Statistics Translated

Also Available

Writing a Proposal for Your Dissertation:

Guidelines and Examples
  Steven R. Terrell
STATISTICS TRANSLATED
A Step-by-Step Guide to Analyzing
and Interpreting Data
SECOND EDITION

STEVEN R. TERRELL

THE GUILFORD PRESS

New York  London
© 2021 The Guilford Press
A Division of Guilford Publications, Inc.
370 Seventh Avenue, Suite 1200, New York, NY 10001
www.guilford.com

All rights reserved

No part of this book may be reproduced, translated, stored in a retrieval

system, or transmitted, in any form or by any means, electronic, mechanical,
photocopying, microfilming, recording, or otherwise, without written
permission from the publisher.

Printed in the United States of America

This book is printed on acid-free paper.

Last digit is print number: 9 8 7 6 5 4 3 2 1

Library of Congress Cataloging-in-Publication Data

Names: Terrell, Steven R., author.

Title: Statistics translated : a step-by-step guide to analyzing and
interpreting data / Steven R. Terrell.
Description: Second Edition. | New York : The Guilford Press, 2021. |
Revised edition of the author’s Statistics translated, c2012. | Includes
bibliographical references and index.
Identifiers: LCCN 2020044990 | ISBN 9781462545407 (paperback) |
ISBN 9781462545414 (hardcover)
Subjects: LCSH: Statistics—Study and teaching.
Classification: LCC HA35 .T47 2021 | DDC 519.5—dc23
LC record available at https://lccn.loc.gov/2020044990
To Dalia, the love of my life for 40 years
 Acknowledgments

First, I would like to thank C. Deborah Laughton, Publisher, Methodology and Statis-
tics, at The Guilford Press for her untiring support, ideas, critique, and insight during
this project; there’s nothing to say except she’s the editor that every author dreams
of. My job was made far easier by her professional and caring staff. I offer a special
thanks to Katherine Sommer, Paul Gordon for the creative cover, and others at Guil-
ford for their tremendous help along the way. William Meyer—all I can say is “Thank
you so much.” The book looks great and I truly appreciate what you’ve done! I am
also most grateful to the following reviewers: Brian Withrow, Department of Crimi-
nal Justice, Texas State University; Tammy Kolb, Center for Education Policy Analy-
sis, Neag School of Education, University of Connecticut; Cyndi Garvan, College of
Education, University of Florida; Chris Ohana, Mathematics and Science Education
Groups, Western Washington University; Robert Griffore, Human Development and
Family Studies Department, Michigan State University; Christine McDonald, Depart-
ment of Communication Disorders and Counseling, School, and Educational Psychol-
ogy, Indiana State University; Val Larsen, James Madison University; Adam T ­ hrasher,
Department of Health and Human Performance, University of Houston; Melissa
Gruys, Department of Management, Wright State University; Lynne Schrum, Elemen-
tary Education Program, George Mason University, retired; Dave Edyburn, School of
Education, University of Wisconsin–Milwaukee; M. D. Roblyer, Instructional Technol-
ogy and Distance Education, Nova Southeastern University, ­retired; Robin A. Barry,
Department of Psychology, University of Wyoming; Andrew H. Rose, Master of S ­ ocial
Work Program, Texas Tech University; Charles M. Super, Center for the Study of Cul-
ture, Health, and Human Development, University of ­Connecticut; and Todd D. ­Little,
Department of Educational Psychology and Leadership, Texas Tech University.
vii
viii | Acknowledgments

Many of my friends played a role in this book without knowing it. When I was
stuck for an idea for a case study, all I had to do was think of what they do in their
professional lives, and the ideas flowed from there. Thank you; you guys are the best.
Finally, thank you to my students at Nova Southeastern University. By using ear-
lier drafts of this book, you provided me the feedback that helped me write what you
needed to know—statistics can be simplified.
 Brief Contents

Introduction: You Do Not Need to Be a Statistician to Understand Statistics! 1

CHAPTER 1 Identifying a Research Problem and Stating Hypotheses 7

CHAPTER 2 Identifying the Independent and Dependent Variables in a Hypothesis 33

CHAPTER 3 Measures of Dispersion and Measures of Relative Standing 60

CHAPTER 4 Graphically Describing the Dependent Variable 87

CHAPTER 5 Choosing the Right Statistical Test 115

CHAPTER 6 The One-Sample t-Test 156

CHAPTER 7 The Independent-Sample t-Test 181

CHAPTER 8 The Dependent-Sample t-Test 217

CHAPTER 9 Analysis of Variance and Multivariate Analysis of Variance 244

ix
x | Brief Contents

CHAPTER 10 The Chi-Square Tests 304

CHAPTER 11 The Correlational Procedures 335

Conclusion: Have We Accomplished What We Set Out to Do? 375

APPENDIX A. Area under the Normal Curve Table (Critical Values of z) 377
APPENDIX B. Critical Values of t 378
APPENDIX C. Critical Values of F When Alpha = .01 379
APPENDIX D. Critical Values of F When Alpha = .05 381
APPENDIX E. Critical Values of F When Alpha = .10 383
APPENDIX F. Critical Values of Chi-Square 385
APPENDIX G. Selecting the Right Statistical Test 386

Glossary 387
Answers to Quiz Time! 397
Index 421
About the Author 433
 Extended Contents

Introduction: You Do Not Need to Be a Statistician to Understand Statistics! 1

A Little Background  1
Many Students Do Not Know What They’re Getting Into  2
A Few Simple Steps  3
Step 1: Identify the Problem  4
Step 2: State a Hypothesis  4
Step 3: Identify the Independent Variable  4
Step 4: Identify and Describe the Dependent Variable  5
Step 5: Choose the Right Statistical Test  5
Step 6: Use Data Analysis Software to Test the Hypothesis  5
So, What’s New in This Edition?   5
Summary  6
Do You Understand These Key Words and Phrases?  6

CHAPTER 1 Identifying a Research Problem and Stating Hypotheses 7

Introduction  7
Step 1: Identify the Problem  7
Characteristics of a Good Problem Statement  8
Finding a Good Research Problem  8
The Problem Is Interesting to the Researcher  9
The Scope of the Problem Is Manageable by the Researcher  9
The Researcher Has the Knowledge, Time, and Resources Needed
to Investigate the Problem  10

xi
xii | Extended Contents

The Problem Can Be Researched through the Collection and Analysis

of Numeric Data  11
Investigating the Problem Has Theoretical or Practical Significance  11
It Is Ethical to Investigate the Problem  12
Writing the Problem Statement  13
Problem Statements Must Be Clear and Concise  13
The Problem Statement Must Include All Variables to Be Considered  14
The Problem Statement Should Not Interject the Researcher’s Bias  15
Summary of Step 1: Identify the Problem  15
Step 2: State a Hypothesis  16
An Example of Stating Our Hypothesis  16
A Little More Detail  17
The Direction of Hypotheses  17
Using Directional Hypotheses to Test a “Greater Than” Relationship  18
Using Directional Hypotheses to Test a “Less Than” Relationship  18
Nondirectional Hypotheses  19
Hypotheses Must Be Testable via the Collection and Analysis
of Data  20
Research versus Null Hypotheses  20
Stating Null Hypotheses for Directional Hypotheses  20
Issues Underlying the Null Hypothesis for Directional Research
Hypotheses 21
Stating Null Hypotheses for Nondirectional Hypotheses  22
A Preview of Testing the Null Hypothesis  23
Where Does That Leave Us?  26
Statistical Words of Wisdom  28
Summary of Step 2: State a Hypothesis  28
Do You Understand These Key Words and Phrases?  29
Quiz Time!  29
Problem Statements  29
Case Studies   30
The Case of Distance Therapy  30
The Case of the New Teacher  30
The Case of Being Exactly Right  31
The Case of “Does It Really Work?”  31
The Case of Advertising  31
The Case of Learning to Speak  32
The Case of Kids on Cruises  32

CHAPTER 2 Identifying the Independent and Dependent Variables in a Hypothesis 33

Introduction  33
Step 3: Identify the Independent Variable  33
Nonmanipulated Independent Variables  34
 Extended Contents | xiii

Another Way of Thinking about Nonmanipulated Independent

Variables 35
Manipulated or Experimental Independent Variables  35
Levels of the Independent Variable  36
Summary of Step 3: Identify the Independent Variable  37
Step 4: Identify and Describe the Dependent Variable  37
Identifying Your Dependent Variable  37
What Type of Data Are We Collecting?  39
Interval Data  40
Data Types—What Is the Good News?  42
Summary of the Dependent Variable and Data Types  43
Measures of Central Tendency  43
The Mean, Median, and Mode—Measures of Central Tendency  45
The Mode  50
Using Statistical Software to Analyze Our Data  51
Summary of the First Part of Step 4: Identify and Describe
the Dependent Variable  56
Do You Understand These Key Words and Phrases?  57
Do You Understand These Formulas?  57
Quiz Time!  57

CHAPTER 3 Measures of Dispersion and Measures of Relative Standing 60

Introduction  60
Measures of Dispersion  60
The Range  61
The Standard Deviation  63
The Variance  67
Measures of Relative Standing  70
Percentiles 71
Computing and Interpreting T-Scores  77
Stanines 78
Putting It All Together  79
Using SPSS for T-Scores and Stanines—Not So Fast!  80
Summary  83
Do You Understand These Key Words and Phrases?  84
Do You Understand These Formulas?  84
Quiz Time!  85

CHAPTER 4 Graphically Describing the Dependent Variable 87

Introduction  87
Graphical Descriptive Statistics  87
Graphically Describing Nominal Data  88
xiv | Extended Contents

Pie Charts  88
Bar Charts  90
Graphically Describing Quantitative Data  92
Scatterplots 92
Histograms 97
Don’t Let a Picture Tell You the Wrong Story!  97
Summary of Graphical Descriptive Statistics  99
The Normal Distribution  99
Things That Can Affect the Shape of a Distribution
of Quantitative Data  101
Summary of the Normal Distribution  112
Do You Understand These Key Words and Phrases?  113
Quiz Time!  113

CHAPTER 5 Choosing the Right Statistical Test 115

Introduction  115
The Very Basics  115
The Central Limit Theorem  116
The Sampling Distribution of the Means  117
Summary of the Central Limit Theorem and the Sampling Distribution
of the Means  124
How Are We Doing So Far?  124
Estimating Population Parameters Using Confidence Intervals  125
The Alpha Value  125
Type I and Type II Errors  127
Predicting a Population Parameter Based on a Sample Statistic Using
Confidence Intervals  129
Pay Close Attention Here  131
Confidence Intervals for Alpha = .01 and Alpha = .10  132
Another Way to Think about z Scores in Confidence Intervals  134
Tying This All Together  134
Be Careful When Changing Your Alpha Values  135
Do We Understand Everything We Need to Know about Confidence
Intervals? 136
Testing Hypotheses about a Population Parameter Based
on a Sample Statistic  137
Making a Decision about the Certification Examination Scores  138
We Are Finally Going to Test Our Hypothesis!  141
Testing a One-Tailed Hypothesis  141
Testing a One-Tailed “Less Than” Hypothesis  144
Summarizing What We Just Said  145
Be Careful When Changing Your Alpha Values  146
The Heart of Inferential Statistics  147
Probability Values  148
 Extended Contents | xv

A Few More Examples  149

Great News—We Will Always Use Software to Compute Our p Value  150
Step 5: Choose the Right Statistical Test  152
You Already Know a Few Things  152
A Couple of Notes about the Table  153
Summary of Step 5: Choose the Right Statistical Test  154
Do You Understand These Key Words and Phrases?  154
Do You Understand These Formulas and Symbols?  154
Quiz Time!  155

CHAPTER 6 The One-Sample t-Test 156

Introduction  156
Welcome to the Guinness Breweries  156
The t Distribution  157
Putting This Together  157
Determining the Critical Value of t 159
Degrees of Freedom  159
Be Careful Computing Degrees of Freedom  161
Let’s Get Back to Our Anxiety Hypothesis  162
Plotting Our Critical Value of t 162
The Statistical Effect Size of Our Example  162
Let’s Look at a Directional Hypothesis  163
Using the p Value  165
Check Your Mean Scores!  165
One More Time  166
Important Note about Software Packages  167
Let’s Use the Six-Step Model!  168
The Case of Slow Response Time  168
The Case of Stopping Sneezing  171
The Case of Growing Tomatoes  174
Summary  176
Do You Understand These Key Words and Phrases?  177
Do You Know These Formulas?  177
Quiz Time!  177

CHAPTER 7 The Independent-Sample t-Test 181

Introduction  181
If We Have Samples from Two Independent Populations, How Do We
Know If They Are Significantly Different from One Another?  182
The Sampling Distribution of Mean Differences  182
Calculating the t Value for the Independent-Sample t-Test 185
Pay Attention Here  185
xvi | Extended Contents

Testing Our Hypothesis  190

The p Value  191
Note on Variance and the t-Test 192
The Statistical Effect Size of Our Example  192
Let’s Try Another Example  193
Remember the Effect Size  195
How Does This Work for a Directional Hypothesis?  195
Reminder—Always Pay Attention to the Direction of the Means!  197
Putting the Independent-Sample t-Test to Work  197
The Case of the Cavernous Lab  197
The Case of the Report Cards  202
The Case of the Anxious Athletes  206
Summary  209
Do You Understand These Key Words and Phrases?  210
Do You Understand These Formulas?  210
Quiz Time!  211
The Case of the Homesick Blues  211
The Case of the Cold Call  211
The Case of the Prima Donnas  212
The Case of the Wrong Side of the Road  213
The Case of Workplace Satisfaction  214
The Case of the Flower Show  215

CHAPTER 8 The Dependent-Sample t-Test 217

Introduction  217
That’s Great, but How Do We Test Our Hypotheses?  217
Independence versus Dependence  218
Computing the t Value for a Dependent-Sample t-Test 218
Testing a One-Tailed “Greater Than” Hypothesis  219
The Effect Size for a Dependent-Sample t-Test 222
Testing a One-Tailed “Less Than” Hypothesis  223
Testing a Two-Tailed Hypothesis  226
Let’s Move Forward and Use Our Six-Step Model  229
Step 1: Identify the Problem  229
Step 2: State a Hypothesis  229
Step 3: Identify the Independent Variable  230
Step 4: Identify and Describe the Dependent Variable  230
Step 5: Choose the Right Statistical Test  231
Step 6: Use Data Analysis Software to Test the Hypothesis  231
The Case of the Unexcused Students  232
Step 1: Identify the Problem  232
Step 2: State a Hypothesis  232
Step 3: Identify the Independent Variable  233
 Extended Contents | xvii

Step 4: Identify and Describe the Dependent Variable  233

Step 5: Choose the Right Statistical Test  233
Step 6: Use Data Analysis Software to Test the Hypothesis  233
The Case of Never Saying Never  234
Step 1: Identify the Problem  235
Step 2: State a Hypothesis  235
Step 3: Identify the Independent Variable  235
Step 4: Identify and Describe the Dependent Variable  235
Step 5: Choose the Right Statistical Test  236
Step 6: Use Data Analysis Software to Test the Hypothesis  236
Just in Case—A Nonparametric Alternative  237
Summary  237
Do You Understand These Key Words and Phrases?  237
Do You Understand These Formulas?  237
Quiz Time!  238
The Case of Technology and Achievement  238
The Case of Worrying about Our Neighbors  239
The Case of SPAM  240
The Case of “We Can’t Get No Satisfaction”  241
The Case of “Winning at the Lottery”   242

CHAPTER 9 Analysis of Variance and Multivariate Analysis of Variance 244

Introduction  244
Understanding the ANOVA  245
The Different Types of ANOVAs  246
One-Way ANOVA  246
Factorial ANOVA  246
Multivariate ANOVA (MANOVA)  247
Assumptions of the ANOVA  248
Random Samples  248
Independence of Scores  248
Normal Distribution of Data  248
Homogeneity of Variance  248
Calculating the ANOVA  249
Descriptive Statistics  249
The Total Variance  251
The Total Sum of Squares  252
The Between Sum of Squares  253
The Within Sum of Squares  253
Computing the Degrees of Freedom  255
Computing the Mean Square  255
Computing the F Value  255
The F Distribution  256
xviii | Extended Contents

Determining the Area under the Curve for F Distributions  257

The p Value for an ANOVA  258
Effect Size for the ANOVA  258
Testing a Hypothesis Using the ANOVA  259
The Case of Multiple Means of Math Mastery  259
The Post-Hoc Comparisons  263
Multiple-Comparison Tests  263
Always Observe the Means!  264
The Case of Seniors Skipping School  265
The Case of Quality Time  268
The Case of Regional Discrepancies  271
The Factorial ANOVA  274
The Case of Age Affecting Ability  275
Interpreting the Interaction p Value  282
The Case of the Coach  284
The Multivariate ANOVA (MANOVA)  287
Assumptions of the MANOVA  288
Using the MANOVA  288
The Case of Balancing Time  292
Summary  295
Do You Understand These Key Words and Phrases?  296
Do You Understand These Formulas?  296
Quiz Time!  296
The Case of Degree Completion  296
The Case of Seasonal Depression  297
The Case of Driving Away  299
The Case of Climbing  300
The Case of Employee Productivity  301

CHAPTER 10 The Chi-Square Tests 304

Introduction  304
The One-Way Chi-Square Test  304
The Factorial Chi-Square Test (the Chi-Square Test
of Independence)  305
Computing the Chi-Square Statistic  306
The Chi-Square Distribution  307
What about the Post-Hoc Test?  309
Working with an Even Number of Expected Values  311
The Case of the Belligerent Bus Drivers  313
The Case of the Irate Parents  316
The Chi-Square Test of Independence  318
Computing Chi-Square for the Test of Independence  319
Computing Expected Values for the Test of Independence  319
 Extended Contents | xix

Computing the Chi-Square Value for the Test of Independence  320

Determining the Degrees of Freedom for the Test of Independence  321
We Are Finally Going to Test Our Hypothesis  321
Using SPSS to Check What We Just Computed  322
The Case of Corporal Punishment  324
Post-Hoc Tests Following the Chi-Square  327
The Case of Type of Instruction and Learning Style  327
Summary  330
Do You Understand These Key Words and Phrases?  330
Do You Understand These Formulas?  330
Quiz Time!  331
The Case of Prerequisites and Performance  331
The Case of Getting What You Asked For  332
The Case of Money Meaning Nothing  332
The Case of Equal Opportunity  333

CHAPTER 11 The Correlational Procedures 335

Introduction  335
Understanding the Idea of Correlations  335
Interpreting Pearson’s r  339
A Word of Caution  340
An Even More Important Word of Caution!  341
A Nonparametric Correlational Procedure  344
The p Value of a Correlation  345
The Case of the Absent Students  346
Another Example: The Case against Sleep  350
The Case of Height versus Weight  353
The Case of Different Tastes  355
Once We Have a Linear Relationship, What Can We Do with It?  358
Linear Regression  358
The Regression Equation  358
Computing the Slope  359
Computing the Intercept  361
Why Wasn’t It Exactly Right?  365
Using the Six-Step Model: The Case of Age and Driving  367
Summary  370
Do You Understand These Key Words and Phrases?  370
Do You Understand These Formulas?  371
Quiz Time!  371
The Case of “Like Father, Like Son”  371
The Case of “Can’t We All Just Get Along?”  372
The Case of More Is Better  373
The Case of More Is Better Still  374
xx | Extended Contents

Conclusion: Have We Accomplished What We Set Out to Do? 375

Statistics in a New Light  375
A Limited Set of Statistical Techniques  375
The Use of Statistical Software Packages  376
A Straightforward Approach  376
At Long Last  376

APPENDIX A. Area under the Normal Curve Table (Critical Values of z) 377
APPENDIX B. Critical Values of t 378
APPENDIX C. Critical Values of F When Alpha = .01 379
APPENDIX D. Critical Values of F When Alpha = .05 381
APPENDIX E. Critical Values of F When Alpha = .10 383
APPENDIX F. Critical Values of Chi-Square 385
APPENDIX G. Selecting the Right Statistical Test 386

Glossary 387
Answers to Quiz Time! 397
Index 421
About the Author 433
 Introduction: You Do Not Need to Be a
Statistician to Understand Statistics!

As I said in the first edition of this book, there aren’t a lot of students who take a
statistics class on purpose; most of them are there only to fulfill the requirements of
their chosen major. It is even rarer to find a student pleasantly anticipating what the
semester will bring; they’ve heard the horror stories, they have no clue how statistics
affect their lives, and many aren’t much interested in finding out! I’ve had a lot of
readers tell me the approach I took in the first edition really helped them. Although
I really appreciate their compliments, I know there are always new opportunities for
learning—both for me and my students! Before I explain what I’ve done to hopefully
make that happen, let me talk a little about how I got to where I am.

A Little Background
I took my first statistics class as a prerequisite to a graduate program. I am the first to
admit that I didn’t understand a thing. The formulas were mysterious, the professor
even more so, and, for the life of me, I couldn’t tell the difference between a t-test and
a teacup. I struggled through, made a passing grade, and went on to graduate school.
As part of my curriculum, I had to take another statistics class. Much like the first,
I couldn’t make heads or tails out of what the teacher was saying, and the C grade I
earned was proof of that. “At least I passed,” I thought. “I will never have to take a
statistics course again!”
Never say never, I soon learned, for I had barely finished my master’s degree when I
decided to continue for the doctorate. Faced with four required research and statistics
classes, I resolved to do my best and try to make it through. I figured I had already

1
2 | STATISTICS TRANSLATED

passed the prior two classes and could prevail again. Besides, I thought, if I couldn’t
do the work, I didn’t deserve the degree.
Although I entered my first doctoral class looking and feeling much like the stu-
dents I described earlier, things soon took a change for the better. During the class I
discovered the teacher was a “real person,” and this “real person” was explaining sta-
tistics in a way I could understand. Not only could I understand what she was saying,
I could understand the largely unintelligible textbook! I came out of the class feeling
great. In fact, not only did I learn what I needed to learn, but I became so interested
that I wound up taking quite a few extra statistics courses.
When I finished my doctorate, I liked what I was doing so much I convinced
the dean where I worked to allow me to teach statistics in our graduate school. I was
excited about teaching this subject for the first time, and, although that was over 25
years and many classes ago, I still love what I am doing. People look at me funny when
I say this, but most of them are like the people I have described earlier. What they do
not know is I’ve learned a few things that make it far easier to use and understand
statistics; I will tell you what they are.

Many Students Do Not Know What They’re Getting Into

The first thing I have discovered is most people do not really know the definition of
statistics. To me, it seems odd they would be sitting in a class where they do not know
what they are about to study, but it occurs all the time. Although we will cover this
topic in great detail later in the book, let me give you a simple definition for now:

„ The science of statistics is nothing more than the use of arithmetic tools to
help us examine numeric data we have collected and make decisions based
on our examination. The first of these tools, descriptive statistics, helps
us understand our data by giving us the ability to organize and summarize
it. Inferential statistics, the second tool, helps us make decisions or make
inferences based on the data.

Second, I’ve found that students greatly overestimate the proficiency in math they
need in order to calculate or comprehend statistics. Whenever I find people worry-
ing about their math skills, I always ask them, “Can you add, subtract, multiply, and
divide? Can you use a calculator to determine the square root of a number?” Often,
their answer is “yes” so I tell them, “That is great news! You have all the math back-
ground you need to succeed!” I’ve found that many people, when I tell them this,
breathe a big sigh of relief!
Next, I’ve discovered that a class in statistics isn’t nearly as bad as people antici-
pate if they approach it with an open mind and learn a few basic rules. Since most
people want to be able to use statistics to plan research, analyze its results, and make
sound statistically based decisions, they want to become consumers of statistics rather
than statisticians per se. As consumers, they want to be able to make decisions based
on data, as well as understand decisions others have made. With that in mind, this
 Introduction | 3

book will help you; all you must do is approach it with an open mind and learn a few,
easily understood steps.
Fourth, I have found that students get confused by all the terminology thrown
around in a statistics course. For instance, for a given statistical problem you could
find the correct answer by using procedures such as the Scheffé test, Tukey’s HSD test,
or the Bonferroni procedure. The truth is, you could use any one of these techniques
and get very similar answers. I was recently at a meeting where a very prominent stat-
istician jokingly said we do this “to confuse graduate students”; there certainly seems
to be a lot of truth in what he said! These types of terms, as well as the ambiguity of
having many terms that mean basically the same thing, seem to intimidate students
sometimes. Knowing that, I like to stick to what students need to know most of the
time. We can deal with the exceptions on an “as needed” basis.
Along these same lines, I’ve discovered many instructors try to cover far too much
in an introductory statistics class, especially when they are dealing with people who
want to be consumers of statistics. Research shows that 30 to 50% of studies published
in prestigious research journals use a small set of common descriptive and inferential
statistics. My experience, however, is that the average student tends to use the same
basic statistical techniques about 90% of the time. Knowing that, we are left with a
manageable problem: we need to be able to identify when to use a given statistical test,
use it to analyze data we have collected, and interpret the results. To help you do that,
I have developed a set of six steps to guide you. Using these steps, you will be able to
work with the statistics you need most of the time.

A Few Simple Steps

You may be asking yourself, “Why are these steps so important? How could something
so important be so easy? Exactly what are these steps?” If so, we can begin to answer
your questions with a short story:
Do you remember back in high school when you took beginning algebra? Can
you remember what you did the first few days of class? If you can’t, let me refresh your
memory.
You spent the first few class sessions learning about the basic rules of algebra.
These included topics such as the commutative property, the associative property, the
distributive property, and other things that seemed really boring at the time. If you
were like me, you probably questioned spending valuable time worrying about rules
that did not seem to apply to algebra. If you were really like me, you didn’t pay atten-
tion since you were saving your energy and attention until it could be used for “real”
math.
Unfortunately, for many of us, our plan backfired. We soon discovered that learn-
ing algebra was a linear process and we had tried to start in the middle of the line;
learning those boring rules was important if we were going to succeed. At that point,
in order to understand what was going on, we found it necessary to go back and learn
what we should have already learned. Some of us went back and learned; some of us
didn’t.
We can apply this analogy to my steps for statistics. In the beginning we will
spend a lot of time discussing these six steps and the fundamentals of statistical deci-
4 | STATISTICS TRANSLATED

sion making. After we get through the basics, we will move into the “real statistics.” For
now, just read through these steps and the brief description of what each step entails.

Identify the Problem

S TEP 1
At the outset, researchers need to understand both the prob-
lem they are investigating and why investigating it is significant.
They may begin by asking questions such as “How do I know
this is a problem?”, “Why is investigating this problem impor-
tant?”, “What causes this type of problem?”, and “What can
possibly be done to address this problem?” In other cases, researchers may not face a
specific problem; rather they may have an opportunity to conduct or expand research
in a given field. Either way, once they have answered those questions, they need to
develop their problem statement. In this step, they are identifying “what” we are going
to investigate; the remaining five steps will help us understand “how” to investigate it.

State a Hypothesis
S TEP 2
A hypothesis is a researcher’s beliefs about what will be found
or what will occur when he or she investigates a problem. For
example, let’s suppose that our school is faced with lower than
average scores on a statewide math test. We have heard that it is
possible that the problem may be based on gender; apparently
males do not perform as well in math as their female classmates, and this may cause
the classes’ overall average to be lower. Since we have considerably more males than
females in our school, we then must ask ourselves, “How can I find out if this is true?”
Or, if we plan on using some type of intervention to increase the achievement of our
male students, we could ask, “What do I think will occur based on what I do?” These
questions lead us to make a statement regarding our beliefs about what will occur. As
I said earlier, stating the hypothesis is the first step in understanding “how” we can
investigate the problem.

Identify the Independent Variable

S TEP 3
When statisticians state a hypothesis, they must identify what
they believe causes an event to occur. The first major part of the
hypothesis is called the independent variable and is the “cause”
we want to investigate. In most instances, the independent vari-
able will contain two or more levels. In our preceding example,
our independent variable was gender, and it had two levels, male and female. In other
cases, we may be investigating a problem that has more than one independent vari-
able; the levels of each independent variable will still be identified in the manner we
have just described.
 Introduction | 5

Identify and Describe the Dependent Variable

S TEP 4
Once researchers identify what they believe will cause an event
to occur, they then must determine what they will use to mea-
sure its effect. This second major part of the hypothesis is called
the dependent variable. The researcher will collect data about the
dependent variable and then use descriptive statistics to begin to
understand what effect, if any, the independent variable had on it. In our example, we
are interested in student achievement—our dependent variable. Again, as we said, a
hypothesis may have more than one dependent variable.

Choose the Right Statistical Test

S TEP 5
Although there are many different inferential statistical tests,
information about the independent variable, the dependent
variable, and the descriptive statistics will, in most instances,
tell the researcher exactly which statistical test to use. We will
learn to compute both the descriptive and inferential statistics
manually and we will also see examples of computer software that can assist us in
becoming good consumers of statistics.

