative
 reference for cell D4. When we copy the IF function to cells F5
through F8, the relative reference in the formula changes from D4 to
D5 through D8, respectively. The co-premium amounts for
management and non-management positions in cells B12 and C12
are applicable to all employees, and, therefore, we use mixed
references to ensure that the two cell addresses remain unchanged
for all formulas in column F. In other words, all formulas in column F
refer to cells B12 and C12. It is important to note that absolute
references (e.g., $B$12 and $C$12) would yield the correct result in
the current example. This, however, is not necessary because we
are not changing columns when copying and pasting the IF
statement.
. Enter the formula =VLOOKUP(E4,A$16:B$20,2,TRUE)*E4 in cell G4
to retrieve the tax rate for the first employee. The VLOOKUP function
in the formula includes three required arguments and one optional
argument:
1. The lookup value (i.e., the salary in cell E4 in this case),
2. A lookup or reference table (ranges of salaries and corresponding
tax rates in cells A16:B20, without the column headings),
3. A column number in the reference table that contains the output (in
the current example, the tax rates are in the second column in the
reference table), and
4. Whether we want to look up a value within a range (TRUE) or find
an exact match (FALSE). The last parameter is optional; if omitted,
the default value is TRUE.
In the reference table, if the salary is at least $60,000 and less than
$70,000, the tax rate is 19%. If the salary is at least $70,000 and less
than $80,000, the tax rate is 22%, and so forth. For a page 619
salary of $100,000 and above, the tax rate is 27%.
These criteria in the reference table are applicable to all five
employees. Verify that the tax amount for the first employee is
$21,371.04 (= 24% × 89,046).
Copy and paste the formula to cells G5 through G8. Because we
use mixed references for the reference table in cells A$16 through
 B$20, these cell addresses remain unchanged when copied from cell
G4 to cells G5 through G8.
. Enter the formula =E4-G4-F4 in cell H4. Copy and paste the formula
to cells H5 through H8. Because we use relative references for these
three cells, when we copy the formula from cell H4 to cells H5
through H8, the three cell addresses change to the corresponding
row number. For example, the formulas in cell H5 and H8 should
appear as =E5-G5-F5 and =E8-G8-F8, respectively. Format cells H4
through F8 with the Accounting Number format by highlighting the
cells and clicking on the Accounting Number format button (with a
dollar sign icon) in the Number group on the Home tab. Verify that
the net earnings of the first employee is $65,874.96.
. Enter the formula =H4/E4 in cell I4. Again, we use relative
references here for the same reason as in step D. Format the value
in cell I4 as a percentage by clicking the Percent (or Percentage)
Style button in the Number group on the Home tab. Copy and paste
the formula from cell I4 to cells I5 through I8. Verify that the net
earnings as a percent of salary for the first employee is 73.98%.
. We calculate the average net earnings of the five employees by
using the AVERAGE function. Enter =AVERAGE(I4:I8) in cell I9.
Format cell I9 as a percentage. Verify that the average net earnings
is 73.64%. Enter Average in cell H9 as a label. Your completed
spreadsheet should look similar to Figure B.2.

FIGURE B.2 Spreadsheet Model for Iniesta Consulting Group

Source: Microsoft Excel

In the spreadsheet model in Figure B.2, Sarah can easily

experiment with a different co-premium amount by changing the
values in cells B12 and C12, and the results in columns F through I
will be updated without requiring modifications to the formulas in
those columns. And because the formulas do not need to be
rewritten, there will be fewer chances of Sarah making mistakes when
evaluating different scenarios.
Moreover, recall that we have to write the formulas only for the first
employee in row 4 and simply copy them to other cells in rows 5
through 8. This is possible because we carefully use mixed and
relative references. This way, formulas are used uniformly for cells
that require a similar calculation, resulting in fewer potential mistakes
and errors.
We also show the calculation results in columns F through I
separately from the data in columns A through E. Sarah can easily
see the net earnings of each employee in column H. By using the
AVERAGE function to summarize the results, she can also conclude
that, on average, the consultants in her group take home
approximately three-quarters (i.e., 73.64%) of their salary.

page 620
Analytic Solver
Analytic Solver (formerly known as XLMiner) is an add-in software
that runs on Microsoft Excel. It offers a comprehensive suite of
statistics and data mining tools, including data wrangling, data
partitioning, and supervised and unsupervised data mining methods.

The Interface
Once Analytic Solver is installed, launch Excel and verify that you can
see the Analytic Solver, Data Mining, and Solver Home tabs. In this
text, we focus primarily on the features in the Data Mining tab. Figure
B.3 shows the features on the Data Mining tab. Examples and
exercise problems in this text are developed based on Analytic Solver
2019. Other versions of Analytic Solver may have a different user
interface and display the output in a different format.

FIGURE B.3 Data Mining Tab in Analytic Solver 2019 for Excel

Source: Microsoft Excel

As shown in Figure B.3, there are six groups of features in the

Data Mining tab. The Data Analysis and Data Mining groups are used
extensively throughout this text. In the Data Analysis group, the
Transform feature contains functions for handling missing data and
transforming continuous and categorical data (see Chapter 2). The
Transform feature also performs principal component analysis (or
PCA), a widely used technique for data reduction (see Section 8.3).
Under the Cluster feature, k-means and hierarchical clustering
functions are part of the unsupervised data mining techniques (see
Chapter 11).
 In the Data Mining group, we use the Partition, Classify, and
Predict features when we examine supervised data mining techniques
(see Chapters 8, 9, and 10). We also use the Associate feature when
we discuss Association Rules (see Section 11.3). Detailed
instructions and examples on how to use these functions in Analytic
Solver are presented in the relevant chapters throughout the text.

Other Excel Add-Ins

In addition to Analytic Solver, there are many other add-in software in
Excel. In this text, we use two additional add-ins that come
preinstalled with Excel. The Analysis ToolPak add-in is used for
statistical analysis in Chapters 3 through 7 and for random number
generation in Chapter 13. The Solver Add-In, developed by the same
company that offers Analytic Solver, is used for linear and integer
programming discussed in Chapter 13. The following instructions
activate these two Excel add-ins.

Activating Analysis ToolPak and Solver Add-

Ins
. In Excel, go to File > Options > Add-Ins.
. In the Manage Excel Add-ins section (toward the bottom of the
screen), click Go . . . .
. On the Add-ins dialog box, check the Analysis ToolPak and Solver
Add-In boxes and click OK. Note: Analytic Solver should already
have been checked by default.
. In Excel, go to the Data tab and verify that the Data Analysis and
Solver command buttons appear in the Analyze (or Analysis) group.
 page 621

APPENDIX C

Getting Started with R

What is R?
R is a powerful computer language that merges the convenience of
statistical packages with the power of coding. It is open source as
well as cross-platform compatible. This means that there is zero cost
to download R, and it can be run on Windows, macOS, or Linux. In
this appendix, we will introduce you to some fundamental features of
R and provide instructions on how to obtain solutions for many of the
exercises in the text.

What is RStudio?
RStudio is a program that makes R easier to use. On its own, R acts
like a programming language and, as such, comes with a minimal
user interface. As standalone software, R shows a single prompt for
you to enter commands; this is called the Console. While everything
we will ever need from R can be done by combining Console
commands with other programs, things can quickly get messy. To
make coding in R easier, we use an integrated development
environment (IDE). IDEs are programs that combine in one place
many common features needed in programming and give them a
graphical user interface.1 In this text, we use an open source version
of an IDE called RStudio, which is very popular among students,
professionals, and researchers who use R.
Installation
Installation of both R and RStudio is straightforward and requires no
special modifications to your system. However, it should be noted
that RStudio does not come with R; therefore, both pieces of
software need to be installed separately. Many R functions for data
analysis are included in free libraries of code written by R users,
called packages. R packages are installed separately and will be
discussed at the end of this Appendix.

Installing R for Windows

. Navigate to https://cloud.r-project.org/.
. Select Download R for Windows.
. Select install R for the first time and download the most current
version for Windows.
. Locate the downloaded file and double-click.
. Select Yes when asked about verifying the software publisher, and
then select the language that you prefer. We select English.
. Follow the instructions in the R Setup window.
For Mac and Linux, follow the instructions at https://cloud.r-
project.org/bin/macosx/, and https://cloud.r-project.org/bin/linux/,
respectively, and choose the appropriate link for your operating
system and version.

page 622

Installing RStudio
. Navigate to https://www.rstudio.com/products/rstudio/.
. Select RStudio Desktop, then select Download RStudio Desktop.
. Scroll down to the Installers for Supported Platforms section, select
the link that corresponds to your operating system, and then select
Open or Run.
. Select Yes when asked about verifying the software publisher.
. Follow the instructions in the RStudio Setup window.

The Interface
Installation should now be complete. You can close all windows and
then double-click on the RStudio icon.
The RStudio interface consists of several panes. By default, three
panes are visible. We will refer to these by the names of the default
tab shown in each: Console, Environment, and Help. We will also
briefly discuss the Source pane, which is hidden until you open it.
Figure C.1 shows what you should see when you open RStudio for
the first time.

FIGURE C.1 The Console, Environment, and Help Panes

Source: R Studio

Console pane: The Console pane is the primary way that you
interact with R. It is here that you input commands (at the >
prompt) and then view most of your output.
Environment pane: Two relevant tabs in the Environment pane
are: Environment and History. A common feature between them is
 the broom icon, which clears the content of each tab. The
Environment tab shows the data, objects, and variables in the
current R session. The History tab provides a list of all console
commands issued in the session.
Help pane (or Files pane): The help section has five tabs. We
discuss two of these here: Help and Plots.
The Help tab is where you can view R documentation (help files).
For example, to learn about the print function, select the Help
tab and then enter print next to the magnifying glass icon. (You
can also view R documentation by entering a question mark
followed immediately by the topic of interest in the Console pane;
so, for this example, you would enter ?print after the prompt.)
The Plots tab is where you can see all graphs and charts. Any
graph or chart can be cleared with the broom icon.
The Source pane: The Source pane is hidden by default in R.
This is where you can write your own scripts. As you will see, most
of what we do in this text can be accomplished by importing a data
set and then using a single command in the Console. Nonetheless,
here is an example of how you would write a simple script:
A. From the menu, select File > New File > R Script
B. In the new window, enter the following:
page 623
print(“This is my first script.”)
print(“This is easy!”)
Save the script with File > Save As. Name your script Script1.
Figure C.2 shows what you should see in the Source pane.

FIGURE C.2 The Source Pane After Writing First Script

 Source: R Studio

Important: Due to different fonts and type settings, copying and

pasting formulas and functions from this text directly into Excel
may cause errors. When such errors occur, you may need to
replace special characters such as quotation marks and
parentheses or delete extra spaces in the functions.
C. Again refer to Figure C.2. Click the Source button from the menu
on the Source pane; this tells R to read and execute the script.
Figure C.3 shows what you should see in the Console pane after
executing your first script.

FIGURE C.3 The Console Pane After Executing First Script

Source: R Studio

R executes complete statements in the order that they appear.

Unique to RStudio, there is also a way to run specific sections of
scripts. This is done by highlighting the desired section of the
script in RStudio and selecting the Run button from the menu on
the Source pane.

Entering Data and Using Functions

Throughout this text, our goal is to provide the simplest way to obtain
the relevant output. Seasoned users of R might argue that there are
“better” approaches than the ones we suggest, but we feel that they
may distract from learning the important concepts.
Like Excel and other statistical packages, R has many built-in
formulas or functions. In the text, we denote all function names in
boldface. Within each function, R also provides various options,
such as labeling the axes of a graph, inserting colors in a chart, and
so on. We will not use every option within a function; rather, we use
those that we feel are most useful and least cumbersome. We
denote all option names in italics.
Most of the time we will be importing data files, as we explain in
the next section. However, suppose we want to use R to perform a
simple calculation. Suppose we want to calculate the mean given the
following data: −4, 0, 6, 1, −3, −4. In order to input these values into
R, we use the c function, which combines the values to form a list;
or, perhaps more mathematically precise, the c function combines
the values to form a vector. We label this data as Example_1 and
use the expression “<-” which is equivalent to the equal sign. We
enter:

Example_1 <- c(−4, 0, 6, 1, −3, −4)

You should see Example_1 listed in the Environment pane. You can
view the data in the Console pane by entering Example_1 after the
prompt. Additionally, you can use the View function and the data will
appear in the Source pane. (Note that R is case sensitive.) We enter:

View(Example_1)

page 624

Another common function is the mean function which we discuss in

more detail in Chapter 3. In order to calculate the mean of the data,
we enter:

> mean(Example_1)
And R returns: −0.6666667.

Importing Data and Using Functions

FILE
Admission

All the data for Business Analytics—Communicating with Numbers

have been stored in Excel spreadsheets. We will assume that you
have stored all the relevant spreadsheets in a Data folder. When we
import a spreadsheet into R, it is referred to as a data frame. A data
frame is a table, or two-dimensional array-like structure, in which
each column contains measurements on one variable and each row
contains one observation, record, or case. A data frame is used for
storing data tables.
We illustrate the mechanics of importing an Excel file using the
Admission data from Chapter 2. The data set contains the student
record number (Student), the college decision on acceptance
(Decision = Admit or Deny), the student’s SAT score, whether the
student is female or male (Female = yes or no), and the student’s
high school GPA (HSGPA). Table C.1 shows a portion of the
Admission data.

TABLE C.1 Portion of the Admission Data

In order to import this data file into R, we select

File > Import Dataset > From Excel, as shown in Figure C.4.2 (The
first time you import data, R might prompt you to add updates.
Simply follow the steps to add the relevant updates.)
FIGURE C.4 Importing the Admission Data into R

Source: R Studio

page 625

See Figure C.5. We select the Browse button and then navigate
to the Admission data in the Data folder. Once we select the
Admission data, we should see the data in the Data Preview dialog
box. In the R instructions in this text, we label all data files as
myData for simplicity and consistency. Because of this, in the Import
Options dialog box, replace Admission with myData. Once you select
the Import button (see the bottom of Figure C.5), you have
successfully imported the data. You can verify this in a number of
ways. For instance, you should now see myData in the Environment
pane under Data, or you can enter View(myData) in the Console pane
and a portion of the data will appear in the Source pane.

FIGURE C.5 Viewing the Admission Data Prior to Importing

Source: R Studio

Note: For an Excel file with multiple worksheets, select the appropriate worksheet from the
Sheet drop-down option.

Suppose we want to calculate the mean SAT score in myData. In

order to select a variable from a data frame, we attach the
expression $VariableName to the name of the data frame. Here, we
enter:

> mean(myData$SAT)

And R returns: 1197.348.

If the variable name in the data frame is more than one word or
numeric (such as year), then it is necessary to enclose the variable
name with single quotation marks. For instance, if the variable name
was SAT score instead of SAT, then we would have entered
mean(myData$‘SAT score’).
Another function that we discuss in Chapter 3 is the summary
function. This function provides various summary measures for all
variables in a data frame. We could enter summary(myData) and R
would return summary measures for all the variables in the myData
data frame, including the categorical variables. Suppose we would
like summary measures only on SAT and HSGPA. In this case, we
attach square brackets to the name of the data frame, and within the
brackets we indicate the columns that should be included in the
calculations using the c function. In order to obtain summary
measures for SAT (third variable) and HSGPA (fifth variable), we
enter:

> summary(myData[,c(3,5)])

Notice in the above command that we enter a comma directly after

the left square bracket. This implies that we are including all 1,230
observations (all 1,230 rows) in the calculations. If for some reason
we only wanted to include the first 100 observations in the
calculations, we would have entered
summary(myData[1:100,c(3,5)]).
Finally, suppose we want to delete the myData data frame. We
use the rm function and enter:

> rm(myData)

You will find that myData no longer appears under Data in the
Environment pane.

page 626

A Note on Line Breaks

The commands that we have outlined here have been relatively
short. There are some instances, however, when the commands get
long and become difficult to read. To mitigate this, we can break up a
command into parts. R will prompt you to finish the command with
plus + symbols in lines following the first line. For example, in
Chapter 3 we discuss a scatterplot. Suppose we want to construct a
scatterplot of SAT against HSGPA for the first 20 observations using
R’s plot function. In addition to constructing the scatterplot, we use
the ylab and xlab options to add titles on the y-axis and the x-axis.
Two entries for constructing a scatterplot are shown below. Entry 1
uses a single line in R (even though two lines are shown on the
page). Entry 2 uses three lines. Both entries result in the same
scatterplot, as shown in Figure C.6.
Entry 1:

> plot(myData$SAT[1:20] ~ myData$HSGPA[1:20],

ylab=“SAT Score”, xlab=“High School GPA”)

Entry 2:

> plot(myData$SAT[1:20] ~ myData$HSGPA[1:20],

+ ylab=“SAT Score”,
+ xlab=“High School GPA”)

FIGURE C.6 Scatterplot of SAT against HSGPA

Packages
Part of what makes R so powerful is its large collection of packages,
or collections of objects not included in the base version. Packages
greatly expand what can be done with R by adding custom functions
and data structures.
To use a package, you must install it and then load it. We use the
caret package, which stands for Classification and Regression
Training, to demonstrate how this is done:

> install.packages(“caret”)
> library(caret)

The install.packages function connects to the official R servers

(CRAN), downloads the specified package(s) and those it depends
on, and installs it. This must be done with each package only once
on each computer used. The library function loads the installed
package(s). Suppose you want to install multiple packages, such as
caret, gains, and pROC. Instead of installing each package
separately, you can use the following command:

> install.packages(c(“caret”,“gains”,“pROC”))

page 627

However, you still need to use the library function separately for
each package as follows:

> library(caret)
> library(gains)
> library(pROC)

Sometimes, R may prompt you to install additional packages. If this

is the case, follow the steps outlined here to download and install
these additional packages.
Note that each package only needs to be loaded once per R
session. Once the package is downloaded and loaded,
documentation for commands it contains can be viewed in R using
the help feature discussed earlier. Documentation files for an entire
package can be viewed online. All information associated with
available packages can be found at https://cran.r-
project.org/web/packages/.
1More formally, this is called a “graphical user interface,” or a GUI. In practice, this means
that charts, graphs, and buttons can be seen and used.
2Note that you can also import a comma- or tab-delimited text file by selecting
File > Import Dataset > From Text (base) or File > Import Dataset >From Text (readr),
respectively.
 page 628

APPENDIX D

Statistical Tables
TABLE D.1 Standard Normal Curve Areas
Entries in this table provide cumulative probabilities, that is, the area under the curve to the
left of −z. For example, P(Z ≤ −1.52) = 0.0643.
Source: Probabilities calculated with Excel.

page 629
Entries in this table provide cumulative probabilities, that is, the area under the curve to the
left of z. For example, P(Z ≤ 1.52) = 0.9357.
Source: Probabilities calculated with Excel.

page 630

TABLE D.2 Student’s t Distribution

Entries in this table provide the values of tα,df that correspond to a given upper-tail area α
and a specified number of degrees of freedom df. For example, for α = 0.05 and
df = 10, P(T10 ≥ 1.812) = 0.05.
 page 631
Source: t values calculated with Excel.
 page 632