Use Data Analysis Software to Test the Hypothesis

S TEP 6
The manual computations, or the results from statistical soft-
ware, will give the researcher the information necessary to test a
hypothesis. Based on this, the decision can be made as to whether
the independent variable had any effect on the dependent vari-
able.

So, What’s New in This Edition?

I’ve found that these rules are straightforward; many students are surprised at how
comfortable and confident they became after reading and applying them. I’m the first
to admit, however, that I continue to learn what might be helpful from students, from
my colleagues, and from the publications I read. Therefore, I’ve expanded this edition
by including more examples of each of the tests, many of which relate to disciplines I
didn’t discuss in the first edition. I’ve also expanded the chapter covering the analysis
of variance (ANOVA) by including a section on the multivariate analysis of variance
(MANOVA), a tool that’s becoming more and more common for student use. Finally,
since new versions of statistical software packages come out frequently, I’ve upgraded
the examples of SPSS in this book to use the latest, greatest version!
6 | STATISTICS TRANSLATED

Summary
As I’ve said, if you understand these six steps and what we are about to discuss,
you can become a proficient consumer of statistics. Let me remind you that this
isn’t going to be an exhaustive overview of statistics. Again, about 90% of the time,
students use a given set of statistics over and over. That is what this text is about—
the common set of statistical tools that you, as a consumer of statistics, need to be
familiar with.
In the next few chapters, we are going to go through each of these steps, one by
one and explain, in detail, what is necessary to accomplish the step. After we have
looked at each step in detail, we will then apply the knowledge we have gained by
analyzing data and making decisions based on our analysis. Believe me, it is going to
be far easier than you think, so let’s get started!

Do You Understand These Key Words and Phrases?

consumer of statistics dependent variable
descriptive statistics hypothesis
independent variable inferential statistics
levels of the independent variable problem statement
CHAPTER 1

Identifying a Research Problem

and Stating Hypotheses

Introduction
The first two steps in the six-step model should help good consumers of statistics
understand exactly what they intend to investigate. As I said earlier, Step 1 focuses on
“what” will be investigated by clearly defining the problem; Step 2 starts the process
of describing “how” it will be investigated. Unfortunately, far too many people try to
get to Step 2 without understanding how it relates to the problem at hand. Because
of that, let’s see what constitutes a good problem statement.

S TEP 1 Identify the Problem

I always caution my students not to get caught up in worrying
about hypotheses, variables, and statistical tests before they
know and can accurately state the problem they want to inves-
tigate. In fact, my very words are “Stop running around with a
solution looking for a problem!” Admittedly, I have also said, “Don’t find a hammer
and then look for a nail,” and “Don’t buy a plane ticket before you know where you
want to go.” The point is, I want them to ask, “Exactly what is the problem I want to
investigate?” prior to ever thinking of “How will I collect and analyze data?”
At this point, many would argue that I’m crossing the line between a research
methods book and one dedicated to statistical decision making. Their reasoning is
that the problem statement leads to the development of a hypothesis only after the
researcher has read the relevant literature, has learned as much as possible about the
problem to be investigated and can develop a sound basis for his or her hypothesis. My

7
8 | STATISTICS TRANSLATED

argument is, if you’re a consumer of statistics, most of the time the problem statement
will be clearly stated for you; if not, in most instances it’s clearly implied.
I’ll admit that this point is arguable, but the bottom line is this: whether we’re con-
sumers of statistics or conducting actual research, we still need to be able to recognize
a good problem statement or develop one when necessary. Either way, keep in mind
that, in Step 1, we are identifying “what” we are going to investigate; the remaining
five steps will help us understand “how” to investigate it. Knowing that, we’re going
to move forward by discussing the characteristics of a good problem statement; we’ll
then learn how to correctly write a problem statement.

Characteristics of a Good Problem Statement

The problem statement represents the specific issue, concern, or controversy that we
want to investigate. A good research problem must meet six criteria:

1. The problem is interesting to the researcher.

2. The scope of the problem is manageable by the researcher.
3. The researcher has the knowledge, time, and resources needed to investigate
the problem.
4. The problem can be researched through the collection and analysis of numer-
ic data.
5. Investigating the problem has theoretical or practical significance.
6. It is ethical to investigate the problem.

Obviously, before we can determine if a problem meets these six requirements,

we first need to learn how to find a problem to investigate; how do we do this? We can
explain the entire process with an example.

Finding a Good Research Problem

Problems that we can investigate are all around us: they come from personal experi-
ence, issues at your workplace or institution, or reading about a topic you are interested
in. Other potential areas for investigation can be found by attending professional
conferences, speaking with experts in the field you are interested in, or replicating
the work of others to better understand or apply the results of research they have con-
ducted.
For example, suppose I am working in a graduate school of psychology and our
students’ passing rate on the state licensure test is below average; that is clearly a prob-
lem I could investigate:

„ Our school has a lower than average passing rate on the state licensure test.
This type of problem is generally called a practical research problem in that it focuses
on an issue within our own school or organization. We may also investigate research-
based problems: those that come from conflicts or contradictions in previous findings
 Identifying a Research Problem and Stating Hypotheses  |  9

or a desire to extend the knowledge about a problem area. For example, let’s say that,
in our readings, we have found conflicting research on the effectiveness of using age-
appropriate popular press material, such as magazines, to teach reading in elemen-
tary schools; we could easily write a problem statement such as:

„ There is conflicting research regarding the use of popular press materials to

teach reading in our elementary schools.

Despite the source of the problem we’re investigating or the type of research it
represents, we must keep the six characteristics of a good problem in mind. As you are
reading these characteristics, keep in mind that we’re not going to look at them in any
order; they all must be met before we can move forward.

The Problem Is Interesting to the Researcher

Many times, I run into students who simply want to finish a project, thesis, or disserta-
tion; they do not care what they work on, as long as they are working on something.
Unfortunately, this approach often fails simply because they aren’t interested enough
to see it through to the end. Far too often this leads to students staying in school lon-
ger than anticipated or dropping out of school before they earn their degree.
For example, I once had a student who was near her wits’ end trying to find a topic
for her senior project. She was a very bright student, so I asked her to work with me on
a project investigating achievement in elementary school science. She jumped at the
chance to work on the project—not necessarily because she was interested but because
she saw it as a way to graduate. Unfortunately, she did not finish the project with me;
she simply didn’t have the energy, effort, and enthusiasm she needed to work on some-
thing that didn’t really interest her. Two good things did come from it, however. She
did graduate after she found a suitable project, and I ultimately finished what I was
working on; both of us met our goals because we were working on something we were
interested in.
On the other hand, several years ago one of my favorite students approached me
with a practical research problem. His school’s passing rate in algebra was abysmal,
and he wanted to investigate the effect of using laptop computers on achievement. I
warned him against choosing this topic because I felt that using technology in a class-
room had already been investigated time and time again. He assured me, however,
that he was approaching the problem from a new perspective, so I agreed to let him
continue. He was so excited and motivated that he finished the best dissertation I
have ever chaired. He was interested in his problem, and his personal goal was to fin-
ish. He did an excellent job and his students’ achievement went up; it was a “win-win”
situation.

The Scope of the Problem Is Manageable by the Researcher

When my students say that they cannot find a problem to work on for their project,
thesis, or dissertation many times I tell them, “It’s not that you cannot find a problem,
there are literally thousands to choose from. Your problem is that you cannot find one
10 | STATISTICS TRANSLATED

with a scope you can manage.” For example, look at the following “problem statement”
presented to me by one of my students:

„ One of the most dynamic affordances of the Internet is the ability to offer
educational programs to students throughout the globe. This has led to
nearly 90% of all universities and colleges in the United States offering
Internet-based courses. While convenient for students, research has shown
that attrition from Internet-based classes is significantly higher than that
from classes taught in a traditional environment.

The problem here is certainly clear; students in Internet-based courses drop

out more often than students in traditional classrooms. This certainly seems to be a
serious problem and one that bears investigation. While important, however, is this
something my student could investigate as written? Of course not! The scope is much
too broad. In this case, it seemed that my student was suggesting an investigation of
Internet-based education throughout the United States. In such cases, I tell my stu-
dents not to attempt to “boil the ocean”; if the issue is so large that it cannot be easily
investigated, focus on narrowing it down into something more specific and manage-
able. This idea is shown in Figure 1.1.
For example, the problem statement above could be easily reworded as a practical
problem statement by simply stating

„ The attrition rate in our school’s Internet-based classes is higher than the
attrition rate from our school’s traditional classes.

By narrowing down the first statement, the student was able to create a problem
statement that could be easily investigated.

The Researcher Has the Knowledge,

Time, and Resources Needed to
Investigate the Problem
This requirement is pretty much a “no
brainer,” isn’t it? It only makes sense that
the researcher meets all these criteria; lack-
ing any of them will ultimately either cause
the research to fail or, at best, add signifi-
cantly to the amount of time it would take
to complete. I’ll give you an example.
Many years ago, I taught a class in
which one of the things we talked about
was brain hemispheric tendencies in learn-
ing. Although the theory has fallen out of
FIGURE 1.1.  Narrowing down and focusing favor in many circles, it hypothesized that
the problem statement. persons with a left-hemispheric tenden-
 Identifying a Research Problem and Stating Hypotheses  |  11

cy in learning processed information in a linear manner, while persons with right-­

hemispheric tendencies learned in a more holistic fashion. This really caught the
attention of one of my students who wanted to investigate achievement in courses
she taught in introductory computer programming. She felt that, since programming
is highly linear and logical, “lefties” would have significantly higher achievement than
“righties.” This viewpoint seemed interesting to me, but when I asked her how much
she knew about learning and cognition, she said the only thing she knew about hemi-
spheric preference was what I talked about in class. Although she did successfully com-
plete her dissertation, it took quite some time for her to get up to speed in the skills
and knowledge she needed. What did she ultimately determine? Interestingly, it didn’t
make any difference which side of the brain a student tended toward; the only thing
that mattered was if the student used their brain at all!

The Problem Can Be Researched through the Collection and Analysis

of Numeric Data
This requirement is also obvious. Since our definition of statistics included “the use
of arithmetic tools to help us examine numeric data,” it only stands to reason that we
cannot collect data from such things as interviews, transcripts, or recordings (these
are commonly referred to as qualitative data). Later in the book we will talk about dif-
ferent types of numeric data, but for now let’s just accept the fact that numeric data
are required and move forward.

Investigating the Problem Has Theoretical or Practical Significance

In looking back at the problems dealing with attrition from Internet-based classes,
and the effects of using laptop computers in algebra classes, we can readily see that
investigating each problem is worthwhile.
In other cases, we might investigate a problem that has theoretical significance. In
my dissertation, for example, I investigated the following problem:

„ Students at Lakeside Elementary School have low levels of achievement on

state-mandated math tests.

In planning my study, I focused on the theory that achievement is highly cor-

related to a student’s intrinsic motivation to learn a subject. Because of this, instead of
simply investigating the achievement problem, I decided to focus on the idea that, if I
could devise a way to increase a student’s intrinsic motivation, math achievement would
increase as well. Based on the theory that timely and informational feedback may
increase intrinsic motivation, I designed and conducted my dissertation research. I
did this by providing elementary school students with weekly graph-based “cause-and-
effect” report cards rather than the traditional report card they received every six to
nine weeks. Unfortunately, my intervention did not work as planned. While this may
sound like bad news, my dissertation chair pointed out two things. First, he agreed
that my study was theoretically sound. He also agreed that my study had been con-
12 | STATISTICS TRANSLATED

ducted in a valid manner; therefore, the study seemed to show that the intervention
didn’t work!
Let’s look at a few more examples and determine if their investigation is either
theoretically or practically significant:

„ Youth who receive their driver’s license at age 16 have a higher number of
accidents than youth who obtain a license at age 18 or older.

This problem would seem to have some practical significance, but there are obvi-
ously two reasons we might not want to investigate it. First, is it really age that affects
the number of accidents, or is it the amount of experience the driver has? Doesn’t it
stand to reason that youngsters who receive their license at 18 would have, percentage-
wise, the same number of accidents as those who started earlier? As beginning driv-
ers, they would have the same level of experience as their 16-year-old peers. Second,
this problem doesn’t warrant a full-scale investigation; it would be easy to simply look
at the percentage of drivers in each age group who have had accidents to determine
whether there is a difference.
Let’s look at one last problem:

„ Nurses whose first job is in a private practice are far more likely to change
jobs within three years than nurses whose first job is in a hospital or other
health care setting.

While this problem seems like it might bear investigating, I would ask the ques-
tion, “Why is it important how often a nurse changes location? Does this have a nega-
tive effect on the facility they are leaving?” In short, this seems like only half of a
practically significant problem statement:

„ Nurses whose first job is in a private practice are far more likely to change
jobs than nurses whose first job is in a hospital or other health care setting.
Replacement and training of new personnel is a burden and the burden has
been shown to negatively affect patient care.

Writing the problem statement in this manner clearly shows its practical signifi-
cance.

It Is Ethical to Investigate the Problem

The history of research is filled with studies that were clearly unethical—withholding
drugs to patients diagnosed with a potentially deadly virus in order to look at the
long-term effects of the disease, to simulating electric shock to investigate a person’s
reaction to authority. It’s obvious to most researchers when a study is not ethical, but,
at most institutions, standards have been established to approve research studies using
human subjects before they are conducted. In most cases, the school, institution, or
business where the research is to be conducted has personnel who will review the pro-
 Identifying a Research Problem and Stating Hypotheses  |  13

posed research to ensure whether it should be approved. As a researcher, if you aren’t

sure, always ask.

Writing the Problem Statement

As you will recall, in Step 1 of our six steps, we’re interested in “what” the researcher
plans to investigate; the remaining steps will describe how it will be investigated. We
have just learned what a problem statement is and have discussed the characteristics
of a good problem statement. The final piece of the puzzle is learning how to write a
good problem statement. Again, some would argue that this crosses the line between
a research methodology book and a statistics book. As I have said, as consumers of
statistics, the problem statement will usually be clearly stated or implied, but we still
need to be able to recognize one when we see it. Knowing that, let’s look closer at the
characteristics of a well-written problem statement.

Problem Statements Must Be Clear and Concise

The first thing to keep in mind as you write a problem statement is that the reader
must understand what the problem is; in order to do that, the problem must be stated
as clearly and concisely as possible. We alluded to this point in an earlier section, but
let’s look at another example to ensure we understand exactly what this means. It’s
common for me to receive a problem statement such as this:

„ Reduced or free lunches and self-esteem.

What does this mean? Is what the author is trying to say clear and concise? In this
case, it is not. In order to state the problem more clearly, the author needs to establish
the relationship between the reduced or free lunches and student self-esteem. It could
be presented in this manner:

„ Students receiving free or reduced-price lunches seem to have lower levels of

self-esteem than classmates not receiving free or reduced-price lunches.

Thinking back, does it meet the six criteria of a good problem? First, we can
only hope that it’s interesting to the researchers and that they have the necessary time,
skills, and resources; if not, why would they want to investigate the problem? Next,
the problem can certainly be analyzed through the collection and analysis of numeric
data, and investigating it has practical significance. It appears that it would be ethical
to investigate the problem, but, again, an institutional review panel could confirm
that for us. Finally, the scope of the problem is manageable since it is limited to one
school. How about this?

„ Epilepsy and diet.

In this case, we do not know much about what the researchers are proposing to
investigate, but it could be reworded to make it clearer:
14 | STATISTICS TRANSLATED

„ Children diagnosed with epilepsy and whose diets contain high levels of fat
and sugar have a greater number of seizures than children diagnosed with
epilepsy who eat healthier diets.

Here again, we’re assuming the researcher is interested in the problem and has
the necessary knowledge, skills, and time. Investigating this problem has both theo-
retical and practical significance, and numeric data can be collected. There is a bit
of an issue with the ethics of investigating this problem in that (1) the subjects are
children and (2) they are being treated for a medical condition; both of these raise a
red flag with review boards and would be closely examined to ensure the health and
well-being of the children while in the study. The scope of the problem, however, seems
to be our biggest concern. Who are the children the author wants to work with? All of
those in the United States? The state they are in? Who? In short, this problem state-
ment should reflect a much narrower scope—one the researcher has access to and can
work with.

The Problem Statement Must Include All Variables to Be Considered

The second criterion for writing a good problem statement is that each of the vari-
ables to be investigated must be included. For example, in the first problem statement
above, we wanted to investigate the relationship between free or reduced lunch prices
(the cause) and self-esteem (the effect). No other variables were included in the prob-
lem statement. In the second statement, we included diet type (the cause) and the
number of seizures (the effect); again, no other variables were mentioned.
If we wanted to investigate the effect of different types of incentives such as high-
er pay, more vacation, or better office conditions on employee morale, would there be
anything wrong with this problem statement?

„ This study proposes to investigate the relationship between different types of

incentives given to employees and levels of motivation.

In this case, we have included a variable that describes the cause—different types
of incentives—but instead of a variable measuring morale, we have included one that
measures motivation—these are two completely different things. Employees could be
highly motivated for fear of losing their jobs, but morale could be very low. While this
problem statement looked good at first glance, remember the first criterion: a prob-
lem statement must be clear and concise.
Finally, is this a good problem statement if we are interested in investigating the
ability of a dog to be trained and whether the dog was purchased from a breeder or
from a pet store?

„ House-training a new puppy can be a difficult task. This study will

investigate whether there is a relationship between where a puppy was
purchased and its ability to be effectively trained.

This statement seems to be OK; the location where the dog was purchased is the
 Identifying a Research Problem and Stating Hypotheses  |  15

cause we want to investigate and the dog’s ability to be trained is the effect. Thinking
back in this chapter, however, are there other issues with this problem statement that
should concern us? I would argue that there is no practical or theoretical significance
in investigating the problem; after all, what difference does it make? When it comes to
purchasing a dog, aren’t there a lot more factors involved in selecting one than their
potential to be trained not to bark? Besides, if the dogs are well trained, who cares
where they came from?

The Problem Statement Should Not Interject the Researcher’s Bias

Here again lies one of the biggest pitfalls that underlie research: far too many people
want to “prove” things or conduct research that supports their personal beliefs or
goals. For example, if I was arguing for the need for more technology in our work-
place, I might write:

„ Management in our company is concerned by low levels of employee morale.

This study will prove that new computers can increase employee enthusiasm.

That is a large assumption on my part, isn’t it? Suppose I do conduct research and
show that there is increased morale and enthusiasm after obtaining new technology?
Have I proven anything? Absolutely not! There are far too many things that affect
morale: the employees themselves, a change in the work requirements or responsibili-
ties because of the new technology, unrealistic administrative expectations and so on.
At the same time, what would happen if levels of morale went down? Would we go
back to Apple, Dell, or Hewlett-Packard and ask for our money back? Of course not;
we do not know if the technology had any effect on what we’re measuring. There are
far too many factors that influence morale to let us assume we can “prove” anything!

Summary of Step 1: Identify the Problem

Understanding and clearly stating the problem you are investigating is the first step
in becoming a good consumer of statistics. As we discussed, the problem statement in
Step 1 tells you “what” you are going to investigate, while the remaining five steps will
describe “how” you will investigate the problem. Keep in mind, while formulating a
good problem statement, you must ensure you meet the following criteria:

1. You must be interested in the problem you are investigating.

2. The scope of the problem you want to investigate must be manageable.
3. You must be comfortable in terms of the knowledge, time, and resources nec-
essary for you to investigate the problem.
4. You must be able to collect and analyze numeric data.
5. There must be some practical or theoretical reason for you to research the
problem.
6. It must be ethical for you to investigate the problem.
16 | STATISTICS TRANSLATED

Once you are sure you have met these criteria and begin writing the actual prob-
lem statement, you must ensure that you are clear and concise, that all variables to be
investigated are included, and that you do not interject your personal bias. Following
all these rules ensures that we have a problem statement we can use to state a hypoth-
esis in Step 2 of our six-step model.

S TEP 2 State a Hypothesis

In this section, we want to accomplish several things. First, we
need to better define a hypothesis and then understand why
and how hypotheses are stated for the problem we are investi-
gating. After that, we will discuss directional and nondirectional
research hypotheses and learn why statisticians always test hypotheses stated in the
null format. After covering these basics, we are going to take a brief look at hypothesis
testing. We will cover hypothesis testing in greater detail later in the book, but this
overview will give us an idea of where we are going.

An Example of Stating Our Hypothesis

Let’s suppose I am the principal of a high school and I have decided to spend the day
reviewing scores from the most recent administration of a standardized science test
called the ABC. Interestingly, as I am reviewing the science scores, a pattern seems to
emerge. Students taking science in the morning appear to consistently score higher
on the ABC than students taking science in the afternoon. Based on my observations,
I surmise there is some type of meaningful relationship between the time-of-day stu-
dents take science and how well they do on the ABC test.
Before moving forward, what have I just identified? The problem of course! Stu-
dents taking science in the morning score higher than students taking science in the
afternoon. It meets all the criteria for being a problem I can investigate. I am inter-
ested in the achievement of my students; I am certainly comfortable with the problem
area and the scope of the problem is manageable. The problem has practical signifi-
cance, it is ethical to investigate, and I can collect numeric data for analysis. It sounds
good so far but what about the way the problem is written? While it seems obvious,
let’s write it out to make sure:

„ Science achievement scores of the morning students are higher than science
achievement scores of the afternoon students.

Yes, this is very clear, and all the variables to be considered are included; we’re
comparing the achievement of students in the morning to that of students in the
afternoon. I have not interjected my bias; I have simply stated what I have observed.
Throughout the rest of this chapter, we’re going to focus on writing hypotheses. While
we won’t state the problem each time, as I said in the last chapter, it will be clearly
inferred.
Given that, I decide to follow up on my observations by going to the library and
 Identifying a Research Problem and Stating Hypotheses  |  17

try to determine if there have been studies investigating the effect of time of day on
achievement. Much to my surprise, I discover a lot of information suggesting that stu-
dents studying science in the morning seem to outperform students taking science in
the afternoon. Based on this finding, my observations are supported; the time of day
a student takes science might influence his or her class achievement.
Now, based on what I have observed and read, I can make a statement about what
I believe is occurring in my school. In other words, I can state a hypothesis since, by
definition, a hypothesis is nothing more than a statement expressing the researcher’s
beliefs about an event that has occurred or will occur. A well-stated hypothesis has
four requirements:

1. It must provide a reasonable explanation for the event that has occurred or
will occur.
2. It must be consistent with prior research or observations.
3. It must be stated clearly and concisely.
4. It must be testable via the collection and analysis of data.

Knowing these four things, I can state a formal hypothesis based on what I have
seen:

„ Students taking science in the morning have higher levels of achievement

than students taking science in the afternoon.

By reading this hypothesis, we can see it meets our four criteria. First, it is stated
clearly and is consistent with both our prior observations and reading. It is a reason-
able explanation for what has occurred, and we are able to collect data for statistical
analysis. Given these things, it appears we have met the criteria—we have a well-stated
hypothesis.

A Little More Detail

Before we move forward, it is a good idea to go into a little more detail about how we
state hypotheses and what we do with them once we state them. We have seen that
there are four basic rules, and we have looked at a good example that follows those
rules. Some of these rules are very straightforward, but there are nuances we should
clear up to avoid possible misunderstandings later.

The Direction of Hypotheses

As we just saw, both my observations and the research suggested that students study-
ing science in the morning have higher achievement than those taking science in the
afternoon. Since it would contradict both my observations and readings, it would not
make sense to hypothesize that the time of day during which a student takes science
has no effect on achievement. Let’s start by looking at how we can use directional
hypotheses to test either a “greater than” or “less than” scenario.
18 | STATISTICS TRANSLATED

Since hypotheses must be stated to reflect our readings or observations, it is nec-

essary to state either directional or nondirectional hypotheses. The use of a directional
hypothesis shows we expect a “greater than” or “less than” relationship in our results.
Nondirectional hypotheses indicate that we expect a difference in our results, but we
do not know if it will be a “greater than” or “less than” relationship. We can use both
types of hypotheses to make comparisons between data we collect and an exact value
(e.g., the average body temperature of a group of patients compared to 98.6°), or we
can use them to compare two or more groups of data we have collected (e.g., the scores
on an achievement test between the students from two separate schools).

Using Directional Hypotheses to Test a “Greater Than” Relationship

In many instances, we might be interested in hypothesizing that one variable about
which we have collected data is greater than an exact value. For example, if we admin-
istered an IQ test to a group of soldiers, we might hypothesize that the average IQ of
our soldiers is greater than the national average of 100. This is a direct comparison of
our collected data to an exact value, so we would use the following hypothesis:

„ The average IQ of our soldiers is greater than 100.

When we are not comparing something against an exact value, we can state
hypotheses that compare data collected from two or more groups. For example, with
the hypothesis we stated concerning the difference in achievement between science
students in the morning and afternoon, we expect the morning group to have higher
scores than the afternoon group. We aren’t stating an exact value; we are just saying
that one group will do better than the other:

„ Students taking science in the morning have higher levels of achievement

than students taking science in the afternoon.

Using Directional Hypotheses to Test a “Less Than” Relationship

We can do the same thing when we are looking at a “less than” scenario. For example,
suppose we are concerned with the rising number of disciplinary problems in our
school and have read studies showing that students who are required to wear uniforms
have fewer disciplinary problems than students in schools where uniforms are not
required. Before requiring our students to wear uniforms, we decide to check this
assertion for ourselves. It would be easy to identify schools in our district where uni-
forms are required and schools where uniforms are not required.
After identifying the schools, it would be a simple matter to collect the necessary
data to investigate the following hypothesis:

„ Children attending schools where uniforms are required have fewer

disciplinary problems than children in schools where uniforms are not
required.
 Identifying a Research Problem and Stating Hypotheses  |  19

We can also make a comparison to an exact value. For example, imagine we are
investigating whether the five-year recidivism rate for inmates released from prisons
in our state is less than the national average of 76%. In order to conduct our study, we
would state the following hypothesis:

„ The recidivism rate of inmates released from prisons in our state is less than
the national average of 76%.

Nondirectional Hypotheses
In the preceding examples, the hypotheses were consistent with the literature or our
experience; it would not make sense to state them in any other manner. What hap-
pens, though, when the prior research or our observations are contradictory?
To answer this question, let’s use the debate currently raging over block sched-
uling as an example. For anyone who hasn’t heard of block scheduling, it involves
students in middle and high schools attending classes every other day for extended
periods. For example, a student might be enrolled in six classes, each of which meets
in a traditional class for one hour each day. In a school with block scheduling, the
student meets for three two-hour classes one day and meets for two hours with each of
the other three classes the next day. Proponents feel that longer periods allow teach-
ers to get into more detail and make for more meaningful class sessions. Opponents
of block scheduling criticize it because they believe many teachers do not know how
to effectively use a longer class period or, as is the case with many math teachers, they
believe students need to be exposed to their subject matter every day.
Instead of both sides of the issue giving us their personal beliefs or feelings, would
not it be a good idea to see what the literature says? The reviews of block scheduling
are just as mixed in the academic research journals. Some articles suggest that block
scheduling negatively affects student achievement, and others seem to show that it
increases student achievement. Knowing this, how do we write our hypothesis? The
answer is simple: we must state a nondirectional hypothesis:

„ There will be a difference in achievement between students taking algebra in

a block schedule and students taking algebra in a traditional schedule.

In this example, the hypothesis implies no direction; instead, we are stating that
we believe a difference in algebra achievement will exist between the two groups. No
“greater than” or “less than” direction is implied.
Just as was the case with the directional hypotheses, we can use a nondirectional
hypothesis to make a comparison to an exact value. Suppose, for example, someone
told me the average retirement age in the United States is 63. In order to investigate
this, I could ask a group of retirees how old they were when they retired. I could then
use that data to test the following hypothesis:

„ The average retirement age of workers in the United States is not 63.

Notice here, again, I have not said “less than” and I have not said “greater than.”
20 | STATISTICS TRANSLATED

All I care about is if a difference exists. Later in the book we will see that different sta-
tistical tests are used when we compare data we have collected against a specific value
or compare data we have collected from two or more groups.

Hypotheses Must Be Testable via the Collection and Analysis of Data

Like a good problem statement, a good hypothesis is one about which we can collect
numeric data. In order to analyze it, we will use inferential statistics; specialized,
mathematical tools developed to help us make decisions about our hypotheses. Know-
ing that, it is imperative that we word our hypotheses in a way that reflects or infers
the collection of some type of data.
For example, in our hypotheses involving the teaching of algebra, it would be
easy to collect final grades and use them to make comparisons between block and
traditional schedules. In our example about IQ, we could test our soldiers and then
compare their average score to the average national score. While these examples are
very straightforward, we need to develop them further.