APPENDIX E

Answers to Selected Exercises

Chapter 1
1.2  35 is likely the estimated average age of the population. It would be impossible to
reach all video game players.
1.4  a. The population consists of all recent college graduates with an engineering degree.
b. No, the average salary was likely computed from a sample in order to save time
and money.
1.6  Answers will vary depending on when data are retrieved. The data are cross-sectional
data.
1.8  The data on the front page of The New York Times are likely to be unstructured in that
they do not conform to a predefined row-column format.
1.10  Answers will vary depending on when data are retrieved. The data are time series
data.
1.12  Structured; cross-sectional data.
1.20  a. Nominal
b. Interval
c. Ratio
1.24  a. Country Happiness GDP
Finland 7.769 45670
Denmark 7.600 57533
Norway 7.544 75295
Iceland 7.494 73060
Netherlands 7.488 48754
b. Country,Happiness,GDP
Finland,7.769,45670
Denmark,7.600,57533
 Norway,7.544,75295
Iceland,7.494,73060
Netherlands,7.488,48754

Chapter 2
2.4  Choice c is the correct definition of a foreign key.
2.6  Choices b, c, and d correctly describe SQL. Choice a describes NoSQL.
2.12  a. 1 of the 10 highest income earners is married and always exercises.
b. 9 individuals are married, exercise often, and earn more than $110,000 per year.
c. 5 values are missing for Exercise, 2 for Marriage, and 3 for Income.
d. 281 individuals are married, and 134 are not.
e. 69 married individuals always exercise, and 74 unmarried individuals never
exercise.
2.18  a. There are no missing values in x2, x3, and x4.
b. The remaining data set has 35 observations. The average values for x3 and x4 are
170.6571 and 21.2286 respectively.
2.24  a. Variables x1, x3, x4, and x5 all have at least one missing value.
b. Observations 15, 24, 33, 34, 43, 63, 66, 78, and 95 have missing values.
c. There are 9 missing values in the data set.
d. 2 missing values of x2 were replaced.
e. After replacing missing values of x3, x4, and x5, their means and median in the new
data set are 3.8832, 1260, and 3.1330 respectively.
f. There are 43 observations remain in the dataset.
2.26  a. There are missing values in the variable Travel Plan. All other variables are
complete.
b. 300 observations were removed.
2.28  a. There are 3 missing values in the data set. In the Yards variable, the missing value
is for observation 25. In the Attempts variable, the missing value is for observation
28. In the Interceptions variable, the missing value is for observation 30. All other
variables and observations are complete.
b. Using the omission strategy, 3 observations were removed.
2.32  a. All variables other than Name and Market Cap have missing values. Observation
20 is missing for the variable Price; observation 173 is missing for variable
Dividend; observations 38 is missing for the variable PE; observation 26 is missing
for variable EPS, observation 98 is missing for the variable Book Value; observation
26, and 154 are missing for the variable 52 week low, observation 46 is missing for
the variable 52 week high, and observation 51 is missing for variable the EBITDA.
There are 9 missing values in the data set.
b. 8 observations were removed.
 c. Missing values for variable PE, EPS, and EBITDA were replaced with 21.915,
3.255, and 1.845, respectively. There were no missing values for Market Cap,
whose median value is 21.17.
2.34  a. Observation 26 has missing values for the Siblings variable, observation 13 is
missing for the variable Height, observation 47 is missing for the variable Weight,
and observations 17 and 51 have missing values for the Income variable. All other
variables are complete. In the data set, there are 5 missing values.
b. 3 observations were removed.
c. Observation 13 has a missing value for the Height variable, and observation 47 has
a missing value for the Weight variable. The missing value for Height was replaced
with its mean, 67.3088, and the missing value for Weight was replaced with its
mean, 150.8971. The missing value for Siblings was replaced with its median, 2,
and the missing values for Income were replaced with its median, 33,361. There
were no missing values for FamilySize, whose median is 4.

page 633
2.38  a. There are 41 observations in the data set. After binning, 4 of the five bins should
have 8 observations, and the remaining bin should have 9 observations.
b. The number of observations with a score of 431 is: 0. The number of observations
with a score of 222 is: 2.
2.40  a. The average of Difference is −1.91.
b. The average PercentDifference is 0.2273.
c. The average of Log is 12.84.
2.42  a. The average difference in year is 29.01.
b. The average month value is 6.
c. Group 3 has the highest number of observations.
2.46  a. Number of customers assigned to Group 4 is 26.
b. Number of customers assigned to Group 3 is 21.
c. Number of customers assigned to Group 2 is 36.
d. The average difference is 124.7.
e. The average percentage difference is 0.5642.
f. The average age of the players is 28.70 years.
g. September is the most frequent birth month.
2.50  a. There are 23 observations in the “Other” category.
b. The average category score for x2 is 3.6200.
c. Three dummy variables should be created.
2.52  a. 15 observations have a category score of 3.
b. There are 3 observations in the “Other” category.
c. Four dummy variables should be created.
2.54  a. The variables LoanType, PropertyType, and Purpose are nominal data because
they do not have naturally ordered categories.
b. Conventional is the most frequent category for LoanType. Multi-family is the most
frequent category for PropertyType. Refinancing is the most frequent category for
Purpose.
c. Three dummy variables should be created. Conventional should be the reference
category of LoanType. Multi-family should be the reference category of
PropertyType. Refinancing should be the reference category of Purpose.
2.56  a. One dummy variable should be created.
b. The average damage score of the cell towers is 0.7959.

Chapter 3
3.2  a. The proportion of the sales for medium-sized shirts was 0.302.
b. Sales of large-sized shirts had the highest frequency and sales of small-sized shirts
had the lowest frequency.
3.4  a. 19.3% of people in the Midwest are living below the poverty level.
b. The South has the highest relative frequency as compared to the other three
regions, which are roughly equal.
3.6  a. A rating of 5 has the highest frequency
b. The higher ratings have the higher frequencies.
3.8  a. Not Religious is the most common response.
b. About 35% responded “Not Religious” which is consistent with the Pew Research
survey.
3.10  a. 125 stocks had returns of at least 10% but less than 20%.
b. The distribution is symmetric.
3.14  a. No. The distribution is not symmetric. It is positively skewed.
b. Over this time period, the stock price was between $50 and $250.
c. The $100 up to $150 interval has the highest relative frequency, which is about
0.44.
3.16  a. 14 customers spent between $700 and $999.
b. 52 customers spent $1,300 or less; 48 customers spent more than $1,300.
3.18  a. The DJIA was more than 26,000 on 44 days in the first half of 2019.
b. The distribution is not symmetric; it is negatively skewed.
3.20  It does not; by using a relatively high value as an upper limit on the vertical axis
($500), the rise in stock price appears dampened.
3.22  a. 202 of the customers were male; 60 of the customers drank wine.
b. 142/202 = 0.7030; 38/68 = 0.5588.
c. Beer is the popular drink at this bar, followed by wine, and then soft drinks. Both
men and women are more likely to choose beer over the other two options.
3.24  a. 5 components from shift 1 were defective; 94 components from shift 2 were not
defective.
b. 6/24 = 0.2500; 13/24 = 0.5417; components constructed during shift 3 seem to be
defective at a higher rate.
c. No, defective rates are not consistent over the shifts.
3.26  a. 120 students are business majors; 68 students study hard.
b. 20/120 = 0.1667; 48/150 = 0.3200; the data suggest that nonbusiness majors are
more likely to study hard, which supports the report.
c. The majority of both business and nonbusiness students do not study hard, but
nonbusiness students are more likely to study hard.
3.28  Negative relationship between obesity and life expectancy.
3.30  No relationship between the returns of A and B, so investing in both would diversify
risk.
3.32  a. Negative relationship between the price of a car and its age.
b. Negative relationship between the price of a car and its mileage.
3.34  Both countries have an upward trend, but China’s begins to stall slightly around 2000.
3.36  a. Age is negatively correlated with price and positively correlated with mileage. Price
is negatively correlated with mileage.
b. Seven cars have mileage greater than 50,000.
c. The negative relationship between age and price is consistent for cars of both
mileage categories.
3.38  a. There are 511 cases of a burglary in a residence.
b. Theft on the street is the most common crime, followed by narcotics on a sidewalk,
then motor vehicle theft on the street.

page 634
3.40  
a. HD median = 77,349

Lowe’s median = 50,208

HD had a higher average revenue.

b. HD s2 = 106,175,080; s = 10,304
Lowe’s s2 = 56,414,506; s = 7,511
HD had a higher dispersion.
3.42  a. mean = 2.3122; median = 2.18
b. Q1 = 2.01; 25% of the states had gas prices below $2.01.
Q3 = 2.55; 75% of the states had gas prices less than $2.55.
 c. s2 = 0.1482; s = 0.3849
3.44  a. Monthly rent is positively skewed, with tails longer than that of a normal distribution.
b. Square footage is positively skewed, with tails longer than that of a normal
distribution.
3.46  
a. Firm A: s2 = 52.02; s = 7.21

Walmart s2 = 319.21; s = 17.87

b. Firm B had the higher average stock price.

c. Firm B had a higher dispersion in stock price.
3.48  a. The Latin America fund is positively skewed, with tails shorter than that of normal
distribution.
b. The Canada fund is negatively skewed, with tails longer than that of a normal
distribution.
3.50  γFinal,Midterm = 0.5518; positive and moderate linear relationship.
3.56  a. Approximately 25% of the observations have values that are less than 200;
Approximately 75% of the observations have values that are less than 550.
b. IQR = 350; Limit: 1.5 × IQR = 525. Q1 − Min = 75 < 525 and Max − Q3 = 750 >
525; thus, there is at least one outlier on the right side of the distribution.
c. The distribution appears positively skewed because the median falls left of center in
the interquartile range and the right whisker is longer than the left whisker.
3.58  a. About 68% of the scores are in the interval [84, 116].
b. About 95% of the scores are in the interval [68, 132]. About 2.5% of the scores are
less than 68.
c. About 16% of the scores are more than 116.
3.60  a. About 95% of the games will last between 2.2 and 3.8 hours.
b. About 16% of the games are longer than 3.4 hours.
c. About 2.5% of the games are shorter than 2.2 hours.
3.62  a. The boxplot suggests that there are no outliers for the Debt variable.

b. s = 124.6086. The smallest and the largest observations for the Debt

variable are 763 and 1,285, respectively.

The z-score for the smallest observation is

The z-score for the largest observation is

Since the absolute value of both z-scores is less than 3, we conclude that there are
no outliers for the Debt variable. This is consistent with the boxplot, which showed
 no outliers.
3.64  a. The Technology boxplot suggests that there are outliers in the upper part of the
distribution.

b. sTech = 36.5637. The smallest and the largest observations in the

data set are −51.09 and 131.75, respectively.

The z-score for the smallest observation is

The z-score for the largest observation is

Since the z-score for the largest observation is greater than 3, we conclude that
there are outliers for Technology. This is consistent with the boxplot, which showed
outliers in the upper part of the distribution.
c. The Energy boxplot suggests that there is an outlier in the lower part of the
distribution.

d. sEnergy = 22.9492. The smallest and the largest observations in

the data set are −54.00 and 52.02, respectively.

The z-score for the smallest observation is

The z-score for the largest observation is

Since the absolute value of both z-scores is less than 3, we conclude that there are
no outliers for Energy. This is not consistent with the boxplot, which showed an
outlier in the lower part of the distribution. Since the boxplot indicates that Energy is
not symmetric, we are better served identifying outliers in this case with a boxplot.

e.

Chapter 4
4.2  a. Not exhaustive because you may not get any offer.
b. Not mutually exclusive because you may get both offers.
4.4  Let event A correspond to “Firm raising an alarm”, and event F to “Fraudulent
Transaction”. We have P(A) = 0.05, P(A|F) = 0.80, and P(F) = 0.01.
P(Ac ∩ F) = 0.002, and therefore, P(F|Ac) = 0.0021
4.6  a. P(A) = 0.70, P(Ac) = 0.30, P(B) = 0.50
b. Not mutually exclusive
 c. A ∩ B
4.10  Let event Ai be “the i-th selected member is in favor of the bonus”.
a. P(A1 ∩ A2) = 0.4286

b.

4.12  Let event A correspond to “Asians”, B to “black”, W to “white”, H to “Hispanic”, and T

to “both parents at home”. We have P(T|A) = 0.85, P(T|W) = 0.78, P(T|H) = 0.70, and
P(T|B) = 0.38
a. P(A ∩ B) = 0; mutually exclusive
P(A) + P(B) < 1; not exhaustive
b. P(Wc) = 0.44
c. P(W ∩ T) = 0.4368
d. P(A ∩ Tc) = 0.015
4.14  Let event C correspond to “Using own car daily”, and event U to “Under 35”. We have
P(C) = 0.62, P(C|U) = 0.38, and P(U) = 0.43.

page 635
a. P(C ∩ U) = 0.1634
b. P(U|C) = 0.2635
4.16  Let event D be “Experience a decline”, and event N be “Ratio is negative”. We have
P(D) = 0.20, P(N|D) = 0.70, and P(N|Dc) = 0.15.
P(N) = 0.26, and therefore, P(D|N) = 0.54
4.18  Let event O correspond to “obese”, W to “white”, B to “black”, H to “Hispanic”, and A
to “Asian”. We have P(O|W) = 0.33, P(O|B) = 0.496, P(O|H) = 0.43, P(O|A) = 0.089,
P(W) = 0.48, P(B) = 0.19, P(H) = 0.26, and P(A) = 0.07.
a. P(O) = 0.3707
b. P(W|O) = 0.4273
c. P(B|O) = 0.2542
d. P(A|O) = 0.0168
4.20  Let F = “Player is fully fit to play”, S = “Player is somewhat fit to play”, N = “Player is
not able to play”, and W = “The Lakers win the game”. We have P(F) = 0.40, P(S) =
0.30, P(N) = 0.30, P(W|F) = 0.80, P(W|S) = 0.60, and P(W|N) = 0.40.
a. P(W) = 0.62
b. P(F|W) = 0.52
4.22  Let event R correspond to “Republican”, D to “Democrat”, I to “Independent”, and S to
“Support marijuana legalization”. We have P(R) = 0.27, P(D) = 0.30, P(I) = 0.43,
P(S|R) = 0.41, P(S|D) = 0.66, and P(S|I) = 0.63.
a. P(S ∩ R) = 0.1107
 b. P(S ∩ D) = 0.1980
c. P(S ∩ I) = 0.2709
d. P(S) = 0.5796
e. P(R|S) = 0.1910
4.26  Let X be the amount spent on a warranty and Y be the revenue earned by the store.
E(X) = $30
E(Y) = $3,600
4.32  a. P(X < 2) = 0.6517
b. P(X < 2) = 0.4580
4.34  a. E(X) = 905; SD(X) = 27.22
b. E(X) = 1430; SD(X) = 31.95
4.36  a. P(X > 2) = 0.5276
b. P(X > 2) = 0.1362
4.38  Let X be the number of designers who show the acceptable design.
a. P(X ≥ 1) = 0.4375 < 0.50; statement is not correct
b. P(X ≥ 1) = 0.5781 > 0.50; statement is correct
4.40  Let X be the number of attendees whom the manager contacts.
a. P(X = 10) = 0.1171
b. P(X ≤ 10) = 0.8725
c. P(X ≥ 15) = 0.0016
4.42  a. P(X ≤ 4) = 0.8153
b. P(X ≥ 3) = 0.5768
4.44  a. μ1 = 6; P(X = 2) = 0.0446
b. μ1 = 6; P(X ≥ 2) = 0.9826
c. μ10 = 60; P(X = 40) = 0.001
4.46  a. μ1 = 2; P(X > 2) = 0.3233
b. μ5 = 10; P(X = 6) = 0.0631
c. μ180 = 360
4.50  a. μ = 304; P(X > 320) = 0.1717
b. μ = 2128; P(X > 2200) = 0.0586
4.52  Let X equal points scored in a game.
a. P(85 < X < 125) = 0.9544
b. P(X > 125) = 0.0228; approximately 2 games (82 × 0.0228 = 1.87)
4.54  Let X represent the mpg rating of passenger cars.
a. P(X ≥ 40) = 0.0384
b. P(30 ≤ X ≤ 35) = 0.4952
c. x = 33.8 + 2.326(3.5) = 41.94
4.56  Let X represent the return on a portfolio.
P(X > 16) = 0.2514 ≠ 0.15; not normal.
4.62  Let X equal the length of time of a football game.
a. P(X < 2.5) = 0.1056
b. P(X < 2.5) + P(X > 3.5) = 0.2112
c. x = 3 − 2.326(0.4) = 2.07

Chapter 5
5.4  
a.

b.

c. Janice; her findings are more likely if a representative sample is used.

5.6  
a.

b.

5.8  a. P(X > 1,000,000) = P(Z > 0.80) = 0.2119

b.

5.10  a. P(X < 90) = P(Z < −0.63) = 0.2643

b.

c. (P(X < 90))4 = (0.2643)4 = 0.0049

5.12  
a. the normal approximation conditions are met because

np = 33 > 5 and n(1 − p) = 67 > 5.

b.

c.

5.14  For n = 50,

For n = 100,
 You would choose 50 balls because with larger sample sizes the standard deviation
of is reduced. The probability of getting 70% green balls is slightly higher with a

smaller sample because of the increased standard deviation.

5.16  
a.

b. 12.5 ± 4.58 or [7.92,17.08]

c. To reduce the margin of error of a 95% confidence level, the program should
increase its sample size.
5.18  
a.

b. 100 ± 4.54 or [95.46,104.54]

5.20  a. 17.25 ± 7.36 or [9.89,24.61]
b. The population is normally distributed.
5.24  1,080 ± 89.30 or [990.70,1,169.30]
The manager is wrong with the new strategy, since 1,200 is not within the 95%
confidence interval.

page 636
5.26  a. microeconomics: [68.74, 74.91]; macroeconomics: [66.16, 74.64]
b. The widths are different because the sample standard deviations for
microeconomics and macroeconomics are different.
5.28  a. For research expenditure: [231.44, 373.49]
b. For duration: [18.40, 22.60]
5.30  a. 0.37 ± 0.011 or [0.359, 0.381]
b. 0.37 ± 0.017 or [0.353, 0.387]
c. The margin of error in part b is greater because it uses a higher confidence level.
5.32  a. 0.47 ± 0.025 or [0.445, 0.495].
b. 0.025
c. 0.038
5.34  a. 0.275 ± 0.037 or [0.238, 0.312]
b. No, because the value 0.30 falls in the interval.
5.38  a. Type I error; the new software is purchased even though it does not reduce
assembly costs.
b. Type II error; the new software is not purchased even though it reduces assembly
costs.
5.40  a. H0: μ ≤ 5; HA: μ > 5
 b. t6 = 0.643; p-value = 0.272. It is necessary to assume that the population is
normally distributed.
c. Because 0.272 > 0.10, do not reject H0. The average waiting time is not more than
5 minutes at the 10% level.
5.42  H0: μ ≤ 6.6; HA: μ > 6.6; t35 = 2.7; p-value = 0.0053. Because 0.0053 < 0.05, reject
H0. At the 5% significance level, he can conclude that the mean increase in home
prices in the West is greater than the increase in the Midwest.
5.44  a. H0: μ = 50; HA: μ ≠ 50
b. t49 = −2.324; p-value = 0.024
c. Because 0.024 < 0.05, reject H0 At the 5% significance level, we conclude that the
average differs from 50.
5.46  a. H0: μ ≤ 65; HA: μ > 65
b. t39 = 2.105; p-value = 0.021.
c. Because 0.021 > 0.01, do not reject H0. At 1% significance level, we cannot
conclude that the average speed is greater than the stated speed limit of 65mph.
5.52  H0: p ≤ 0.60; HA: p > 0.60; z = 1.04; p-value = 0.1492. Because 0.1492 > 0.01, do not
reject H0. At the 1% significance level, we cannot conclude that more than 60% of
seniors have made serious adjustments to their lifestyle.

Chapter 6
6.2  a. The positive sign for the Poverty coefficient is as expected; the slope coefficient for
Income is not as expected.

b.

6.4  a. Female employees

b. Female employees without an MBA
c. No
6.6  
a.

b. 3.35
6.8  
a.

b. As disposable income increases by $1, consumption is predicted to increase by

$0.8465.
c. $31,992.75
6.12  a. Assuming electricity is used for cooling (but not for heating), we expect electricity
cost to be higher in months with a higher average outdoor temperature, higher in
months with more working days, and higher in months in which more tons of
product are produced.

b.

c. $28,279.05.
6.14  
a.

b. For every additional $1,000,000 in drive-through sales, net profits are predicted to
increase by $112,300, holding counter sales constant.
c. 0.7647; or net profits are predicted to be $764,700 when counter sales are
$6,000,000 and drive-through sales are $4,000,000.
6.16  
a.

b. 1.48 startups
c. A $1 million increase in research expenditure results in a predicted increase in the
number of startups by 0.0087, holding everything else constant. Approximately

$114.94 million in additional research expenditures would be

needed to have 1 additional predicted startup, everything else being the same.
Note that $114.94 × 0.0087 equals (approximately) 1.
6.18  
a. R2 = 0.8299
b. 0.1701
6.22  Model 2 is a better fit since it has a smaller standard error and a higher adjusted R2.
We cannot use R2 for comparison because the models have different numbers of
predictor variables.
6.24  a. se = 0.2144.
b. 0.7573; 75.73% of the variability in net profits is explained by the model.

c.

6.32  
a.
 b. Because 0.077

< 0.10, reject H0. At the 10% significance level, the two explanatory variables are
jointly significant.
c. H0: β1 = 0; HA: β1 ≠ 0; p-value = 0.738. Because 0.738 > 0.10, do not reject H0. At
the 10% significance level, P/E is not significant in explaining Return.
H0 : β2 = 0; HA: β2 ≠ 0; p-value = 0.025. Because 0.025 < 0.10, reject H0. At the
10% significance level, P/S is significant in explaining Return.
6.34  
a.

b.

page 637
Because 0.006 < 0.05, reject H0. At the 5% significance level, an
extra second of Time decreases Watches by more than 0.02.
6.36  
a.

b. because 0.026 < 0.10, reject H0. At the 10%

significance level, the predictor variables are jointly significant in explaining the
electricity costs.
c. At the 10% significance level, the average temperature is significant in explaining
the electricity costs, the number of work days is not significant, and the tons
produced is not significant.
6.42  
a. at the 5% significance level, BMI

influences salary.
b. White college-educated man: 38.01; non-white college-educated man: 33.52.
6.44  
a.

b. 14.13 minutes
c. 85.30%
d. At the 5% significance level, the predictor variables are jointly significant. At the 5%
significance level, each predictor variable is individually significant.
6.46  
 a.

b. As the number of bays increases by 1 bay the number of vehicles served per month
is predicted to increase by 23.5056, holding other variables constant; as the
population in a 5-mile radius increases by 1,000 people, the number of vehicles
served per month is predicted to increase by 0.5955, holding other variables
constant. When changing to convenient interstate access, the number of vehicles
served per month is predicted to increase by 84.5998, holding other variables
constant. When changing to the winter, the number of vehicles served per month is
predicted to increase by 77.4646, holding other variables constant.
c. At the 5% (and 10%) significance level, the predictor variables are jointly significant
in explaining the number of vehicles served per month. At the 5% significance level,
Access and Winter are significant, but Garage Bays and Population are not
significant. At the 10% significance level, Access, Winter, and Garage Bays are
significant, but Population is not significant.
d. 0.8739
e. 361.34 vehicles
6.48  
a.

b. $1,057,485.47
c. At the 5% significance level, all the quarters are significantly different from the
fourth quarter.
d. At the 5% significance level, we cannot conclude that the sales differ between
quarters 2 and 3.
6.50  Create dummy variables for White, Black, Asian, and Hispanic students using the
Ethnicity variable.

a.

Asian student: 1654; Hispanic student: 1389

b. At the 5% significance level, Asian students score differently on the SAT than
Hispanic students.
c. At the 5% significance level, white students score lower than Asian students.
6.52  a. The residuals fan out when plotted against x.
b. Constant variability; with changing variability, the estimators are unbiased, but not
efficient. Another problem is that the t tests and F test are not valid. It is common to
use robust standard errors for conducting significance tests.
6.54  The scatterplot shows that a simple linear regression model is not appropriate as
GPA is positively related to Hours at lower levels, but negatively related at higher
 levels of Hours.
6.56  a. Perfect multicollinearity, because Study + Sleep + Leisure = 24. The proposed
model cannot be estimated.
b. Drop the Sleep variable.
6.58  a. Experienced (older) employees are likely to have more variability in salaries
because not all employees reach the same level of success over time.
b. The residuals fan out when plotted against experience, confirming the changing
variability problem.
6.60  There does not appear to be an issue with correlated observations, as the residuals
do not show any pattern around the horizontal axis.
6.62  
a.

The predictor variables are jointly significant. At the 5% significance level,

Advertising is significant, but Temp and Discount are not significant.
b. The correlation matrix reveals excessive correlation among the predictor variables
which suggests multicollinearity. It might be best to do nothing about
multicollinearity because none of the predictor variables appears to be redundant.
Also, although Temp and Discount are found to be individually insignificant, the
estimated model has high explanatory power.
c. The residuals appear to be (reasonably) randomly dispersed across the domains of
each predictor variable. Changing variability does not seem to be serious.

Chapter 7
7.8  
a.

predictor variables are jointly and individually significant at the 5%

level.
b. 80% (Just Woman); 55% (Woman and Hispanic Man)
7.10  
a. Rural + 20.4745 College + 49.8701

(Rural × College); predictor variables are jointly and, except for College, individually
significant at the 5% level.
b. 160.64 (Rural); 153.10 (Urban)

page 638
7.12  
a.
 Train + 0.9785 (Exper × Train)

b. The model with the interaction term is preferred because it has a higher adjusted
R2 and the interaction variable is significant at the 10% level.
c. 12.46 (10 years); 8.79 (20 years)
d. Less experienced employees benefit more from the training program (reduced pick
errors) than more experienced employees.
7.14  
a. 1227.6866 Years + 36655.1363 Grad − 0.5227

(Profit × Grad) − 193.1612 (Years × Grad)

b. Predictor variables are jointly significant at the 5% level.
c. Profit, Years, and Grad are individually significant at the 5% level, but the
interaction variables are not.
d. $101,855
7.16  
a. HSGPA + 1.6688 White − 0.0011 (SAT

× HSGPA) − 0.0010 (SAT × White); predictor variables are jointly and, except for
the second interaction term, individually significant at the 5% level.
b. White: 3.23 (SAT = 1200); 3.41 (SAT = 1300); 3.59 (SAT = 1400)
Non-White: 2.81 (SAT = 1200); 3.10 (SAT = 1300); 3.38 (SAT = 1400)
7.18  
a. 319.0062 Grad + 219.1327 Debt +

2643.2320 City + 0.0185 (Cost × Grad); the partial effect of Cost increases with
Grad and the partial effect of Grad increases with Cost.
b. $42,666 (Cost = $20,000); $46,398 (Cost = $30,000); $50,130 (Cost = $40,000)
c. $43,706 (Cost = $20,000); $51,148 (Cost = $30,000); $58,589 (Cost = $40,000)
7.26  
a. 0.0004 Agesq − 0.0810 Female + 0.0788

NPS + 0.22142/2)
b. 47.91
c. $91,903 (Males); $84,750 (Females)
7.28  
a.

(Quadratic)

b. Higher Adjusted R2 for the quadratic model.

c. 20.59
7.30  
a.

For every 50 second increase in the time, there is a decrease of about one watch.

b. 5.7493 Male

For every 10% increase in the time, there is a decrease of about 1.55 watches.

c. Logarithmic Model because of its higher R2 (0.6437 > 0.6335)

7.32  
a. 271.95(76) = $16,655.57

b. $15,912.73

c. Exponential Model because of its higher R2 (0.7099 > 0.6935).

7.34  
a.

b. Linear Model because of its higher R2 (0.7001 > 0.6497)

7.48  
a.

b. At the 10% level, Age is significant, but Gender is not; neither is significant at the
5% level.
c. 0.8410 (Male); 0.5970 (Female)
7.52  
a. 0.4343

b. 0.4282

7.54  
a. Linear: Age − 0.2019 Religious
 Logistic:

Both variables are significant at the 5% significance level.

b. 74% (linear); 74.5% (logistic)
c. 0.6516 (Religious); 0.8410 (Not religious)
7.60  a. Choose Model 2 because its RMSE is slightly lower (37.2535 < 37.7394).
b. Choose Model 2 because its average RMSE is lower (38.5315 < 39.8582).
7.64  a. Choose Model 2 because its RMSE is lower (273.2006 < 305.1602).
b. Choose Model 1 because its average RMSE is lower (362.9936 < 365.2085)
7.66  a. Choose Model 1 because its RMSE is lower (7.3696 < 7.8195).
b. Choose Model 2 because its average RMSE is lower (6.7026 < 6.9891).
7.70  a. Both models have the same accuracy rate of 68%; so, choose any.
b. Choose Model 2 because its average accuracy rate is higher (75% > 74%).

Chapter 8
8.2  a. The Euclidean distance between observations 1 and 2 = 421.0001
b. The Manhattan distance between observations 1 and 2 = 421.3300

page 639
c. s1 = 2369.6638; s2 =

0.3958; s3 = 0.2211.
The normalized Euclidean distance between observation 1 and 2 = 0.7902
d. x1 min = 1346, x2 min = 2.67, x3 min = 0.01, Range1 = 7587, Range2 = 1.16,
Range3 = 0.61.
The min-max standardized Euclidean distance between observations 1 and 2 =
0.269.
8.6  a. The Euclidean distance between observation 1 and 2 = 0.469
The Euclidean distance between observation 1 and 3 = 0.8307
The Euclidean distance between observation 2 and 3 = 1.1705

s1 = 0.4774; s2 = 0.2368; s3 = 0.3114.

The normalized Euclidean distance between observation 1 and 2 = 1.5533
The normalized Euclidean distance between observation 1 and 3 = 1.8437
The normalized Euclidean distance between observation 2 and 3 = 2.8573
b. The Manhattan distance between observations 1 and 2 = 0.8
 The Manhattan distance between observations 1 and 3 = 1.1000
The Manhattan distance between observations 2 and 3 = 1.5000
x1 min. = 4.3; x2 min = 2.9; x3 min = 0.9; Range1 = 1.9; Range2 = 0.8; Range3 =
1.0
The min-max standardized Manhattan distance between observations 1 and 2 =
0.7329
The min-max standardized Manhattan distance between observations 1 and 3 =
0.7461
The min-max standardized Manhattan distance between observations 2 and 3 =
1.0789
8.8  a. Matching coefficients:
Observations 1 and 2 = 1/4
Observations 1 and 3 = 0/4
Observations 1 and 4 = 3/4
Observations 2 and 3 = 2/4
Observations 2 and 4 = 0/4
Observations 3 and 4 = 1/4
b. Most similar: Observations 1 and 4 have matching coefficients of 3/4.
Least similar: Observations 1 and 3, and 2 and 4 have matching coefficients of 0/4
8.12  a. Matching coefficients:
Observations 1 and 2 = 0/3
Observations 1 and 3 = 1/3
Observations 1 and 4 = 1/3
Observations 1 and 5 = 1/3
Observations 2 and 3 = 2/3
Observations 2 and 4 = 2/3
Observations 2 and 5 = 2/3
Observations 3 and 4 = 1/3
Observations 3 and 5 = 1/3
Observations 4 and 5 = 1/3
b. Most similar: Vehicles 2 and 3, 2 and 4, and 2 and 5 have a matching coefficient of
2/3. Vehicles 2 and 3 are both sedans without all-wheel drive while 2 and 4 are both
manual without all-wheel drive, and 2 and 5 are both manual sedans.
Least similar: Vehicles 1 and 2 have a matching coefficient of 0/3. Vehicle 1 is an
SUV with all-wheel drive, and automatic transition; vehicle 2 is a sedan without all-
wheel drive, with a manual transition.
8.14  a. Matching coefficients:
Students 1 and 2 = 1/4
Students 1 and 3 = 0/4
 Students 1 and 4 = 1/4
Students 1 and 5 = 2/4
Students 2 and 3 = 3/4
Students 2 and 4 = 2/4
Students 2 and 5 = 1/4
Students 3 and 4 = 3/4
Students 3 and 5 = 2/4
Students 4 and 5 = 3/4
b. Most similar: Students 2 and 3 and Student 4 and 5 have a matching coefficient of
3/4. Student 2 and 3 are both male undergraduate students who are not statistic
minors. Student 2 is a math major, while Student 3 is not. Student 4 and 4 are both
female non-math majors and non-statistic minors. Student 4 is an undergraduate
student while Student 5 is a graduate student.
Least similar: Students 3 and 1 have a matching coefficient of 0/4. Student 3 is a
male undergraduate non-math major, non-statistic minor; Student 1 is a female
graduate math major, statistics minor.
8.16  a. Matching coefficients:
Observations 1 and 2 = 2/4
Observations 1 and 3 = 0/4
Observations 1 and 4 = 2/4
Observations 1 and 5 = 2/4
Observations 2 and 3 = 2/4
Observations 2 and 4 = 2/4
Observations 2 and 5 = 0/4
Observations 3 and 4 = 2/4
Observations 3 and 5 = 2/4
Observations 4 and 5 = 2/4
b. Jaccard’s coefficients:
Observations 1 and 2 = 2/4
Observations 1 and 3 = 0/4
Observations 1 and 4 = 2/4
Observations 1 and 5 = 1/3
Observations 2 and 3 = 1/3

page 640
Observations 2 and 4 = 2/4
Observations 2 and 5 = 0/4
Observations 3 and 4 = 1/3
Observations 3 and 5 = 0/3
Observations 4 and 5 = 1/3
 c. Most similar: The pairs of observations 1 and 2, 1 and 4, and 2 and 4 each have a
matching coefficient of 2/4 and Jaccard’s coefficients of 2/4.
Least similar: The pairs of observations 1 and 3, and 2 and 5 have matching
coefficients of 0, and Jaccard’s coefficients of 0.
Comparing Matching and Jaccard’s coefficients: Jaccard’s coefficients are all less
than or equal to their corresponding matching coefficients.
8.18  

Actual Class Predicted Class 1 Predicted Class 0

Class 1 3 2

Class 0 2 3

8.22  Misclassification rate = 0.02

Accuracy rate = 0.98
Sensitivity = 0.9820
Precision = 0.9750
Specificity = 0.9782
8.24  a. Lift (1:5) = 0.8571
b. Lift (1:8) = 1.0714
8.32  a. At cutoff value = 0.25

Actual Class Predicted Class 1 Predicted Class 0

Class 1 18 1

Class 0 45 36

At cutoff value = 0.5

Actual Class Predicted Class 1 Predicted Class 0

Class 1 14 5
 Actual Class Predicted Class 1 Predicted Class 0

Class 0 12 69

At cutoff value = 0.75

Actual Class Predicted Class 1 Predicted Class 0

Class 1 7 12

Class 0 2 79

b. At cutoff value = 0.25

Misclassification rate = 0.46
Accuracy rate = 0.54
Sensitivity = 0.9474
Precision = 0.2857
Specificity = 0.4444

At cutoff value = 0.5

Misclassification rate = 0.17
Accuracy rate = 0.83
Sensitivity = 0.7368
Precision = 0.5385
Specificity = 0.8519

At cutoff value = 0.75

Misclassification rate = 0.14
Accuracy rate = 0.86
Sensitivity = 0.3684
Precision = 0.7778
Specificity = 0.9753
c.
 d. Lift (1:20) = 3.4211
e. Lift (1:50) = 1.6842
8.36  a. Using the covariance matrix, the weights for the first principal component scores
are: w1 = −0.8074, w2 = −0.5900. The first principal component accounts for
86.1636% of the total variability of the data.
b. Using the correlation matrix, the weights for the first principal component scores
are: w1 = −0.7071, w2 = 0.7071. The first principal component accounts for
85.3173% of the total variability of the data.
c. The numbers are different because the raw data are on slightly different scales.
Hence, when we use the covalence matrix, the scale of the data influences its
weight when calculating the principal components. When we use the correlation
matrix, we remove the effects of scales on the calculated weights.
8.42  a. z11 = 0.45, z21 = 0.2684, z12 = 0.5, and z22 = 0.1132
b. The first principal component score of the first observation, PC1,1 = −0.5227
c. The second principal component score of the second observation, PC2,2 = 0.1251
8.46  a. There are 9 variables in the original data set.
 b. The third principal component accounts for 13.16% of the variance in the data.

page 641
c. We need to retain 3 principal components to accounts for at least
85% of the total variance in the data.