Research versus Null Hypotheses

In each of the preceding examples, we have stated what is called a research hypothesis
or, as it is sometimes called, an alternate hypothesis. The logic behind the name is
very straightforward. First, it is stated based on prior research or observations, and,
second, it is what the researcher wants to investigate. In order to address potential
problems, we will cover later, it is imperative that we state a null hypothesis for every
research hypothesis we state. The null hypothesis is nothing more than the exact
opposite of the research hypothesis and, like everything else, can best be explained
with an example.

Stating Null Hypotheses for Directional Hypotheses

If we look back at our example in which we wanted to administer an intelligence test
to our soldiers, we hoped to show that our soldiers have higher than average levels of
intelligence. In order to do that, we stated the following directional research hypoth-
esis:

„ The average IQ of our soldiers is greater than 100.

Obviously if something is not greater than 100, it could be less than 100 or exactly
equal to 100. Since, in our research hypothesis, we are only interested in the “greater
than” scenario, we can state our null hypothesis in this manner:

„ The average IQ of our soldiers is not greater than 100.

It is just as easy to develop a null hypothesis for comparing the achievement of

students taking science in the morning to students taking science in the afternoon.
Our directional research hypothesis was:
 Identifying a Research Problem and Stating Hypotheses  |  21

„ Students taking science in the morning have higher levels of achievement

than students taking science in the afternoon.

In this case, the opposite of “higher levels” could either be less than or equal to:

„ Students taking science in the morning do not have higher levels of

achievement than students taking science in the afternoon.

Using our example concerning school uniforms, our research hypothesis was

„ Children attending schools where uniforms are required have fewer

disciplinary problems than children in schools where uniforms are not
required.

Here we are doing exactly the opposite of the prior two research hypotheses;
now we are using a “less than” research hypothesis. Given that, we will state our null
hypothesis in the following manner:

„ Children attending school where uniforms are required do not have fewer
disciplinary problems than children attending school where uniforms are not
required.

Finally, if we were investigating whether the average number of years a sailor stays
in the navy is greater than 20, we could state:

„ Sailors in the navy stay longer than 20 years.

Because we are only worried about the greater than scenario, our null hypothesis
would read:

„ Sailors in the navy do not stay longer than 20 years.

Issues Underlying the Null Hypothesis for Directional Research Hypotheses

Stating the null hypothesis for a directional hypothesis is a thorny issue. Although
the manner we just used is technically correct, it is common practice to write the null
hypothesis for a directional hypothesis to simply say that no difference exists. For
example, we just stated the following research hypothesis:

„ Children attending schools where uniforms are required have fewer

disciplinary problems than children in schools where uniforms are not
required.

Its null counterpart was:

22 | STATISTICS TRANSLATED

„ Children attending schools where uniforms are required do not have fewer
disciplinary problems than children in schools where uniforms are not
required.

In thinking about this null hypothesis, it is stating that one of two things will
occur:

1. Children attending schools where uniforms are required have exactly the
same number of disciplinary problems as children in schools where uniforms
are not required.
2. Children attending schools where uniforms are not required will have a great-
er number of disciplinary problems than students in schools where uniforms
are required.

Given this, as well as other issues we will discuss shortly, it is better to state the
null hypothesis for the directional hypothesis in the following manner:

„ There will be no difference in the number of disciplinary problems between

children attending schools where uniforms are required and children
attending schools where uniforms are not required.

Stating the null hypothesis in this manner ignores the “greater than” or “less
than” condition stated in the research hypothesis by saying that no difference exists;
doing so better reflects what we are trying to hypothesize. No difference would mean
we have an equal number of disciplinary problems in the two groups. If we subtracted
the average of one group from the average of the other group, the answer would be
zero or “null.” This may not seem logical right now, but it is the correct form for sta-
tistical decision making, and we will be stating it in this manner for the remainder of
this book.

Stating Null Hypotheses for Nondirectional Hypotheses

In our example where we compared achievement between students in a block schedule
to students in a regular schedule, we had the following research hypothesis:

„ There will be a difference in achievement between students taking algebra in

a block schedule and students taking algebra in a traditional schedule.

Stating the null hypothesis for a nondirectional hypothesis is very logical; the
exact opposite of “there will be a difference” is “there will be no difference.” Putting
that into a null format, you would write:

„ There will be no difference in achievement between students taking algebra

in a block schedule and students taking algebra in a traditional schedule.
 Identifying a Research Problem and Stating Hypotheses  |  23

We could do the same thing with the nondirectional research hypothesis if we

wanted to compare the age of our students to a known value:

„ The average age of my students is not equal to 42.

This one is not as straightforward but think about it this way. Since we have the
words “is not equal to” in the research hypothesis, the opposite of that is “is equal to.”
We can put our null hypothesis into the correct form by stating:

„ The average age of my students is equal to 42.

Again, in this case, if the average age of my students is 42 and the average age
of students in the population is 42, when I subtract one from the other, the answer is
zero. Zero obviously corresponds to null, again the name of the type of hypothesis.

A Preview of Testing the Null Hypothesis

Later in the book, we are going to discuss, in detail, the concepts underlying testing a
null hypothesis. There are, however, a few basics we need to cover before we get there.
In order to do so, let’s suppose we are investigating math achievement between boys
and girls and want to test the following null hypothesis:

„ There will be no difference in math ability between boys and girls.

This would be easy to test; we could administer tests to the students and then look
at the difference in scores between the boys and the girls, then determine if there is
a difference between the two groups. That would be the correct thing to do, right?
Unfortunately, it is not that simple and leads us to briefly talk about sampling error, one
of the basic building blocks of inferential statistics.

Sampling Error
Any time we are comparing data we have collected to an exact value or to another set
of data we have collected, there is only a very small chance they would be exactly the
same. This, of course, could wreak havoc with our decision making unless we recog-
nize and control for it. Let’s look at a couple of examples to help get a better under-
standing of this problem.
Suppose before we collect any data, we already know that the boys and girls have
exactly the same ability level in math. This means, theoretically, that if we gave them a
math exam, the average score for the boys should be exactly equal to the average score
for the girls. While this could occur, there is a very high probability that the average
scores of the two groups will not be exactly the same. This is because, although math
knowledge would be the highest contributor to an individual score, many, many other
factors might affect individual scores. Factors such as how well students slept the night
before an exam, their mental attitude on the day of the exam, whether they had break-
fast, or even how well they guessed on some answers might affect their score. Because
24 | STATISTICS TRANSLATED

of this, we could give the students the same exam repeatedly and two things would
occur. First, the average score for both groups would change slightly each time, and
second, in only a very few instances would the scores for the two groups be exactly
equal. The fact that the overall average changes daily is caused by sampling error, as is
illustrated in Table 1.1.

TABLE 1.1.  Ability Level for Boys and Girls in Math

Day Boys average Girls average Overall average
1 88 90 89
2 82 88 85
3 84 90 87
4 85 85 85

The same type of problem exists when we take samples of data from a large group.
For example, suppose we have a group of 900 students and we are interested in find-
ing the average reading ability for the entire group (we call all possible participants in
a group the population and any value we know about the population is called a popula-
tion parameter). Giving a reading examination to all 900 of them is possible but testing
a representative sample would be far easier (a sample is a subset of the population, and
any value we know about it is called a sample statistic). Theoretically, if you randomly
pick a sample of adequate size from a population, any measure you take of the sample
should be about the same as the measure of the entire population. In this case, if we
randomly chose a group of 30 students and tested them, their average reading ability
should reflect that of the entire population. Unfortunately, again, our results will be
affected by sampling error; let’s use an example to demonstrate this.
Let’s use a traditional range of zero to 100 for our reading scores. Suppose, how-
ever, that I already know the average reading ability is 70 for the 900 students I want
to survey. I also know that the average score of 70 represents 300 students who scored
50 on the exam (we will refer to them as Group A), 300 who scored 70 on the exam
(Group B), and 300 who scored 90 on the exam (Group C). All our data are shown in
Table 1.2.

TABLE 1.2.  Average Reading Scores for Three Groups

Group N Average
A 300 50
B 300 70
C 300 90
Population total 900 70

This means that, when I select 30 students as my random sample, I hope to get a
set of scores with an average of 70, exactly that of the population. Because of the ran-
dom nature by which I choose the students, however, the sample average will usually
not be exactly equal to the population average. Most of the time, the average score of
my sample will be about 70, but at other times I might randomly select more students
 Identifying a Research Problem and Stating Hypotheses  |  25

with lower scores, resulting in an average less than 70. In other samples, a higher num-
ber of students with scores greater than 70 would result in a sample average larger
than the population average. In short, the luck of the draw would affect the average
of my sample; sometimes it would be exactly equal to the population average, but, in
many instances, it would be higher or lower.
This idea of sampling error is further demonstrated in Table 1.3. In this table, I
have created five samples by randomly selecting 30 scores from the population. In the
first sample, I have 8 scores of 50 from Group A, 11 scores of 70 from Group B, and
11 scores of 90 from Group C. This gives me a sample average of 72, slightly higher
than the population average of 70. In the second sample, however, you can see that I
have more scores from the A and B groups than I did in the first sample; this causes
my sample average to fall below 70. The number of values in each group are about the
same in the third and fourth sample, but the average goes even lower in the fifth sam-
ple; many scores from the A group have pulled our sample average down to 64. Again,
this will even out over time, with most of the sample averages clustering around 70;
with enough samples it would equal the population mean of 70 exactly. Despite that,
it’s clear that randomly selecting samples can cause misleading results.

TABLE 1.3.  Five Samples of Average Reading Scores for Three Groups
A B C
Sample (50) (70) (90) Average
1  8 11 11 72
2 12 12  6 66
3  9  9 12 72
4 10 10 10 70
5 14 11  5 64

In order to make sure we’re getting the hang of this, look at Table 1.4. In the
first column we have a population of scores where the average is 10. If I take random
samples of five of the scores, shown in the remaining four columns, I compute sample
averages ranging from 7 to 13. Again, this shows that random sampling can dramati-
cally affect our calculations.

TABLE 1.4.  The Effect of Random Sampling on Average Scores

Population scores Sample 1 Sample 2 Sample 3 Sample 4
5, 6, 7, 8 5 11  5  8
9, 10, 11, 12 6 12  7  9
13, 14, 15 7 13  9 10
8 14 11 11
9 15 13 12
Population Sample Sample Sample Sample
average = 10 average = 7 average = 13 average = 9 average = 10
26 | STATISTICS TRANSLATED

Where Does That Leave Us?

Given these examples, whether we’re comparing two samples to one another or one
sample to an exact value, one of two things is true. First, any difference we find could
be due to chance; we call this sampling error. Second, differences could be representa-
tive of a true difference between the groups and not be caused by sampling error. In
the following examples, we will see the problem this creates for a statistician.
First, using our example of science students before and after lunch, let’s imagine
we have a population of 1,000 students and randomly select 50 students from both the
morning and afternoon groups. After our calculations, we might find that the aver-
age score for the morning students was 89% and the average score for the afternoon
students was 88%. Based on these scores, would we be comfortable in saying the two
groups are really different, or are they different due to chance? It seems that although
the morning students’ scores are slightly higher than the afternoon students’ scores,
the difference is so small it could be due to sampling error.
Right now, all we have are the average scores from the two groups, so we cannot
make a sound decision; we will need to use inferential statistics to help us do that. We
will then find that either the research hypothesis appears to be true and the difference
is not due to sampling error, or that the research hypothesis appears to be false and
the difference is due to sampling error. Because statisticians cringe when they hear
words such as “true” and “false,” let’s look at a better way of describing our results.

Rejecting and Failing to Reject Our Hypothesis

The key phrase in the preceding paragraph was “the difference is due to sampling
error” since it reflects the uncertainty of making decisions when sampling is involved.
Because, when we are comparing values from a sample, we are never 100% sure of
our decision, we cannot use words such as “true,” “false,” or “prove” as they indicate
certainty. Since sampling error is involved, we instead say we reject or fail to reject our
null hypothesis. This means we believe the groups are different and we reject the null
hypothesis, or they aren’t different, and we fail to reject the null hypothesis. When
we reject our null hypothesis, we are supporting the research hypothesis that corre-
sponds to it. Similarly, when we fail to reject our null hypothesis, we are not support-
ing the research hypothesis. Table 1.5 shows the relationship between the null and
research hypotheses.

TABLE 1.5.  Actions Regarding the Null and Research Hypotheses

Action regarding the null hypothesis Action regarding the research hypothesis
When we reject the null hypothesis, Support the research hypothesis; what we
the groups appear to be different. Any have hypothesized appears to be accurate.
differences appear to be due to reasons Any differences appear to be due to
other than sampling error. reasons other than sampling error (i.e., the
differences are real).
When we fail to reject the null hypothesis, Fail to support the research hypothesis;
the groups appear not to be different. Any what we have hypothesized does not appear
differences appear to be due to sampling to be accurate. Any differences appear to
error. be due to sampling error.
 Identifying a Research Problem and Stating Hypotheses  |  27

Going back to our example, suppose we now found the average grade for students
taking science in the morning was 89% and the average score for students taking sci-
ence in the afternoon was 83%. Although we would need to verify these findings using
the appropriate statistical test, we would feel safe in saying the morning students had
higher achievement than the afternoon students and it wasn’t due to sampling error.
Knowing this, we would reject the null hypothesis; the difference between the groups
does not seem to be due to chance. By saying we reject the null hypothesis, we are say-
ing we support our research hypothesis. We must take it a step further though.

All Hypotheses Must Include the Word “Significant”!

In order to remember that we are looking for differences that are “real” and not due
to sampling error, we must reword our hypotheses in a way that reflects that. Chang-
ing our hypotheses is easy; all we do is include the word “significant.” For our research
hypothesis dealing with the difference between morning and afternoon students, this
means that we would write the following:

„ Students taking science in the morning have significantly higher levels of

achievement than students taking science in the afternoon.

Our null hypothesis would read:

„ There will be no significant difference in levels of achievement between

students who take science in the morning and students who take science in
the afternoon.

For our example of children wearing uniforms versus those children not wearing
uniforms, our research hypothesis would be

„ Children attending schools where uniforms are required have significantly

fewer disciplinary problems than children in schools where uniforms are not
required.

Our null hypothesis would read

„ There will be no significant difference in the number of disciplinary

problems between children attending schools where uniforms are required
and children attending schools where uniforms are not required.

By adding the word “significant” to our hypotheses, we are supporting our desire
to ensure that any difference we find when we are analyzing data is “real” and not due
to chance. This idea is shown in Table 1.6.
28 | STATISTICS TRANSLATED

TABLE 1.6.  Actions Regarding the Null and Research Hypotheses When
Differences Are Significant
Action regarding the null hypothesis Action regarding the research hypothesis
When we reject the null hypothesis, the When we reject the null hypothesis, we
groups appear to be significantly different. support the research hypothesis; what we
Any differences appear to be due to reasons have hypothesized appears to be accurate.
other than sampling error. Any differences appear be “real” and due to
reasons other than sampling error.
When we fail to reject the null hypothesis, When we fail to reject the null hypothesis,
the groups appear not to be significantly we fail to support the research hypothesis;
different. Any differences appear to be due what we have hypothesized does not appear
to sampling error. to be accurate. Any differences appear to
be due to sampling error.

Statistical Words of Wisdom

Everything we just said can be summed up in one paragraph:

„ When you are calculating inferential statistics, always test the null
hypothesis. When you test the null hypothesis, you are not trying to “prove”
anything. You are simply trying to determine one of two things. First, if
your results cause you to reject your null hypothesis, you will support your
research hypothesis. Second, if your results cause you to fail to reject the null
hypothesis, you fail to support your research hypothesis.

These are statistical “words of wisdom.” The task of learning and understanding
statistics will be much easier if you do not forget them.

Summary of Step 2: State a Hypothesis

Stating the research or alternate hypothesis for a study you want to conduct is a
straightforward process. After identifying the problem you want to investigate, you
must state the relationship you predict will emerge. The prediction is based on either
your observation of past performance or on other sources such as literature you have
read. When you have determined the prediction you want to make, you need to state
it as both a null hypothesis and a research hypothesis. The research hypothesis will be
either directional or nondirectional, depending on what the literature or your experi-
ence tells you.
Testing your hypothesis means you will do one of two things. You will either reject
the null hypothesis and support the research hypothesis, or you will fail to reject the
null hypothesis and not support the research hypothesis. As you will see in the next
two chapters, using statistics to test the hypothesis will require you to identify and
describe the independent and dependent variables stated in the hypothesis. This will
help you determine the exact statistical test you will need to use to test the hypothesis.
 Identifying a Research Problem and Stating Hypotheses  |  29

Do You Understand These Key Words and Phrases?

alternate hypothesis directional hypothesis
ethical research fail to reject the null hypothesis
hypothesis nondirectional hypothesis
null hypothesis parameter
population practical significance
problem statement qualitative data
quantitative data rejecting the null hypothesis
research hypothesis sample
sampling error significantly different
statistic theoretical significance

Quiz Time!

Before we move forward, let’s make sure we have got a good handle on what we have covered
by working on the problem statements and case studies below. Once you are finished with both
sections, you can check your work at the end of the book.

Problem Statements
Evaluate each of the problem statements below; are all six criteria for a good problem state-
ment met? Are they clear, do they include all variables, and do they avoid interjecting personal
bias? If the problem statement does not meet the criteria, what is wrong?

1. Student absenteeism at East Highland High School.

2. Teachers in many older public schools are concerned with the toxic effects of
mold and asbestos they believe may be present in their buildings. Teachers in this
study will locate, inspect, and report on types of potential health hazards in their
schools.

3. Due to ever-increasing concerns regarding toxins and other pollutants in public

water supplies, many at-home vegetable growers are becoming leery of watering
their plants with other than distilled water.

4. The tremendous growth of the beef industry in Central Iowa has left many smaller,
family-owned ranches facing the possibility of bankruptcy. This study will investi-
gate the attitudes of small ranchers toward three options: going into bankruptcy,
merging with larger ranching conglomerates, or attempting to stay lucrative by
developing specialty beef products aimed at niche markets.

5. Is there a difference in efficacy between different treatment therapies on levels of

clinical depression?
30 | STATISTICS TRANSLATED

6. Many elementary school teachers find it difficult to instill an interest in science

and math into younger students. Research has shown that student engagement can
be positively affected when students are interested in each topic and understand the
relevance of it. This study will focus on the effects of lesson plans based on differ-
ent types of radiation and the ability to use them to measure interplanetary space.

7. This study will investigate average miles per gallon between three grades of gaso-
line in imported cars.

8. Medical journals have repeatedly discussed the relationship between prenatal

lifestyle and the health of newborn infants. This study will investigate the effect of
different diets on weight gain during pregnancy.

9. A comparison of crime rates in urban, suburban, and rural areas.

Case Studies
State both the research and null hypothesis for each of the following scenarios:

The Case of Distance Therapy

When a person feels he needs some type of counseling, he typically enlists the services of a
local therapist. By working directly with the client, the therapist can address the problems and
concerns with which the client is trying to cope. What happens when a person has time or loca-
tion constraints and cannot find a professional to work with?
Some people feel the answer to this dilemma is “distance therapy.” Much like technology-
based distance education, the therapist can work with a client using Internet-based chat ses-
sions, phone consultations, and email. While many people debate the efficacy of this approach,
very little research has been conducted on it; no one really knows if either counseling method
works better or worse than the other.
Given what we know, suppose we wanted to determine if it takes therapists the same
number of sessions to deal effectively with depression cases in a traditional versus a distance
approach. How would we state our research and null hypotheses?

The Case of the New Teacher

A new teacher, straight out of college, was complaining to one of her friends. “I have the hard-
est time with my classes. I send home notes asking for parent–teacher conferences, and very
few of the parents ever respond. I do not know what to do.” Her friend, a teacher at a school in
a very prominent area of town, told her “Didn’t you know that is to be expected? Your school
is in an area of town with a lower socioeconomic base. In order to make ends meet, many par-
ents work more than one job, or work extended hours. Because of that, sometimes it’s difficult
 Identifying a Research Problem and Stating Hypotheses  |  31

for them find time to meet with you. Since we are in a wealthier part of town, we experience
exactly the opposite. When we send out a note asking to meet with them, most of our parents
contact us within a day or two.” The young teacher, not having much experience, then told her
friend: “I cannot believe that; parents are parents, and they should all want to hear what the
teacher says about their child. I would like to collect data to help determine if there really are
fewer parent–teacher conferences in my school than in yours.” In order to do that, how would
she state her null and research hypotheses?

The Case of Being Exactly Right

A local chemical processing plant has been built to handle production waste from local indus-
tries. The purpose of the plant is to effectively clean the waste to ensure nothing hazardous
is emitted into the atmosphere. Their equipment must run at exactly 155° for the process to
work correctly; anything significantly higher or lower will result in the inability to guarantee
effective cleanliness. If engineers working at the plant were developing software to control the
process, what research and null hypotheses would they work with?

The Case of “Does It Really Work?”

With the rise of distance education, more courses are being offered in an Internet-based for-
mat. In addition to the traditional general studies courses such as English and history, some
colleges are starting to offer courses that have been traditionally lab-based. Knowing that,
the biological sciences faculty at one small college approached the dean about developing an
online anatomy and physiology course. The dean was receptive but cautioned the teachers that
they would have to ensure the distance education students would have the same experiences
and learn as much as their on-campus peers. After developing and using the system for one
semester, the faculty wanted to demonstrate to the dean that their efforts were successful.
Using final exam scores as a measure of achievement, how would they state their null and
research hypotheses?

The Case of Advertising

The Vice-President of Human Resources of a manufacturing plant was worried about the low
number of college graduates applying for administrative jobs within the company. “I cannot
understand it,” she told a close friend. “We are spending several thousand dollars a month in
advertising, and we are getting very few applicants.” Her friend, the owner of a marketing con-
32 | STATISTICS TRANSLATED

sulting firm, asked several questions about the content of her advertisements and then asked
where they were published. The Vice-President quickly answered, “They are in the newspaper,
of course; isn’t that where all job announcements are placed?” The consultant explained to
the Vice-President that many job applicants look for positions advertised on social media sites
such as LinkedIn, Twitter, and Facebook. Many young people, she explained, do not bother to
read the newspaper, and they certainly do not look for jobs there.
The Vice-President decided it couldn’t hurt to try advertising on the Internet and was
interested in seeing if the number of applicants really increased. If she wanted to test the effec-
tiveness of the advertisement by comparing the number of applicants in the month prior to the
advertisement to the number of applicants in the month after the advertisement was placed on
the Internet, how would she state her research and null hypotheses?

The Case of Learning to Speak

Teachers of foreign languages seem to fall into two groups. The first group believes that a full
understanding of the formal grammar of a language is important before one even tries to
speak it. The other group believes in “language immersion”; students naturally learn the gram-
mar of a language when they are forced into a setting where they have no choice. If we wanted
to compare the two methods to see if there was a significant difference in language skills after
one year of instruction, how would we state the null and research hypotheses?

The Case of Kids on Cruises

The cruise industry is becoming more popular in the United States with many cruise lines
catering to their potential customers’ desires by offering such things as a wider variety of on-
board entertainment, different types of dining experiences, and a larger variety of tours while
in port. The CEO of one company, looking to grow her cruise line, noted that many of the
line’s older passengers sought cruises where children were not allowed. When she presented
the idea of prohibiting kids to her board of directors, some board members agreed with her,
but others said children should be welcome on every cruise. If the CEO wanted to compare
the preferences of her older customers, what are the research and null hypotheses she would
investigate?
CHAPTER 2

Identifying the Independent and

Dependent Variables in a Hypothesis

Introduction
Now that we understand how to state our hypothesis, let’s take a closer look at it.
By doing so, we will find much of the information that will ultimately help us select
the appropriate statistical tool we will use to test it. Let’s refer to one of the research
hypotheses we stated in Chapter 1 to get a better feel for what we are going to do:

Students taking science in the morning have significantly higher achievement

than students taking science in the afternoon.

In this hypothesis, we can see that we are looking at the relationship between the
time of day students take science and their performance in the course. Although we
ultimately must rely on statistical analysis to confirm any relationship, for now we
can think of our independent variable as the “cause” we are interested in investigat-
ing and our dependent variable as the “effect” we want to measure. In this chapter
we’ll do several things; first we’ll talk about identifying the independent and depen-
dent variables, and from there we’ll move to a discussion of different data types and
how we can statistically describe them.

S TEP 3 Identify the Independent Variable

In the preceding example, identifying the independent variable
is easy; we are interested in the time of day a student takes sci-
ence. It’s important to understand, however, that independent

33
34 | STATISTICS TRANSLATED

variables can be developed in two different ways. First, if we are interested in deter-
mining the effect of something that occurs naturally and does not require intervention
by the researcher or statistician, they are called nonmanipulated independent variables
(these are sometimes called quasi-independent variables). In cases where the researcher
must actively manipulate a variable by randomly creating groups or assigning objects
to a class, the independent variable is manipulated or experimental.

Nonmanipulated Independent Variables

In our hypothesis where we stated that students taking science in the morning have
significantly higher achievement than students taking science in the afternoon, we
didn’t assign students to a science class; we looked at the existing classes. This is a per-
fect example of a nonmanipulated independent variable since the researcher played
no part in determining which subjects were in either group.
What is the independent variable in the next hypothesis?

„ Male employees miss significantly more days of work than female employees.

The independent variable is gender and, since we do not assign employees to one
gender or the other, it isn’t manipulated. We are interested in seeing the relationship
between gender (the independent variable or cause) and the number of workdays
missed.
How about this hypothesis?

„ First-year college students have significantly lower grade point averages

than students in other classes.

We are not assigning individual students to the various classes; we are using their
current status. In this case, “year in college” is the nonmanipulated independent vari-
able, and we are trying to determine its effect on grade point average.
What is the independent variable in this hypothesis?

„ People with pets in their home have significantly lower levels of depression
than people who do not have pets in their home.

In this case, pet ownership is the independent variable; either a person has a pet
at home, or they do not. The dependent variable is a person’s level of depression.
What is the independent variable in the following hypothesis?

„ Canadian citizens take significantly more vacation days each year than
citizens of the United States.

Here our dependent variable is the number of vacation days taken each year;
our independent variable is country of citizenship: the United States or Canada. As
you can see, the citizens are already living in either Canada or the United States; this
means the independent variable is nonmanipulated.
 Identifying the Independent and Dependent Variables  |  35

Another Way of Thinking about Nonmanipulated Independent Variables

There is another way of developing a nonmanipulated independent variable that is
somewhat tricky because it sounds like the researcher is intervening. Suppose, for
instance, we had a researcher who was interested in investigating the following hypoth-
esis:

„ Low-achieving math students spend significantly less time on homework

than high-achieving math students.

This is pretty much a “no-brainer” hypothesis, but it serves to illustrate the point.
Because we are going to compare two groups, low achievers and high achievers, our
independent variable must be achievement. Our only problem is, how do we deter-
mine which students we are going to label as low achieving and which we are going to
label as high achieving?
This can be easily done. Suppose, for instance, we have the results of a recent
administration of a standardized math test. We can label those as falling into the
lower 50% as low achievers and those scoring in the upper 50% as high achievers.
While this may sound like we are manipulating the independent variable, we really
are not. Although we have decided the parameters of group membership, the students
naturally fell into either the low-achieving or high-achieving group by their own per-
formance.
It is important to understand we can divide the students by whichever standard
we choose, if it logically makes sense. For example, we could have also divided the
students into four ability groups or arranged them so that anyone in the top 30% was
labeled as a high achiever and everyone else was labeled as low achievers. Regardless
of how we do it, setting up the independent variable in this manner is justified.

Manipulated or Experimental Independent Variables

In other cases, the researcher must actively assign participants to one group or anoth-
er in order to test the stated hypothesis. For example, if we wanted to test the effect of
computer-based instruction versus lecture-based instruction on accounting students,
we might state the following hypothesis:

„ Students taking accounting in a computer-based format will have

significantly higher achievement than students taking accounting in a
lecture-based format.

In this case, we have stated a directional research hypothesis and, in order to test
it, we could randomly assign participants to one group or the other. Since we would be
actively intervening, this is a manipulated, or as it is sometimes called, an experimen-
tal independent variable. Remember, though, if the two groups had already existed, it
would be a nonmanipulated independent variable.
Now, suppose we are interested in looking at the effect of taking an after-lunch
nap on the behavior of children in kindergarten. Further, suppose that we have six
36 | STATISTICS TRANSLATED

classes and decide to randomly assign three classes to the “nap” group and three class-
es to the “no nap” group. In this case, we are dealing with existing classrooms, but we
are still randomly assigning them to either
the “nap” or “no nap” group. This meets
the criteria for a manipulated independent
variable, and our progress so far is shown
in Figure 2.1.