Chapter 9
9.2  Analytic Solver:
a. Optimal value of k: 5
b. Accuracy = 50%; Specificity = 0.3846; Sensitivity = 0.6667; Precision = 0.4286.
c. The model is not effective due to low overall predictive accuracy and low AUC
value (0.5085).
R:
a. Optimal value of k: 4
b. Accuracy = 0.5814; Specificity = 0.6190; Sensitivity = 0.5455; Precision = 0.6.
c. The model is not effective due to low overall predictive accuracy and low AUC
value (0.5411).
9.4  Analytic Solver:
a. Optimal value of k: 9
b. Misclassification rate: 18.33%
c. Accuracy = 67.5%; Specificity = 0.8438; Sensitivity = 0; Precision = 0.
d. Accuracy = 77.5%; Specificity = 0.9375; Sensitivity = 0.125; Precision = 0.333.
e. The model is only slightly effective due to relatively low overall predictive accuracy
and AUC value (0.5313).
R:
a. Optimal value of k: 6
b. Misclassification rate: 0.2163
c. Accuracy = 0.75; Specificity = 0.9375; Sensitivity = 0; Precision = 0.
d. Accuracy = 0.6; Specificity = 0.625; Sensitivity = 0.5; Precision = 0.25.
e. The model is only slightly effective due to relatively low overall predictive accuracy
and AUC value (0.6406).
9.6  Analytic Solver:
a. Optimal value of k: 9
b. Accuracy = 91.5%; Specificity = 0.9564; Sensitivity = 0.4545; Precision = 0.4839;
AUC = 0.8839.
c. The KNN classifier is better at classifying non-target class cases than classifying
target class cases.
d. The plots show that the KNN classifier performs better than the baseline model.
Both the lift curve and ROC curve lie above the diagonal line. The AUC value is
0.8839 suggesting an effective model.
R:
 a. Optimal value of k: 10
b. Accuracy = 0.9036; Specificity = 0.9623; Sensitivity = 0.3902; Precision = 0.5424;
AUC = 0.9118.
c. The KNN classifier is better at classifying non-target class cases than classifying
target class cases.
d. The plots show that the KNN classifier performs better than the baseline model.
Both the lift curve and ROC curve lie above the diagonal line. The AUC value is
0.9118 suggesting an effective model.
9.8  Analytic Solver:
a. Optimal value of k: 9
b. Accuracy = 79.17%; Specificity = 0.8889; Sensitivity = 0.7333; Precision = 0.9167.
c. AUC = 0.8037; The predictive performance measures and graphs suggest that the
KNN classifier is effective in classifying the data.
d. Predicted admission outcome for the first new applicant: 1(Admit)
R:
a. Optimal value of k: 7
b. Accuracy = 0.8298; Specificity = 0.8421; Sensitivity = 0.8214; Precision = 0.8846.
c. AUC = 0.828; The predictive performance measures and graphs suggest that the
KNN classifier is effective in classifying the data.
d. Predicted admission outcome for the first new applicant: 1(Admit)
9.10  Analytic Solver:
a. Optimal value of k: 3
b. Accuracy = 56.14%; Specificity = 0.625; Sensitivity = 0.48; Precision = 0.5.
c. AUC = 0.6031; The predictive performance measures and graphs suggest that the
KNN classifier is slightly effective in classifying the data.
d. Predicted outcome for the first new consumer: 0 (Do not respond)
e. Accuracy = 54.39%; Specificity = 0.5313; Sensitivity = 0.56; Precision = 0.4828.
R:
a. Optimal value of k: 6
b. Accuracy = 0.6814; Specificity = 0.7681; Sensitivity = 0.5455; Precision = 0.6.
c. AUC = 0.7085; The predictive performance measures and graphs suggest that the
KNN classifier is moderately effective in classifying the data.
d. Predicted outcome for the first new consumer: 0 (Do not respond)
e. Accuracy = 0.6106; Specificity = 0.4203; Sensitivity = 0.9091; Precision = 0.5.
9.12  Analytic Solver:
a. Optimal value of k: 3
b. Misclassification rate: 33.33%
c. Accuracy = 58.33%; Specificity = 0.2857; Sensitivity = 0.8438; Precision = 0.5745.
d. No. Part of the lift curve lies below the baseline.
e. Lift value of the leftmost bar of the decile-wise lift chart: 1.25
 f. AUC = 0.6908
g. The predictive performance measures and graphs suggest that the KNN classifier
is slightly effective in classifying the data.

page 642
R:
a. Optimal value of k: 9
b. Misclassification rate: 0.3072
c. Accuracy = 0.6167; Specificity = 0.6154; Sensitivity = 0.6176; Precision = 0.6774.
d. Yes. The entire lift curve lies above the baseline.
e. Lift value of the leftmost bar of the decile-wise lift chart: 1.5
f. AUC = 0.7216
g. The predictive performance measures and graphs suggest that the KNN classifier
is moderately effective in classifying the data.
9.14  Analytic Solver:
a. Optimal value of k: 9
b. Misclassification rate: 27.78%
c. Accuracy = 63.89%; Specificity = 0.6154; Sensitivity = 0.6522; Precision = 0.75.
d. The plots show that the KNN classifier performs better than the baseline model.
Both the lift curve and ROC curve lie above the diagonal line. The AUC value is
0.7508.
e. The predictive performance measures and graphs suggest that the KNN classifier
is moderately effective in classifying the data.
R:
a. Optimal value of k: 9
b. Misclassification rate: 0.2459
c. Accuracy = 0.7042; Specificity = 0.5862; Sensitivity = 0.7857; Precision = 0.7333.
d. The plots show that the KNN classifier performs better than the baseline model.
Both the lift curve and ROC curve lie above the diagonal line. The AUC value is
0.7697.
e. The predictive performance measures and graphs suggest that the KNN classifier
is moderately effective in classifying the data.
9.16  Analytic Solver:
a. Optimal value of k: 7; Predicted outcome for the first new consumer: N (Do not
install solar panels)
b. Misclassification rate: 36.55%
c. Accuracy = 66.30%; Specificity = 0.7371; Sensitivity = 0.5548; Precision = 0.5912.
d. The plots show that the KNN classifier performs better than the baseline model.
Both the lift curve and ROC curve lie above the diagonal line. The AUC value is
 0.6929.
e. The predictive performance measures and graphs suggest that the KNN classifier
is moderately effective in classifying the data.
R:
a. Optimal value of k: 8; Predicted outcome for the first new consumer: N (Do not
install solar panels)
b. Misclassification rate: 0.3818
c. Accuracy = 0.6569; Specificity = 0.7535; Sensitivity = 0.5088; Precision = 0.5737.
d. The plots show that the KNN classifier performs better than the baseline model.
Both the lift curve and ROC curve lie above the diagonal line. The AUC value is
0.7022.
e. The predictive performance measures and graphs suggest that the KNN classifier
is moderately effective in classifying the data.
9.18  Analytic Solver:
a. Accuracy = 75%; Specificity = 0.7; Sensitivity = 0.7917; Precision = 0.76.
b. Lift value of the leftmost bar of the decile-wise lift chart: 0.9167; This implies that by
selecting the top 10% of the validation cases with the highest predicted probability
of belonging to the target class, the naïve Bayes model would identify 0.9167 times
as many target class cases as if the cases are randomly selected.
c. AUC = 0.725
d. Predicted outcome for the first three new observations: Yes, No, Yes
R:
a. Accuracy = 0.7674; Specificity = 0.6; Sensitivity = 0.913; Precision = 0.7241.
b. Lift value of the leftmost bar of the decile-wise lift chart: 1.17; This implies that by
selecting the top 19% of the validation cases with the highest predicted probability
of belonging to the target class, the naïve Bayes model would identify 1.17 times as
many target class cases as if the cases are randomly selected.
c. AUC = 0.7196
d. Predicted outcome for the first three new observations: Yes, No, Yes
9.20  Analytic Solver:
a. Accuracy = 69.23%; Specificity = 0.25; Sensitivity = 0.8889; Precision = 0.7273.
b. No, the lift curve does not lie above the baseline entirely.
c. AUC = 0.5417
d. Predicted outcome of the five new observations: 1, 0, 0, 1, 1
e. Accuracy = 46.15%; Specificity = 0.25; Sensitivity = 0.5556; Precision = 0.625; No,
the lift curve does not lie above the baseline entirely; AUC = 0.4236; The naïve
Bayes model that uses all four predictors performs better.
R:
a. Accuracy = 0.5; Specificity = 0.1; Sensitivity = 0.75; Precision = 0.5714.
b. No, the lift curve does not lie above the baseline entirely.
c. AUC = 0.6188
 d. Predicted outcome of the five new observations: 1, 1, 1, 0, 1
e. Accuracy = 0.6154; Specificity = 0.1; Sensitivity = 0.9375; Precision = 0.625; No,
the lift curve does not lie above the baseline entirely; AUC = 0.6344; The naïve
Bayes model that uses only x1 and x2 performs better.
9.22  Analytic Solver:
a. Accuracy = 67.09%; Specificity = 0.7143; Sensitivity = 0.6; Precision = 0.5625.
b. Yes, the entire lift curve lies above the baseline.
c. AUC = 0.8051

page 643
d. The predictive performance measures and graphs suggest that
the naïve Bayes classifier is effective in classifying the data.
R:
a. Accuracy = 0.7051; Specificity = 0.7273; Sensitivity = 0.6765; Precision = 0.6571.
b. Yes, the entire lift curve lies above the baseline.
c. AUC = 0.7871
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is moderately effective in classifying the data.
9.24  Analytic Solver:
a. Observation 1: x1 = 1, x2 = 2, x3 = 1; Observation 2: x1 = 1, x2 = 2, x3 = 1.
b. Accuracy = 68.75%; Specificity = 0.7826; Sensitivity = 0.4444; Precision = 0.4444.
c. AUC = 0.6787
d. Accuracy = 37.5%; Specificity = 0.1304; Sensitivity = 1; Precision = 0.3103.
R:
a. Observation 1: x1 = 1, x2 = 2, x3 = 1; Observation 2: x1 = 1, x2 = 2, x3 = 1.
b. Accuracy = 0.6774; Specificity = 0.8571; Sensitivity = 0.3; Precision = 0.5.
c. AUC = 0.4548
d. Accuracy = 0.2903; Specificity = 0.1429; Sensitivity = 0.6; Precision = 0.25.
9.26  Analytic Solver:
a. Observation 1: x1 = 1, x2 = 1, x3 = 2; Observation 2: x1 = 1, x2 = 1, x3 = 1.
b. Accuracy = 48.57%; Specificity = 0.8421; Sensitivity = 0.0625; Precision = 0.25.
c. AUC = 0.3664
d. Accuracy = 41.43%; Specificity = 0.1316; Sensitivity = 0.75; Precision = 0.4211.
R:
a. Observation 1: x1 = 1, x2 = 1, x3 = 2; Observation 2: x1 = 1, x2 = 1, x3 = 1.
b. Accuracy = 0.5797; Specificity = 1; Sensitivity = 0; Precision = NaN.
c. AUC = 0.5384
d. Accuracy = 0.5507; Specificity = 0.4750; Sensitivity = 0.6552; Precision = 0.4750.
9.28  Analytic Solver:
 a. Accuracy = 75%; Specificity = 0.5263; Sensitivity = 0.92; Precision = 0.7188.
b. AUC = 0.8189
c. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is effective in classifying the data.
R:
a. Accuracy = 0.7907; Specificity = 0.5882; Sensitivity = 0.9231; Precision = 0.7742.
b. AUC = 0.759
c. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is moderately effective in classifying the data.
9.30  Analytic Solver:
a. Accuracy = 70.39%; Specificity = 0.1639; Sensitivity = 0.9421; Precision = 0.7186.
b. Lift value of the leftmost bar of the decile-wise lift chart: 1.22. This implies that by
selecting the top 10% of the validation cases with the highest predicted probability
of belonging to the target class, the naïve Bayes model would identify 1.22 times as
many target class cases as if the cases are randomly selected.
c. AUC = 0.6223
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is only slightly effective in classifying the data.
R:
a. Accuracy = 69.35%; Specificity = 0; Sensitivity = 1; Precision = 0.6935.
b. Lift value of the leftmost bar of the decile-wise lift chart: 1.20. This implies that by
selecting the top 11% of the validation cases with the highest predicted probability
of belonging to the target class, the naïve Bayes model would identify 1.20 times as
many target class cases as if the cases are randomly selected.
c. AUC = 0.6549
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is only slightly effective in classifying the data.
9.32  Analytic Solver:
a. Accuracy = 81.4%; Specificity = 0.9551; Sensitivity = 0.6593; Precision = 0.9305.
b. Lift value of the leftmost bar of the decile-wise lift chart: 1.92; AUC = 0.8298.
c. Predicted outcome for the first three individuals: 0 (Do not volunteer), 0, 1
(Volunteer)
R:
a. Accuracy = 0.8184; Specificity = 0.9712; Sensitivity = 0.6517; Precision = 0.9541.
b. Lift value of the leftmost bar of the decile-wise lift chart: 1.98; AUC = 0.8481.
c. Predicted outcome for the first three individuals: 0 (Do not volunteer), 0, 1
(Volunteer)
9.34  Analytic Solver:
a. Observation 1: Age = 1, Hours = 2; Observation 2: Age = 1, Hours = 1.
b. Accuracy = 48.75%; Specificity = 0.4079; Sensitivity = 0.5595; Precision = 0.5109.
c. AUC = 0.4865
 d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is not effective in classifying the data.
R:
a. Observation 1: Age = 1, Hours = 2; Observation 2: Age = 1, Hours = 1.
b. Accuracy = 0.566; Specificity = 0.28; Sensitivity = 0.8214; Precision = 0.5610.
c. AUC = 0.5562
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is not effective in classifying the data.
9.36  Analytic Solver:
a. Accuracy = 66.25%; Specificity = 0.375; Sensitivity = 0.95; Precision = 0.6032.
b. Yes, the entire lift curve lies above the baseline.

page 644
c. AUC = 0.8009
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is effective in classifying the data.
R:
a. Accuracy = 0.6835; Specificity = 0.6944; Sensitivity = 0.6744; Precision = 0.725.
b. Yes, the entire lift curve lies above the baseline.
c. AUC = 0.7565
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is moderately effective in classifying the data.
9.38  Analytic Solver:
a. Accuracy = 71.25%; Specificity = 0.75; Sensitivity = 0.6923; Precision = 0.8372.
b. The plots show that the naïve Bayes classifier performs better than the baseline
model. Both the lift curve and ROC curve lie above the diagonal line.
c. AUC = 0.8310
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is effective in classifying the data.
R:
a. Accuracy = 0.7975; Specificity = 0.5667; Sensitivity = 0.9388; Precision = 0.7797.
b. The plots show that the naïve Bayes classifier performs better than the baseline
model. Both the lift curve and ROC curve lie above the diagonal line.
c. AUC = 0.802
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is effective in classifying the data.
9.40  Analytic Solver:
a. Observation 1: Age = 2, Education = 2, Hours = 1; Observation 2: Age = 2,
Education = 1, Hours = 1.
b. Accuracy = 87.5%; Specificity = 0.854; Sensitivity = 0.8929; Precision = 0.8772.
c. AUC = 0.8531
 d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is effective in classifying the data.
R:
a. Observation 1: Age = 2, Education = 2, Hours = 1; Observation 2: Age = 2,
Education = 1, Hours = 1.
b. Accuracy = 0.7864; Specificity = 0.7556; Sensitivity = 0.8103; Precision = 0.8103.
c. AUC = 0.7395
d. The predictive performance measures and graphs suggest that the naïve Bayes
classifier is moderately effective in classifying the data.

Chapter 10
10.2  a. Possible split points for days: {191, 208.5, 217.5, 229, 249.5, 287.5}
b. Possible split points for precipitation: {24.9, 35.55, 38, 40.95, 48.25, 58.2}
10.4  a. Gini index for root node: 0.2888
b. Gini index for age < 45.5: 0.1958
c. Gini index for age ≥ 45.4: 0.3648
d. Gini index for the split: 0.2803
e. Rules generate from this split: If age < 45.5, then the probability of a diabetes
diagnosis is 11%. If age ≥ 45.5, then the probability of a diabetes diagnosis is 24%.
10.6  a. C
b. B
c. A
10.8  a. Minimum error: 0.1917; Number of decision nodes: 2.
b. Number of leaf nodes in the best-pruned tree: 3; Number of leaf nodes in the
minimum error tree: 3.
c. Predictor variable: x2; Split value: 63.5.
d. Accuracy = 83.75%; Sensitivity = 0.9762; Specificity = 0.6842; Precision = 0.7736.
e. Yes. The entire lift curve lies above the baseline.
f. AUC = 0.8365
g. Predicted response values of the new observations: 1, 1, 0, 1, 1; Class 1 probability
of the first observation: 0.7887.
10.10a. Number of leaf nodes: 8; Predictor variable: x2; Split value: 62
b. cp value = 0.0075188; Number of decision nodes: 2
c. No
d. Number of leaf nodes: 3
e. Accuracy = 0.8235; Sensitivity = 1; Specificity = 0.625; Precision = 0.75
f. Lift value of the leftmost bar of the decile-wise lift chart: 1.42. This implies that by
selecting the top 71% of the validation cases with the highest predicted probability
 of belonging to the target class, the model would identify 1.42 times as many target
class cases as if the cases are randomly selected.
g. AUC = 0.8125
h. Predicted response values of the new observations: 1, 1, 0, 1, 1; Class 1
probability of the first observation: 0.7604
10.12a. Number of leaf nodes: 6; Predictor variable: CreditCard; Split value: “Yes” (whether
or not the customer has credit card debt).
b. cp value = 0.0357143; Number of splits: 2.
c. No
d. Number of leaf nodes: 3
e. Accuracy = 0.7119; Sensitivity = 0.8696; Specificity = 0.6111; Precision = 0.5882.
f. The model is moderately effective because 1) the lift value of the leftmost bar of the
decile-wise lift chart = 1.51, 2) the lift curve lies above the diagonal line, and 3) AUC
= 0.7506
g. Probability of the first new customer: 0.0714; Probability of the second new
customer: 0.2593
10.14a. Number of leaf nodes of the best-pruned tree: 3; Predictive variable: Income; Split
value: 135,000; Number of leaf nodes of the minimum error tree: 5.
b. Accuracy = 97.25%; Sensitivity = 0.8182; Specificity = 0.9864.

page 645
c. The model is effective because 1) the lift value of the leftmost bar
of the decile-wise lift chart = 8.18, 2) the lift curve lies above the diagonal line, and
3) AUC = 0.9352.
d. Probability of the first community member: 0.016; Probability of the second
community member: 0.9589.
10.16a. Number of leaf nodes: 3; Predictor variable: Age; Split value: 35; Rule: If Age is at
least 35, then the individual is more likely to go to church.
b. cp value = 0.00233508; Number of splits: 27; Number of leaf nodes: 28.
c. Yes; Number of splits: 18; cp value = 0.00291886.
d. Accuracy = 0.5317; Sensitivity = 0.4482; Specificity = 0.6118; Precision = 0.5256.
e. AUC = 0.5467; The model is not effective because 1) the lift value of the leftmost
bar of the decile-wise lift chart = 1.05, 2) the lift curve lies only slightly above the
diagonal line, and 3) AUC = 0.5467.
f. 50%
10.18a. Number of leaf nodes: 14; Rule: If the individual is 44 or older with an income of
less than 15000, the individual is likely to download the mobile banking app.
b. cp value = 0.0085714; Number of splits: 4.
c. No
d. Accuracy = 0.5503; Sensitivity = 0.68; Specificity = 0.4189; Precision = 0.5426.
e. Lift value of the leftmost bar of the decile-wise lift chart: 1.02. This implies that by
selecting the top 30% of the validation cases with the highest predicted probability
 of belonging to the target class, the model would identify 1.02 times as many target
class cases as if the cases are randomly selected.
f. AUC = 0.5359
g. Number of new customers who are likely to download: 12; Probability of the first
customer downloading: 0.6125.
10.20a. Number of leaf nodes in the best-pruned tree: 2; Number of leaf nodes in the
minimum error tree: 2; Rules: If GPA is less than 2.65, then the student is not likely
to graduate within 4 years. If GPA is at least 2.65, then the student is likely to
graduate within 4 years.
b. Accuracy = 90.25%; Sensitivity = 1; Specificity = 0.7045; Precision = 0.8730.
c. The lift curve lies above the baseline.
d. AUC = 0.8523
e. Students 2 and 3
10.22a. Predictor variable: Income; Split value: 64.
b. cp value = 0.0122180; Number of splits: 2.
c. No
d. Accuracy = 0.8151; Sensitivity = 1; Specificity = 0.6071; Precision = 0.7412.
e. The model is effective because 1) the lift value of the leftmost bar of the decile-wise
lift chart = 1.4, 2) the lift curve lies above the diagonal line, and 3) AUC = 0.8036.
f. Probability of the first gamer: 0.7725; Probability of the second gamer: 0.0714.
10.24a. Possible split values for x1: {179.5, 252, 289}
b. Possible split values for x2: {66, 92.5, 105.5}
c. Possible split values for x3: {6.25, 9.75, 14.25}
d. MSE of the partition x1 = 252:

25

72.25

48.625

e. MSE of the partition x2 = 92.5:

19.5)2] = 20.25

50.625
 f. MSE of the partition x3 = 14.25:

(14 − 20.3333)2 + (32 − 20.3333)2] = 68.2222

0 = 51.1667

g. Because split x1 = 252 generates the lowest MSE, the best split is on x1, and the
best split point is x1 = 252.
10.26a. The split x1 = 7.5 will generate the least MSE:

(25 − 49.5)2] + (1/3) × (86 − 86)2 =

400.1667
b. The split x2 = 22.5 will generate the least MSE:

(1/2) × [(86 − 80)2 + (74 − 80)2] = 24

c. The split x3 = 15 will generate the least MSE:

(25 − 49.5)2] + (1/3) × (86 − 86)2 =

400.1667
d. Because the split x2 = 22.5 has the lowest MSE, the best split is on x2, and the
best split value is 22.5.
e. Rules: If x2 < 22.5, then y = 25; if x2 ≥ 22.5, then y = 49.5.
10.28a. Minimum MSE: 27.8093; Number of decision nodes: 2
b. Number of leaf nodes in the best-pruned tree: 3; Number of leaf nodes in the
minimum error tree: 3.
c. Predictor variable: x5; Split value: 7.58; Rules: If x5 < 7.58, then y = 36.13; If x5 ≥
7.58, then y = 19.57.
d. RMSE = 5.8746; MAD = 4.5801.
e. Predicted value of the first observation: 15.47
f. Minimum = 15.47; Maximum = 36.1267; Average = 21.15.
10.30a. Number of leaf nodes: 8
b. Predictor variable: x5; Split value: 11; Rules: If x5 < 11, then y = 32; If x5 ≥ 11, then
y = 19.
c. Tree with the lowest cross-validation error: 21st tree; Number of splits: 28.
d. Yes; cp value = 0.027056.
 e. Number of leaf nodes: 7
f. ME = −0.0074; RMSE = 5.6235; MAE = 3.9277; MPE = −3.9765; MAPE = 18.8877.

page 646
g. Minimum = 13.47; Maximum = 27.25; Average = 18.90.
10.32a. Minimum MSE: 46.1824; Number of decision nodes: 2.
b. Number of leaf nodes of the best-pruned tree: 2; Number of leaf nodes of the
minimum error tree: 3
c. Predictor variable: x3; Split value: 102.53; Rules: If x3 < 102.53, then y = 299.18; If
x3 ≥ 102. 53, then y = 312.81.
d. RMSE = 7.9182; MAD = 6.7371.
10.34a. Number of leaf nodes of the best-pruned tree: 7; Number of leaf nodes of the
minimum error tree: 7; Rules: If CreditCard = No, then TravelSpend = 2764.35; If
CreditCard = Yes, then TravelSpend = 1502.48.
b. RMSE = 1011.3150; MAD = 740.4973
c. Predicted annual travel spending of the first customer: $1,747.83; Predicted annual
travel spending of the second customer: $2,128.47
10.36a. Predictor variable: SQFT; Split value: 2221; Rules: If SQFT < 2221, then Price =
970,000; If SQFT ≥ 2221, then Price = 1,700,000.
b. cp value = 0.01878; Number of splits: 9.
c. Yes; cp value = 0.023827.
d. Number of leaf nodes: 6
e. ME = −51189.13; RMSE = 467558.6; MAE = 332553; MPE = −34.90362; MAPE =
53.6871; Over-Predict because of the negative ME; The model is not very effective
due to the large prediction errors.
f. Predicted price of the first house: $829, 762.6; Predicted price of the second house:
$1,190,534.5.
10.38a. Number of leaf nodes: 4
b. RMSE = 27.2155; MAD = 21.4680.
c. Mean = 221.947; Median = 234.857.
10.40a. Predictor variable: Generation; Split value: 17; Rules: If Generation < 17, Price ≥
12, then Sales = 7.7; If Generation < 17, Price < 12, the Sales = 13; If Generation ≥
17, then Sales = 18.
b. cp value = 0.018209; Minimum validation error = 0.27671; Number of leaf nodes: 5.
c. Yes; cp value = 0.019442.
d. Number of leaf nodes: 4
e. ME = 0.0523; RMSE = 3.1054; MAE = 2.4323; MPE = −4.4364; MAPE = 17.9348.
f. Predicted per capita electricity sales: 16.6075
10.42a. Number of leaf nodes in the best-pruned tree: 11; Number of leaf nodes in the
minimum error tree: 11.
b. Predictor variable: games_started; Split value: 186.
 c. RMSE = 4,636,301.156; MAD = 2,856,216.137.
d. Average predicted salary = 8,174,731.90
10.44 Analytic Solver:
a. Accuracy = 84.375%; Sensitivity = 1; Specificity = 0.8214.
b. AUC = 0.9375
c. Predicted value: 1; Class 1 probability: 0.5.
R:
a. Accuracy = 0.7742; Sensitivity = 0.5; Specificity = 0.84.
b. AUC = 0.87
c. Predicted value: 1; Class 1 probability: 0.56.
10.46  Analytic Solver:
a. Accuracy = 84.375%; Sensitivity = 1; Specificity = 0.8214.
b. AUC = 0.9241
c. Most important predictor: x1
d. Predicted value: 1; Class 1 probability: 0.5.
R:
a. Accuracy = 0.8387; Sensitivity = 0.5; Specificity = 0.92.
b. AUC = 0.9
c. Most important predictor: x4
d. Predicted value: 0; Class 1 probability: 0.43.
10.48  Analytic Solver:
a. Accuracy = 73.75%; Sensitivity = 0.7674; Specificity = 0.7072.
b. Lift value of the leftmost bar of the decile-wise lift chart: 1.8605
c. Predicted value: 1; Class 1 probability: 1.
R:
a. Accuracy = 0.8608; Sensitivity = 0.8250; Specificity = 0.8974.
b. Lift value of the leftmost bar of the decile-wise lift chart: 1.98
c. Predicted value: 1; Class 1 probability: 1.
10.50  Analytic Solver:
a. Accuracy = 84.50%; Sensitivity = 0.8144; Specificity = 0.8738; AUC = 0.8864.
b. Accuracy = 83%; Sensitivity = 0.8351; Specificity = 0.8252; AUC = 0.8865; Most
important predictor: Characters.
c. 53.85% of the cases are predicted as spams.
R:
a. Accuracy = 0.8291; Sensitivity = 0.7864; Specificity = 0.8750; AUC = 0.8924.
b. Accuracy = 0.8442; Sensitivity = 0.8252; Specificity = 0.8646; AUC = 0.8912; Most
important predictor: Hyperlinks.
c. 53.85% of the cases are predicted as spams.
10.52 Analytic Solver:
a. Accuracy = 66.67%; Sensitivity = 0.6341; Specificity = 0.7097.
b. Accuracy = 62.50%; Sensitivity = 0.6098; Specificity = 0.6452.
c. The bagging tree shows more robust performance because of higher AUC value
(0.7352 vs. 0.6943).
R:
a. Accuracy = 0.7324; Sensitivity = 0.7857; Specificity = 0.6552.
b. Accuracy = 0.7042; Sensitivity = 0.8095; Specificity = 0.5517.
c. The bagging tree shows more robust performance because of higher AUC value
(0.7878 vs. 0.757).

page 647
10.54 Analytic Solver:
a. Accuracy = 70.39%; Sensitivity = 0.9421; Specificity = 0.1639; AUC = 0.5560.
b. Accuracy = 70.39%; Sensitivity = 0.9421; Specificity = 0.1639; AUC = 0.5560; Most
important predictor: CollegeParent.
c. Both models have the same performance.
R:
a. Accuracy = 0.7186; Sensitivity = 0.9402; Specificity = 0.2172; AUC = 0.5968.
b. Accuracy = 0.7186; Sensitivity = 0.9402; Specificity = 0.2172; AUC = 0.5968; Most
important predictor: GPA.
c. Both models have the same performance.
10.56 Analytic Solver:
a. Accuracy = 71.25%; Sensitivity = 0.6562; Specificity = 0.75; AUC = 0.6911.
b. The single-tree model is more robust due to higher accuracy, sensitivity, and AUC
values.
c. Probability of the first customer having plans to travel: 0.1; Probability of the second
customer having plans to travel: 0.
R:
a. Accuracy = 0.7215; Sensitivity = 0.5806; Specificity = 0.8125; AUC = 0.7547.
b. The bagging ensemble model is more robust due to higher accuracy, specificity,
and AUC values.
c. Probability of the first customer having plans to travel: 0.01; Probability of the
second customer having plans to travel: 0.14.
10.58 Analytic Solver:
a. Accuracy = 54.95%; Sensitivity = 0.7996; Specificity = 0.3078; AUC = 0.5613; Most
important predictor: Income.
b. The random trees ensemble mode is slightly more robust due to higher accuracy,
sensitivity, and AUC values.
 c. 100% of the individuals in the data set are likely to go to church.
R:
a. Accuracy = 0.5223; Sensitivity = 0.4632; Specificity = 0.5788; AUC = 0.5382; Most
important predictor: Age.
b. The random trees ensemble mode is slightly more robust due to higher accuracy,
sensitivity, specificity, and AUC values.
c. 33.33% of the individuals in the data set are likely to go to church.
10.60 Analytic Solver:
a. Accuracy = 84.38%; Sensitivity = 0.8523; Specificity = 0.8333; AUC = 0.9123.
b. The boosting ensemble model is more robust due to higher AUC value.
c. Probability of the first gamer making in-app purchases: 0.3945; Probability of the
second gamer making in-app purchases: 0.2082.
R:
a. Accuracy = 0.8428; Sensitivity = 0.8571; Specificity = 0.8267; AUC = 0.9135.
b. The boosting ensemble model is more robust due to higher accuracy, specificity,
and AUC values.
c. Probability of the first gamer making in-app purchases: 0.4908; Probability of the
second gamer making in-app purchases: 0.4199.

Chapter 11
11.2  Analytic Solver: At most 6 clusters if the minimum distance between clusters is 5; At
most 10 clusters if the minimum distance between clusters is 3.
R: At most 6 clusters if the minimum distance between clusters is 5; At most 10
clusters if the minimum distance between clusters is 3.
11.4  Analytic Solver:
a. Number of observations in the largest cluster (Cluster 1): 42; Average value of x4 of
the largest cluster (Cluster 1): 141.1191.
b. Number of observations in the largest cluster (Cluster 2): 22; Average value of x4 of
the largest cluster (Cluster 1): 157.3636.
c. Number of observations in the largest cluster (Cluster 2): 21; Average value of x4 of
the largest cluster (Cluster 1): 158.1905.
R:
a. Number of observations in the largest cluster (Cluster 1): 42; Average value of x4 of
the largest cluster (Cluster 1): 141.1191.
b. Number of observations in the largest cluster (Cluster 2): 22; Average value of x4 of
the largest cluster (Cluster 1): 157.3636.
c. Number of observations in the largest cluster (Cluster 2): 21; Average value of x4 of
the largest cluster (Cluster 1): 158.1905.
11.6  Analytic Solver: At most 3 clusters if the minimum distance between clusters is 0.8.
 R: At most 3 clusters if the minimum distance between clusters is 0.8.
11.8  At most 3 clusters if the minimum distance between clusters is 0.8. The average value
of x5 is 13.14.
11.10 Analytic Solver:
a. At most 10 clusters if the minimum distance between clusters is 0.8; Number of
transfers in the largest cluster (Cluster 1): 30.
b. At most 10 clusters if the minimum distance between clusters is 0.8; Number of
students on the Dean’s list in the largest cluster (Cluster 1): 35.
R:
a. At most 5 clusters if the minimum distance between clusters is 0.8; Number of
transfers in the largest cluster (Cluster 3): 12.
b. At most 3 clusters if the minimum distance between clusters is 0.8; Number of
students on the Dean’s list in the largest cluster (Cluster 1): 31.
11.12 Cluster 1: 142 players; Cluster 2: 1 player; Cluster 3: 1 player.
11.14 Analytic Solver:
a. Yes. Variables are measured using different scales.

page 648
b. Cluster 1, which includes 31 countries, has the highest average
GNI per capita (19439.6774), Cluster 3, which includes 3 countries, has the 2nd
highest average GNI per capita (1106.6667), and Cluster 2, which includes 4
countries, has the lowest average GNI per capita (392.5).
R:
a. Yes. Variables are measured using different scales.
b. Cluster 1, which includes 31 countries, has the highest average GNI per capita
(19439.6774), Cluster 3, which includes 3 countries, has the 2nd highest average
GNI per capita (1106.6667), and Cluster 2, which includes 4 countries, has the
lowest average GNI per capita (392.5).
11.16a. Compared to Cluster 1, customers in Cluster 2 tend to be older (average age of
41.4706 vs. 24.5385), have higher income (average income of 73117.6471 vs.
28307.6923), have larger families (household size of 3.8824 vs. 2.0769), and spend
more on pizzas annually (average spending of 1011.5882 vs. 290.4615).
b. Compared to Cluster 1, Customers in Cluster 2 are more likely to be married (15 vs.
6) and own houses (15 vs. 0).
11.18 Analytic Solver:
a. Number of cities in the largest cluster (Cluster 1): 28.
b. Average January average temperature: 31.6286; Average April average
temperature: 51.7429; Average July average temperature: 73.8357; Average
October average temperature: 55.1821
R:
a. Number of cities in the largest cluster (Cluster 1): 33.
 b. Average January average temperature: 33.1455; Average April average
temperature: 53.1030; Average July average temperature: 74.7727; Average
October average temperature: 56.3939.
11.20 Analytic Solver:
a. Yes. Variables are measured using different scales.
b. At most five clusters are generated if the minimum distance between clusters is 10.
c. Number of community areas in the largest cluster: 33; Average median household
income of the largest cluster (Cluster 2): 45228.0606.
R:
a. Yes. Variables are measured using different scales.
b. At most two clusters are generated if the minimum distance between clusters is 10.
c. Number of community areas in the largest cluster: 43; Average median household
income of the largest cluster (Cluster 2): 45449.44.
11.22a. Number of individuals in the largest cluster (Cluster 1): 28.
b. Cluster 1: low weight, low income; Cluster 2: non-Christian; Cluster 3: high income;
Cluster 4: non-white, large family.
11.24 Analytic Solver:
With k = 2, size of the larger cluster (Cluster 1): 23; Average distance for the larger
cluster (Cluster 1): 0.7895.
With k = 3, size of the largest cluster (Cluster 3): 24; Average distance for the
largest cluster (Cluster 3): 0.8002.
With k = 4, size of the largest cluster (Cluster 4): 21; Average distance for the
largest cluster (Cluster 4): 0.6147.
R:
With k = 2, size of the larger cluster (Cluster 2): 27; Average distance for the larger
cluster (Cluster 2): 0.8629.
With k = 3, size of the largest cluster (Cluster 3): 19; Average distance for the
largest cluster (Cluster 3): 0.5980.
With k = 2, size of the largest cluster (Cluster 3): 16; Average distance for the
largest cluster (Cluster 4): 0.5807.
11.26a. Average silhouette width: 0.54
b. Average silhouette width: 0.34
c. Average silhouette width: 0.33
11.28a. Average silhouette width: 0.66
b. Average silhouette width: 0.54
c. Average silhouette width: 0.58
11.30a. Average silhouette width: 0.26
b. Average silhouette width: 0.39
c. Average silhouette width: 0.46
11.32 Analytic Solver:
 a. Size of the largest cluster (Cluster 2): 27; Cluster center values of the largest
cluster: Comp = 0.6922, Att = 0.7019, Pct = 0.3269, Yds = 0.6784, Avg = 0.2571,
Yds/G = 0.3333, TD = 0.5974, Int = 0.4804.
b. Size of the largest cluster (Cluster 1): 14; Cluster center values of the largest
cluster: Comp = 0.6531, Att = 0.5442, Pct = 0.7095, Yds = 0.8063, Avg = 0.9606,
Yds/G = 0.8719, TD = 0.8139, Int = −0.2428.
c. The clustering structures are similar for players in both data sets. The largest
clusters in the two clustering structures contain better performing players with more
touchdowns.
R:
a. Size of the largest cluster (Cluster 1): 16; Cluster center values of the largest
cluster: Comp = −1.3499, Att = −1.3296, Pct = −1.1689, Yds = −1.3121, Avg =
−0.7293, Yds/G = −0.9251, TD = −1.1840, Int = −0.8474.
b. Size of the largest cluster (Cluster 3): 13; Cluster center values of the largest
cluster: Comp = 0.3080, Att = 0.3344, Pct = 0.1511, Yds = 0.1018, Avg = −0.3667,
Yds/G = −0.6465, TD = 0.2621, Int = 0.6625.
c. The clustering structures are similar for players in both data sets. The first cluster
includes the lowest performing quarterbacks, the second cluster includes the
highest performing quarterbacks, and the third cluster includes the medium
performing quarterbacks.

page 649
11.34 Analytic Solver:
a. Size of the largest cluster (Cluster 2): 56; Cluster with the highest number of
homeruns: Cluster 3.
b. Size of the largest cluster (Cluster 1): 30; Cluster with the highest number of
homeruns: Cluster 3.
R:
a. Size of the largest cluster (Cluster 1): 63; Cluster with the highest number of
homeruns: Cluster 1.
b. Size of the largest cluster (Cluster 2): 22; Cluster with the highest number of
homeruns: Cluster 1.
11.36a. Yes. Variables are measured using different scales.
b. Cluster characteristics and GNI:
Cluster 1: High population growth, low % of female population, high % of male
population. GNI = 17930.
Cluster 2: Relatively high population growth, high fertility rate, high birth rate. GNI
= 270.
Cluster 3: Medium on all indicators. GNI = 18663.2258.
Cluster 4: Low population growth, high % of female population, low % of male
population, low fertility rate, how birth rate. GNI = 4975.
11.38 Cluster characteristics:
Cluster 1: Food items that are high in calories, total fat, potassium, carbohydrate,
dietary fiber.
Cluster 2: Food items that are high in saturated fat, cholesterol, sodium, and
protein, but low in carbohydrate, dietary fiber, and sugar.
Cluster 3: Food items that are high in sugar but low in calories, total fat, saturated
fat, potassium, and protein.
11.40a. No. All measures are percentage growth.
b. Size of the largest cluster (Cluster 1): 6; The cluster that has the highest average
growth in GDP per capita: Cluster 2.
11.42 Analytic Solver:
a. Cluster characteristics:
Cluster 1: Medium on all indicators.
Cluster 2: Youngest, lowest income, lowest usage, and shortest tenure.
Cluster 3: Older, medium income, high usage, and medium tenure.
Cluster 4: Medium age, high income, high usage, and longer tenure.
b. Percent of unsubscribers: Cluster 1 (43.90%), Cluster 2 (93.94%), Cluster 3 (0%),
Cluster 4 (0%); Cluster with the highest percent of unsubscribers: Cluster 2.
R:
a. Cluster characteristics:
Cluster 1: Relatively younger, lower income, lower usage, and shorter tenure.
Cluster 2: Youngest, lowest income, lowest usage, and shortest tenure.
Cluster 3: Oldest, highest income, highest usage, and longest tenure.
Cluster 4: Relatively older, higher income, higher usage, and longer tenure.
b. Percent of unsubscribers: Cluster 1 (46.15%), Cluster 2 (93.94%), Cluster 3 (0%),
Cluster 4 (0%); Cluster with the highest percent of unsubscribers: Cluster 2.
11.44a. Lift ratio of the top rule ({b} => {c}): 1.2277
b. The support count of 11 implies that 11 of the transactions include both {b} and {c}.
c. Lift ratio of the top rule ({g,e} => {f}): 2.5
d. The confidence of 100 implies that 100% of the transactions that include {g, e} also
include {f}.
11.46a. The least frequent item: d
b. Lift ratio of the top rule ({e} => {c}): 1.3333
c. Lift ratio of the top rule ({c,d}=>{e}: 2.6667
11.48.a. Number of rules generated: 10; Top rule: {c,g} => {e}; Lift ratio of the top rule:
2.3602
b. Number of rules generated: 2; Top rule: {e} => {c}; Lift ratio of the top rule: 1.5633
11.50a. The least frequent item: e
b. Lift ratio of the top rule ({e} => {c}): 1.4006; The lift ratio implies that identifying a
transaction with item set {e} as one which also contains item set {c} is 40.06%
 better than just guessing that a random transaction contains {c}.
11.52a. Lift ratio of the top rule ({horror} => {action}): 1.6369; The lift ratio implies that
identifying someone who watches a horror movie as one who also is going to watch
an action movie is 63.69% better than just guessing that a random individual is
going to watch an action movie.
b. Lift ratio of the top rule ({drama, horror} => {action, comedy}: 2.6190; The lift ratio
implies that identifying someone who watches drama and horror movies as one
who also is going to watch action and comedy movies is 161.90% better than just
guessing that a random individual is going to watch action and comedy movies.
11.54 Lift ratio of the top rule ({watermelon} => {cherry}): 1.3285; The lift ratio implies that
identifying someone who purchases watermelon as one who also is going to
purchase cherry is 32.85% better than just guessing that a random individual is
going to purchase cherry.
Lift ratio of the 2nd top rule ({watermelon} => {banana}): 1.1752; The lift ratio
implies that identifying someone who purchases watermelon as one who also is
going to purchase banana is 17.52% better than just guessing that a random
individual is going to purchase banana.
Lift ratio of the third top rule ({orange} => {apple}): 1.0417; The lift ratio implies that
identifying someone who purchases orange as one who also is going to purchase
apple is 4.17% better than just guessing that a random individual is going to
purchase apple.
11.56a. Most frequently downloaded song: Here Comes the Sun
b. Lift ratio of the top rule ({All You Need Is Love, Here Comes the Sun, Yellow
Submarine} => {A Day in the Life}): 1.7730; The lift ratio implies that identifying
someone who downloads All You Need Is Love, Here Comes the Sun, and Yellow
Submarine as one who also is going to download A Day in the Life is 77.30% better
than just guessing that a random individual is going to download A Day in the Life.

page 650
11.58a. Number of rules generated: 11
b. Crime most likely to be committed in the department store: Theft; Crime most likely
to be committed on the sidewalk: Narcotics; Crime most likely to be committed in
apartments: Battery.
11.60a. Number of rules generated: 8
b. Lift ratio of the top rule ({horror} => {action}): 1.2766; The lift ratio implies that
identifying someone who uses Snapchat as one who also uses Instagram is
27.66% better than just guessing that a random individual uses Instagram.
11.62a. Number of rules: 4
b. Lift ratio of the top rule ({HELOC} => {Checking}): 1.2214; The lift ratio implies that
identifying someone who has a HELOC account as one who also has a checking
account is 22.14% better than just guessing that a random individual has a
checking account.
Chapter 12
12.2  a. 127.8867
b. 146.0634
c. The exponential smoothing model is preferred because it leads to the smallest
value for MSE (551.54 < 649.95), MAD (19.94 < 20.40), and MAPE (14.03 <
14.46).
12.4  a. The 3-period moving average model is preferred because it leads to the smallest

values for MSE = 0.0009, MAD = 0.0244, and MAPE = 2.0832%;

b. The α = 0.6 model is preferred because it leads to the smallest values for MSE =
0.0007, MAD = 0.0213, and MAPE = 1.6115%;

12.8  
a.

b.

12.10
a. 1124.83d3 + 189.4201t; positive

trend; the revenue consistently higher in the first three quarters, especially the 2nd
quarter.

b.

12.12
a. 0.2330d3 + 0.0549d4 + 1.3323d5 + 0.5542d6

− 2.5573d7 − 3.6466d8 − 3.6025d9 − 3.2695d10 − 1.9032d11 + 0.3337t; positive

trend; consumer sentiment consistently higher in the first six months of the year.

b.

12.14a. The quadratic trend seems to fit the data well.

b.

The quadratic model is preferred because it has the higher adjusted R2 (0.9550 >
0.8349)
12.16
a.

0.0007t3

b. The cubic model is preferred because it has the higher adjusted R2 (0.9886 >
0.9741); 3,197.89

12.20
a.

b.

12.22
a. − 10589.2436d3 + 4219.4278t +

27.9857t2

b.

12.26
a.

b. The quadratic trend model is preferred because it leads to the smallest value for
MSE, MAD, and 0.0164 × 592 = 126.07

12.28
a.
 b. The exponential trend model is preferred because it leads to the smaller value for

MSE, MAD, and MAPE;

12.30
a. Linear: 924.0796d2 − 671.7664d3 + 229.3586t

b. The exponential trend model is preferred because it leads to the smallest value for
MSE, MAD, and MAPE;

12.32
a. 30045.4236d2 + 9114.5188d3 +

9731.0611d4 + 19083.6035d5 + 17153.7458d6 + 22072.4882d7 + 20695.8306d8 −

959.0271d9 + 13745.7153d10 − 5750.7424d11 + 360.2576t

b. The exponential trend model is preferred because it leads to the smallest value for
MSE, MAD, and MAPE;

12.36 The computer-generated model is preferred.

Analytic Solver:

12.38 The computer-generated model is preferred.

 page 651
12.40 The multiplicative model is preferred.

12.42 The additive model is preferred.

Chapter 13
13.2  Excel: The range of simulated observations is 13.8218.
R: The range of simulated observations is 13.69922.
13.6  Excel: Mean = 8.9700, Standard deviation = 1.3961, Maximum = 12, Minimum = 4
R: Mean = 8.9980, Standard deviation = 1.4609, Maximum = 12, Minimum = 5
13.10 Excel: Average demand = 11.0259 units, Standard deviation = 4.2025
R: Average demand = 11.1400 units, Standard deviation = 4.0186
13.16 Excel:
a. The likelihood of not meeting a weekly demand is 50.2000%. On average, the
opportunity cost is $8.5890 per week.
b. On average, the cost of having too many eggs is $8.5958 per week.
R:
a. The likelihood of not meeting a weekly demand is 46.60%. On average, the
opportunity cost is $8.9915 per week.
b. On average, the cost of having too many eggs is $8.39456 per week.
13.20 Excel: The likelihood of overstocking is 66.20% (331 out of 500 simulations), and the
likelihood of understocking is 33.80% (169 out of 500 simulations).
R: The likelihood of overstocking is 65.40% (327 out of 500 simulations), and the
likelihood of understocking is 34.60% (173 out of 500 simulations).
13.22 Excel: The average number of riders unable to get on the roller coaster is 7.06.
R: The average number of riders unable to get on the roller coaster is 7.755.
13.24a. Optimal solution: x1 = 20, x2 = 5, and z = 55
b. Both constraints are binding, and there is no slack.
c. Constraint 1: Shadow price = 0.3846, Range of feasibility = 93.3333 to 245.0000
Constraint 2: Shadow price = 0.1538, Range of feasibility = 32.8571 to 86.2500
d. Range of optimality: x1 = 1.7143 to 4.5000, x2 = 1.3333 to 3.5000
13.26a. Optimal solution: x1 = 3.9290, x2 = 6.4286, and z = 111.4286
 b. Both constraints are binding, and there is no surplus.
c. Constraint 1: Shadow price = 1.4286, Range of feasibility = 48 to 160
Constraint 2: Shadow price = 0.14286, Range of feasibility = 35.0000 to 116.6667
d. Range of optimality: x1 = 4.0000 to 13.3333, x2 = 9 to 30.
13.30a. Maximum revenue = $2,833.3333 by making 33.3333 liters of Chardonnay wine
and 22.2222 liters of Blanc de Blancs champagne.
b. Both grapes and processing time constraints are binding.
c. Grapes constraint: Shadow price = $5.8333, Range of feasibility = 300 to 480 kgs
Processing time constraint: Shadow price = $3.3333, Range of feasibility = 125 to
200 hours
d. Range of optimality: Chardonnay wine = $37.50 to $60.00, Blanc de Blancs
champagne = $41.25 to $66.00.
13.32 Consuming 2.1739 cups of milk and 2.7536 cups of cereal will meet the daily
requirements of calcium and vitamin D, at a minimum cost of $1.2134.
13.38a. x11 = 5, x12 = 60, x21 = 45, x22 = 0
b. z = 115.
13.42a. The lowest shipping cost is $238,000.
b. Manufacturer 1 should ship 6,600 batteries to Plant 2 and 3,400 batteries to Plant
3. Manufacturer 2 should ship 5,000 batteries to Plant 1 and 3,000 batteries to
Plant 3.
 page 652

INDEX

A
Absolute references, Excel, 614
Accuracy rate
binary choice (classification) models, 293–294
classification performance measure, 335, 336
Addition rule, 143
Adidas, 297
Adjusted R2, 225–227
Agglomerative categorical variable
multiple categories, linear regression and, 216–217
transforming numerical to, 398–399
Agglomerative clustering (AGNES algorithm), 478–480
defined, 477
dendrogram and, 479–480
Gower’s coefficient and, 485–487
linkage methods of, 478
mixed data, 485–488
numerical/categorical variables, 478–485
using Analytic Solver/R, 481–455
AGNES algorithm. See Agglomerative clustering
Airbnb, 535
Airline Passengers Rights, Health, and Safety, 164
Alpha (stock), 233
Alternative hypothesis, 194–196
Amazon (AMZN), 113, 535
Amazon.com, 531–532
Amazon Prime, 95, 133
American Heart Association, 262
Analysis of variance (ANOVA). See ANOVA
Analysis ToolPak, Excel. See Microsoft Excel Analysis ToolPak
ANOVA
linear regression model and, 224
table for regression, 230
Antecedent, association rule analysis, 502
Apple, Inc., 107, 147, 519–521, 534, 540, 541, 545, 548, 553
A priori method, association rule analysis, 503
Arithmetic mean, central location measure, 115
Artificial intelligence (AI), data mining and, 318
Ashenfelter, Orley, 287
Assigning probabilities, 141–142
Association, measures of
correlation coefficient, 123–124
covariance, 123
Association rule analysis
antecedent/consequent, If-Then, 502
A priori method, 503
assessing rules, 504–505
confidence of the association rule, 503–504
defined/uses of, 502
If-Then logical statements, 502
items/item set, 502
lift ratio, 504
perform, Analytic Solver/R, 506–508
support of the association rule, 503
Assumptions
common violation of OLS, 243–249. See also Linear regression model, detect/remedy
common violations of
independent, 390
linear regression model and, 241–243
Average, central location measure, 115
Average linkage method, agglomerative clustering and, 478
Average Man, The (Daniels), 564

B
Bagging, ensemble tree models, 454, 455, 456
Banerjee, Sunil, 25
Bankrate.com, 205
Bar chart, frequency distribution, categorical variable, 83–85
Bayes’ Theorem, total probability rule and, 148–151
BellKor, 6
Bell-shaped distribution, 129–133
Benchmark category, 213
Bernoulli process, 156–157
Best-pruned tree, 414
Beta (stock), 233
Big data
characteristics /veracity of, 12
defined, 12
structured/unstructured v., 13
writing with, 26–27, 77–78, 133–135, 172–173, 207–208, 253–254, 309, 364, 404–405,
469–471, 514, 556–557
See also Data; Visualize data
Billboard’s Hot 100, 26–27
Binary choice (classification) models, 288–289
accuracy rate of, 293–294
linear probability regression model, 289–290
logistic regression model, 290–293
Binary variable, 72–73
Binding constraints, 583
Binning, transforming numerical data, 58–62
Binomial distribution, 156–159, 564
Binomial random variable, 157
Blackboard, 14
Blacklists, 369
Bloomberg, 269
Boosting, ensemble tree models, 455, 456
Boston Red Sox, 25
Box-and whisker plot, 127–129
Boxplot, detecting outliers, 127–129
Bubble plot, visualize data, 107–108, 112
Bureau of Economic Analysis, 20
Bureau of Labor Statistics (BLS), 9, 20
Business analytics
defining/importance, 4
fashion retailers and, 6
gaming industry and, 6
healthcare and, 7
online subscriptions and, 6
sports and, 6–7
stages of, 4–5
Business intelligence (BI), 5
Business Week, 21

C
Cambridge Analytica, 7
Capital budgeting, integer programming optimization, 596–599
Car crash data set, 608–609
Case Study
analyze admission/enrollment decisions, predictor variables, 310–312
Billboard’s Hot 100, 26–27
car accidents that result in deaths, 514–516
cost, and optimization, 604–606
evaluate forecasting models, housing starts, 556–557
grading scales, 172–173
housing prices, comparing two areas, 134–135
income inequality, 207–208
performance measurement, 364–365
predicting college acceptance, naïve Bayes’ method, 405–407
predicting personal income, 469–471
predictive model for price of house, 254–256
Castle, Matt, 285
Categorical data, 15, 16
matching coefficient/Jaccard’s coefficient, similarity measures, 327–329
transforming, 69–75
Categorical variable, visualize relationships
bar chart, 83–85
contingency table, 96–97
frequency distribution for, 82–83
scatterplots, 105–106
page 653
Categorizing probabilities, 141–142
Category reduction, 69–72
Category scores, create, 74–75
Caterpillar, Inc., 237
Causal models, time series forecasting and, 533–534
Centers for Disease Control and Prevention (CDC), 164, 240
Central limit theorem (CLT)
for the sample mean, 181–182
for the sample proportion, 183
Central location, measures of, 115
 defined, 114
mean, types of, 115
median, 115
mode, 115
percentile, 118–119
Centroid method, agglomerative clustering, 478
Chance magazine, 287
Changing variability, detect/remedy, 246–247
Chartered Financial Analyst (CFA), 191–192, 298–299, 386–387
Charts, visualize data, guidelines, 91–92. See also Visualize data
Cheesecake Factory, 191
Chicago Tribune, 163
Children’s Online Privacy Protection Act, 174
CineMatch, 6
Classical linear regression model, 241–243
Classical probability, 142
Classification and regression trees (CART)
advantages/limitations, 370
analysis applications, 368
defined/uses, 412
full tree, 413–415
minimum error tree/best-pruned tree, 414
nodes of, 412
pure subset, 413
Classification models
k-nearest neighbor (KNN), 370–385
outcomes, summarizing, 368
performance charts for, 340–344
performance evaluation, 335–337
performance measures for, 335–336
Classification performance diagrams
cumulative lift chart, 340–341
decile-wise lift chart, 341
receiver operating characteristic (ROC) curve, 342–344
Classification summary table, k-nearest neighbor (KNN) and, 376
Classification trees
decision tree, building, 419–421
develop, Analytic Solver/R, 422–433
Gini impurity index and, 416–419
pruning, 421
split points, identify possible, 415–416
Cloud computing, 113
Cluster analysis
defined, 477
k-means, 492–493
Cluster plot, 496
CNBC, 147, 207, 298
CNN, 113, 163
CNNMoney, 542
Cobb-Douglas production function, 288
Coefficient of determination, R2, 224–225, 283
Coefficients, quadratic regression model, 273
College admission data set, 609–610
College Investor, The, 280
College Scorecard, 211, 235–236
Color, visualizing data and, 109–112
Combining events, 141
Competing hypothesis, 195
Complement of event, 141
Complement rule, 142–143
Complete linkage method, agglomerative clustering, 478
Composite primary key, 34
Computer-generated test statistic and p-value, 233
Conditional probability, 144–145
Confidence interval
population mean μ, 186–189, 188
population proportion p, 190–191
tdf values and probabilities, locating, 186–188
Confidence of the association rule, analysis, 503–504
Confusion matrix
computing performance measures from, 335
k-nearest neighbor (KNN), 380
use Excel to obtain, 337–338
Consequent, association rule analysis, 502
Constraints
linear programming (LP), optimization with, 583–584
optimization and, 578, 579
Consumer information, data mining and, 368
Contingency table, categorical variable, relationship between two, 96–100
Continuous random variable, 154
Continuous uniform distribution, 165
Continuous variable, 15
Correlated observations, detect/remedy, 247–249
Correlation coefficient, 123–124
Counting, sorting data, 39–44
Covariance, measure of association, 123
CRISP-DM, data mining process, 319–320
Cross-Industry Standard Process for Data Mining (CRISP-DM), 320. See CRISP-DM
Cross-sectional data, 9
Cross-validation methods, regression models
holdout method, 301–304
k-fold method, 300, 306–307
overfitting and, 300–301
using R, 547
Cubic trend model, 538–539
Cumulative distribution function, 154
Cumulative lift chart, classification, 340–341, 344
Cut-off values, selecting, 338–340
Cyclical component, time series forecasting process, 520