Levels of the Independent Variable

Although we have already alluded to it,
we need to develop an understanding of
the relationship between the levels of the
FIGURE 2.1.  Types of independent variables. independent variable. The levels are noth-
ing more than the values that the indepen-
dent variable can represent. Using our example of science students taking class in the
morning and in the afternoon, we will be able to understand this better.
If someone were to ask us what our independent variable in the study is, we would
tell them it is the time of day instruction is delivered. In this case, we have identified
two levels: “morning” and “afternoon.” Since we are using two preexisting groups,
they are nonmanipulated.
If we were to look at the study where we divided students into a computer group
and a lecture group, we would be actively intervening in the process. The independent
variable could be called “type of instruction.” Since we actively placed students in
groups, it is a manipulated or an experimental variable and has two levels: students
in the lecture group and students in the computer group. This means, of course,
that we need to expand our prior picture,
as shown in Figure 2.2. When looking at
Figure 2.2, we should not assume that all
independent variables have only two levels.
Theoretically, an independent variable can
have an infinite number of levels but seeing
an independent variable with more than
four or five levels is unusual. Although we
won’t discuss it too often in this book, we
can also have more than one independent
variable. As we will see later, the number
of independent variables helps us to select
FIGURE 2.2.  Levels of the independent vari- the statistical test we will employ to test our
able. hypothesis. For now, let it suffice to say, if
we have more than one independent vari-
able, we will be using multivariate statistics.
Hypotheses having only one independent variable may qualify as univariate statistics,
but that will depend on the number of dependent variables involved. That will be dis-
cussed in Step 4 of our six-step model.
 Identifying the Independent and Dependent Variables  |  37

Summary of Step 3: Identify the Independent Variable

The independent variable is the “cause” we want to investigate when we state our
hypotheses. As we saw, variables are nonmanipulated (quasi-independent) if we are
investigating a cause that occurs naturally; they are manipulated (experimental) if we
are actively involved in assigning group or treatment membership. Whether they are
manipulated or not manipulated, independent variables must have at least two levels
(we will see one slight exception to that rule later in the book), and we are interested
in determining the “effect” that each of these levels has on our dependent variable.
Hypotheses with only one independent variable generally use univariate inferential
statistics, while those with more than one independent variable are tested using mul-
tivariate statistics. All this information, along with the ability to identify and describe
dependent variables, will help us pick the correct statistical tool to use in testing our
hypotheses.

S TEP 4 Identify and Describe the Dependent Variable

Unlike Steps 1–3, it will take us several chapters to cover every-
thing we need to in order to accomplish Step 4. The remainder
of this chapter focuses on identifying the dependent variable,
determining the type of data it represents, and computing mea-
sures of central tendency. The next two chapters discuss measures of dispersion and
relative standing as well as ways by which we can graphically display our data. Using
that information, we’ll be able to move to Step 5, choosing the right statistical test for
our hypothesis. For now, however, let’s look at the dependent variable.

Identifying Your Dependent Variable

As we have already said, the dependent variables represent the “effect” we want to
measure. This might include such things as academic achievement, sales, grade point
average, weight, height, income, or any of the other myriad things about which we like
to collect data. The cause leading to this effect, of course, is the independent variable
and its levels. We can add this at Figure 2.3 to make this idea even clearer.
We can better understand this concept by looking at the following research
hypothesis:

FIGURE 2.3.  The relationship between independent and dependent variables.

38 | STATISTICS TRANSLATED

„ Businesses that advertise only on television will have significantly higher

sales than businesses that advertise only in print or only on the radio.

What is the independent variable in this hypothesis? Remember, when we are

identifying the independent variable, we are looking for the “cause” we want to inves-
tigate. In this case, we could label it “advertising.” What are the levels of the indepen-
dent variable? In this case we have three: (1) television, (2) radio, and (3) print. The
dependent variable, or effect, is what we are measuring. In this case, it is the total
amount of sales for the businesses in each category represented by the three levels of
the independent variable. In short, we are trying to investigate the effect of the adver-
tising method (the cause) on the amount of sales (the effect).
Using that same process, identify the independent variable, its levels, and the
dependent variable in the following hypothesis:

„ Customer satisfaction for airlines that charge customers for checking their
bags will be significantly lower than customer satisfaction of airlines that do
not charge for bags that are checked in.

In this case, we are trying to investigate cause and effect by looking at the differ-
ence in customer satisfaction toward airlines that charge for baggage and those that
do not. We could label our independent variable “airline,” and there are two levels:
those airlines that charge customers to check their baggage and those that do not. The
dependent variable is customer satisfaction.
Again, in some cases we may have more than two levels of the independent vari-
able:

„ There will be a significant difference in the amount of rainfall in each of the

four seasons.

Here, our independent variable is “season,” and there are four levels; fall, win-
ter, spring, and summer. Our dependent variable would require that we measure the
amount of rainfall during each of the four seasons.
Let’s look at one more hypothesis before moving forward:

„ There will be a significant difference in the number of absences and

dropouts between students in a rural school district, students in a suburban
school district, and students in an urban school district.

In this case, the independent variable is “school district” and there are three lev-
els: rural, suburban, and urban. What about the dependent variable? Wait a minute:
this looks odd, doesn’t it? We seem to have two dependent variables: absences and
dropouts. How do we handle this? This is a very common problem. As I’ve said, just
as is the case of multiple independent variables, any time we are dealing with more
than one dependent variable we will use multivariate inferential statistics to test our
hypotheses. We won’t concern ourselves with that idea just yet, so let’s move forward;
 Identifying the Independent and Dependent Variables  |  39

our next task involves understanding the type of data our dependent variable repre-
sents.

What Type of Data Are We Collecting?

There are four distinct types of data that we must concern ourselves with: nominal
data, ordinal data, interval data, and ratio data. Each of these data types has a unique set
of properties, and understanding those properties is key in determining the statistical
tool we will use to test our hypothesis.

Nominal Data
Nominal data, sometimes called categorical or discrete, are data that are measured in
categories. For example, if we are interested in determining the number of males and
females in our class, the data value “gender” is nominal in nature. In doing this, we
only need to count the number of persons falling into each category. The important
thing to remember is that the categories are mutually exclusive because the items
being counted can only fall into one category or another. For example, a person can-
not be counted as both a male and a female. Other examples of nominal data include
ethnicity, college class, or just about any other construct used to group persons or
things. In Table 2.1, we can see where we have asked 100 people the political party to
which they belong. Each of the parties represents a distinct category; a person cannot
belong to more than one.

TABLE 2.1.  Nominal (Categorical) Data Representing Political Party

Republican Democrat Independent Other
40 35 20 5

In Table 2.2, we have asked 500 students to tell us the primary method they use
to get to school. There are five possible choices and, since we have asked for only the
primary method of transportation, a student could only choose one. Since all we want
to know is how many students fall into each group, it meets the definition of nominal
data.

TABLE 2.2.  Nominal (Categorical) Data Representing Primary Mode of

Transportation
Public
Drive Carpool transportation Walk Bike
210 105 85 62 38

Ordinal Data
The second type of data is called ordinal data or, as it is sometimes called, rank data.
When we use ordinal data, we are interested in determining what is important or
significant while investigating non-numeric concepts such as happiness, discomfort,
40 | STATISTICS TRANSLATED

or agreement. For example, I recently took my car to the dealership for scheduled
maintenance. Within a day of picking it up, I received an email asking me to rank their
service using a form like the one in Table 2.3. You can see that a scale ranging from
“Very Dissatisfied (1)” to “Very Satisfied (5)” was used. I was asked to assign a ranking
to each of the statements shown therein.
There are things about ordinal data that are essential to remember. First, the
answers are subjective; there is no one standard for the choices. You could have the
same service performed, and pay the same cost, as another customer. The rank you
give is your opinion; you might be satisfied with the price you paid while another
customer is very dissatisfied. Because of that, even if we assign numeric values to the
scores (e.g., Very Dissatisfied = 1 and Satisfied = 4), we can’t compute an average level
of satisfaction for the service performed simply because the scores might not be rela-
tive to one another. Again, all the answers are subjective. In short, the properties of
the actual variable are not known. You will see, later in the book, that we can use other
statistical tools to look at the mid-point of the data distribution.

TABLE 2.3.  Ordinal (Rank) Data Representing Satisfaction with Service

Department
Very Very
Dissatisfied Dissatisfied Neutral Satisfied Satisfied
Service Personnel X
Explanation of X
Service Needed
Satisfied with X
Service Performed
Timeliness of X
Service
Price of Service X

Interval Data
Interval data are the first of two types of data that are called quantitative or continuous.
By this, we mean that a data value can hypothetically fall anywhere on a number line
within the range of a given dataset. Test scores are a perfect example of this type of
data since we know, from experience, that test scores generally range from 0 to 100
and that a student taking a test could score anywhere in that range.
As suggested by its name, the number line shown in Table 2.4 represents the pos-
sible range of test scores and is divided into equal increments of 10. Given that, the
10-point difference between a grade of 70 and a grade of 80 is the same as the differ-
ence between a grade of 10 and a grade of 20. The differences they represent how-
ever are not relative to one another. For example, if a person scores 80 on the exam,
their grade is four times larger than that of a person who scores 20 on the exam. We
cannot say, however, that the person who scores 80 knows four times more than the
person who scores 20; we can only say that the person with the 80 answered 80% of
the questions correctly and that the person scoring 20 answered 20% of the questions
correctly. We also cannot say that a score of zero means they know nothing about the
 Identifying the Independent and Dependent Variables  |  41

subject matter, nor does a score of 100 indicate absolute mastery. These scores only
indicate a complete mastery or ignorance of the questions on the examination.

TABLE 2.4.  Test Scores

0 10 20 30 40 50 60 70 80 90 100

The point is that even though the score data are measured in equal intervals on
this type of scale, the scale itself is only a handy way of presenting the scores with-
out indicating the scores are in any way related to one another. Another example of
interval-level data is temperature, as shown in Table 2.5.

TABLE 2.5.  Interval Data for Temperatures

Fahrenheit
School temperature Comment
University of Buffalo Zero This doesn’t mean there is no temperature
in Buffalo; this is just an arbitrary value
where the temperature in Buffalo falls on
the scale. There could be a value below zero.
University of Virginia 25 This doesn’t mean that Virginia has a
temperature and Buffalo doesn’t. The two
temperatures are relative to the scale and
nothing else.
University of Alabama 50 Birmingham is not twice as warm as
at Birmingham Virginia. Just like above, this is simply where
the temperature for Birmingham falls on the
scale.
University of Texas at 100 El Paso is not four times as hot as Virginia,
El Paso nor is it twice as hot as Birmingham. Having
lived there, I know this is actually not that
hot!

Ratio Data
Ratio data are also classified as quantitative or continuous data. Ratio data differ from
interval data because they do have an absolute zero point and the various points on
the scale can be used to make comparisons between one another. For example, weight
could be measured using an interval scale because we know that the relative difference
between 20 pounds and 40 pounds is the same as the relative difference between 150
pounds and 300 pounds; in both cases the second value is twice as large as the first.
There are, however, three important distinctions between interval and ratio data.
First, a value of zero on a ratio scale means that whatever you are trying to mea-
sure doesn’t exist; for example, a swimming pool that is zero feet deep is empty. Sec-
ond, ratio data allow us to establish a true ratio between the different points on a
scale. For example, a person who owns 600 shares of a company has twice as many
shares as a person owning 300 shares; the same person would own six times as many
shares as a person owning 100 shares. Third, this added degree of precision allows us
to use ratio scales to measure data more accurately than any of the other previously
mentioned scales. Other examples of ratio-level data include distance and elapsed
time; one of my favorite examples, annual income, is shown in Table 2.6.
42 | STATISTICS TRANSLATED

TABLE 2.6.  Ratio Data for Income

Job Annual income Comment
Student $0.00 This student is making absolutely nothing—he has
a lack of income. That’s OK; he can study now and
earn money later.
Textbook author $25,000 Believe me; we are only in this for the love of
writing. This is a very optimistic estimate!
College professor $50,000 This is another profession where you better not
be in it for the money. At the same time, look at
the bright side; you’re making twice as much as a
textbook author.
Medical doctor $250,000 This is better. Ten times as much as the author and
five times as much as the college professor. Did I
mention it takes longer to get a PhD than it does to
get an MD?
Chairman of $1,000,000,000 This is where we want to be, a billion dollars. The
Microsoft college professor’s salary times 20,000! It is even
4,000 times what the medical doctor makes! For
those of you majoring in education and psychology,
forget it. It is probably too much trouble keeping up
with that much money anyway.

Data Types—What Is the Good News?

The first bit of good news is that we are making a lot of progress in our six-step model.
We can identify an independent variable and its levels, as well as identify a dependent
variable and the type of data we are collecting. This is shown in Figure 2.4.
There is one more bit of good news that will help us in our journey toward becom-
ing a proficient consumer of statistics. Once we can identify the different types of
data, there is one general rule that you need to keep in mind that will really help
narrow down the type of statistical tool you’re going to use. Generally speaking, if
the data we collect are quantitative, we will use parametric statistics when we begin our
inferential decision making. These include many of the tools you may have already
heard of, the t-test and the analysis of variance for example. If the data are nominal
or ordinal, we will use nonparametric statistics to help us make decisions. Except for the
chi-square test, you may have never heard of other nonparametric tests since we tend
to use quantitative data more frequently. Again, these are general rules, and we will
see minor exceptions to them later. For now, we can feel comfortable with the sum-
marization provided in Table 2.7.

FIGURE 2.4.  The dependent variable and types of data that may be collected.
 Identifying the Independent and Dependent Variables  |  43

TABLE 2.7.  Relationship between Data Type and Appropriate Statistical Tests
Parametric or
Type of data Quantitative? Nonparametric statistics?
Nominal (categorical) No Nonparametric
Ordinal (rank) No Nonparametric
Interval Yes Parametric
Ratio Yes Parametric

Summary of the Dependent Variable and Data Types

The dependent variable in a hypothesis represents the “effect” we are trying to measure.
In order to do so, we have four types of data—nominal, ordinal, interval, and ratio; the
last two of these can be thought of together as quantitative data. Each of these data
types has unique characteristics and understanding these characteristics will ulti-
mately help us identify the appropriate statistical tool we need to test our hypotheses.
For now, however, we’ll learn how to “describe” our data numerically and graphically.
This will include discussing common measures such as the average of a dataset but
will also include things—like the normal distribution—that some of us may have never
heard of. Like everything we have talked about so far, though, these concepts are easy
to understand and are integral in our quest to be good consumers of statistics.

Measures of Central Tendency

Measures of central tendency are just that—tools that help us determine the “center”
or midpoint of data we have collected. This is essential because it not only allows us to
better understand one set of data, but it also lets us begin comparing two or more sets
of data to each other. As we go along, I will show you the basic formulas that underlie
numeric descriptive statistics; learning to compute these “by hand” will help you bet-
ter understand and apply the results. Later in the chapter we’ll see how to do the same
calculations the easy way—by using a computer! For now, let’s create a dataset to use
an example.
Table 2.8 shows data from a sample of students in a typical classroom (remember,
we call all possible participants in a group the population and a subset of the popula-
tion is a sample). Our sample represents things a teacher might be interested in know-
ing.

TABLE 2.8.  Dataset for Nominal, Ordinal, Interval, and Ratio Data
Student Math
number Gender Ethnicity Attitude Math Exam Absences
1 F H 5 92 7
2 F NHW 3 91 4
3 M AA 3 90 0
4 M NHW 3 89 8
5 M AA 5 88 3
6 F NHW 1 87 3
7 F H 2 86 3
44 | STATISTICS TRANSLATED

As you can see by looking at the headers, the contents of the columns are self-
explanatory, but let’s take a closer look and decide what type of data they represent.
First, look at the column labeled Gender. Since the valid values for this field are M
for male students and F for female students, what type of data are we collecting? That
is easy; since the value of the field represents a category a given student falls into,
we know it is nominal-level data. What about the field called Ethnicity? It is also an
example of nominal data since we generally say a student falls into only one ethnic
category. In this case, the categories are AA (African American), NHW (non-Hispanic
White), and H (Hispanic).
The next column, Math Attitude, contains ordinal data. Students are asked to
rate their attitude toward math on a scale of 1-Dislike, 2-Somewhat Dislike, 3-Neutral,
4-Somewhat Like, and 5-Like. Keeping in mind that this is subjective and represents
each student’s opinion, the values are not relative to one another. In Table 2.9, I’ve
added subtitles under each of the variable names showing their data type; I’ve also
added a second ordinal value, College Importance, and ask students to rate their
interest in going to college on a scale of 1-No Interest, 2-Undecided, and 3-Interested.
Again, this is a given student’s personal opinion; the factors that influence one stu-
dent’s interest in college may not be relevant to another student.

TABLE 2.9.  Expanded Dataset for Nominal, Ordinal, Interval, and Ratio Data
Math College
Student Gender Ethnicity Attitude Math Exam Absences Importance
number (Nominal) (Nominal) (Ordinal) (Interval) (Ratio) (Ordinal)
1 F H 5 92 7 1
2 F NHW 3 91 4 2
3 M AA 3 90 0 3
4 M NHW 3 89 8 2
5 M AA 5 88 3 2
6 F NHW 1 87 3 2
7 F H 2 86 3 3

Now, what type of data does “math exam” represent? In order to answer that
question, let’s first assume that achievement scores range from the traditional 0 to
100. Given that, we can narrow down our choices to interval or ratio level data. The
next question is, “Does a math exam have an absolute zero point?” From our previous
discussion, we know the answer is “no.” A student may make a zero on the exam, but
that does not mean she knows absolutely nothing about math; nor does 100 indicate
absolute mastery of the topic. We also cannot make comparisons or ratios using the
data; a person scoring 80 on a test does not necessarily know twice as much about
math as a person who scores 40; they simply scored twice as much on that given exam.
Knowing that, the data are interval in nature.
Finally, let’s look at the field titled Absences. We know these data are quantitative
in nature, but this time an absolute value of zero is possible; a student may have never
missed class. We can also make meaningful comparisons between the values; a stu-
dent who has been absent four times has been absent four times as many as a student
who has been absent once; this means the data are measured on a ratio scale. Remem-
 Identifying the Independent and Dependent Variables  |  45

ber, however, it is generally not important to know if the data are either interval or
ratio level. Just knowing whether the data we are collecting are quantitative is enough.

The Mean, Median, and Mode—Measures of Central Tendency

A measure of central tendency helps determine the center of a dataset. We have
already mentioned one of the most common measurements, the mean or average, but
it is not the only measure of this type. Two other closely related measures of central
tendency are the median and the mode. Before talking about them, let’s take a closer
look at how we compute the mean.

The Mean or Average

To begin computing a mean score, let’s start by totaling the Math exam data from
Table 2.9; this leaves us with 623. To determine the mean score, we simply divide this
value by the total number of scores, 7, giving us an average of 89. If you ever want to
impress your friends, however, there is a technical formula for calculating the mean.
Before we get into it, however, there might be several characters in the formula that
you might not recognize; let’s go through them first.
Sx
In the formula below, the symbol x =(referred to as “x-bar”) stands for the actual
n
mean value that will be calculated. The upper-case Greek letter sigma (i.e., S) means
to sum everything following it. In this case, the x represents the individual values in
our sample, so the top line of the formula says, “add up all of the values in the sample.”
The lower-case n on the bottom stands for the number of items in the sample; here we
have collected data for seven students. Putting this all together, the formula for the
mean of a sample is:

Sx
x=
n
Now, we can use the values in our table above to fill in the formula:

92 + 91 + 90 + 89 + 88 + 87 + 86
x=
7
This gives us:

623
x=
7
By dividing 623 by 7, we can see that the average math exam score is 89, just as
we computed above:

x = 89

Remember, we very rarely know the population mean, but, on any of the rare
occasions where you might know all the values in a population, we can compute it for
a population using the same formula. The only difference is that we would substitute
the lower-case Greek letter mu (i.e., μ) for x and an upper-case N as the number of
members in the population. The equation would then look like this:
46 | STATISTICS TRANSLATED

Sx
m=
N
Before we move forward, let me point out something here that might easily be
overlooked. Did you notice that we used a lower-case x to refer to the value we com-
puted for a sample and then we used a Greek letter to refer to that same value for
a population? That idea holds true throughout. If we are calculating a value for a
population, we’ll use Greek letters; for samples, we’ll use English letters. The ability to
recognize and understand these seemingly insignificant things is a big part of becom-
ing a good consumer of statistics!

WE DO NOT CALCUL ATE THE ME AN FOR NOMINAL OR ORDINAL DATA

Calculating the average for quantitative data is commonplace, but it does not make
sense to calculate it for nominal or ordinal data. For example, can we calculate the
average gender? The real question is, “Why would we want to calculate an average
gender? A person is either a male or a female!” Despite already knowing that, let’s try
calculating the mean for gender and see where it takes us.
First, we already know that we must have numeric values to compute the mean of
data we have collected. In this case, since gender is not represented by numeric values,
let’s go so far as to replace all the values of M and F with numeric values of 1 and 2,
respectively; we can see this in Table 2.10.

TABLE 2.10.  Nominal Data Transformed to Numeric Values

Gender Numeric gender
F 2
F 2
M 1
M 1
M 1
F 2
F 2

We could then use these values to enter into our equation for a sample mean:

2 + 2 +1+1+1+ 2 + 2
x=
7
Using these data, we would compute an average of 1.57. Is this meaningful to us?
No, because it doesn’t tell us anything meaningful about our data.
The same holds true if we try to calculate the mean for ordinal data. For example,
let’s use the data for College Importance from Table 2.9 and calculate the mean. Just
as before, we would add all the values of importance leaving us with 15. When we
divide that by the 7, the number of values in the dataset, it leaves us with an average of
2.14. Keeping in mind that each student’s attitude is subjective, this value doesn’t tell
 Identifying the Independent and Dependent Variables  |  47

us anything about the overall group. A college importance ranking of 2 by one student
could mean something entirely different than a ranking of 2 by another student.

The Median
The median is the score that falls in the exact center of a sorted dataset and can be
used with ordinal, interval, or ratio data. Because there is no commonly agreed upon
symbol for either the population or sample median, we will use “Mdn” in this book.
To see an example of a median, let’s use the absences from our original table shown
again in Table 2.11; if we sorted them, lowest to highest, we could see that the 3 is right
in the middle.

TABLE 2.11.  Absences Ratio Data

Absences
0
3
3
3
4
7
8

Let’s expand our findings using the median score of the math exam. To do that,
we would sort the scores, from highest to lowest, as shown in Table 2.12. From that,
we can see that the median score belongs to a non-Hispanic White male with a math
exam grade of 89.

TABLE 2.12.  Median Ethnicity Based on Class Rank

Ethnicity Gender Class rank Math exam
H F 1 92
NHW F 2 91
AA M 3 90
NHW M 4 89
AA M 5 88
NHW F 6 87
H F 7 86

In cases where you have a larger dataset and you cannot readily see the median,
there are formulas to help us locate the median of a dataset. There are two ways to
do this; the formula you use depends on whether you have an even or odd number of
values in your set of data.
48 | STATISTICS TRANSLATED

LOCATING THE MEDIAN OF A DATASE T WITH AN ODD NUMBER OF VALUES

In order to find the midpoint of a dataset with an odd number of values, add 1 to the
total number of items in the dataset and then divide the total by 2. Using the data
above showing ethnicity and math exam information, we can find the midpoint of the
data using the following steps:

Total number in dataset + 1

1. Midpoint =
2
7 +1
2. Midpoint =
2
8
3. Midpoint =
2

4. Midpoint = 4

Remember, this simply means that 4 is the midpoint of the dataset; we must go
to that point in the table to determine the actual median value. For example, in
this case, if you look at the Math exam column in the table and count 4 up from the
bottom or 4 down from the top, you will see, based on the math scores, that the actual
median value is 89.

LOCATING THE MEDIAN OF A DATASE T WITH AN E VEN NUMBER OF VALUES

When we had an odd number of values in our dataset, it was easy to find the middle;
it is the value where there are just as many data values above it as there are below it. In
cases where there is an even number of values, our logic must change just a bit. For
example, let’s use the data from Table 2.9 and add information to the bottom, shown
in Table 2.13, about a student who scored 50 on the math exam.

TABLE 2.13.  Ranked Math Exam Data

Student number Math exam
1 92
2 91
3 90
4 89
5 88
6 87
7 86
8 50

Looking at this table, we can see that there is no middle record, so we cannot
readily point out the median of the dataset. In order to get around that, we have to
 Identifying the Independent and Dependent Variables  |  49

use the two records that fall in the middle of the dataset—in this case the two records
with the scores of 89 and 88. In order to find the median we would add these values
together and take the average; doing so would give us a median of 88.5.
We can use the same formula as before to find our midpoint but will wind up with
a fractional number:

Total number in dataset + 1

1. Midpoint =
2
8 +1
2. Midpoint =
2
9
3. Midpoint =
2

4. Midpoint = 4.5

The resulting answer, 4.5, shows the position in the dataset of the median. Since
there are no fractional values in our table, it’s necessary to average the two numbers
in the fourth and fifth position. That again means you use 88 and 89. The resultant
average, just like before, is 88.5; this is our median. Another thing we need to notice
is that, while the test scores are ranked highest to lowest, the actual scores are not rel-
evant to one another. For example, look at the difference between the students ranked
first and second; we can see a difference of one point; the difference between the
seventh and eighth is 36 points. This means that, although the data are ranked, the
differences are not relative and computing a mean score tells us nothing about rank.
The median is also a good tool to use if our data are quantitative but are badly skewed;
this simply means that the values on one side of the mean are more spread out than
the values on the other side of the mean. We will talk about skewness in detail later,
but here is an example to help us get a basic understanding. Suppose we are interested
in moving to a neighborhood where a realtor tells us there are five homes and the aver-
age home price is $100,000. Upon hearing that, we decide to visit the neighborhood
where we find there are, indeed, five houses. Four of the houses are worth $10,000
each and one is worth $460,000. This is shown in Table 2.14.

TABLE 2.14.  Home Values in Order of Cost

House 1 House 2 House 3 House 4 House 5 Average
$10,000 $10,000 $10,000 $10,000 $460,000 $100,000

While we might be disappointed with what the realtor told us, it was true; the
average home in the neighborhood is worth $100,000. However, since there are many
more values below the mean than there are above the mean, it would have been better
if the realtor had told us that the median value was $10,000. That would have been a
more meaningful indicator of what the average home in the neighborhood is worth.
50 | STATISTICS TRANSLATED

WE DO NOT CALCUL ATE THE MEDIAN FOR NOMINAL DATA

The median is the best measure of central tendency for ordinal-level data, as well as
badly skewed quantitative data, but we cannot use it with nominal data. For example,
Table 2.15 presents a list showing the highest college degree earned by employees of a
marketing company. In this case, we can see that the midpoint in the list represents
a person with an MBA, but what does that tell us? Absolutely nothing; it tells us the
center point of an ordered list and that is it. As far as we know, everyone on the list has
an MBA.
TABLE 2.15.  Position of Median Degree Earned
Degree earned
BS
BS
BS
BS
MBA
MBA
MBA
MS
PhD

The Mode
Up to this point we have talked about the mean and the median; both can be used
to help us examine the central tendency of a dataset. As we saw, the mean cannot be
used with nominal and ordinal data, and the median cannot be used with nominal
data. Fortunately, we have another tool, the mode, which can be used with all types
of data. The mode is nothing more than the value that occurs most often in a dataset.
Like the median, there is no one standard symbol, so we will use Mo.
If we look at the Absences field from our original dataset (shown as Table 2.16),
we can see that our mode is 3 since it occurs most often; we can further see that F is
our modal value for gender since we have four females and three males.

TABLE 2.16.  Determination of Modal Values for Absences and Gender

Student
number Gender Ethnicity Math exam Class rank Absences
1 F H 92 1 7
2 F NHW 91 2 4
3 M AA 90 3 0
4 M NHW 89 4 8
5 M AA 88 5 3
6 F NHW 87 6 3
7 F H 86 7 3
 Identifying the Independent and Dependent Variables  |  51

The idea of the mode is straightforward, but we must be aware of a small issue
that arises when we have an equal number of any given value. For example, suppose
we added another student, whose ethnicity is AA, to the data; what then would our
mode be for Ethnicity? Since we would then have three NHWs, three AAs, and two
Hs, it appears we do not have a mode since no one value occurs more often than the
others. Since NHW and AA occur an equal number of times, we have two modes; we
call this a bimodal dataset. In other instances, you might find three, four, five, or even
more modes in a dataset. Instead of getting complicated and trying to use a different
name for each distribution, we just refer to any dataset having more than one mode
as multimodal.
So, after all of that, we have three measures of central tendency: the mean, the
median, and the mode. Table 2.17 helps clarify exactly when to use each.