D
Daimler AG, 319
Daniels, Gilbert S., 564
Data
cross-sectional, 9
defining, 8
file formats, 20–24
measurement scales and, 16–19
sample/population, 8–9
sources of, 20–21
structured/unstructured, 1012
time series, 10
variables and, 15–16
writing with, 604–606
See also Big data; Data entries; Visualize data
Database
defined, 32
NoSQL and, 37
SQL, retrieving data and, 35–36
Database management system (DBMS), 32
Data breach, Marriott International, 7
Data inspection, 39–44
Data management
defining, 32
software systems, 32
wrangling, 32
Data mart, 37–38
page 654
Data mining
consumer information and, 368
CRISP-DM process for, 319–320
data partitioning, 332–333
defining/methods of, 318
performance evaluation, supervised, 334
SEMMA methodology, 320–321
supervised, 321. See also Supervised data mining
unsupervised, 322. See also Unsupervised data mining
Data modeling, entity relationship diagram (ERD) and, 33–35
Data partitioning, 332–333
GINI impurity index and, 417–418
k-nearest neighbor (KNN) method and, 372–378
naïve Bayes’ method and, 391
recursive, 420
Data partitioning, model selection and, 544–547
cross-validation with time series, 544
defined/uses for, 544–546
Data preparation
missing values, handling, 46–50
subsetting, 50–53
Data retrieval, 35–36
Data sets (exercises)
car crash, 608–609
college admissions, 609–610
house price, 610
longitudinal survey data, 610–611
NBA, 612–613
tech sales reps, 613
Data transformation, categorical, 69–75
category reduction, 69–72
category scores, 74–75
dummy variables, 72–73
Data transformation, numerical, 57–66
binning, 58–62
 rescaling, 66
Data warehouse, 36–38
Data wrangling, 32, 79
DBMS packages, 36
Decile-wise lift chart, classification, 341, 344–345
Decision tree
building, 419–421
minimum error/best-pruned, 414
nodes of, 412
pruning, 421
See also Classification and regression trees (CART)
Decision variables, linear programming (LP) and, 578, 579
Definite events, 140
Degrees of freedom, 186
Delimited file format, 21, 22
Dendrogram, agglomerative clustering, 479–480
Dependent events, independent events v., 145
Descriptive analytics, 4, 5
Deterministic relationship, variables, 212
DIANA divisive analysis. See Divisive clustering, DIANA
Dimension reduction, unsupervised learning, 322
Dimension table, 37–38
Discrete probability distribution, 154
generate random observations of, 564–567
Monte Carlo simulation and, 564
Poisson distribution, 159–160
Discrete random variable, 154–155
Discrete uniform distribution, 156
Discrete variable, 15
Discriminate analysis
advantages/limitations, 370
defined, 369
Dispersion, measures of
defined, 120–121
interquartile range (IQR), 120
mean absolute deviation (MAD), 120
range, defined, 119
Sharpe ratio, 120
variance/standard deviation, 120
Divisive clustering, DIANA, 477
Dow Jones Industrial Average (DJIA), 16, 17, 20, 95
Dual price, 583
Dummy variables
avoiding trap of, 216
category reduction and, 72–73
exponential trend model with seasonal, 539
interaction of two, regression model, 260–262
linear trend model with seasonal, 531–533
numerical variable, interaction between, 262–264
predictor variable and, 212–213
seasonal, quadratic trend model and, 540
Duracell, 8

E
Economic Policy Institute, 207
Economist, The, 21
E-mail filters, 369
Empirical probability, 142
Empirical rule, z-scores and, 130
Ensemble tree models
bagging, 454, 455, 456
boosting, 455, 456
defined/uses for, 454
develop using Analytic Solver/R, 456–465
random forest, 455, 456
Ensemble tree models, advantages/limitations, 370
Entity-relationship diagram (ERD), 33–35
Equal Employment Opportunity Commission (EEOC), 164
Error rate, classification performance measure, 335
Error sum of squares (ESS), 478
Error sum of squares (SSE), 214–215
Estimate, 178
competing models, 302, 303
linear probability/logistic regression models, 288–296
linear regression model, 215–216
maximum likelihood estimation (MLE) and, 290
nonlinear relationship, regression models for, 272–284. See also Regression models,
nonlinear relationships
OLS method, 214–215
regression models, interaction variables and, 260–267. See also Regression model,
interaction variables
 standard error of, 223–224
Estimation
confidence interval for μ and p, 186
confidence interval/interval estimate, 185–186
population mean μ, confidence interval, 186–189, 188
population proportion p, confidence interval, 190–191
Estimator, 178
Euclidean distance, measure similarities with numerical data, 323–324
Events, probability and, 140–141
Excel. See Microsoft Excel; Excel Analysis ToolPak; Microsoft Excel formula option;
Microsoft Excel Solver, Simplex method
Excess kurtosis, 122
Excluded variables, detect/remedy, 249
Exhaustive events, 140
Expected value
standard error of the sample mean, 179
standard error of the sample proportion, 182
Experiment, 140
Exponential regression model, 278–282
Exponential smoothing, time series forecasting
advanced methods, 548–555
Holt exponential smoothing method, 549–552
Holt-Winters, 552–555
page 655
moving average, 528
simple, 525–528
Exponential trend model, nonlinear regression, 535–539
eXtensible Markup Language (XML), 21–24
Extensions of total probability rule, Bayes’ Theorem and, 150–151

F
Facebook, 7, 10, 12, 15, 193, 387, 439
Fact table, 37–38
Fahrenheit scale, 18
False negative outcome (FN), 368
False positive outcome (FP), 368
Fashion retailers, data analytics and, 6
Federal Reserve Bank, 238, 542
Federal Reserve Economic Data, 20–21
Fidelity Gold Fund, 208
Fidelity Growth and Value mutual funds, 125
Fidelity Growth Index, 81, 86
Fidelity Select Automotive Fund, 208
Fidelity’s Technology and Energy mutual funds, 133
File formats, data, 20–21
delimited, 22
eXtensible Markup Language (XML), 22–23, 24
fixed width, 21
HyperText Markup Language (HTML), 23, 24
JavaScript Object Notation (JSON), 23–24
markup languages, 23–24
summary, 24
Financial data, 11
Fisher, Ronald, 498
Fixed-width format, 21
Flaw of Averages, The (Savage), 563
Forbes, 21, 25, 184, 403
Forecasting process, time series
causal models and, 533–534
cross-validation with, 544
defining/components and uses for, 520–521
exponential smoothing, moving average, 528
model selection criteria, 521–522
moving average smoothing, 523–525
nonlinear regression models, trend/seasonality, 535–541. See also Nonlinear regression
models, trend/seasonality and
performance measures, 522
qualitative/quantitative methods of, 521
simple exponential smoothing, 525–528
smoothing techniques for, 522–523
trend/seasonality, linear regression models for, 529–534. See also Linear regression
models
Foreign key (FK), 33
Formulas, Excel, 614–616
Fortune, 21
Frequency distribution
categorical variable, 82–83
numerical variable, 86–87
Full tree, 413
Functions
Analytic Solver (Excel/R), 46–50
Cobb-Douglas production, 288
 cumulative distribution, 154
MS Excel, 615
R/RStudio, entering/importing data, 623–625
software. See Microsoft Excel entries; R; R lpsolver; RStudio for program functions in
bold print

G
Gallup, 299
Gallup-Healthways Well-Being Index, 209
Gaming industry, data analytics and, 6
Gap, Inc., 6
Gartner.com, 12
General Electric (GE), 10
ggplot2 package, R and, 113
Gini impurity index, 416–419
Goodness-of-fit measures
adjusted R2, 225–227
coefficient of determination R2, 224–225
models for, 226–227
standard error of the estimate and, 223–224
validation and, 227
Google, 7, 113, 193, 369
Google Chrome, 23
Google Docs, 22
Gower, John C., 485
Gower’s coefficient, agglomerative clustering and, 485–487
Grammar of Graphics, 113
Graphs, visualize data, guidelines, 91–92. See also Visualize data
Great Depression, 208
Great Recession, 208, 289
Grouping object, cluster analysis and, 476

H
Happiness Index Report (UN), 27
Harper, Bryce, 25
Harrah’s Entertainment, 6
Healthcare, business analytics and, 7
Heat map, visualize data, 109–112
Hierarchical cluster analysis, 485–488
agglomerative clustering, numerical/categorical variables, 478–485
defined, 477
dendrogram and, 479–480
divisive, 477
hidden patterns in, 480
Histogram, numerical variable, visualize, 87–91
H&M, 6
Hold exponential smoothing method, recursive equations and, 548–549, 552
Holdout method
cross-validation, 301–304
defined, 300
linear regression model, Analytic Solver/R, 305–306
Holt exponential smoothing, 548–552
Holt-Winters exponential smoothing, 552–555
Home Depot, 15, 125
House price data set, 610
HTML tags, 23
Human-generated data, 11–12
HyperText Markup Language (HTML), 21, 23, 24
Hypothesis testing
competing hypothesis, 195
defining/types of, 194
four-step procedure, p-value approach, 200
null/alternative, 194–196
one-tailed/two-tailed, 195
population mean μ, 197–200
population proportion p, 203–204
test statistic for μ, 198
Type I/Type II errors, 196–197

I
IBM, 318
IBM DB2, 32
If-Then rules
association rule analysis, 502
classification and regression trees (CART), 412
classification tree, 424
decision nodes and, 444
 decision tree methodology, 414
support of the association rule, 503
page 656
Impossible events, 140
Independent assumption, 390
Independent events, dependent events v., 146
Indicator variable, 72–73
Inflation, data on, 21
Information, data as, 8
Input parameters, linear programming (LP), 578, 579
Inspecting data, 39–44
Instagram, 15
Integer programming, optimization with
capital budgeting, 596–599
transportation problems, 599–603
Interaction effect, regression model, 260
Interaction variables, regression model
dummy variable/numerical variable, 262–264
two dummy variables, 260–262
two numerical variables, 264–267
Interior node, decision tree, 412
Internet Addiction Test (IAT), 362
Interquartile range (IQR), 120
Intersection of two events, 141
Interval estimate, 185–186
Interval scale, 18
Inverse transformation, 168
iPhone, 14, 107
Items, association rule analysis, 502
Item set, association rule analysis, 502

J
Jaccard’s coefficient, categorical data similarity measures, 327–329
JavaScript, 24
JavaScript Object Notation (JSON), 21, 23–24
Johnson & Johnson (J&J), 234, 253
Joint probability, 143–144
Jones, Linda, 25
Journal of Happiness, 127, 244
K
Keyword filtering, 369
k-fold cross-validation method, 300, 306–307
k-means cluster analysis
defining/uses of, 492–493
perform with Analytic Solver, 494–495
perform with R, 495–497
k-nearest neighbors (KNN) method
advantages/limitations, 370
confusion matrix, 380
cumulative gains chart, 382
cumulative lift chart, 383
data partitioning, Analytic Solver, 372–378
decile-wise lift chart, 383
defined, 368
memory-based-reasoning, 371
observations and, 371–372
performance charts, 377
predicted probabilities, 381
ROC curve, 384
scoring results, 385
similarity measures and, 322
10-fold cross-validation, R, 378–385
underlying distribution and, 372–373
uses for, 370
Knowledge, data as, 8
Kurtosis coefficient, 122

L
Laplace smoothing, 390
Leaf node, decision tree, 412, 413
Left-tailed test, 199–200
Leptokurtic, 122
Lift ratio, association rule analysis, 504
Linear probability regression model, 289–290
Linear programming (LP), optimization with
binding constraints, 584
constraints in, 583
defined/uses, 577–578
 dual price, 584
formulating, 579–581
four components of, 579
nonbinding constraint, 583
objective function and, 578, 579
shadow price, 584
solve/interpret problems, 581–593
Linear regression model
adjusted R2, 225–227
ANOVA and, 224, 230
assumptions, 241–243
categorical variables, multiple categories, 216–217
coefficient of determination, R2, 224–225
components of, 212–216
defined, 214
dummy variable, avoiding trap of, 216
estimate, 215–216
estimating with Excel or R, 218–219
goodness-of-fit measures and, 223–224
log-transformed v., 282–284
multiple, 213–214
OLS estimators, 214–215
regression analysis and, 212
reporting regression results, 235
simple, examples of, 213
tests of significance and, 229–334. See also Tests of significance
trend forecasting, 530–531
Linear regression model, detect/remedy common violations of
changing variability, 246–247
correlated observations, 247–249
excluded variables, 249
multicollinearity, 244–246
nonlinear patterns, 243–244
Linear relationship, visualize, 101
Linear trend model
estimating with seasonality, R, 533
seasonality and, 531–533
time series forecasting, 530–531
Line chart
defined, 112
 numerical variables, 108–109
LinkedIn, 10
Logarithmic regression model, 278–279
Logarithmic transformation, 276–277
Logistic regression model, 290–293
advantages/limitations, 370
college admissions/enrollment, 310–312
estimate, 294–296
Log-log regression model, 277–278
Log-transformed regression model, comparing linear and, 282–284
Longitudinal survey data set, 610–611
Los Angeles Angels, 25
Los Angeles Lakers, 152
Lowe’s Companies, Inc., 125, 534
lpSolve package, R. See R lpSolve package

M
Machine-learning, data mining and, 318
Machine-structured data, 11–12
Macintosh, 22
Major League Baseball, 7, 362, 499
Management information systems (MIS), 261
page 657
Manhattan distance, measure similarities with numerical data, 323–324
Manhattan Project at Los Alamos National Laboratory, 563
Marginal effect, quadratic regression model (x on y), 273
Margin of error, 186
Markup languages, 22–24
Marriot International, 7
Matching coefficient, categorical data similarity measures, 327–329
Mathematical transformations, 62–66
Maximization function, 578
Maximum likelihood estimation (MLE), 290
Mean, as central location measure, 115
Mean absolute deviation (MAD), 120, 300, 346, 368, 522
Mean absolute error (MAE), 346
Mean absolute percentage error (MAPE), 300, 346, 522
Mean error (ME), 346
Mean percentage error (MPE), 346
Mean squared error (MSE), 300, 522
Measurement scales
nominal, 16–17
ordinal, 17–18
types of, 18
Measures of association, 123
Measures of dispersion, 120–121. See also Dispersion, measures of
Measures of shape, 121–122
Median, as central location measure, 115
Memory-based reasoning, 371
Microsoft Access, 32, 36
Microsoft Corp., 193
Microsoft Excel (exercise functions)
add-ins, 620
Analysis ToolPak/Solver add-ins, 620
Analytic Solver, 620
Analytic Solver, R and, functions, 46–50
bar chart, frequency distribution, 84–85
binning, transforming numerical data (Analytical Solver, R and), 58–62
binomial probabilities, 161
bubble plot, 107
category reduction (Analytical Solver, R and), 70–72
category scores, 74–75
confidence interval for μ, 189
confusion matrix/performance measures, 337–338
construct residual plots with, 250–251
correlation coefficient, 124
counting/sorting data, functions, 40–43
cumulative lift chart, 344
decile-wise lift chart, 344–345
dummy variable (Analytic Solver, R and), 73
formulas, 614–616
generate random observations, continuous probability distribution, 568
heat map, 110–111
histogram, numerical variable, 89–90
linear regression model, estimating, 218
line chart, numerical variable, 108–109
mathematical transformation, functions, 64–65
Monte Carlo simulation, formulate/develop, 570–573
moving averages, exponential smoothing, 528
normal distribution, 170
numerical variable, visualize relationship, 102
 percentile, quartile/summary, 119
Poisson probabilities, 162
prediction performance measures, 347–348
random observations, discrete probability distribution, 565
regression models, compare linear and log-transformed, 283–284
relative/absolute/mixed references, 614–615
ROC curve, 345
scatterplot, categorical variable, 105–106
spreadsheet model, developing, 617–619
stacked column chart/contingency table, pivot table fields, 91–100
summary data, functions, 51–53
testing for μ, 202
Microsoft Excel Analysis ToolPak (exercise functions)
generate random observations, continuous probability distribution, 568
how to add-in, 620
random observations, discrete probability distribution, 565–566
random seed, number generation, 570
summary measures, 117
Microsoft Excel Analytic Solver (exercise functions)
agglomerative clustering, 481–483
association rule analysis, 506–508
binning, transforming numerical data, 58–62
category reduction, R and, 70–72
classification using five predictor variables, naïve Bayes’ method, 392–394
decision tree, develop, 422–425
defining, 620
develop prediction tree, 443–445
dummy variable, create (R and), 73
ensemble tree models, develop, 456–459
estimate logistic regression model, 294–296
holdout method, logistic regression model, 305
Holt exponential smoothing method, 550
Holt-Winders exponential smoothing, 553
how to add-in, 620
interface of, 620
k-means analysis, 494–495
k-method, 373–378
missing values, R and, 46–50
principal component analysis, perform, 355–359
scoring options, 375
standard data partition, 374
Microsoft Excel formula option, calculate mean/median, 116–117
Microsoft Excel Solver, Simplex method (exercise functions)
capital budgeting, integer programming optimization, 597–598
formulate/solve, linear programming problem, 584–589, 590–591
integer programming optimization, transportation problems, 601–602
Microsoft Excel XLMiner, 620
Microsoft Windows, 22
Microsoft Word, 22
Minimization function, 578
Minimum error tree, 414
Misclassification rate
classification performance measure, 335, 336
k-nearest neighbor (KNN) and, 376
Missing values, data preparation, 46–50
Mixed data, agglomerative clustering and, 485–488
Mixed references, Excel, 614–615
Modeling data, 33–35
Model selection
adjusted R2, 225–227
coefficient of determination, 224–225
data partitioning and, 544–547
standard error of the estimate, 223–224
time series forecasting, 521–522
Moneyball, 7
Money Magazine, 288
Monte Carlo simulation
defined/uses of, 563–564
discrete uniform distribution and, 564
formulate/develop, 569–575
generate random observations, continuous probability distribution, 567–569
Poisson distribution and, 564
random observations, discrete probability distribution, 564–567
Mortgage Banking Association, 171
Moving average smoothing
exponential, 528
time series forecasting, 523–525
page 658
Multicollinearity, linear relationship detect/remedy, 244–246
Multiple linear regression model, 213–214, 232
Multiplication rule, probability, 145
Mutually exclusive events, 140
Myers-Briggs Personality assessment, 77, 82, 96–97
MySQL, 32

N
Naïve Bayes’ Theorem
advantages/limitations, 370
applications of, 369–370
classification using five predictor variables, Analytic Solver, 392–394
defining/uses of, 390–391
partitioning data, R, 394–398
predicting college acceptance, 405–407
scoring results, 394
smoothing/Laplace smoothing, 390
transforming numerical into categorical values, 398–399
National Association of Business Economists (NABE), 205
National Association of Colleges and Employees, 261
National Association of Securities Dealers Automated Quotations (NASDQ), 16, 17
National Basketball Association (NBA), 9, 364–365, 612
National Football League, 361, 499
National Health and Nutrition Examination Study, 152
National Institute of Health, 262
National Longitudinal Surveys (NLS), 469–470, 610–611
National Public Radio (NPR), 228, 542
NBA data set, 612–613
NBCnews.com, 206
NCR, 319
Negative linear relationship, 213
Negatively skewed distribution, 88
Netflix, 6, 535
Net promoter score (NPS), 77–78
Neural networks, 369–370
New York City Open Data, 501, 513
New Yorker, The, 240, 312
New York Stock Exchange (NYSE), 16, 17, 528
New York Times, The, 6, 15, 21, 240, 274, 542
Nielsen ratings, 164
Nike, 297
No linear relationship, 213
Nominal scale, 16–17
Noncausal models, time series forecasting, 533–534. See also Forecasting process, time
series
Nonlinear patterns, detection/remedy, 243–244
Nonlinear regression models, trend/seasonality and
exponential trend model, 535–538
polynomial trend model, 538–539
seasonality, 539–541
Nonlinear relationship
regression models for, 272–284. See also Regression models, nonlinear relationships
visualize, 101
Nonzero slope coefficient, test for, 233–334
Normal distribution
defined, 165
standard, 165–167
transformation, normal random variables, 167–169
Normalization, numerical data, 325–327
Normally distributed population, sampling from, 180–181
Normal random variables, transformation, 167–169
NoSQL, 37
Notepad, 22
Null hypothesis, 194–196
Numerical data, 15, 16
binning, transforming, 58–62
Euclidian/Manhattan distance, measure similarities with, 323–324
standardizing/normalizing similarity measures with, 325–327
transforming, 57–66
transforming, categorical values, 398–399
Numerical variable
agglomerative clustering and, 478–485
interaction of two, regression model and, 264–267
regression model, numerical variable interaction, 262–264
Numerical variable, visualize relationships
bubble plot, between three, 107–108
frequency distribution, 86–87
histogram, 87–91
line chart, 108–109
scatterplots, 101–102

O
Oakland Athletics, 6–7
Objective function, linear programming (LP) and, 578, 579
Observations
binning, categorization of and, 58
generate random, discrete probability distribution, 564–567
generate random observations, continuous probability distribution, 568–569
identify, 322. See also Similarity measures
k-nearest neighbor (KNN) method and, 371
measurement scales and, 18
missing values, 49–50
similarity measures for. See Similarity measures
OHRA, 319
Omission strategy, missing values, 46–50
One-tailed hypothesis test, 195, 199–200
Online subscription services, data analytics and, 6
Opinion Today, 193
Optimization algorithms, prescriptive analytics and, 562
constraints, 578, 579
defined, 577, 578
integer programming, 596–603. See also Integer programming, optimization with
linear programming and, 577–593. See also Linear programming, optimization with
Oracle, 25, 32, 36
Oracle database, 113
Orbitz.com, 9
Ordinal scale, 17–18
Ordinary least squares (OLS) method, estimating, 214–215
classical linear regression model, assumptions and, 241–243
common violations of assumptions of, 243–249. See also Linear regression model,
detect/remedy common violations of
Outcomes
false negative/false positive, 368
summarizing, classifications and, 368
true negative/true positive, 368
Outliers
boxplots and, 127–129
defined, 115
z-scores and, 129–132. See also Z–scores
Overfitting
cross-validation method, 300–301
performance evaluation and, 334
Oversampling, performance evaluation, 333
P
Pairwise observations, 323
Parameter, population, 178
Parameters, linear programming (LP) and, 578, 579
Partitioning data. See Data partitioning
Pattern recognition, unsupervised learning, 322
Paypal, 535
p-value approach
four-step process, hypothesis testing, 200
sample mean, alternative hypothesis, 199
page 659
Peck, Art, 6
Percentile, quartile functions, 118–119
Performance charts, classification
cumulative lift chart, 340–341
decile-wise lift chart, 341
Excel to obtain, 344–345
receiver operating characteristic (ROC) curve, 342–344
Performance evaluation
classification models and, 335–337
cut-off values, selecting, 338–340
data partitioning, 332–333
oversampling, 333
performance charts, classification, 340–344
prediction and, 345–348
supervised data mining, 334
using Excel to obtain confusion matrix/performance measures, 337–338
Performance measures
case study, 364–365
prediction, Excel, 347–348
time series forecasting, 522
using Excel to obtain confusion matrix and, 337–338
Pew Research Center, 93, 139, 152, 184, 208–209, 259
Philadelphia Phillies, 25
Platykurtic, 122
Point estimator, 178
Point-of-sale data, 11
Poisson distribution, 159–160, 564
Poisson process, 159
Poisson random variable, 159
Polynomial trend model, nonlinear regression model time series forecasting
 cubic trend model, 538–539
quadratic trend model, 538–539
Population
advantages/disadvantage, 8–9
mean, parameter, 115
See also Sampling distributions
Population mean μ
confidence interval, 186–189
hypothesis testing, 197–200
Population parameter, 178
Population proportion p
confidence interval, 186
hypothesis testing for, 203–204
test statistic for, 203
Positive linear relationship, 213
Positively skewed distribution, 88
Positive predictive value, classification performance measure, 335–336
Posterior probability, 148–149, 390
Precision, classification performance measure, 335–336, 336
Predicted probabilities
k-nearest neighbors (KNN), 381–382
linear probability regression models, 291
Prediction, performance evaluation for, 345–348
Prediction performance measures, 346
Prediction tree, 441–443. See also Regression tree
Predictive analytics, 4, 5
Predictor variables, 212–213
predictive model for price of house, case study, 254–256
regression trees and, 441–443
Preparing data, 45–53
Prescriptive analytics, 4–5, 562
Price, David, 25
Primary key (PK), 33
Principal component analysis (PCA)
applying to data set, 353
defining/uses for, 352
perform with Analytic Solver/R, 355–359
rotating data, 353–355
Prior probability, 148–149, 390
Prison Legal News, 298
Probabilistic simulation, 563
Probability
assigning, 141–142
defined, 140
events and, 140–141
locating tdf values, 186–188
rules of, 142–146
total probability rule, Bayes’ Theorem and, 148–151
Probability distribution, 154
Programming languages, 24
Pruning, decision tree, 421
Psychology Today, 285, 288
Pure subset, CART, 413
p-value, 198
Python, 24

Q
Quadratic regression model, 272–276
Quadratic trend model, 538–540
Qualitative forecasting, time series forecasting, 521
Qualitative variables, 15, 16
Quantitative method, time series forecasting, 521
Quantitative variables, 15, 16
Quarterly data, 539–540

R
R-exercise functions
agglomerative clustering, 483–455
association rule analysis, 508–510
bar chart, frequency distribution, 85
binomial probabilities, 161
boxplots, detecting outliers, 128–129
bubble plot, 107–108
confidence interval for μ, 190
correlation coefficient, 124
cross-validation of regression models with, 547
cumulative lift chart, 383
decile-wise lift chart, 383
decision tree, develop in, 425–433
 defining, 621
develop prediction tree, 445–449
ensemble tree models, develop, 449–465
entering/importing data, 623–625
estimate logistic regression model, 296
estimating linear trend model with seasonality, 533
generate random observations, continuous probability distribution, 568–569
ggplot2 package, 113
Gower’s coefficient, agglomerative clustering, 486–487
heat map, 111–112
histogram, numerical variable, 90–91
holdout method, logistic regression model, 306
Holt exponential smoothing method, 550–552
Holt-Winder exponential smoothing method, 553–555
installing for windows, 621
k-means cluster analysis, 495–497
k-nearest neighbor (KNN), 10-fold cross-validation, 378–385
linear regression model, estimating with, 218–219
line breaks, 626
line chart, numerical variable, 109
mean/median, summary function, 117–118
Monte Carlo simulation, formulate/develop, 573–575
normal distribution, 170
numerical variable, visualize relationship, 102
packages, 626–627
partitioning data, naïve Bayes’ method, 394–398
percentile/quartile/summary, 119
page 660
Poisson probabilities, 162
principal component analysis, perform, 355–359
random observations, discrete probability distribution, 566–567
regression models, compare linear and log-transformed, 283–284
residual plots, construct, 250–251
ROC curve and, 384
RStudio, 621, 622
scatterplot, categorical variable, 106
scoring results, 385
stacked column chart/contingency table, 100
testing for μ, 202
See also R lpSolve package; RStudio
Random component, time series forecasting process, 520
Random forest, ensemble tree models, 455, 456
Random variable
defined, 153
discrete, summary measures, 154–155
discrete probability/cumulative distribution and, 154
Poisson, 159
types of, 154
Ratio scale, 18
Real-world business scenarios, see Simulation algorithms
Recall, classification performance measure, 335
Receiver operating characteristic (ROC) curve
classification, 342–345, 368
k-nearest neighbor (KNN) and, 384
Recursive equations
Holt exponential smoothing and, 548–549
Holt-Winters exponential smoothing, 542–555
Recursive partitioning process, 420
Reference category, 213
Regression analysis
defined, 212
reporting results, 235
Regression model
cross-validation methods, 301–307. See also Cross-validation methods, regression
model
time series data, using R, 547
Regression model, interaction variables
dummy variable/numerical variable, 262–264
numerical variables, between two, 264–267
two dummy variables, 260–262
Regression model, nonlinear relationships
comparing linear/log-transformed, 282–284, 283–284
exponential, 278, 279–282
logarithmic, 278–279
logarithms, 276–277
log-log regression model, 277–278
quadratic regression model, 272–276
semi-log, 278
Regression sum of squares (SSR), 224–225
Regression trees
defining/uses for, 440–441
develop prediction tree, Analytic Solver/R, 443–449
split points, identify, 441–443
Relational database, 32
Relationship(s)
defining, 33
deterministic/stochastic, between variables, 212
entity diagram, 33–35
error rate/model complexity, 334
nonlinear, regression models for, 272–284. See also Regression models, nonlinear
relationships
simple linear regression model and, 213
Relationship between two variables
contingency table, categorical variable, 96–97
stacked column chart, categorical variables, 97–100
Relative references, Excel, 614
Religious Landscape Study, 93
Rescaling, transforming numerical data, 66
Residual, 214
Residual plots
against changing variability, 246–247
computing, 242
construct using Excel/R, 250–251
correctly specified model, 242
nonlinear patterns, 243–244
residuals against time t, 247–248
Response variable, 212
Retail customer data, 31
Retrieving data, 35–36
Reuters, 184
Reward-to-variability ratio, 120
Right-tailed test, 199–200
Risk-adjusted stock return, 234
R lpSolve package (exercise functions)
capital budgeting, integer programming optimization, 599
formulate/solve, linear programming problem, 589–590, 592
integer programming optimization, transportation problems, 602
See also R; RStudio
ROC curve. See Receiver operating characteristic (ROC) curve
Root mean squared error (RMSE), 300–301, 346, 368
Root node, decision tree, 412, 419
Rotating data, principal component analysis (PCA), 353–355
RStudio-exercise functions
defining, 621
 entering data/using functions, 623–624
installing for windows, 622
interface, 622–623
Rules of probability, 142–146

S
Safari browser, 23
Sample, Explore, Modify, Model, and Assess (SEMMA), 320. See SEMMA methodology
Sample data, advantages/disadvantage, 8–9
Sample mean
central limit theorem (CLT) for, 181–182
sampling distributions of, 178–180
standard error of, 179
statistic, 115, 178
Sample proportion, 178
central limit theorem (CLT) for, 183
sampling distribution of, 182–183
Sample regression equation, 214
Sample regression line, 214
Sample space, 140
Sampling distributions
central limit theorem (CLT) for the sample man, 181–182
defining/types of, 178
normally distributed population, 180–181
of sample mean, 178–180
sample proportion of, 182–183
Samsung, 107
Samuel, Arthur, 318
Savage, Sam L., 563
SB Nation, 171
Scatterplots
categorical variable and, 102, 105–106
defined, 112
Schneider, Robert, 25
Scholastic Aptitude Test (SAT), 228
Scoring a record, 368
Seasonal component, time series forecasting process, 520
exponential trend, dummy variables and, 539
nonlinear trend model with, 539–541
Seasonality, linear trend model and forecasting, 531–533
page 661
SeekingAlpha.com, 534
Semi-log regression model, 278
SEMMA methodology, data mining, 320–321
Sensitivity, classification performance measure, 335, 336
Shadow price, 583
Shape, measures of
kurtosis coefficient, 122
skewness coefficient, 121–122
Sharpe, William, 120
Sharpe ratio measures, 120
Significance level, 199
Silhouette plot, 496
Similarity measures
defining/uses of, 322–323
matching coefficient/Jaccard’s coefficient, categorical data, 327–329
numerical data, 323–324
standardization/normalization, numerical data, 325–327
Simple exponential smoothing, time series forecasting, 525–528
Simple linear regression model, 213
Simulation algorithms
defining/uses for, 562
Monte Carlo real-world business scenarios, 562–575. See also Monte Carlo simulation
prescriptive analytics and, 562
probabilistic/stochastic, 563
Simulation algorithms, prescriptive analytics and, 562
Single linkage method, agglomerative clustering, 478
Skewness coefficient, 88, 121–122
Smoothing, 390
exponential, advanced methods of, 548–555. See also Exponential smoothing
Holt exponential, 548–552
Holt-Winders exponential, 552–555
moving average, 523–525
moving averages, and exponential, 528
simple exponential, time series forecasting, 525–528
time series forecasting, 522–523
Snapchat, 15
Social media, 15, 317
Software
database management systems (DBMS), 32
Excel. See Microsoft Excel entries
 ggplot2 package, R, 113
R, See also R lpSolve package; RStudio
Tableau, 113
text editing, 22
Sorting, counting data, 39–44
Sources of data, 20–21
Spam filters, 369
Specificity, classification performance measure, 336
Split points
CART and, 415–416
identify, regression trees, 441–443
Sports, data analytics and, 6–7
Spotify, 528–529
Spreadsheet model, develop in Excel, 617–619
S&P’s Case-Shiller home price index, 542
SPSS, 319
SQL Server, 32
Stacked column chart, categorical variables, relationship between two, 97–100
Standard deviation, measure of dispersion, 120
Standard error of the estimate, 223–224
Standard error of the sample mean, 179
Standard error of the sample proportion, expected value, 182
Standardization, numerical data, 325
Standard Normal Curve Areas, 628–629
Standard normal distribution, 165–167
Standard normal table, 165
Standard transformation, random normal variables, 167–169
Star schema, 37–38
State of American Well-Being--2017 State Well-Being Rankings, 209
Statistic, 115, 178
Stochastic relationship, variables, 212
Stochastic simulation, 563
Structured data, 10–11
file formats for, 21
unstructured/big data v., 13
Structured Query Language (SQL), data retrieval and, 35–36
Student’s t distribution, 186, 630–631
Subjective probability, 142
Subsetting, data preparation, 50–53, 79
Summary measures, 50–51. See also individual measures
association, 123–124
 central location, 114–119
discrete random variable, 154–155
dispersion, 119–121
shape, 121–122
Sunnyville Bank, 411, 438
Supervised data mining, 321. See also Data mining; Unsupervised data mining
advantages/limitations of, 369
comparing techniques, 370
defining/uses for, 368
k-nearest neighbors (KNN) method, 370–385
naïve Bayes’ method, 390–399
oversampling, 333
performance evaluation, 334
Support of the association rule, association rule analysis, 503
Support of the item or item set, association rule analysis, 503
Symmetric distribution, 88, 129–133

T
Tableau data visualization, 113
Tabular data, 22
Target, 126
t distribution, 186, 630–631
tdf distribution, 186
Techfoogle.com, 185
Tech sales reps data set, 613
TeraData, 319
Test for a nonzero slope coefficient, 233–334
Test of individual significance, 231–232
Test of joint significance, 229–331
Tests of significance
computer-generated test statistic and p-value, 233
individual, 231–332
joint, 229–331
nonzero slope coefficient, 233–334
Test statistic and the p-value, 233
Test statistic for μ, 198
Text analytics, 12
TextEdit, 22
Text files, data format, 21
Time series data, 10
cross-validation with, 544–546
forecasting process for. See Forecasting process, time series
tdf values and probabilities, locating, 186–188
Total probability rule, Bayes’ Theorem and, 148–151
defined, 148–149
extensions of, 150–151
Total sum of squares (SST), 224–225
Training data set, 368
Training set, 300–301
Transforming data
categorical, 69–75
normal random variables, 167–169
numerical, 57–66
page 662
Transportation problems, integer programming optimization, 599–603
Trend component, time series forecasting process, 520, 530–531
cubic trend model, 538–539
exponential trend model, nonlinear regression model, 535–538
exponential with dummy variables, seasonal, 539
polynomial trend model, 538–539
quadratic, seasonal dummy variables, 540
Trout, Michael, 25
TrueCar, 542
True negative outcome (TN), 368
True positive (TP) outcome, 368
Trump, Donald, 7
Twitter, 10, 11
Two-tailed hypothesis test, 195, 199–200
Type I errors, 196–197
Type II errors, 196–197

U
Ulam, Stanislaw, 563
Unbiased estimator, 179
Uncertainty, random variable and, 153–154
Unconditional probability, 144
Under Armour, 15, 297–298, 309
Union of two events, 140–141, 143–145
Uniqlo, 6
United Nations, 27
Unstructured data, 11–13
Unsupervised data mining
agglomerative clustering, mixed data, 485–488
defined, 322. See also Data mining; Supervised data mining
hierarchical cluster analysis, 476–485
USA Today, 21
U.S. Bureau of Labor Statistics, 469–470
U.S. Census Bureau, 10, 15, 21, 55, 93, 95, 152, 163, 469–470, 542
U.S. Department of Agriculture (USDA), 475
U.S. Department of Health and Human Services, 268
U.S. Energy Information Administration, 529
U.S. National Cancer Institute, 7
U.S. News and World Report, 25
U-shaped cost curve, 272–276

V
Validation set, 300–301
Value, big data, 12
Values, handling missing, 46–50
Vanguard Balanced Index Fund, 170
Variable(s)
categorical. See Categorical variable
decision, 578, 579
defined, 15
dummy, category reduction and, 72–73
mathematical transformations, 64–65
numerical. See Numerical variable
predictor, 212
relationship between two, visualize, 96
response, 212
transforming numerical data, 58–62
types of, 15
Variance, measure of dispersion, 120
Variance inflation factor (VIF), 245
Variety, big data and, 12
Velocity, big data and, 12
Veracity, big data, 12
Vertical axis, charts/graphs, guidelines, 91–92
Visualize data
advanced options, ggplot2, Tableau, 112–113
bubble plot, numerical variable, relationship between three, 107–108
categorical variable, 82–85
charts/graphs, guidelines, 91–92
cluster plot, 496
heat map, 109–112
line chart, numerical variable, 108–109
numerical variable, 86–91
numerical variable, relationship between two, 101–102
relationship between two variables, 96
scatterplot, categorical variable, 105–106
silhouette plot, 496
Tableau, 113
Volume, big data and, 12
von Neumann, John, 563

W
Wall Street Journal, The, 21, 147, 177, 203, 519
Walmart, 126
Ward, Joe H., 478
Ward’s method, agglomerative clustering, 478
Washington Post, The, 104, 147
Washington Post-Kaiser Family Foundation, 152
Websites, data providing, 20–21
Whitelists, 369
World Bank, 113, 355, 361, 499, 501, 536
World development indicator data, 21
World Economic Forum, 501
World Health Organization, 238, 270, 271
Wrangling, data, 32, 79
Writing with data, 26–27, 77–78, 133–135, 207–208, 253–254, 309, 364, 404–405, 469–
471, 514, 556–557, 604–606

Y
Yahoo finance, 15
Young, Kimberly, 362
YouTube, 11
Z
Zara, 6
Zhang, Chun, 25
Zillow.com, 14, 254, 499
z-scores, detecting outliers
calculating, 131–132
empirical rule and, 130
z-table, 165