TABLE 2.17.  Relationship between Data Type and Measure of Central Tendency
Nominal Ordinal Interval Ratio
Mean No No Yes Yes
Median No Yes Yes Yes
Mode Yes Yes Yes Yes

Using Statistical Software to Analyze Our Data

First, before we go any further, let me remind you that, in the Introduction, I told you
this book is designed to help you become a good consumer of statistics. I also said we
would be using examples from statistical software to make our lives easier. Knowing
that, there are many good software packages we could use and, if we used the same
data, we would get the same answers. You will find that some software is easier to use
than others, there are many different abbreviations that mean the same thing from
package to package and some are more sophisticated than others. I’ve found that
IBM SPSS (i.e., Statistical Package for the Social Sciences) is the easiest to use and
comprehend. Because of that, we’ll look at how to use the software as well as examples
of output from it to explain the concepts we are covering. Let’s start with Table 2.18
showing data collected from 10 patients.

TABLE 2.18.  Patient Data

Gender Health Perception Anxiety Exercise (minutes)
F 3 100  0
M 2  55 60
F 1  80 10
F 1  25 45
F 2  65 30
M 3  90 15
M 3 100 45
F 1  40 50
F 2  10 40
F 2   0 60
52 | STATISTICS TRANSLATED

Within the table you see four columns; the first column shows valid values for
Gender—F and M. Knowing that this is nominal data, we could easily compute a mode
of 7 since there are seven females represented in the table. The second column repre-
sents a patient’s self-perception of their health. Values range from 1 (Poor) to 3 (Very
Good). From our earlier discussion, we know it is appropriate to compute both the
mode and median for ordinal data; here the mode is 2 since it occurs four times. We
could sort the patients based on the self-perception of their health and, since our data-
set contains an even number of cases, we would average the fifth and sixth values to
determine the median; in this case, it is also 2.
The third column, labeled Anxiety, represents a person’s score, from zero to 100,
based on a survey they completed. Knowing that, is this an interval or ratio-scale item?
We can determine that by answering two questions. First, does anxiety have an abso-
lute zero point? In this case, it does not. Although some of the patients have a score of
zero, it doesn’t mean they do not experience anxiety in their lives; it only means they
have no anxiety as measured by the survey or scale used as a measurement instru-
ment. Second, is a person with a score of 80 twice as anxious as a person with a score
of 40? They are not; the scores only represent each person’s anxiety level relative to the
scale. Now, knowing the answers to both of those questions leads us to determine that
the data are on an interval scale. Because of that, we can compute a mean of 56.5, a
median of 60, and a mode of 100.
Next, we have the number of minutes that each person exercised daily. Using this
data, can ratios be made? Yes, a person who exercised for 60 minutes exercised twice
as long as a person who exercised for 30 minutes. There is also an absolute zero—if
the person did no exercise at all, their score would be zero. Knowing both things, we
can feel safe in saying that the data are ratio level. As we did with interval data, we can
compute a mean, in this case, 35.5 and a median of 42.5. We can also see that the data
are bimodal since we have two occurrences each of 45 and 60.
Everything we have computed so far is shown in Table 2.19.

TABLE 2.19.  Central Tendency Based on Data Type

Gender Perception Anxiety Exercise
Mean 56.5 35.5
Median 2 60 42.5
Mode F 2 100 45 and 60

Entering the Data into SPSS

Entering our data into SPSS is a straightforward process. It’s based on a spreadsheet
format, much like EXCEL or any of the other spreadsheets we have all used at one
time or another. When we start SPSS, we will see the Data View screen shown in Fig-
ure 2.5. We will ultimately use this screen to enter the data we want to analyze but, for
now, let’s click on the Variable View button at the bottom left of the screen. This will
take us to the next screen shown in Figure 2.6, where we can define each variable we
will use for each of the values in our dataset.
Each variable in our dataset will be defined within a row of the table (Figure 2.7).
 Identifying the Independent and Dependent Variables  |  53

FIGURE 2.5.  SPSS Data View spreadsheet.

FIGURE 2.6.  SPSS Variable View spreadsheet.

54 | STATISTICS TRANSLATED

FIGURE 2.7.  SPSS Variable View spreadsheet with variables entered.

We will use the Name column to identify each variable and then use the remainder
of the columns in that row to define the characteristics of that variable. The first row
represents Gender. It is a string variable Type with a maximum Width of one char-
acter; obviously there are no decimal points. The label “Gender” will be used when
referring to this variable on output produced by SPSS. In this case, by clicking on the
actual variable name, using the Variable Labels dialog box, we have created two Val-
ues. If the actual data value is F, it will show as “Female” on any reports we create; a
value of M will show as “Male.” We have no entry in the Missing column, which means
the system will not automatically enter values if a data value is missing; Column and
Align indicate that there are eight spaces, with the value left justified, on any output
we create. The measure, of course, is nominal, and the Role label indicates this field
will be used as input.
The next variable, Perception, represents a person’s self-perception of their health
and is a numeric value between 1 and 3, no decimal points are allowed. The variable
Label is “Health Perception” and there are three possible values: 1 (Poor), 2 (Good),
and 3 (Very Good). The Missing, Column, and Align categories work the same as
for the Gender value, but, in this case, since the patients are ranking themselves, the
Measure is ordinal.
The remaining two columns show levels of Anxiety and the amount of Exercise.
In this case, the data are quantitative, either interval or ratio; SPSS simply labels them
both as “scale” data. The remaining columns are self-explanatory. From here, we need
to analyze our data, so let’s go back to the Data View screen and enter the data from
above.
In Figure 2.8 you can see that we’ve entered the data from Table 2.18, so we can go
ahead with our analysis. First, as seen in Figure 2.9, we’ll select the Analyze command,
Descriptive Statistics and Frequencies. This will be followed by the screen shown in
Figure 2.10. In this case, in the Frequencies: Statistics box we will select Mean, Medi-
 Identifying the Independent and Dependent Variables  |  55

FIGURE 2.8.  Data View spreadsheet with data values entered.

an, and Mode, followed by Continue; we will then select OK in the left box. SPSS will
then produce the output shown in Figure 2.11.
It’s readily apparent that the values shown in this box are exactly equal to those
we computed earlier. Three things bear mentioning, however. First, as we just said,
Perception is ordinal but since the values are numeric, SPSS computes the mean,
median, and mode. Second, we can see there is a lowercase “a” next to the Mode for
Exercise. This only occurs when the data are multimodal, and the software displays
the lowest value; this is explained in the note below the table. Finally, since Gender

FIGURE 2.9.  Selecting the Descriptives and Frequencies commands on the SPSS Data View
spreadsheet.
56 | STATISTICS TRANSLATED

FIGURE 2.10.  Using the Frequencies command on the Data View spreadsheet.

is nominal, none of the descriptive statistics other than N (i.e., the number of values)
are produced.

Summary of the First Part of Step 4: Identify and Describe

the Dependent Variable
In this section we have covered the first part of Step 4; we learned to identify indepen-
dent variables and their levels, dependent variables, and the different types of data
and, finally, we learned how to compute measures of central tendency both manually
and using SPSS. In the next chapter, we’ll continue with Step 4 of our six-step model
by looking at measures of dispersion and relative standing. Since we already know how
to locate the center of a dataset, these tools will help us determine how the data are
spread out around the central point. For now, let’s make sure we completely under-
stand everything we have covered in this chapter.

Statistics

Health
Gender Perception Anxiety Exercise

N Valid 10 10 10 10

Missing 0 0 0 0

Mean 2.00 56.5000 35.5000

Median 2.00 60.0000 42.5000

Mode 2 100.00 45.00a

a. Multiple modes exist. The smallest value is shown

FIGURE 2.11.  Descriptive statistics generated by the Frequencies command.

 Identifying the Independent and Dependent Variables  |  57

Do You Understand These Key Words and Phrases?

categorical data continuous data
dependent variable independent variable
interval data levels of the independent variable
manipulated (experimental) measures of central tendency
  independent variable mean
nonmanipulated mode
  (quasi) independent variable median
multivariate nonparametric statistics
nominal data ordinal data
numeric descriptive statistics quantitative data
parametric statistics rank data
ratio data univariate

Do You Understand These Formulas?

Mean of a population:
Sx
m=
N
Mean of a sample:
Sx
x=
n

Midpoint for a sample:

Total number in dataset + 1
2

Quiz Time!

Let’s start by looking at the following hypotheses. Read each one and then identify the inde-
pendent variable and its levels; then explain why the levels are either manipulated or nonma-
nipulated. Following that, identify the dependent variable and the type of data it represents.

1. There will be a significant difference in motivation scores between students in

online programs and students in traditional programs.

 2. There will be a significant difference in the number of males and females working
as computer programmers in corporate America.

 3. There will be a significant difference in the number of females in computer science
classes between students in the United States, France, and Russia.
58 | STATISTICS TRANSLATED

 4. There will be a significant difference in weight gained during their freshman year
between students who live at home and students who live away from home.

 5. There will be a significant difference in first-year salaries between graduates of Ivy
League schools and graduates of state universities.

 6. Administrative assistants who work in cubicles are significantly less productive
than administrative assistants who work in enclosed offices.

 7. Primary-care patients who are treated by an osteopathic physician will have sig-
nificantly fewer health problems than primary-care patients who are treated by an
allopathic physician.

 8. Truck drivers who check their tire pressure frequently will have significantly higher
miles-per-gallon than truck drivers who do not check their tire pressure frequently.

 9. Insurance companies that use computer-dialing services to call prospective clients
will have a significantly lower number of sales than insurance companies that use
live agents to call prospective clients.

10. The rankings of favorite sporting activities will be significantly different between
Mexico, the United States, and Canada.
Let’s check to make sure you understand how to compute the basic measures of central ten-
dency using the data in Table 2.20. Let me warn you, be careful! Some of these answers are not
as obvious as they might appear! You will be able to check your answers at the end of the book.

1. What is the average height?

 2. What is the median class rank?
 3. What is the mode of shoe size?
 4. What is the average class?
 5. What is the median height?
 6. What is the mode of height?
 7. What is the median shoe size?
 8. What is the mode class rank?
 9. What is the mode of class?
10. What is the median class?
11. What is the average class rank?
12. What is the average shoe size?
 Measures of Dispersion and Relative Standing  |  59

TABLE 2.21.  Data for Quiz Time!

Height Class Class rank Shoe size
60 FR  1  8
66 FR 10  7
69 SO  5  8
66 SR 13 11
69 JR  6  8
68 JR 12 12
64 FR  2  7
68 SO 14  9
70 SR  8 12
69 JR  3 11
70 FR 15  9
73 SR  6 13
72 SR  9  8
65 SO 11  8
72 SO  4 13
CHAPTER 3

Measures of Dispersion
and Measures of Relative Standing

Introduction
In this chapter we’ll continue with Step 4 of our six-step model by learning how to
compute and understand measures of dispersion and measures of relative standing.
When we measure dispersion, we are simply looking at how spread out our dataset
is. Measures of relative standing allow us to determine how far given values are from
each other as well as from the center of the dataset. As usual, we’ll learn how to
compute everything manually, but then we’ll use SPSS to show us how easy it really
is.

Measures of Dispersion
Measures of dispersion are only used with quantitative data and help us answer the
question, “How spread out is the dataset we plan to use?” For example, I might tell
you I own paintings that average about $300 each; I might also tell you that I paid as
little as $25 for some of my paintings and as much as $1,000 for others. Based on that,
you could say that the amount I’ve paid for paintings is widely dispersed around the
mean. In fact, if we subtract $25 from $1,000, we can see that the amount I’ve paid for
pictures could fall anywhere between these two values; in this case a range of $975.
Computing the range is the most common of the measures of dispersion, but we’ll
also look at the standard deviation and variance. All three measures are critical for the
accurate computation of inferential statistics.

60
 Measures of Dispersion and Relative Standing  |  61

The Range
As we just saw, the range tells us how far apart the largest and smallest values in a
dataset are from each other; it is determined by subtracting the smallest value in the
dataset from the largest value in the data. For example, in Table 3.1, when we subtract
25 from 75, we have a range of 50; we call this our computed range. In this case, we also
have a mean score of 50.

TABLE 3.1.  Range of Data Values from 25 to 75

25 30 35 40 45 50 55 60 65 70 75

While this is very easy to compute, we must be careful in its interpretation when we
are interested in the dispersion of the data. Remember, the purpose of the measures
of dispersion is to understand how the data are spread out around the center of the
dataset. Using the range to help us understand this is very imprecise and can be very
misleading. For example, what does a range of 50 tell us? What can we learn about how
the data are dispersed around the center of the data by computing it? Again, let’s use
the data in Table 3.1 and look at some of the issues we might encounter.
We already know that both our mean and range are 50. In this case, all we can
tell by using the range as a measure of dispersion is that the data are evenly spread
out around the average. This can be seen very clearly; the smallest value is 25 points
below the mean, the largest value is 25 points above the mean, and the data values are
spaced at equal increments of 5 points each.
We must be careful, however, because we can often compute the same range for
two datasets that have the same mean score and range, but their dispersion will be
completely different. For example, as in the previous example, the lowest value in
Table 3.2 is 25 and the largest is 75; this leaves us with a range of 50.

TABLE 3.2.  Range of Data Values from 25 to 75 Clustered around the Mean
25 42 43 44 45 50 51 52 53 54 75

In this case, although we can see that the range is somewhat large, most of the
values are clustered very closely around the mean. This isn’t to say that the range can-
not be used effectively; it is perfectly acceptable for a given task while remembering
its limitations. For example, investors might want to look at the range of values that a
stock sold for on a given day, or a college professor might be interested in knowing the
overall spread of test scores for an examination. In both cases the results are meaning-
ful, but the practical application of the range is often limited.

The Possible Range of the Data

There is a manner by which the range can be used in a very meaningful fashion. As we
noted above, the computed range is the result of subtracting the smallest number in
the dataset from the largest number. We can also establish a possible range that’s noth-
ing more than the entire range of values a dataset might include. For example, the
possible range on the Graduate Management Admissions Test (GMAT) runs from 200
62 | STATISTICS TRANSLATED

to 800 and, since there are 100 senators in the United States Congress, the possible
range of the number of “yes” votes on any piece of legislation would be 0 to 100. We
can also establish our own possible range by simply deciding the boundaries we are
looking for; we’ll use shopping mall sales pitches as an example.
We have all experienced those people who approach you in the mall asking for
you to try one product or the other; if you’re like me, you pretend you don’t notice
them and walk away as quickly as possible! Other people, who I guess are more will-
ing to part with their hard-earned money than I am, stop to talk and try out the new
product. In this case, suppose our marketers are interested in knowing the appeal of
a new cologne to males in the 20 to 40 age bracket; when we subtract 20 from 40, we
have a possible range of 20. While that isn’t terribly interesting by itself, we can use the
possible range and the computed range together to help in decision making.
For example, if the employee of the marketing company asked each male trying
the new product their age, the marketing company could use the data to compute the
range of mall patrons who tried the product. In this case, let’s say the youngest person
who rated the product was 20 and the oldest was 32; this would result in a computed
range of 12. If we were to compare our computed range of 12 to our possible range
of 20, we would see that our computed age range is somewhat spread out since it
accounts for 60% of the possible range. At the same time, since our sample only repre-
sents the opinions of the younger men, in order to really understand the appeal of the
cologne to their targeted age group, the marketing staff would need to get opinions
from the remaining 40%, all of whom would fall in the 33 to 40 age range. We can see
this entire idea demonstrated in Table 3.3.

TABLE 3.3.  Comparison of 60% Computed Range to the Possible Range

Computed range = 12 (60%)
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
Possible range = 20 (100%)

Let’s use the same possible range but, this time, the youngest person trying the
product was 28 and the oldest was 34; this would give us a computed range of 6. Since
our possible range is 20, we’ve only accounted for 30% of our possible range—not a
very wide dispersion. In order to get responses from the entire possible range, the
marketers would need to focus their efforts on males in the 20 to 27 age range as well
as those in the 35 to 40 age range. As you can see, range only tells us is how spread out
the data are and what percentage of the possible range can be accounted for by the
computed range. This is shown in Table 3.4.

TABLE 3.4.  Comparison of 30% Computed Range to the Possible Range

Computed range = 6 (30%)
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
Possible range = 20 (100%)
 Measures of Dispersion and Relative Standing  |  63

The Standard Deviation

In many cases we need to determine exactly how far a given value is from the center of
the dataset. The standard deviation allows us to measure the distance from the mean
of a dataset in terms of the original scale of measurement. In other words, if we have
a dataset consisting of a range of values, height for example, we can use the standard
deviation to tell us how far a given height is, in terms of original scale of measurement
(e.g., inches or feet) from the mean of the dataset. We will use the standard deviation
as an integral part of several of the inferential statistics we will learn to compute later
in the book, but, for now, the concepts underlying the standard deviation are easy to
understand.
First, since the standard deviation is computed in terms of the original scale,
there’s an easy way to think about it. The standard deviation simply approximates the
average distance any given value in the dataset is away from the mean of the dataset.
Let’s use the heights of nine employees in Table 3.5 to demonstrate this idea.

TABLE 3.5.  Range of Heights Used to Compute a Standard Deviation

Height in inches
64 65 66 67 68 69 70 71 72

Here the data are sorted and range from the smallest value on the left (i.e., 64
inches), to the largest value on the right (i.e., 72 inches). The mean, 68, falls right in
the center of the dataset. Although we won’t go into the calculations now, the stan-
dard deviation for this data is 2.74; again, this approximates the average distance each
person’s height is from the average height. How do we apply this to the overall idea of
dispersion? Simple: the larger the standard deviation, the more spread out the heights
are from the mean; the smaller the standard deviation, the closer each score is to the
mean.
In order to better demonstrate this, in Table 3.6 let’s change the data we just used
so that the values above and below the mean are more widely spread out; I have illus-
trated this with the arrows moving away from the mean in both directions. While the
mean remains the same, the standard deviation increases to 5.48. This tells us that the
approximate average distance any value in the dataset is from the mean is larger than
when the data values were grouped closer to the mean.

TABLE 3.6.  Range of Heights to Compute a Larger Standard Deviation

Height in inches
60 62 64 66 68 70 72 74 76
Mean

The exact opposite happens if we group the data closer to the mean of the data-
set. In Table 3.7, the mean is again 68, but the values are much more closely clustered
around the mean. This results in a standard deviation of 1.58, much smaller than the
original standard deviation of 2.74.
64 | STATISTICS TRANSLATED

TABLE 3.7.  Range of Heights to Compute a Smaller Standard Deviation

Height in inches
66 66 67 67 68 69 69 70 70
Mean

Making Estimations and Comparisons Using the Standard Deviation

Unlike the range, we can use the standard deviation to determine how far each value
in a dataset is from the mean; we don’t have to simply rely on concepts such as “small”
and “large” when talking about dispersion. Later in the book we’ll use other statistical
tools and concepts to develop this idea further, but, at this point, all we need is the
standard deviation to make comparisons. For example, if we are interested in compar-
ing the number of sick days our employees take on a yearly basis, we might use the
data in Table 3.8.

TABLE 3.8.  Using the Standard Deviation to Compare Values in One Set of Data
Annual sick days
2 5 5 6 6 6 7 7 10
average

Herein, the average employee calls in sick 6 days each year. If we computed the
standard deviation, we would find that the average number of sick days taken is 2.12
days away from the mean. For clarity’s sake, we’ll round that off to 2 meaning that the
average number of days employees call in sick ranges from 4 (the mean of 6 minus
one standard deviation) to 8 (the mean of 6 plus one standard deviation). Anything
outside of this is not in the average range. In this case, the person who used only two
sick days is below average but is still acceptable. Obviously, this is a good thing unless
they are coming to work sick when they should be staying at home! On the other end,
we might want to talk to the person taking 10 days a year; this might be a bit excessive.
We can also make comparisons of datasets using the standard deviation. For
example, suppose we are a new fan of professional football in the United States and
want to know as much about it as possible. One thing that interests you is the equality
between the two divisions—the National Football Conference (NFC) and the Ameri-
can Football Conference (AFC). Let’s use Table 3.9 to represent the total points scored
by each team during a season in order to compare the two conferences. Yes, I know
there are more than nine teams in each conference, but this page is only so wide!

TABLE 3.9.  Football Scores Used to Compare Two Sets of Data

NFC 304 311 315 320 325 340 345 350 360
AFC 100 105 298 305 340 410 510 642 710

As shown in Table 3.10, the NFC teams score, on average, 330 points per season
while the AFC teams score 380 points per season; their range is 56 and 610, respec-
 Measures of Dispersion and Relative Standing  |  65

tively. At this point, we might decide to head to Las Vegas; it’s apparent that the
AFC teams are much better than the NFC teams—maybe there is money to be made!
Before we bet our hard-earned cash, however, let’s look a little deeper into how the
team scores are dispersed within the two leagues. I’ve already computed the mean,
average range, and standard deviation for each league. Let’s use this to compare their
performance.

TABLE 3.10.  Using the Standard Deviation to Compare Values

between Football Conferences
Conference Average points Standard deviation Average score range
NFC 330 19.4 310.6–349.4
AFC 380 213.11 166.9–593.11

First, the standard deviation for the NFC teams is 19.4 and 213.11 for the AFC
teams. This means, on average, the NFC teams score between 310.6 and 349.4 points
per season while the AFC teams score, on average, between 166.9 and 593.11 points
per season. What does this tell us? The large standard deviation means the AFC
scores are very spread out—there are some very good AFC teams and some underper-
forming AFC teams. NFC teams, while they have a lower overall average, are more
evenly matched. What does this mean? When an AFC team plays an NFC team, do
we always bet on the AFC team because of their higher overall average? No, it would
probably be better to look at the individual teams when we are placing our wagers!

Computing the Standard Deviation for a Sample

The basic logic of the standard deviation is straightforward. We’ll use the SPSS soft-
ware later in the book to make more precise measurements, but, for now, let’s move
forward by learning how to compute it. Do you remember earlier in the book where
I said all you need is basic math for most statistical calculations? Even though the fol-
lowing formula may look a bit complicated, that’s still my story and I’m sticking to it!

Σ ( x − x )2
s=
n −1
Although there are quite a few symbols in this equation, it is actually very easy to
compute. First, let’s make sure we know what we are looking at:

1. s stands for the sample standard deviation.

2. n is the number of values in our sample.
3. x is each individual value in our sample.
4. x is the mean of our sample.
5. S means to add together everything following it.
6. means to take the square root of the value under it.
66 | STATISTICS TRANSLATED

Now that we understand all of the symbols, let’s go step by step through the compu-
tation; after that, we’ll use it with our data to see what the actual standard deviation is:

1. (x − x )2 means to subtract the average score of the entire dataset from each
of the individual scores; this is called the deviation from the mean. Each of
these values is then squared.
2. S is the Greek letter sigma and indicates we should add together all of the
squared values from Step 1.
3. n – 1 is the number of values in the dataset minus one. Subtracting one from
the total number helps adjust for the error caused by using only a sample of
data as opposed to the entire population of data.

Let’s go back and verify the computations for the NFC scores above by going
through Table 3.11.

TABLE 3.11.  Computing the Standard Deviation for Football Scores

Deviation from the Deviation from the
Mean of NFC mean mean squared
NFC scores scores (x − x ) ( x − x )2
304 330 304 – 330 = –26 –26 * –26 = 676
311 330 311 – 330 = –19 –19 * –19 = 361
315 330 315 – 330 = –15 –15 * –15 = 225
320 330 320 – 330 = –10 –10 * –10 = 100
325 330 325 – 330 = –5 –5 * –5 = 25
340 330 340 – 330 = 10 10 * 10 = 100
345 330 345 – 330 = 15 15 * 15 = 225
350 330 350 – 330 = 20 20 * 20 = 400
360 330 360 – 330 = 30 30 * 30 = 900
x = 330 S=0 S = 3012

In the first column, we see each of the NFC scores in our sample. When we sub-
tract the sample mean, shown in the second column, from each of those values, we see
the difference shown in column three; this is our deviation from the mean. Notice, if
the observed value is less than the mean, the difference will be negative. The fourth
column shows the result when we square each of the values in the third column
(remember Math 101; when we square a negative value, it becomes a positive value).
The sum of all the squared values, 3012, is shown at the bottom of column four.
Remembering that n – 1 is the total number of values in our dataset, minus 1, we can
now include all these values in our formula:

3012
s=
8
The square root of 3012 is 376.5; when we divide that by 8, we have computed a
standard deviation of 19.4—believe it or not, it’s really that easy! We will verify what
we’ve done using SPSS in only a few pages, but first let’s discuss a minor inconvenience.
 Measures of Dispersion and Relative Standing  |  67

WHY DID WE DO WHAT WE JUST DID?

When we first began our discussion of the standard deviation, we said we could think
of it as the average distance that each value is from the mean of the dataset. Given
that, when we look back at columns three and four in the table above, the obvious
question becomes, “If we are interested in knowing the average distance each value
in our dataset is from the mean, why do we even need the fourth column? Why don’t
we just add the deviations from the mean and divide them by the number of values in
the sample?”
Unfortunately, it’s not that easy. As you can see in the column labeled “Deviation
from the Mean,” if we computed the sum of these values, the answer would be zero.
Because the sum of x − x is an integral part of our equation, that means the standard
deviation would always be zero. We avoid that by squaring each of the values; this,
of course, means that any deviation from the mean less than zero becomes a positive
value. After squaring those values, we then divide the sum of those values by n – 1,
giving us an approximation of the average distance from the mean.
At this point, we run into yet another problem! Since we have squared each of the
values, we are no longer measuring the standard deviation in terms of the original
scale. In this case, instead of a scale representing the points scored by each division,
we have the squared value of each division. In order to transform this value back into
the scale we want, all we need to do is take the square root of the entire process; that
gets us back to where we want to be.

Computing the Standard Deviation for a Population

As you’ve probably guessed, if we can compute the standard deviation for a sample, we
can do the same for a population. The formula, essentially, remains the same; there
are only three small differences:

S ( x − m)2
σ=
N
As you can see, since we are now computing a population parameter, we have
changed the symbol for the standard deviation to the lower-case Greek letter sigma
(i.e., σ). We have also included the Greek letter mu (i.e., μ), which represents the mean
of the population. Since we are no longer worried about the inherent error caused by
sampling, we can drop the idea of subtracting 1 from our denominator and replace
it with an uppercase N. This means that we are using all the values in the population.
The overall concept, however, is the same. The population standard deviation simply
shows the dispersion of the dataset when we know all the possible values in the popu-
lation.

The Variance
The variance is a third tool that is used to show how spread out a dataset is and inter-
preting it is easy. If the variance value is large for a given dataset, the data values are
more spread out than if the variance value is small. You are probably thinking, “That
sounds too simple, what’s the catch?” The good news is there’s not a catch. Let’s look
68 | STATISTICS TRANSLATED

at the formula for calculating the variance of a sample and then put some numbers
into it to show you what I mean:

S ( x − x )2
s2 =
n −1
If it seems that we’ve seen some of this before, you’re right. Other than a couple
of minor changes, this is nothing more than the formula for the standard deviation
for a sample. We are now using s2, the symbol for the sample variance, instead of s,
which is the symbol for the sample standard deviation. The fact that we are squaring
the s value only means that we are no longer taking the square root of our computa-
tions, because of that, we must remove the square-root symbol. In other words, if we
know the standard deviation, all we need to do is multiply it by itself to compute the
variance. For example, since we know the standard deviation for our NFC football
scores is 19.4, all we need to do is multiply that by itself leaving a variance of 376.4.
That means, of course, that if we know the variance of a set of data, its square root is
the standard deviation. The obvious question, then is, “Why do we even need to com-
pute the variance since it’s just another measure of dispersion? Can’t we just use the
standard deviation in its place?”
The answer to those questions is simple; in most cases we can use the standard
deviation in our statistical inferential decision making. Certain statistical tests, how-
ever, use the variance as the tool for measurement.

Interpreting the Variance

Before we start talking about how to interpret the variance, let’s create a dataset to
work with. In Table 3.12, the top row of numbers contains values ranging from 3 to 9;
the average for this data is 6, the standard deviation is 2.16, and the variance is 4.67.
The second row in the table shows each value’s deviation from the mean. For example,
the deviation of the value 3 below the mean of 6 is –3 while the deviation of 9 above
the mean is +3. The rightmost column shows that the sum of all deviations from the
mean is zero; that stands to reason simply because of the way the mean is calculated.
The third row shows how far each value is from the mean in terms of the standard
deviation. For example, the value –3 is –1.39 standard deviations below the mean.
We determined this by dividing the observed value by the standard deviation; for
example, we can see that –3/2.16 = –1.39. Again, since the standard deviation is based
on the mean, the sum of those values, shown in the rightmost column of Table 3.12,
is zero.

TABLE 3.12.  Computing Distances from the Mean with a Small Amount of
Dispersion
Average = 6
Standard deviation = 2.16
3 4 5 6 7 8 9 Variance = 4.67
Distance from –3 –2 –1 0 1 2 3 Sum = 0
the mean
Standard deviations –1.39 –.926 –.463 0.00 .463 .926 1.39 Sum = 0
from the mean
 Measures of Dispersion and Relative Standing  |  69

In order to interpret these data, you must look at them from several perspectives.
First, just as is the case with the range and standard deviation, the greater the
variance, the more spread out the data are. Unlike the range and standard deviation,
however, the variance can be smaller than, equal to, or larger than the actual range of
the data. In this case, our range (6) is greater than our variance (4.67); this indicates
that the values are not very spread out around the mean, but that’s not always the case.
For example, suppose the following data represent the number of times an hour
our neighbor’s dogs bark; our imaginary neighbor has seven dogs, therefore we have
seven values. If we computed the range and variance for the data, we would find that
the variance of 1108.33 is much larger than the range of 85. As shown in Table 3.13,
this means the number of barks per dog is highly dispersed around the mean of 50;
some dogs bark quite a bit, others much less so.

TABLE 3.13.  Computing Distances from the Mean with a Large Amount of
Dispersion
Average = 50
Standard deviation = 33.29
5 15 30 60 70 80 90 Variance = 1108.33
Distance from –45 –35 –20 10 20 30 40 Sum = 0
the mean
Standard deviations –1.35 –1.05 –.601 .300 .601 .901 1.20 Sum = 0
from the mean

Second, the variance can never be negative simply because it is the standard
deviation squared. Unlike the situation when we use the standard deviation while
computing inferential statistics, this means we will not be canceling out like values
above and below the mean of a dataset. This will allow us, later in the text, to use
specific statistical tests based on the idea of comparing variance between two groups
to test a hypothesis. For now, however, just remember that the larger the variance, the
more spread out the data values are away from the mean.
Third, we must remember that both the standard deviation and variance measure
dispersion around the center of the dataset, but there’s a distinct difference. Since the
variance is the squared standard deviation, this means it is no longer representative of
the original scale. For example, if we weigh a group of students and compute an aver-
age of 100 and a standard deviation of 10, it means a person weighing 90 pounds would
be one standard deviation, measured in pounds, below the mean. The variance of the
dataset, since it’s the standard deviation squared (i.e., 100 in this case), simply shows
relative position from the center of the dataset and it is not in the original scale. Again,
all we know is that a larger variance value means that the data are more spread out.

The Variance of a Population

Again, since we can compute the variance for a sample, we can compute the varia-
tion for a population if we know all the population parameters. The only differences
between the formula for a population and the formula for a sample are three different
symbols:
S ( x − m)2
σ2 =
N
70 | STATISTICS TRANSLATED

The symbol for a population’s variance is σ2 (i.e., sigma squared). In order to

compute it, we subtract the population mean mu (μ) from each of the observed val-
ues; we then compute and square the sum of those values. Following that, we divide
that sum by the number of values in the population (N) giving us the variance for the
population.

Using the SPSS Software to Verify Our Measures of Dispersion

Let’s use our SPSS software again to make sure our calculations are correct. Since
we already know how to use the Variable View screen, let’s go directly to Data View,
shown in Figure 3.1; you can see that we have already entered the data under a variable
named Score and selected Analyze, Descriptive Statistics, and Frequencies.
Once Frequencies is selected, as you can see in Figure 3.2, you can specify the
variables you want to analyze, as well as indicate which values you want SPSS to gener-
ate. In the Dispersion box, we are selecting range, standard deviation, and variance.
When we select Continue and OK, we get the output shown in Figure 3.3; as you
can see, our standard deviation, variance, and range all match what we computed
manually when examining the NFC scores.

Measures of Relative Standing

After learning to identify the center of a dataset and determine the dispersion of the
values within the dataset, we now need to discuss measures of relative standing. These
tools will allow us to look at the relationship between individual values within our data
and a given measure of central tendency. The five most common measures of relative
standing are percentiles, quartiles, z scores, T-scores, and stanines.

FIGURE 3.1.  Selecting the Frequencies command on the Data View spreadsheet.
 Measures of Dispersion and Relative Standing  |  71

FIGURE 3.2.  Using the Frequencies command to generate the standard deviation, variance,
and range.

Percentiles
Many of us are used to hearing the word percentile, especially when discussing test
scores, ability, or aptitude. For example, we might hear that Bob, one of our friends,
scored in the 80th percentile on a physical fitness test. What does that mean though?
Is this good or bad? While this sounds good, we wouldn’t really know anything about
Bob’s physical fitness unless we looked at his actual score on the physical fitness test.
His percentile score is simply the per-
centage of people scoring below that
Statistics value. With that definition, all we know
is that Bob scored higher than 80% of
Scores
the others taking the same fitness test.
N Valid 9 Bob might be in bad physical shape, but
he’s better off than 80% of the others!
Missing 0 Let’s look at a detailed example.
Suppose we have a group of 11
Std. Deviation 19.40361 newly hired attorneys in our company
Variance 376.500 and, while they all passed the bar exam,
there were some very low scores. One
Range 56.00 of the young attorneys, a real go-get-
ter, knew his score of 79 was fairly low
FIGURE 3.3.  Output from the Frequencies com- and was trying to determine a way to
mand. “stand out in the crowd.” Having heard
72 | STATISTICS TRANSLATED

that statisticians are wizards at making bad things look better, the young attorney
approached me and told me his dilemma. While I disagreed with the “wizard” part, I
did tell him there’s another way he could look at his score. He could compare himself
to the rest of the newly hired attorneys and determine, using a percentile score, where
he ranked in relation to them. Although he would still have exactly the same score on
the bar exam, his percentile score might make him look better than he really is.
Knowing that, let’s use the following scores to represent the bar scores of all our
newly hired attorneys. Not knowing any better, I will just assume the scores on the bar
exam range from 0 to 100, with our group of attorneys having scores ranging from 70
to 100 (Table 3.14).

TABLE 3.14.  Bar Examination Scores for New Attorneys

Bar scores for new attorneys
100 78 79 76 75 74 75 74 72 70 70

In order to begin computing percentiles, the first thing we need to do is arrange

the scores in Table 3.14 into descending order; this is shown in Table 3.15. This gives
us the data we need to determine our young attorney’s percentile score by using the
following formula:

Percentile = [(Below + 1/2 Same)/N] * 100

TABLE 3.15.  Bar Scores Ranked from Highest to Lowest

100
79
78
76

Bar scores 75
highest 75
to lowest
74
74
72
70
70

In this equation, “Below” represents the number of scores less than the one we are
interested in; “Same” means all values equal to the one we are interested in, and N, as
usual, means the total number of values in our dataset. Notice, N is upper-case here;
the 11 scores we have refer to the entire population of newly hired attorneys.
In this case, our attorney’s score is 79; there are nine values less than 79 and only
one greater than 79. We can substitute those values, along with N, into our formula,
and a few simple calculations will lead us to our percentile score:
 Measures of Dispersion and Relative Standing  |  73

1. Percentile = [(9 + 1/2 1)/11] * 100

2. Percentile = [(9 + .5)/11] * 100
3. Percentile = (9.5/11) * 100
4. Percentile = .863 * 100
5. Percentile = 86.3

At this point, it’s necessary to drop anything to the right of the decimal point;
we don’t round it off, but rather just drop it. This means things are looking better for
our young friend. Although his actual score was 79, he represents the 86th percentile.
This means his score is greater than 86% of the people taking the bar exam. This
could possibly make him look better in the eyes of his new employer; that’s what he is
betting on.

When It Comes to Percentiles, No One Is Perfect

When I tell my students there is no such thing as a 100th percentile, I always get a
blank stare from some of them. I know they are thinking, “Wait, if someone scored
100 on a test where that was the highest score possible, shouldn’t they be in the 100th
percentile?” Let’s answer their question by imagining a set of scores ranging from 1 to
100 with only one occurrence of each. If we wanted to compute the percentile for the
only person scoring 100, we would have 99 scores below it, 100 would only occur once,
and there would be 100 values. Let’s substitute those values into our equation:

1. Percentile = [(99 + 1/2 1)/100] * 100

2. Percentile = [(99 + .5)/100] * 100
3. Percentile = (99.5/100) * 100
4. Percentile = .995 * 100
5. Percentile = 99.5
6. Percentile = 99th

Again, some people find this perplexing; the person made a perfect score but
was in the 99th percentile; what happened? It’s easy when you think about our defini-
tion. What is the only score that is greater than 99% of the other scores? Easy, it’s 100.
There’s no way that a person could outscore 100% of people taking the bar exam; that
means they would have to outscore themselves. Impossible!

The Median as a Percentile

Although you didn’t know it at the time, we’ve already seen another example of a
percentile, but we called it the median. From our earlier definition and what we know
now, another name for the median is the 50th percentile. We can further divide our
dataset into four quartiles, each representing 25% of the dataset. The first quartile, or
Q1, is the 25th percentile, Q2 is the 50th percentile (or median), and Q3 is the 75th
percentile.
74 | STATISTICS TRANSLATED

We can make this clearer with an example, but, before we do, let me warn you that
there are at least two different ways to compute quartiles. Statisticians tend to disagree
over which one is most representative; for our purposes, we’ll use these guidelines:

X Our definition of the median, represented by Q2, will remain the same; it’s
the 50th percentile.
X Quartile 1, represented by Q1, is the median of the lower half of the data, not
including the value Q2.
X Quartile 3, or Q3, is the median of the upper half of the data, again not
including the value Q2.

Let’s demonstrate starting with Table 3.16. Therein we have an odd number of
values in our dataset, so our median (Q2) is 5.

TABLE 3.16.  Computing Q2 for an Odd Number of Values

1 2 3 4 5 6 7 8 9
Median
Q2

As we just said, we are going to compute Q1 as the median of all values less than
the median; remember, this is the same as the 25th percentile. In this case, the median
will be the average of 2 and 3, or 2.5. Q3 will be computed in the same manner; it is
the median of all values above the median; in this case the average of 7 and 8, or 7.5.
These are marked in Table 3.17.

TABLE 3.17.  Computing Q1, Q2, and Q3 for an Odd Number of Values
1 2 3 4 5 6 7 8 9
5 is the
50th
2.5 is 25th percentile 7.5 is the 75th
percentile or Q1 (i.e., the percentile or Q3
median or
Q2)

In this table, we had an odd number of values; Table 3.18 shows us a dataset with
an even number of values

TABLE 3.18.  Data for an Even Number of Values

10 15 20 30 40 46 52 58 60 62

Knowing we have an even number of values, the median, Q2, is the average of the
middle two numbers; remember that 10 is the number of values in the dataset:

10 + 1
Midpoint =
2
 Measures of Dispersion and Relative Standing  |  75

Our results would show the median is in the 5.5th position. In this case, the median,
Q2, is 43 [i.e., (40 + 46)/2 = 43]. This would also mean that Q1 is the median of all
values to the left of 43. Remember, the value 40 is in our original data, and it is less
than 43, so we must include it. Given that, we have an odd number of values; 10, 15,
20, 30, and 40; we simply identify the one in the middle, 20, as our 25th percentile or
Q1. In order to compute Q3, the 75th percentile, we would use 46, 52, 58, 60, and 62
and locate the middle value. Here Q3 is equal to 58. All of this is shown in Table 3.19.

TABLE 3.19.  Computing Q1, Q2, and Q3 for an Even Number of Values
10 15 20 30 40 46 52 58 60 62
20 is 43 is the 50th 58 is
the 25th percentile the 75th
percentile (i.e., the median percentile
or Q1 or Q2) or Q3

Computing and Interpreting z Scores

Percentiles help us understand the relative standing of a value when compared to
all other values in a dataset. In many instances, however, statisticians want to know
exactly how far a given value is away from the mean or from another value in the data-
set. We saw earlier that we can determine these differences in terms of the standard
deviation, but that was somewhat tedious. Let’s both simplify and expand that concept
by using the standard deviation to compute a z score:

x−x
z=
s
Here we are using the symbols for the sample mean a sample standard deviation
(i.e., s) within the formula. Let’s use the data in Table 3.20. We have a mean of 10, and
to make life simple, let’s use a standard deviation of 1 to calculate the z score for an
observed value of 12.

TABLE 3.20.  Data for Computing z Scores

7 8 9 10 11 12 13

We can compute the z score in three steps:

12 − 10
1. z =
1
2
2. z =
1

3. z = 2

In this case, our z score is 2; this means our value, 12, is two standard deviations
above the mean. We might also find that the z score we compute is negative. For exam-
76 | STATISTICS TRANSLATED

ple, using the same standard deviation and mean score, along with an observed score
of 7, we would compute a z score of –3. This means our observed score is 3 standard
deviations below the mean of the dataset. We can see this as follows:

7 − 10
1. z =
1
−3
2. z =
1

3. z = –3

Both of these examples are shown in Table 3.21.

TABLE 3.21.  z Scores

Value 7 8 9 10 11 12 13
z score –3 –2 –1 0 +1 +2 +3

In most instances, our z score will not be a whole number; it’s quite common to
compute a fractional z score. For example, let’s use a mean of 15, a standard deviation
of 3, and an observed value of 20:

20 − 15
1. z = 3
5
2. z = 3

3. z = 1.67

This would indicate that the observed value, 20, is 1.67 standard deviations above
the mean.

Using z Scores for Comparison

Besides telling us the relative position of a value within a dataset, the z score can be
used to compare values in two different datasets. An example will make this clear.
Suppose a student made 65 on the math section of an achievement test and an
81 on the reading section of the same test. The student’s parents, upon seeing these
scores, might think their child was doing well in reading but not as well in math. As we
learned in our discussion of percentiles, just knowing the raw score on exams of this
type is sometimes not enough; we need to know how well the student did in relation to
the other students taking the exam. We could do this by breaking the score down into
percentiles, but that would involve sorting all the students’ scores and then determin-
ing all the different percentile values. Instead of doing this, we can use the z score to
report the child’s scores in both math and reading.
 Measures of Dispersion and Relative Standing  |  77

In order to do that, let’s use a sample mean of 50 and a standard deviation of 5

for the math test. Using the child’s score of 65, we would compute a math z score of
+3; that means the child scored 3 standard deviations above the mean. Using a sample
mean score of 90 and a standard deviation of 3, our z score for the reading section
would be –3 [i.e., (81–90)/3]; this means the child scored 3 standard deviations below
the mean. By looking at this we could tell, contrary to the impression made by the raw
scores, our student was well above average in math but well below average in reading.
This is shown in Table 3.22.

TABLE 3.22.  Comparison of z Scores and Raw Scores

z = –3 z = –2 z = –1 z=0 z = +1 z = +2 z = +3
Math 35 40 45 50 55 60 65
Reading 81 84 87 90 93 96 99

When we get into the section of graphical descriptive statistics, we will revisit the
subject of z scores. We will find that we can use a predefined table to assign actual
percentile values to our z scores, thereby eliminating the problem of sorting all the
test scores and then determining overall percentile values.

Computing and Interpreting T-Scores

Sometimes trying to explain z scores relative to the average of a group is rather dif-
ficult. For example, some people just can’t comprehend what you’re trying to say when
you tell them a student has a score of +2 or a –3.07. To address this problem, we can
use the T-score, which is nothing more than the z score presented in another form.
Before we learn how to compute it, let me point out something that confuses some
people. Later in the book we are going to learn to use a t-test, and it will involve com-
puting a t value; they aren’t related to what we are computing here. In this case, we
have an upper-case T; the t-test will use a lower-case, italicized t. In order to compute
the T-score, use the following formula:

T = (z * 10) + 50

You can see that all we need to do to compute T is to multiply our z value by 10
and then add 50 to it. If we used the z score of +1 we just computed using the math
score, the T-score for that same value would be 60; when our z score was –3, our
T-score would be 20 (i.e., –30 + 50).
When you compute the T-score, the average will always be 50 with a standard
deviation of 10. For example, in the prior example we had an average math score of
50. This means that if we had an observed value of 50, the z score would be zero. It
also means that when we are computing the T-score using a z score of zero, the answer
is still zero. If we add zero to 50, the mean of our T-score distribution is 50. This is
shown in Table 3.23.
78 | STATISTICS TRANSLATED

TABLE 3.23.  Comparison of the Mean, z Score, and T-Score

Observed value Mean z score T-score
50 50 0 50

Now, let’s use our observed score of 65. Since it is three standard deviations above
the mean, our z score is +3; this means our T-score is 80. If we have an observed value
of 35, our z score is –3; this translates into a T-score of 20. Both examples are shown
in Table 3.24.

TABLE 3.24.  Using the Mean, z Score, and T-Score to Compare Values
in Two Datasets
Observed Standard
Mean value deviation z score T-score
50 65 5 +3 80
50 35 5 –3 20

At this point, if we wanted to show these test scores to a parent, we could explain
that the T-scores range from 20 to 80, with an average of 50. This would provide them
with information in a format they are more used to seeing.

Stanines
Stanine (short for “standard nine”) scores divide the set of values we are looking at
into nine groups. Stanine scores are frequently used in education, especially to report
standardized test scores and to compare groups of students. Computing a stanine
score is very easy; simply multiply the z score by 2 and add 5:

Stanine = (z * 2) + 5

For example, let’s use a dataset with a mean score of 40 and a standard deviation
of 8. If we have an observed value of 56, our z score is 2.

56 − 40
1. z =
8
16
2. z =
8

3. z = 2

When we insert our z score of 2 into the stanine formula above, the result is a
stanine score of 9:

1. Stanine = (2 * 2) + 5
2. Stanine = 9
 Measures of Dispersion and Relative Standing  |  79

Here we need to check to determine if the result is fractional; if so, round it up

to the next number. If the value is greater than 9, you just use 9; this only occurs in a
very few cases.
In other cases, you might be working with a z score that’s not even. For example,
we’ll leave our observed value as 56 and our mean as 50, but let’s change our standard
deviation to 7; this would leave us with a z score of .857.

56 − 50
1. z =
7
6
2. z =
7

3. z = .857

We could then substitute our z score into the stanine formula:

1. Stanine = (.857 * 2) + 5
2. Stanine = 6.714

This would result in a stanine score of 6.714 which, when we removed everything
to the right of the decimal point, would result in an actual score of 6. By looking at
this example, educators can see that a student with a high stanine number is doing
better than a student with a lower stanine number and, in many instances, will use
these scores for ability grouping. For example, it is common to see children with stan-
ine scores of 1 through 3 included in a low-ability group, students with scores of 4
through 6 in an intermediate group, and students with the highest scores placed in an
advanced group.

Putting It All Together

In order to get a better feel for the relationship between each of the measures of rela-
tive standing, look at Table 3.25. You can see that we have taken the mean and stan-
dard deviation of a dataset and, using an observed value, computed the z score, the
T-score, and the stanine score for that value. As you examine this table, three things
will become obvious:

1. When an observed value is greater than the mean, the z score will be greater
than zero. That will cause the T-score to be greater than 50 and the stanine
to be greater than 5.
2. When an observed value is equal to the mean, the z score will equal zero. This
will cause the T-score to equal 50 and the stanine to equal 5.
3. When an observed value is less than the mean, the z score will be less than
zero. This will cause the T-score to be less than 50 and the stanine to be less
than 5.

This is all seen in Table 3.25.

80 | STATISTICS TRANSLATED

TABLE 3.25.  Comparing Standard Deviations, z Scores, T-Scores,

and Stanines
Observed Standard Stanine
Mean value deviation z score T-score score
20 15 5 –1 40 3
80 80 5 0 50 5
55 60 5 1 60 7

Using SPSS for T-Scores and Stanines—Not So Fast!

Up to this point, we’ve been able to verify our results using the SPSS software. Unfor-
tunately, at this point, we run into somewhat of a snag. First, SPSS will not easily
compute T-scores or stanines, but we are OK with that. A good consumer of statistics
needs only to be able to recognize and interpret those values. The second part of
our dilemma is that, while SPSS will compute a z score, it is presented in spreadsheet
format; it is easy to do but presents the results in a manner we aren’t expecting. Know-
ing that, the following SPSS spreadsheet contains the same values we used earlier to
compute percentiles and quartiles; you can see I have selected Analyze, Descriptive
Statistics, and, because it gives us more detailed output, I have chosen the Explore
option, shown in Figure 3.4.
Once Explore is selected, we can choose the variables we want to analyze and
the statistics we want to compute; in this case, in the Display field at the bottom of
the Explore box I’ve chosen to show Statistics only; in the Explore Statistics box, I’ve
selected Descriptives and Percentiles (Figure 3.5).
Pressing Continue and OK will show us the descriptive statistics, but let’s focus
on the percentiles in Figure 3.6. As I said earlier, statisticians love to argue about
mundane topics and there’s no one agreed-upon way to compute quartiles. Because

FIGURE 3.4.  Using the Explore option of the Descriptive statistics command.
 Measures of Dispersion and Relative Standing  |  81

FIGURE 3.5.  Using the Explore option to calculate descriptive statistics and percentiles.

of that, SPSS uses two different computations: weighted average and Tukey’s hinges.
Although I didn’t label it as such, in our manual calculations we used Tukey’s formula.
Because of that, you can see that the values we computed for the median, as well as the
first and second quartiles (i.e., the 25th and 75th percentiles), match exactly. Besides
this formula, what else was Tukey famous for? If you said that he coined the word
“software,” you’re right!
Finally, let’s see how SPSS computes z scores for us. Let’s start with the same
dataset but change the variable name to z score to avoid confusion; we’ll then select
Analyze and Descriptive Statistics (Figure 3.7).

Percentiles

Weighted
Average(Definition 1) Tukey's Hinges

Median and
Percentiles Median and Quartile Quartile

5 10.0000

10 10.5000

25 18.7500 20.0000

50 43.0000 43.0000

75 58.5000 58.0000

90 61.8000

95 .

FIGURE 3.6.  Output from the Frequencies command showing percentiles.

82 | STATISTICS TRANSLATED

Again, we’ll see the Descriptives screen, but this time we will do something a bit
differently. First, within the Descriptives box, we’ll check the box in the lower left that
says “Save standardized values as variables”; then we’ll select any of the statistics we
want in the Options box. In this case, we’ve asked for the mean score and the standard
deviation (Figure 3.8).
Once we select Continue and OK, SPSS will do two things. First, it will compute
and display a mean score of 39.3 and a standard deviation of 19.45964 for the variable
we called z score; this is shown in Figure 3.9.
More importantly, when we return to the Data View screen, shown in Figure 3.10,
we’ll see that SPSS has saved the actual computed z score for each value in the column
labeled Z score.
In short, it has done nothing more than put a “z” in front of whatever we named
the original variable. For example, we can see that the z score for 62 is 1.16652. This
means that 62 is 1.16652 standard deviations above the mean of 39.3. We can check
that with a bit of basic math:

62 − 39.3
1. z =
19.45964
22.7
2. z =
19.45964

3. z = 1.17

FIGURE 3.7.  Using the Descriptive statistics option.

 Measures of Dispersion and Relative Standing  |  83

FIGURE 3.8.  Using the Descriptive statistics option to compute the mean and standard
deviation.

Once again, we have shown that our calculations match those of the software, so
let’s wrap up this chapter.

Summary
In this chapter, we continued talking about Step 4, learning to understand how spread
out our dataset is, using measures of dispersion; we also learned how to determine
the relative position of a value within a dataset when compared to a given measure of
central tendency. We’ll continue with Step 4 of our six-step model by using graphical
tools to examine our data. Just like everything else we have talked about up to this
point, these tools are essential in deciding which inferential statistical tool we will use
to analyze our data.

Descriptive Statistics

N Mean Std. Deviation

z Score 10 39.3000 19.45964

Valid N (listwise) 10
FIGURE 3.9.  Output from the Descriptive statistics option showing the mean and standard
deviation.
84 | STATISTICS TRANSLATED

FIGURE 3.10.  Computed z scores created as a new column in the Data View spreadsheet.

Do You Understand These Key Words and Phrases?

computed range deviation from the mean
measures of dispersion measures of relative standing
percentile possible range
quartile range
standard deviation stanine
T-score variance
z score

Do You Understand These Formulas?

Computed range:
Largest observed value minus the smallest observed value

Percentile:
[(Below + 1/2 Same)/N] * 100

Population standard deviation:

S ( x − m )2
σ=
N
Population variance:
S ( x − m )2
σ2 =
N
 Measures of Dispersion and Relative Standing  |  85

Possible range:
Largest possible value minus the smallest possible value

Sample standard deviation:

S ( x − x )2
s=
n −1

Sample variance:
S ( x − x )2
s2 =
n −1

Stanine:
Stanine = (z * 2) + 5

T-score:
T = (z * 10) + 50

z score for a sample:

x−x
z=
s

Quiz Time!

Table 3.26 shows a dataset representing the ages of employees working for three different com-
panies. Assuming a minimum working age of 18 and mandatory retirement at 65:

TABLE 3.26.  Ages for Employees Working for Three Companies

Company A 35 25 44 62 55 22 31 41 36 65
Company B 47 37 25 24 22 31 33 35 27 36
Company C 20 40 40 35 42 47 50 52 54 60

1. What’s the possible age range of each company?

 2. What’s the possible age range of all companies combined?
 3. What’s the computed age range of each company?
 4. What’s the computed age range of all companies combined?
 5. What does the computed versus actual age range tell you about the dispersion com-
puted in Questions 1–4?

 6. What’s the standard deviation of age in each company?

86 | STATISTICS TRANSLATED

 7. What does the standard deviation tell you about the dispersion of age in each com-
pany?

 8. What’s the standard deviation of age in all companies combined?

 9. What’s the variance of age in each company?
10. What’s the variance of age when all companies are combined?
11. What does the variance of age when all companies are combined tell you about
their dispersion?
Let’s check to make sure we completely understand these important topics.

1. You can see in Table 3.27 that there are 10 sets of mean scores, observed scores,
and standard deviations. For each of these, compute the z score and T-score. Follow-
ing that, you should rank the mean values and compute the stanines.

TABLE 3.27.  Data for Computation of z Scores, T-Scores, Stanines, and

Ranked Mean Scores
Observed Standard Ranked
Mean value deviation z score T-score Stanine mean score
30 33 2.00
48 52 5.00
55 54 3.00
71 77 7.00
14 8 2.70
23 35 5.00
61 48.6 2.90
100 114 6.33
81 78.5 1.55
47 60.00 12.0

2. Intelligence tests have shown that the average IQ in the United States is 100 with a
standard deviation of 15. What z score would enable students to qualify for a pro-
gram where the minimum IQ was 130?

3. Let’s assume that average income of families in the United States is $22,000; we
also know that families whose annual income is less than 2 standard deviations
below the mean qualify for governmental assistance. If the variance of the incomes
in America is $9,000, would a family making $19,500 a year qualify?

4. We’ve found that the average number of states that citizens of the United States
have visited is 10. If we have a sample standard deviation of 3, what is the z score for
a person who has visited 14 states?, 5 states?, 20 states?
CHAPTER 4

Graphically Describing
the Dependent Variable

Introduction
As we have already said, there are two ways of describing data—numerically and
graphically. We spent the last two chapters talking about the numeric methods: mea-
sures of central tendency, measures of dispersion, and measures of relative standing.
In this chapter we will focus on basic graphical tools, many of them already familiar
to you. Let me say, right off the bat, that we are only going to look at examples of
graphical statistics that were generated using our SPSS software. While many may
disagree, I think creating the graphs “by hand” is a waste of time. I learned to do that in
a basic statistics class years ago, and I’ve not had to do it since. We have the software
to create the graphs for us; why not use it?

Graphical Descriptive Statistics

What is the old saying about one picture being worth a thousand words? You will find
how true this really is once you learn how to represent datasets graphically. Like many
other topics in statistics, you will find there are a number of ways we could graphically
describe our data; many of them are dependent on the type of data we have collected.
In this chapter, we’re going to focus on a few commonly used techniques. As men-
tioned earlier, we will use SPSS to create our graphs for us.

87
88 | STATISTICS TRANSLATED

Graphically Describing Nominal Data

Before we get started, let’s create an example of a nominal dataset. In Table 4.1 we can
see the classification for 36 college students. In the table, FR stands for freshman, SP
for sophomore, JR for junior, and SR for senior. By counting each occurrence, we can
see the table has entries for 6 freshmen, 18 sophomores, 6 juniors, and 6 seniors. If we
counted the values, we would see that the modal value is SP, as it occurs in the data
more often than the other values. Since the data are nominal, it wouldn’t make sense
to try to compute the median or mean.

TABLE 4.1.  Class Data

FR FR FR FR FR FR SP SP SP SP SP SP
SP SP SP SP SP SP SP SP SP SP SP SP
JR JR JR JR JR JR SR SR SR SR SR SR

Although a table like this one is a perfectly acceptable way to describe our data
graphically, it doesn’t concisely show what we want to see. In order to try to get a better
understanding, we could break it down further by creating a frequency table (Table
4.2) showing the count of each value as well as the percentage of the total that each
frequency represents.

TABLE 4.2.  Frequency Distribution of Class Data

Cumulative
Class Frequency Percentage percentage
FR  6 16.7 16.7
SP 18 50.0 66.7
JR  6 16.7 83.3
SR  6 16.7 100.00
Total 36 100.00

Figure 4.1 shows how we could use SPSS to verify this by entering the data into
the SPSS spreadsheet, then select Analyze, Descriptive Statistics, and Frequencies.
Selecting Frequencies brings us to the next step (Figure 4.2). By selecting Class as our
Variable, checking the box titled “Display Frequency Tables” in the lower left, and
clicking on OK, we get the output shown in Figure 4.3; as expected, it matches what
we computed manually in Table 4.2.

Pie Charts
Another graphical tool that we’ve all seen used quite often is a pie chart. It is noth-
ing more than a picture that looks like a pie, with each slice representing how often
a given value occurs in the dataset. Using the same data from the frequency table, if
you look closely at Figure 4.4, you will see options for many types of graphs. We will
look at many of these graphs in the following pages but, for now, select Graph, Legacy
Dialogs, and Pie. After selecting Pie, the dialog box shown in Figure 4.5 will appear.
We want our Slices to represent the number of cases and we have defined our cases as
Class. Entering OK results in the pie chart shown in Figure 4.6.
 Graphically Describing the Dependent Variable  |  89

FIGURE 4.1.  Using the Frequencies command.

FIGURE 4.2.  Using the Frequencies command for class (nominal) data.
90 | STATISTICS TRANSLATED

Class

Cumulative
Frequency Percent Valid Percent Percent
Valid FR 6 16.7 16.7 16.7

JR 6 16.7 16.7 33.3

SP 18 50.0 50.0 83.3

SR 6 16.7 16.7 100.0

Total 36 100.0 100.0
FIGURE 4.3.  Output from Frequencies command for class data.

Doesn’t this really make a difference? It becomes immediately clear, with just
a glance, that there are as many sophomores as there are all the other classes put
together. It is also obvious that there are an equal number of students in each of the
other three classes.
This verifies what we just saw in our frequency table. The percentage for fresh-
men, juniors, and seniors is the same (i.e., 16.7% each), and the number of sopho-
mores is the same as the other three groups combined (i.e., 50%). Obviously, the table
showing the frequency count is easy to read, but if you want to get your point across,
the picture is far more dynamic.

Bar Charts
Another easy way to present data is using a bar chart. As we saw when we looked at the
box where we indicated we wanted to create a pie chart, there were options for creat-
ing several other types of graphs. Rather than present all those input screens again,
Figure 4.7 shows us what SPSS would generate if we had selected Bar instead of Pie.
On the bar chart, the actual count for each of the variables is displayed; you can

FIGURE 4.4.  Using the Pie command to create a pie chart.

 Graphically Describing the Dependent Variable  |  91

FIGURE 4.5.  Setting up the Pie command to use class (nominal) data.

tell that by looking at the values on the left side of the graph. We can verify what we
saw on the pie chart and in the descriptive statistics; the number of freshmen, juniors,
and seniors is the same, while the number of sophomores is equal to the number in
each of the other three groups combined.
Had we wanted to, we could have created a bar chart that showed the percent-
ages each of these values represented. If we had, the values on the left side of the
chart would have ranged from 0 to 100
and the values for each group would
have represented their percentage out
of 100 rather than their actual count. Of
course, the relative sizes of the bars for
each group would have been the same.
As was the case with the pie chart, this
bar chart really makes the data “come
to life.”
There is one additional thing to
note here. Later in this chapter we’ll talk
about a histogram, and you will see that
it’s very similar to the bar chart. The dif-
ference between the two is one of scale;
the bar chart is used to present nominal
data, while the histogram will be used
FIGURE 4.6.  A pie chart created from the class when we are working with quantitative
data. data.
92 | STATISTICS TRANSLATED

Graphically Describing Quantitative

Data
Although it is possible to use pie charts,
frequency tables, and bar charts to dis-
play quantitative data, it is far more
appropriate to use other tools we’re
about to discuss. Before we do, how-
ever, let’s create a dataset we can use as
we’re creating the new graphs. In Table
4.3, we’re showing data for a company
that began paying tuition for staff devel-
opment 10 years earlier. The company
is now noticing that, as the amount of
FIGURE 4.7.  A bar chart created from the class
data. money spent each year goes up, the
number of resignations goes up as well.

TABLE 4.3.  Data Showing a Positive Relationship

between Tuition and Resignations
Year Tuition spent Resignations
 1  $32,000 20
 2  $38,000 25
 3  $41,000 28
 4  $42,000 26
 5  $50,000 25
 6  $60,000 30
 7  $57,000 40
 8  $70,000 45
 9  $80,000 45
10  $90,000 90

Management is interested in seeing if a relationship exists between these two vari-

ables. They are concerned that some employees might be using the tuition for a better
education only to prepare themselves for a move to another company where they will
make more money! Later in the book, we will use one of two numeric tools, Pearson’s
or Spearman’s correlation coefficient, to understand this better, but, for now, a scat-
terplot will give us a good feel for what is happening.

Scatterplots
A scatterplot is used to show the relationship between paired values from two datasets.
In order to create the scatterplot, we will prepare the SPSS spreadsheet shown in Fig-
ure 4.8 to include the three variables.
Creating a scatterplot takes several steps, so let’s go through it carefully to ensure
 Graphically Describing the Dependent Variable  |  93

FIGURE 4.8.  Data for year, tuition spent, and resignations in the Data View spreadsheet.

we understand each one. First, although not shown, we would select Graphs, Legacy
Dialogs, and Scatter/Dot. Then, as shown in Figure 4.9, we’ll ask for a Simple Scatter-
plot. We then can select Tuition to be on the x-axis (i.e., the horizontal axis) and Res-
ignations to be on the y-axis (i.e., the vertical axis). Notice, in Figure 4.10, that we’re
not including Year because it is not pertinent to the scatterplot itself. After clicking on
OK, SPSS creates the scatterplot shown in Figure 4.11.
The left and bottom sides of the box are labeled with the possible range of Resig-
nations and Tuition. Since our data didn’t indicate to the software what the possible
range is, SPSS created it by including a small amount above and below the actual
values. The possible range of values for Resignations goes from about 15 to 100, while
the possible range of Tuition goes from about $30,000 to $90,000.
SPSS then paired the variables, by year, and plotted them on the chart. For exam-
ple, in the first year of the program, there were 20 resignations and the company
spent $32,000 on tuition reimbursement. To plot that point on the chart, SPSS went
out to the right on the x-axis to a point equivalent to $32,000 and then up the y-axis to
a point equivalent to 20. The intersection of these two values is marked with a small
circle. This process is then repeated for each of the sets of values.
Although it is not necessary, SPSS can plot a line of best fit through the data points.
This line, shown in Figure 4.12, shows the trend of the relationship between the plotted
variables. Later in the book we will compute the actual correlation coefficient between
the two variables, but just by looking at Figure 4.12, we can see that the company
might have a reason to be concerned. It does seem that the amount of tuition money
spent and the number of resignations are related.
This plot represents a positive relationship; as one value goes up, the other goes
up. If the line of best fit was exactly 45 degrees, we would know that the relationship
between the two sets of variables is perfect in that, if we knew one value, we could
accurately predict the other. This doesn’t occur in most instances, however; we will
see lines that look nearly perfect, and, in other instances, the line will be nearly flat.
For example, using the data in Table 4.4, we could create a scatterplot and see
exactly the opposite. As one value goes up, the corresponding value tends to go down;
94 | STATISTICS TRANSLATED

FIGURE 4.9.  Selecting the Scatterplot command.

FIGURE 4.10.  Setting up the Scatterplot command to use the tuition and resignations data.
 Graphically Describing the Dependent Variable  |  95

FIGURE 4.11.  Output from Scatterplot command.

FIGURE 4.12.  Positive relationship between resignations and tuition paid.

96 | STATISTICS TRANSLATED

this is known as a negative relationship. We can use these data to create Figure 4.13.
Here, the line of best fit goes slightly down from left to right; this indicates we have a
somewhat negative relationship. Let’s change the data in Table 4.4 to that in Table 4.5.
SPSS shows a flat, but slightly positive, line of best fit in Figure 4.14.

TABLE 4.4.  Data Showing a TABLE 4.5.  Data Representing a

Negative Relationship between Moderately Negative Relationship
Resignations and Tuition between Resignations and Tuition
Tuition Resignations Tuition Resignations
$32,000 40 $32,000 20
$38,000 35 $38,000 35
 $41,000 28  $41,000 58
 $42,000 25  $42,000 25
 $50,000 20  $50,000 22
 $60,000 22  $60,000 72
 $57,000 15  $57,000  4
 $70,000 25  $70,000 15
 $80,000 10  $80,000 90
 $90,000 30  $90,000  5

FIGURE 4.13.  Negative relationship between resignations and tuition paid.

 Graphically Describing the Dependent Variable  |  97

FIGURE 4.14.  A slightly positive relationship between resignations and tuition paid.

Histograms
A histogram is really nothing more than a bar chart used to chart quantitative data.
We’ll use the data in Table 4.6, collected from our coworkers, when we asked them
how many minutes they usually take for lunch. Using these data, SPSS would produce
the histogram shown in Figure 4.15.

TABLE 4.6.  Minutes Taken for Lunch

30 35 35 40 40 40 45 45 45 45 50 50
50 50 50 55 55 55 55 55 55 60 60 60
60 60 65 65 65 65 70 70 70 75 75 80

In a histogram, the numbers across the bottom of the chart represent the actual
values in our dataset; it begins with 30 on the left and goes to 80 on the right. The
values on the left side of the chart are the number of times a given value appears in
the dataset. As you can see, the value 30 appears once in the dataset, and the bar rep-
resenting it stops at the line showing the value of 1. We could double-check ourselves
by seeing that the value 50 occurs 5 times in both the table and the bar chart. We’ll
use the histogram extensively in the next chapter to discuss the normal distribution, but
for now, let me give you a word of warning about graphical statistics.

Don’t Let a Picture Tell You the Wrong Story!

Like we said, a picture tells a thousand words; unfortunately, it’s also been said that
there are three kinds of lies—lies, damn lies, and statistics. Given that, it probably
98 | STATISTICS TRANSLATED

FIGURE 4.15.  Output from the Histogram command.

won’t surprise you that statistics have been used, in many cases, to try to mislead, mis-
inform, or manipulate an unsuspecting victim. This seems to be especially true when
graphical statistics are used.
For example, let’s imagine a local environmental group denouncing the effect
of humans on the natural habitat of brown bears in the Rocky Mountains. As an
example, they compare the life expectancy of the brown bears to their cousins above
the Arctic Circle. They first explain that the average life expectancy of all bears ranges
from 0 to 30 but adamantly believe that the following graph clearly shows that brown
bears have a far shorter life expectancy than polar bears. Based on this, they demand
the government spend more money on environmental conservation; their results are
shown in Figure 4.16.
Although the environmentalists certainly seem to have a case, do you notice any-
thing suspicious about this graph? If not, look on the left side of the chart. As you can
see, the possible values range from 23 to 27. This means that the bars showing average
life expectancy are not actually proportionate to the possible range of 0 to 30; instead
they are proportionate to a much shorter possible range. This makes the difference
between the average life expectancy of brown bears and the average life expectancy of
polar bears look much more dramatic than it really is. If we were to plot the average
life expectancy using the true possible range, the chart would look like that shown in
Figure 4.17.
That makes quite a bit of difference, doesn’t it? While the first graph might sup-
port the environmentalists’ concerns, the second graph shows there is really very little
difference in the average life span of brown and polar bears. This is not to say that we
should not worry about a shorter life expectancy of brown bears, but we should worry
 Graphically Describing the Dependent Variable  |  99

about graphs like this where the scale

has been changed; it happens more
often than you would believe!
Let’s look at another example. In
this case, United States Olympics offi-
cials are desperately begging for more
money. After all, they claim, their ath-
letes won the fewest medals in the recent
summer games; this is shown in Figure
4.18. In this case, it appears that the
Olympic Committee may have a valid
concern; the United States does have
the smallest “piece of the pie.” Unfor-
FIGURE 4.16.  Histogram showing the relation- tunately, however, the committee seems
ship between ages of bears. to be trying to pull the wool over our
eyes. By looking at the list of “countries”
shown in the upper-right corner we can see that, instead of being compared to other
individual countries, the United States is being compared to each continent and the
multiple countries they represent; it’s no wonder it looks like the United States is
lagging behind! Again, while this might sound like an outlandish example (i.e., our
Olympic Committee would never do this), I’ve seen graphs far more misleading than
this being sold as the truth. Always keep a careful eye out for misleading statistics;
they’re everywhere. Being able to identify them and understand how they are being
misused will not only make you a hero among your friends, but it will also help you
become a better consumer of statistics.

Summary of Graphical Descriptive Statistics

To this point, we have covered the very basic graphical tools for describing numeric
data; we are used to seeing pie charts, bar charts, histograms, and scatterplots. News
reports, financial documents, and adver-
tisements, for example, are quick to use
these tools to make sure their audi-
ence “gets the point.” We are going to
move forward now and use a histogram
to demonstrate the idea of the normal
distribution. As I’ve said about so many
things, this will really help us as we move
forward with the idea of inferential deci-
sion making.

The Normal Distribution

As you’ve seen, there are quite a few tools
FIGURE 4.17.  Adjusted histogram showing the for graphically describing numeric data.
relationship between ages of bears. While each of the topics we’ve covered is
100 | STATISTICS TRANSLATED

important, at this point, we need to start

tying everything together. Let’s start by
discussing the relationship between a
histogram and a normal distribution.

The Histogram and the Normal

Distribution
In many instances, if we plot data values
from a relatively large sample or popula-
tion, the data form a “mound” or “bell”-
shaped distribution clustered about the
center of the dataset. For example, we
saw that earlier with the data describing
the number of minutes employees take
FIGURE 4.18.  Olympic medals won.
for lunch. While all data distributions
are not shaped accordingly, when they
are, we call it a normal distribution. In
datasets that are normally distributed, the mean, median, and mode are equal, and
the remaining values are symmetrically distributed around the mean. When that hap-
pens, we wind up with the distribution we saw earlier; it’s shown here again in Figure
4.19.
We can demonstrate the idea of a normal distribution in Figure 4.20 by tracing
a line over the tops of each of the bars. When we do, we can see that our data distri-
bution is, in fact, bell-shaped. In this case, we have a normal distribution. The mean,

FIGURE 4.19.  Histogram of a normal distribution.

 Graphically Describing the Dependent Variable  |  101

FIGURE 4.20.  Histogram showing the bell-shaped distribution.

median, and mode are all equal, and the other values are distributed symmetrically
around the center. It would be nice to have a normal distribution every time we collect
data; that would allow us to use a discrete set of statistical procedures, called paramet-
ric statistics, to analyze our data. You may already be familiar with the names of many
of these statistical tests: the t-test and the analysis of variance (ANOVA), for example.
Although using these parametric tools will be the primary focus of the latter part of
this book, we must keep two things in mind.
First, remember that histograms are used to plot quantitative data. As we talked
about earlier, when we’re working with nominal or ordinal-level data, we’re forced to
use a less powerful set of tools called nonparametric statistics. Like the parametric sta-
tistics, you may already be familiar with some of them; the chi-square test, for example,
is the most widely recognized.
Second, the use of parametric statistics is based on quantitative data that are
normally distributed. In most cases the distribution doesn’t have to be perfectly bell-
shaped since the inferential statistics are powerful enough to allow for some latitude
within the data. In certain instances, however, a researcher might find a set of quan-
titative data where the distribution is so distorted that parametric statistics will not
work. Right now, we will not get into the underlying reasons, so let’s just leave it at
this. In this book, when we have nominal or ordinal data, we will use nonparametric
statistics; if we have quantitative data, we will use parametric statistics.

Things That Can Affect the Shape of a Distribution of Quantitative Data

Although it would be nice, a perfect normal distribution rarely happens. Think about
it—if we’re looking at the heights of students in an elementary class, we could compute
102 | STATISTICS TRANSLATED

the average. How often, however, would the heights less than average and the heights
greater than average be exactly equally distributed? You might have a few more short-
er students than taller students or vice versa; either way it would affect how the distri-
bution would look. This type of distortion is very common and very rarely leads to hav-
ing a data distribution where both sides of the curve are perfectly symmetrical. This is
not a big concern in most instances because, like I said, the inferential tests we will use
can compensate for problems with the shape of a distribution up to a certain degree.
Knowing that, let’s look at a few things that can affect the way a distribution looks.

Skewness
Skewness means that our data distribution is “stretched out” to one side or the other
more than we would expect if we had a normal distribution. When we have more val-
ues greater than the mean, we say the distribution is positively skewed or skewed to the
right. When we have more values than expected less than the mean, we say that it is
negatively skewed or skewed to the left. Let’s look at each.

POSITIVE SKEWNESS

Let’s add these values—90, 90, 90, 100, 100, and 105—to the lunch-time data we have
been using. This means we have more data points on the high end of the scale than we
do on the lower end. In Figure 4.21, you can see that this causes the normal curve to
be far more stretched out on the right side than it is on the left side. When this occurs,
we say the dataset is positively skewed or skewed to the right.
In a dataset that is positively skewed, the mean score will be larger than the medi-

FIGURE 4.21.  Positive skewness.

 Graphically Describing the Dependent Variable  |  103

an score. Since we already know how to use SPSS to create graphs, let’s just look at
the SPSS printout shown in Figure 4.22; in it you can see that the mean is 60.83 while
the median is 57.5. Larger degrees of skewness would cause this difference to become
even greater. Besides just looking at the data distribution and the difference between
the mean and median, we also have statistical formulas to help us to determine the
degree of skewness. For example, the output below includes a field labeled “skewness.”
Its value, either positive or negative, indicates how skewed the data distribution is; the
further away from zero it is, the greater the skewness is in one direction or the other.
In this case, we have a skewness statistic of .734. Since this value is greater than zero, it
verifies that the dataset is positively skewed. This only becomes problematic, however,
if the value is greater than +2. Anything larger would tell us that the distribution is so
skewed that it will interfere with our ability to use parametric inferential statistics. In
those cases, we would have to use a nonparametric alternative.
Before we move on, let me point out a couple of other things in Figure 4.22 you
might have already noticed about the SPSS output. First, for some reason, it doesn’t
include the modal value. This is not usually an issue as we can easily determine it by
looking at the graphical data distribution, and there are also other commands in SPSS
that will compute the mode for you. Second, the printout includes many descriptive
statistics that we haven’t discussed up to this point. A lot of these statistics, such as the
Minimum and Maximum, are common sense; others will be covered later in the text.

NEGATIVE SKEWNESS

To demonstrate negative skewness, let’s add these values to the original lunch-time
dataset: 5, 10, 10, 20, 20, and 20. By doing that, we can see in Figure 4.23 that the dis-
tribution is more spread out on the left side than it is on the right.
In Figure 4.24, you can see the mean, 49.17, is less than the median, 52.50. As you

Statistic Std. Error

Scores Mean 60.8333 2.85599

95% Confidence Interval for Lower Bound 55.0655

Mean Upper Bound 66.6011
5% Trimmed Mean 60.0926
Median 57.5000

Variance 342.581

Std. Deviation 18.50895

Minimum 30.00
Maximum 105.00
Range 75.00

Interquartile Range 21.25

Skewness .734 .365

Kurtosis .041 .717

FIGURE 4.22.  Descriptive statistics showing positive skewness.

104 | STATISTICS TRANSLATED

FIGURE 4.23.  Negative skewness.

might have expected, the relationship between the mean and the median is opposite
that of the positively skewed distribution. The more negatively skewed the distribution
is, the smaller the mean value will be when compared to the median value.
The negative skewness in the diagram is supported by a skewness index of –.734
which, although it is negative, is not less than –2 (i.e., it is between –2 and 0). As was
the case with a positive skewness value less than +2, this means the normality of the

Statistic Std. Error

Scores Mean 49.1667 2.85599

95% Confidence Interval Lower Bound 43.3989

for Mean Upper Bound 54.9345
5% Trimmed Mean 49.9074
Median 52.5000

Variance 342.581

Std. Deviation 18.50895

Minimum 5.00
Maximum 80.00
Range 75.00

Interquartile Range 21.25

Skewness -.734 .365

Kurtosis .041 .717

FIGURE 4.24.  Descriptive statistics showing negative skewness.

 Graphically Describing the Dependent Variable  |  105

data distribution is not significantly affected; while this distribution is not perfectly
normal, it is still close enough to use parametric statistics.

Kurtosis
Kurtosis refers to problems with the data distribution that cause it to look either more
peaked or more spread out than we would expect with a normal distribution. Platykur-
tosis occurs when we have more values than expected in both ends of the distribution,
causing the frequency diagram to “flatten” out. Leptokurtosis happens when we have
more values in the center of the dataset than we might expect, causing the distribution
to appear more peaked than it would in a normal distribution.

PL AT YKURTOSIS

Once again, let’s modify the original dataset we used. This time, however, instead of
adding values to either end of it, let’s remove one each of the values 50, 55, and 60 in
order to help us understand platykurtosis. If we look at our descriptive statistics in
Figure 4.25, we can see that the mean and median both equal 55, just as we saw in the
normal distribution in Figure 4.19. In this case, however, the data values are not sym-
metrical around the mean. Instead we can see in Figure 4.26 that the normal curve is
plotted slightly above the higher value. That means our dataset is flatter than would
be expected if the distribution was perfectly bell-shaped.
As was the case with skewness, SPSS will calculate an index of kurtosis which, as
seen in Figure 4.26, is –.719. This value, since it is negative, verifies what we saw in the
graph. Much like the skewness numbers, however, the kurtosis number would have to
be less than –2.00 to significantly affect the normality of the dataset.

Descriptives

Statistic Std. Error

Scores Mean 55.0000 2.21906

95% Confidence Interval for Lower Bound 50.4799

Mean Upper Bound 59.5201
5% Trimmed Mean 55.0000
Median 55.0000

Variance 162.500

Std. Deviation 12.74755

Minimum 30.00
Maximum 80.00
Range 50.00

Interquartile Range 20.00

Skewness .000 .409

Kurtosis -.719 .798

FIGURE 4.25.  Platykurtosis descriptive statistics.

106 | STATISTICS TRANSLATED

FIGURE 4.26.  Platykurtosis.

LEPTOKURTOSIS

In other instances, there may be more values around the center of the distribution
than would be expected in a perfectly normal distribution. This causes the distribu-
tion to be taller in the center than we would expect; this condition is called leptokurto-
sis. In Figure 4.27, we again have the mean, median, and mode equal to 55 (remember,
we have to look at the actual diagram in Figure 4.28 to determine the mode), but
the distribution is not normal; you can see that we have added several occurrences
of the value 55 as well as taken away some of the lower and higher values. As odd as

Std.
Statistic Error

Scores Mean 55.0000 1.78174

95% Confidence Lower Bound 51.3829

Interval for Mean Upper Bound 58.6171
5% Trimmed Mean 55.0000
Median 55.0000

Variance 114.286

Std. Deviation 10.69045

Minimum 30.00
Maximum 80.00
Range 50.00

Interquartile Range 10.00

Skewness .000 .393

Kurtosis .334 .768

FIGURE 4.27.  Leptokurtosis descriptive statistics.

 Graphically Describing the Dependent Variable  |  107

FIGURE 4.28.  Leptokurtosis.

it looks, however, the index of kurtosis for this distribution is only .334, which is still
well within the range allowed by parametric statistics; these descriptive statistics are
shown in Figure 4.27.

The Empirical Rule

Does the histogram in Figure 4.29 look familiar? It should; we have already used it to
demonstrate a normal curve. This time, however, I’ve added some of the computed
inferential statistics to the side of the chart. We can see that the mean is 55 and we
have computed a standard deviation of 12.25. We can use these two values, along with
the empirical rule, to better understand the dispersion of our data.
The first thing the empirical rule tells us is, if the distribution is approximately
normally distributed, at least 68% of the values in the dataset fall within ±1 standard
deviation of the mean. Using the descriptive statistics on the right side of Figure 4.29
means we know that just over two-thirds of the scores fall somewhere between 42.75
and 67.25. The rule goes on to say that about 95% of the scores will fall within ±2
standard deviations of the mean, 30.5 to 79.5, and nearly all the scores (99.7%) will
fall between ±3 standard deviations of the mean, 18.25 to 91.75. This also tells us that
anything falling outside of these boundaries is an anomaly; values less than 18.25 or
greater than 91.75 occur less than 30% of the time. We will go into more detail about
this later, but for now let’s look at an example of what we have just described.
You may not know it, but the average height for a male in the United States is 5
feet 9 inches (69 inches); the standard deviation for this distribution is 3 inches. This
means that about 68% of men are somewhere between 66 and 72 inches tall (i.e., 69 ±1
standard deviation), about 95% of men are somewhere between 63 and 75 inches tall
(i.e., 69 ±2 standard deviations), and nearly all men are somewhere between 60 and 78
inches tall (i.e., 69 ±3 standard deviations). The remaining 3% of men are either less
108 | STATISTICS TRANSLATED

FIGURE 4.29.  Normal distribution.

than 60 inches or taller than 78 inches. Knowing that explains why we tend to really
notice men that are very, very short or very, very tall; there are not too many of them.

LE T’S TIGHTEN THIS UP A LIT TLE

You probably noticed that, in the last section, I kept saying “about” while referring
to a certain percentage from the mean. I did that to get you used to the concept of
the empirical rule; now let’s look at the exact percentages. What the empirical rule
actually says is that, in a mound-shaped distribution, at least 68.26% of all values fall
within ±1 standard deviation from the mean, at least 95.44% of all values fall between
the mean ±2 standard deviations and nearly all (99.74%) values fall between the mean
and ±3 standard deviations. Using our example of height, this now means only .26%
of all men fall outside the mean ±3 standard deviations; they’re rarer than we thought!
Let’s look at Figure 4.30. There you can see that I used SD for standard deviation
instead of a symbol. I did this because the empirical rule applies to both mound-
shaped samples and populations, so it would not be appropriate to use either of the
symbols discussed earlier.
It makes sense, by looking at this figure, that we can see percentage differences
between the mean and any value above or below the mean. Although it is not as clear
conceptually, we can also look at percentages between any two values in the distribu-
tion, be they both above the mean, both below the mean, or one on either side of
the mean. For example, we know that 68.26% of all values fall within ±1 standard
deviation of the mean. It naturally follows that half of these values, 34.13%, would fall
between the mean and +1 standard deviation, and half of the values, 34.13%, would
fall between the mean and –1 standard deviation. You can see that in Figure 4.31.
If we wanted to determine the difference between the mean and –3 standard devi-
 Graphically Describing the Dependent Variable  |  109

ations, we would go through the same

process. As shown in Figure 4.32, we
know that the region between ±3 stan-
dard deviations is 99.74%; if we divide
that in half, we know that 49.87% of the
values lie in that range.
We can also measure areas under
the curve that include values above and
below the mean. For example, what per-
centage of values lies between 1 stan-
dard deviation below the mean and 3
standard deviations above the mean?
All we would have to do is add the per-
centage of values between the mean and
FIGURE 4.30.  The empirical rule. –1 standard deviation, 34.13%, to the
percentage of all values that lie between
the mean and +3 standard deviations,
49.87%. We would wind up with 84% of the values; this can be seen in Figure 4.33.
Understanding the relationship between standard deviations and the percentage
of values they represent allows us to tie together something we learned earlier. For
example, if we have a z score of +1, we know this represents one standard deviation
above the mean. Given that we also know the percentage of values between the mean
and a z score of +1 is 34.13%, the same would hold true if we had a z score of –1. If we
had a z score of +2, it would represent 47.72% of the values between it, and the mean
and a z score of –3 would include 49.87% of the values between it and the mean.
This works well if you have a whole number of standard deviations, but what hap-
pens if you have a fractional value? We can deal with that situation using Table 4.7, the
area under the normal curve table or, as it is sometimes called, a critical values of z table;
this table is also shown in Appendix A.
Although the table may look confusing, don’t be deceived; it is actually very easy
to use. Look at the leftmost column. These are the first two digits of any z score (i.e.,
the whole number and the first decimal position). In order to find the percentage of

FIGURE 4.31.  Mean ±1 standard deviation. FIGURE 4.32.  Mean ± 3 standard deviations.
110 | STATISTICS TRANSLATED

TABLE 4.7.  Area under the Normal Curve (Critical Value of z)

z score 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.0 .0000 .0040 .0080 .0120 .0160 .0199 .0239 .0279 .0319 .0359
0.1 .0398 .0438 .0478 .0517 .0557 .0596 .0636 .0675 .0714 .0753
0.2 .0793 .0832 .0871 .0910 .0948 .0987 .1026 .1075 .1103 .1141
0.3 .1179 .1217 .1255 .1293 .1331 .1368 .1406 .1443 .1480 .1517
0.4 .1554 .1591 .1628 .1664 .1700 .1736 .1772 .1808 .1844 .1879
0.5 .1915 .1950 .1985 .2019 .2054 .2088 .2123 .2157 .2190 .2224
0.6 .2257 .2291 .2324 .2357 .2389 .2422 .2454 .2486 .2517 .2549
0.7 .2580 .2611 .2642 .2673 .2704 .2734 .2764 .2794 .2823 .2852
0.8 .2881 .2910 .2939 .2967 .2995 .3023 .3051 .3078 .3106 .3133
0.9 .3159 .3186 .3212 .3238 .3264 .3289 .3315 .3340 .3365 .3389
1.0 .3413 .3438 .3461 .3485 .3508 .3531 .3554 .3577 .3599 .3621
1.1 .3643 .3665 .3686 .3708 .3729 .3745 .3770 .3790 .3810 .3830
1.2 .3849 .3869 .3888 .3907 .3925 .3944 .3962 .3980 .3997 .4015
1.3 .4032 .4049 .4066 .4082 .4099 .4115 .4131 .4147 .4162 .4177
1.4 .4192 .4207 .4222 .4236 .4261 .4265 .4279 .4291 .4306 .4319
1.5 .4332 .4345 .4357 .4370 .4382 .4394 .4406 .4418 .4429 .4441
1.6 .4452 .4463 .4474 .4484 .4495 .4505 .4515 .4525 .4535 .4545
1.7 .4554 .4564 .4573 .4582 .4591 .4599 .4608 .4616 .4625 .4633
1.8 .4641 .4649 .4645 .4664 .4671 .4678 .4686 .4693 .4699 .4706
1.9 .4713 .4719 .4726 .4732 .4738 .4744 .4750 .4756 .4761 .4767
2.0 .4772 .4778 .4783 .4788 .4793 .4798 .4803 .4808 .4812 .4817
2.1 .4821 .4826 .4830 .4834 .4838 .4842 .4846 .4850 .4854 .4857
2.2 .4861 .4864 .4868 .4871 .4875 .4887 .4881 .4884 .4887 .4890
2.3 .4893 .4896 .4998 .4901 .4904 .4906 .4909 .4911 .4913 .4916
2.4 .4918 .4920 .4922 .4925 .4927 .4929 .4931 .4932 .4934 .4936
2.5 .4938 .4940 .4941 .4943 .4945 .4946 .4948 .4949 .4951 .4952
2.6 .4953 .4955 .4956 .4957 .4959 .4960 .4961 .4962 .4963 .4964
2.7 .4965 .4966 .4967 .4968 .4969 .4970 .4971 .4972 .4973 .4974
2.8 .4974 .4975 .4976 .4977 .4977 .4978 .4979 .4979 .4980 .4981
2.9 .4981 .4982 .4982 .4983 .4985 .4984 .4985 .4985 .4986 .4986
3.0 .4987 .4987 .4987 .4988 .4988 .4989 .4989 .4989 .4990 .4990

values that lie between the mean and a given z score, go down this column until you
find the row where the value corresponds to your z score. For example, if we wanted to
verify that the percentage for a z score of 1 is 34.13%, go down that column until you
get to the row containing 1.0. Immediately to the right of that, you will see the value
.3413; this is the decimal notation for 34.13%. All you must do is move the decimal
point in the table two places to the right to get the percentage. In this case, you can
see we are in the column where the heading is 0.00; this means we have a z score of
 Graphically Describing the Dependent Variable  |  111

exactly 1.00. If we had a z score of 1.04,

we would have to go to the right, in the
same row, until we get into the column
under the header .04. In that case, the
percentage would be .3508 or 35.08%.

Decimal Values versus Percentages

Although most z tables are shown in
decimal format, many people find it
easier to work with percentages instead
of decimals, so that is how we will do
it throughout the book. Again, all you
FIGURE 4.33.  The area between –1 standard must do to change the decimal values to
deviation and +3 standard deviations. percentages is move the decimal point
two spaces to the right. You can also
multiply the decimal value by 100 and get the same result.

Determining the Area under a Curve between Two Values

Now, let’s work more with the table to make sure we have the idea. What is the area
under the curve between the mean and a z score of 2.44? Go down the leftmost col-
umn until you get to the row labeled 2.4 and then to the right until you are in the
column under .04. Go to the right across the row until you see the value is .4927 or
49.27%. What about the area between the mean and a z score of .03? Go over the row
labeled 0.00 until you get to the column labeled .03. The value for the area under the
curve is .0120 or 1.2%.
Finally, how about a z score of –1.79? Although the z score is negative, the table
works the same. In this case, we would go down the leftmost column until we get to
1.7 and then across that row until we come to the column under .09; we can see the
value there is .4633 or 46.33%.
In certain cases, we find it necessary to determine the percentage of data values
that fall between two given z scores; we call this “determining the area under a curve.”
The general procedure is to use our table to determine the percentage for two z scores
you are interested in and then, by addition or subtraction, determine the percentage
of values between them. A couple of examples make this clear.
Let’s use z scores of 1.10 and 2.21. We can see the percentage value for a z score
of 1.1 is 36.43% and the z score of 2.21 is 48.64%. Since both values are positive, we
subtract the smaller value from the larger value and find a difference of 12.21%. In
other words, 12.21% of the values under the curve lie between a z score of 1.1 and a z
score of 2.21. This is shown in Figure 4.34.
Now, let’s assume we have a z score of –2 and a z score of +2.2, what is the percent-
age between them? Using the table, we see the value for 2.2 is 48.61% and the value
for 2 is 47.72%. Unlike the prior example, here we have one positive value and one
negative value. In instances such as this you must add the two percentages together.
In this case, when we add the two values together, we find that 96.33% of the values
 112 | STATISTICS TRANSLATED

FIGURE 4.34.  The area between z scores of 1.10 FIGURE 4.35.  The area between z scores of –2
and 2.21. and +2.2.

under the curve are contained in the range from a z score of –2 and a z score of +2.2.
This is shown in Figure 4.35.
Finally, let’s look at a case where both z values are negative; let’s use –1 (34.13%)
and –3 (49.87%). In this case, since both values are on the same side of the mean, we
must again subtract the smaller number from the larger. We can plot them on our
graph, shown in Figure 4.36, and we find that 15.74% of values fall between our two
z scores.

Summary of the Normal Distribution

Understanding the normal distribution is the key to inferential testing since many of
the classic statistical tests assume that the data are normally distributed. As we saw,
quantitative data that result in a perfect normal distribution have the same mean,
median, and mode and the data values are symmetrically distributed on both sides of
these measures of central tendency.
In instances where we have more values greater than the mean than expected,
we see that the distribution is positively skewed, whereas datasets with more scores
below the mean than expected are
negatively skewed. In other instanc-
es, our data may be skewed on both
ends, resulting in a normal distribu-
tion that is flatter than expected; we
say this distribution is platykurtotic.
Distributions that have more values
in the center of the dataset than ex-
pected are called leptokurtotic. All
these aberrations can be controlled
for by the robust nature of many
FIGURE 4.36.  The area between z scores of –1 of the parametric statistical tests.
and –3.
 Graphically Describing the Dependent Variable  |  113

Sometimes, however, the distribution is so skewed or kurtotic that nonparametric

tests must be used when evaluating our hypotheses. Keep in mind, however, that this
is the exception and not the rule.
Finally, we used z scores to help determine the distance between two data points
in the normal distribution and tied everything together by studying the empirical rule.
All of this together, especially the understanding of distances under the curve, will be
critical when we begin testing hypotheses based on data we have collected.

Do You Understand These Key Words and Phrases?

area under the normal curve table bar chart
empirical rule frequency distribution table
graphical descriptive statistics histogram
kurtosis leptokurtosis
line of best fit negatively skewed
normal distribution pie chart
platykurtosis positively skewed
scatterplot skewed left
skewed right skewness

Quiz Time!

Since we said that we always use a computer to help us when we are graphically describing our
dependent variable, we’re in somewhat of a jam. Knowing we can’t do that, let’s use the follow-
ing questions to understand conceptually where we are. When you are finished, check your
answers at the end of the book.

1. If I wanted a quick and easy way to present the frequency of values in a dataset
containing nominal data, which graphical tool(s) would I use? Why?

 2. Using math data on the y-axis and the reading data on the x-axis, as shown in Table
4.8, what trend would the line of best fit show us if we plotted the data on a scatter-
plot?

TABLE 4.8.  Math and Reading Scores

Math score 80 60 75 50 77 45 100
Reading score 20 100 75 90 59 65 95

 3. If we plotted the salary data in Table 4.9 on the y-axis and years of education on the
x-axis, what would our plot look like?
114 | STATISTICS TRANSLATED

TABLE 4.9.  Years of Education and Salary

Education 10 12 14 14 16 16 18
Salary in thousands 20 30 40 37 59 55 70

 4. Explain the difference between histogram and measures of tendency. Can this be
used with nominal data? Why? Quantitative data? Why?

 5. What are the requirements for a dataset to be called a perfect normal distribution?

 6. If a dataset had a mean of 45 and a median of 40, would it be skewed? Why?
 7. If we plot a dataset with a mean of 60 and a median of 55, would it be either
platykurtotic or leptokurtotic? Why?

 8. Does a dataset with an equal mean and median always mean the data are normally
distributed? Why?

 9. Do we always have to use nonparametric statistics when we’re working with datasets
that are skewed or kurtotic?

10. Discuss the relationship between the mean, median, and skewness. How would we
know if it is problematic for inferential decision making?

11. On each row in Table 4.10 there are two z values. Using these, determine the per-
centage of the curve for each and then calculate the difference between the two.

TABLE 4.10.  Values for Quiz Time Question 11

Area under the Area under the
z value 1 curve for z 1 z value 2 curve for z 2 Difference
 3.01  2.00
–2.50  3.00
–1.99 –0.09
 1.50 –1.50
CHAPTER 5

Choosing the Right Statistical Test

Introduction
We are to the point where we will choose the statistical test to investigate our hy-
pothesis. Before we can, though, we must pass through a roadblock. In order to
move forward, we need to go into a lot more detail about hypothesis testing. First, we
will discuss the overall concept of inferential decision making and learn to manually
compute a few of the basic statistical tests necessary for hypothesis testing; we will
also use SPSS to verify our results. Once we have done that, we will be ready to move
forward, select the test we need, and go on to Step 6.

The Very Basics

As we know, inferential statistics are used to help us make decisions based on data we
have collected. Before we move forward, however, let’s think about the word “inferen-
tial” for a minute. What does it really mean? First, it is easy to see that it is based on the
word “infer”; that means we are going to decide based on evidence we collect. Since, in
statistical decision making, we are generally using a sample of data from a population,
this means we are drawing a conclusion based on limited evidence. As we saw earlier,
that also means there is always a risk of making an incorrect decision based on chance
or random error. What does that mean to us? Simple: inferring something is a long
way from proving something! As we said way back in our words of wisdom, we never
try to prove anything with inferential decision making.
Therefore, when we use sample data to make inferences about a population, we
want to answer one or both of the following questions:

115
116 | STATISTICS TRANSLATED

1. Can we estimate population parameters based on sample data? For example,

if we know a sample statistic such as the mean, can we estimate the mean of
the population?
2. Can we use sample data to test a belief about the population the sample was
drawn from? For example, if we hypothesize that a population value, such as
the mean, is equal to a given value; can we use the sample data to support
that assertion?

In order to start answering these questions, let’s start by talking about the central
limit theorem and the standard error of the mean.

The Central Limit Theorem

A theorem is nothing more than a statement or an idea that has been demonstrated or
proven to be true. For example, we use the Pythagorean theorem in geometry all the
time; we use it to compute the length of an unknown side of a right triangle if we know
the length of the other two sides. The theorem has been computed and verified for
hundreds of years, so it is apparent that Pythagoras knew what he was talking about!
In developing the central limit theorem, French mathematician Pierre-Simon
Laplace proved that if you take samples from a population, compute the mean of each
sample, and then plot those means on a histogram, the distribution will be approxi-
mately normal. In fact, if you plot enough samples, the distribution will be normally
distributed, and the mean of the distribution will be equal to the mean of the popu-
lation from which the samples were drawn. I am sure we can all appreciate the work
Pierre put into this, but what does it mean? Let’s use an example to help understand it.
Suppose you are teaching a section of Psychology 101 and give a final exam to all
50 of your students. Based on the results of the exam, you are interested in estimating
the mean score of the population of all 1,000 students at your university who are tak-
ing the same course. While you are thinking about how to do it, you suddenly realize
four things:

1. If you knew the final exam scores for the population of all students in your
university, you would not have a problem. You could easily add up the scores
for all 1,000 students in the population, divide that by 1,000, and wind up
with the average score for the population.
2. Your class of 50 students is only a sample of the population of 1,000 students,
and your class’s average score may or may not be exactly equal to the mean
score of the population.
3. Based on what we already know, the mean scores from the other samples (i.e.,
classes of 50 students) are going to be exactly equal to the population mean
or, much more likely, they are going to fluctuate around the mean.
4. The more samples you select and plot, the closer the overall mean of the
samples will be to the mean of the population. If you take an infinite number
of samples, the overall mean of the samples will be exactly equal to the popu-
lation mean.
 Choosing the Right Statistical Test  |  117

In other words, if we are going to attempt to make inferences about a population

parameter based on a sample statistic, it becomes readily apparent that understand-
ing the fluctuation of the values is the key to making such decisions. The central limit
theorem helps us to do exactly that.

The Sampling Distribution of the Means

In order to demonstrate this idea, imagine that we could take an infinite number
of random samples from the population of Psychology 101 students. Each of these
samples would include 50 students, the same as the number in your class. If we calcu-
lated the mean of each of these samples and plotted them on a histogram, something
interesting would occur; the greater the number of mean scores we plotted, the more
the plot would start looking like a normal distribution.
For example, suppose you use the values in Table 5.1 to represent the mean scores
from your first 10 samples of 50 students. Since we have already seen how to compute
graphical descriptive statistics, I have gone ahead and plotted the following data; when
you look at the histogram, you can see that the distribution looks very spread out.
First, the curve that is superimposed on it is very flat, and, second, although it is not
shown, the standard deviation of 4.49 is fairly large given the narrow range of 14. As
we saw earlier, this also means the data are skewed on both ends of the distribution.

TABLE 5.1.  Data for 10 Samples

85 87 87 92 92
92 92 95 97 99

As we can see in Figure 5.1, because the distribution is flatter than a normal
distribution, it is also platykurtotic. Let’s continue by using the data in Table 5.2 to
represent the mean scores for 24 samples.

FIGURE 5.1.  Histogram for 10 samples.

118 | STATISTICS TRANSLATED

TABLE 5.2.  Data for 24 Samples

86 86 87 87 89 90
91 91 91 92 92 92
92 92 92 93 93 93
94 95 97 97 98 98

In Figure 5.2, by using a larger number of mean scores, the distribution is starting
to look different. The curve is more bell-shaped, and the standard deviation (3.46) is
getting smaller. Since the standard deviation is a measure of dispersion, the data dis-
tribution is less spread out, less skewed, and less platykurtotic.
Now, let’s use the set of 49 sample means shown in Table 5.3, plot them, and see
what happens. This is really starting to make a difference isn’t it? We can see that the
curve is becoming very bell-shaped and the standard deviation, in this case, 2.86, is
getting smaller (Figure 5.3).

TABLE 5.3.  Data for 49 Samples

86 87 87 88 88 88 89
89 89 89 90 90 90 90
90 91 91 91 91 91 91
92 92 92 92 92 92 92
93 93 93 93 93 93 94
94 94 94 94 95 95 95
95 96 96 96 97 97 98

We could keep adding data values and plotting them on our distribution but, suf-
fice it to say, the process we are following is shown in Figure 5.4.
By looking at the distribution of sample means in Figure 5.4, you can see it looks

FIGURE 5.2.  Histogram for 24 samples.

 Choosing the Right Statistical Test  |  119

FIGURE 5.3.  Histogram for 49 samples.

like a normal distribution. It is, but unlike the normal distributions we have seen so
far, this one has three special qualities.

1. The distribution has a special name. Since we are plotting the mean scores
from repeated samples, it is called the sampling distribution of the means.
2. The mean of the sampling distribution of the means is called the mean of means
and is shown as m x . The more samples you plot on a histogram, the closer
the mean of means gets to the mean of the population from which we take
the samples.
3. When we compute a standard deviation for our sampling distribution of the
means, we must keep in mind the error inherent in the sampling process.
Because of that, instead of calling it the standard deviation, we now call it the
standard error of the mean or SEM or σ x .

FIGURE 5.4.  Creating a distribution of sample means.

120 | STATISTICS TRANSLATED

Figure 5.5 gives us the overall feel for the central limit theorem.

Three Important Notes

At this point, three things must be stressed: sample size, the number of samples, and
the shape of the population distribution.
First, in the definition we said that a sample size needs to be sufficiently large.
While there is some disagreement as to what “sufficiently large” means, it is generally
agreed that a sample size of 30 is large enough.
Second, I used the word “repeated” to describe the number of samples taken from
the population. As we said earlier, measures from samples, such as the mean, cluster
around the same measure of the population. They create a bell-shaped distribution
with a mean equal to that of the population, only if we have an adequate number of
samples. We demonstrated that with our three examples:

1. When we plotted only 10 sample means, we saw a flat, spread out distribu-
tion.
2. When we plotted 24 sample means, we saw a distribution that looked more
like a normal distribution.
3. When we plotted 49 sample means, we had a nice, bell-shaped distribution
with most scores clustering around the mean of the population. Many of the
scores were exactly equal to the mean, but most were a little higher or a little
lower; this we know is due to sampling error. Had we continued plotting
mean scores, eventually we would have a perfectly normal distribution with a
mean of means exactly equal to the population mean.

Third, although it was implied that the distribution of scores for the entire popu-
lation of PSY101 students was normally distributed, it does not have to be a normal
distribution for the central limit theorem to apply. If enough samples are taken and

FIGURE 5.5.  Comparison of the mean between a population and a sampling distribution of
means.
 Choosing the Right Statistical Test  |  121

the samples are large enough, when the means of the samples are plotted, a normal
curve will emerge.

The Empirical Rule Applies to the Sampling Distribution of the Means

Having a normal distribution emerge as a result of the central limit theorem brings
with it some good news; the empirical rule applies. However, since we are plotting the
means from multiple samples, instead of measuring distances from the mean using
the standard deviation, we will use the standard error of the mean (SEM). As a result,
we can safely say that at least 68.26% of all sample means fall within plus or minus (±)
one SEM of the mean, 95.45% will fall between the mean (±) two SEMs, and almost
every sample mean (99.73%) will fall between the mean (±) three SEMs. At this point
you might be thinking, “This sounds good, but how can we compute the standard
error of the mean? I do not have access to the entire population and cannot take an
infinite number of samples and plot their means.” Luckily, we have a very easy formula
that allows us to do that:
σ
σx =
n
We have seen most of this before. The symbol on top of the equation is sigma
(i.e., s), the symbol for the population standard deviation; below that is the square
root of our sample size. The standard error of the mean, the value we are computing,
is represented by a lower-case sigma followed by the symbol for a sample mean
(i.e.,  σ x ) .
Instead of using the actual population standard deviation, just to make it easier to
understand, let’s use a nice whole number, in this case 14. Along with our sample size
of 49, we can easily compute the SEM:

14
1. σ x =
49
14
2. σ x =
7

3. σ x = 2

Keep in mind that we probably will not know the population standard deviation;
we simply used 14 in this case to illustrate what we are doing. If the population stan-
dard deviation is not known, we can use the sample standard deviation in its place.
When we do this, it is no longer called the standard error of the mean; instead we
call it the sample standard error of the mean. This means our formula changes slightly
because we have to replace the symbol for the population standard deviation (i.e., s)
with the symbol for a sample standard deviation, that is, s, as well as replace the
symbol for the population SEM (i.e., σ x ) with that of the sample standard error of
the mean (i.e., σ x):
122 | STATISTICS TRANSLATED

s
sx =
n
You will find that the sample standard error of the mean is a very good estimate
of what we would have computed had we known the population value.

Before We Go Any Further

Since we will use these formulas and terms extensively in the next section, let’s put
together a reference list (Table 5.4) we can easily refer to.

TABLE 5.4.  Symbols Related to the Sampling Distribution of the Means

Sampling distribution The resulting distribution when repeated samples are drawn
of the means from a population and their mean is calculated and plotted.
Given enough samples, this will result in a perfect normal
distribution.
σx Mean of the sampling distribution of the means; also known
as the “mean of means.”
s σ xrepresents the sample standard error of the mean;
sx = s represents the sample standard deviation, and n is the
n number of data values. This is the same as the standard
deviation in a normal distribution of values except in this case
we are using a sampling distribution of means.
σ σ x represents the population standard error of the mean,
σx =
n sigma (i.e., s) represents the population standard deviation,
and n is the number of data values. This is the same as the
standard deviation in a normal distribution if you know all of
the values in a population.

Putting It All Together

Using the information from our example, we now have enough information to start
putting things together. If we computed the mean score of our 49 samples, we would
find that it is 92. From our discussion, we know that the mean of the sampling dis-
tribution of the means (i.e., the mean of means) will be very close to the mean of the
population (i.e., 92); in this case I have made them exactly equal for the sake of sim-
plicity. In the preceding equation, when we used 14 as the population standard devia-
tion and 49 as our sample size, we calculated a SEM of 2. We know that the empirical
rule applies, and we can see everything shown together in Figure 5.6.
If the empirical rule still applies, this means our idea of z scores still applies. We
can compute the z score using the same basic formula, but we must change the values
to reflect those of the sampling distribution of the means, not those of a sample. We
can see this in the following formula:

z=
( x − mx )

σx
We already know that the mean of means (i.e., m x ) is 92. We also know, from our
example, that the population standard error of the mean (i.e., σ x ) is 2. If we have an
 Choosing the Right Statistical Test  |  123

observed sample mean that is one sample SEM above the mean (i.e., 94), then our
z score is 1:
(x − mx )
1. z =
σx

2. z =
(94 − 92)
2

3. z =
( 2)
2

4. z = 1

Let’s try that again and see if we

completely understand what we are
doing. Suppose the mean of our sam-
pling distribution of means is 50 and the
mean of the sample we are interested in
FIGURE 5.6.  The empirical rule. is 42. In this case, we do not know the
population SEM, but the sample SEM is
3; what is our z score? Let’s go through
the steps:

1. z =
(42 − 50)
3

2. z =
( −8)
3

3. z = –2.67

In this case, what does a z score of –2.67 tell us? Simple: the sample mean we are
interested in is 2.67 SEMs below the mean of the sampling distribution of the means
(i.e., the mean of means).
Now, let’s really start tying this together. Based on what we learned earlier, what
is the area under the curve between the mean and our z value? If we go to the back
of the book and look at the area under the normal curve table, we can see that our z
value is equivalent to .4962. This tells us that 49.62% of the values under the curve lie
between 42 and 50. Remember, even though our observed value is negative, that does
not affect the percentage we are interested in; a z score of 2.67 above the mean would
result in the same percentage under the curve. All we are interested in is measuring
the distance between two points.
124 | STATISTICS TRANSLATED

Summary of the Central Limit Theorem and the Sampling Distribution

of the Means
The concepts in this section are very straightforward. Creating a sampling distribu-
tion of the means involves nothing more than selecting a sufficiently large number of
samples from a population, calculating their means and then plotting those means on
a histogram. Given enough samples, this results in a normal distribution with a mean
equal to that of the population (i.e., the mean of means). Since the sampling distribu-
tion is normally distributed, the empirical rule still applies, and you can use z scores to
determine any given area under the curve. You will find these tools will be very useful
in the next sections when we begin talking about testing a hypothesis. Table 5.5 recaps
everything we have talked about in this chapter and will help you as we move forward.

TABLE 5.5.  The Process for Creating a Sampling Distribution of the Means
An infinite number of
random samples with
Original means computed and Sampling distribution
data distribution Population plotted of the means
Type of data Quantitative Quantitative
Shape of Any shape Bell-shaped
distribution
Measure of central Mean Mean equal to the
tendency population mean (i.e.,
the mean of means)
Measure of Standard Standard error of the
dispersion deviation mean
Measure of Empirical rule Empirical rule applies
relative standing applies

How Are We Doing So Far?

This is a rather long chapter filled with a lot of useful information. Given that, let’s not
wait until the end of the chapter to check our progress. Work through these exercises
and, as usual, check your work in the back of the book.

1. For each of the following sample means, compute the z score using the descrip-
tive statistics given in Figure 5.7. Note that the column labeled “Mean Statistic” is
the mean of the sampling distribution of the means and the column labeled “Std.
Deviation Statistic” is the population standard error of the mean.

a. 109
b. 77
c. 113
d. 101
e. 95
 Choosing the Right Statistical Test  |  125

Descriptive Statistics

N Mean Std. Deviation Variance

Statistic Statistic Std. Error Statistic Statistic

Score 100 98.0000 .81650 8.16497 66.667

Valid N 100
FIGURE 5.7.  Descriptive statistics for computing z scores.

f. 88
g. 96
h. 90
2. In Table 5.6, using a SEM of 2 and a population mean of 92, complete Table 5.6.

TABLE 5.6.  Computing the Distance between z Scores

Area under the Area under the Area between
Value 1 z1 curve for z 1 Value 2 z2 curve for z 2 z 1 and z 2
90.0 95.0
89.0 90.0
91.5 93.5
92.0 96.0

Estimating Population Parameters Using Confidence Intervals

Now that we understand the central limit theorem and the sampling distribution of
the means, it is time to put them to work for us. The first thing we will do is learn to
estimate a value in a population (i.e., a population parameter) based on data from a
sample (i.e., a sample statistic). Before we do that, however, we need to discuss a very
important technical term.

The Alpha Value

As we saw earlier, any time we are dealing with a sample of data we must realize
that our sample represents only one of the many samples that could be drawn from
a given population. Because of that, we are never exactly sure if statistics calculated
from sample data are equal to their counterpart parameter in the population data.
For example, if we are looking at a distribution of sample means, 50% of the sample
means are going to be at or below the population mean and 50% of the sample means
are going to be at or above the population mean. Because of sampling error, our
sample mean could be anywhere in that range (Figure 5.8).
Knowing this, what we, as consumers of statistics, must determine is “what is our
126 | STATISTICS TRANSLATED

FIGURE 5.8.  The population mean and the mean of the sampling distribution of means.

cutoff point?” If we are trying to test a hypothesis, how do we decide which sample
statistics are different owing to sampling error and which are significantly different
from the population parameter?
First, we accept the fact that any decision we make may be wrong; after all, we are
dealing with probability. Luckily, we have some leeway at this point. We are able, by
stating an alpha value, to state a priori (i.e., beforehand) the degree of risk we are will-
ing to take when deciding. The alpha value, often abbreviated using the Greek letter
α, is sometimes called the level of significance. Its use is probably best explained using
an example.
Suppose we are interested in determining if the average final exam score for the
students in our Management class (i.e., our sample) is significantly different from the
average score of other Management students throughout the university (i.e., the popu-
lation of Management students). In order to determine this, we would have to test the
following null hypothesis:

„ There is no significant difference in exam scores between our Management

students and other Management students throughout the university.

Now, let’s imagine we have computed an average score of 92 for our students and
found the population average to be 80. Because the difference between these mean
scores is somewhat large, it appears we should reject our null hypothesis (i.e., there
appears to be a significant difference between the two groups). Before we make that
assumption, though, let’s think it through.
We have already said that statistics from a sample, the mean for example, cluster
around the same measure in the population. Knowing that, the mean score of our
class (i.e., our sample) is nothing more than one of the multitude of mean scores fluc-
tuating around the population mean. In our case, all we know about our sample mean
is that it is different from the population mean. This difference is either due to sam-
pling error, or it is so large it cannot be attributed to chance; it represents a significant
difference between the sample mean and the population mean.
 Choosing the Right Statistical Test  |  127

If the first is true, we will fail to reject our null hypothesis—although there is a
difference, it is due to chance and is not significant. If the second is true, then we
will reject our null hypothesis—it is unlikely we can attribute our results to chance;
they represent a significant difference between our sample mean and the population
mean. The question that remains, then, is, “This sounds good on the surface but how
do we know when the sample mean, and the population mean are different due to
chance and when they are significantly different from one another?” In answering
that question, I have some bad news and some good news. The bad news is, since we
are dealing with a sample of data, we are never 100% sure if the differences are due
to chance or represent a significant difference. The good news is, by using our alpha
value we can control the risk of making the wrong decision.
Our alpha value allows us to do four things:

1. We can use it to indicate the percentage of time we are willing to make an

incorrect decision based on the sample data we have collected. We have to
admit that sampling error is going to occur, but we are going to set limits
within which we can say that a sample statistic is significantly different from
the population parameter due to chance (i.e., sampling error) and times when
it is truly significantly different due to reasons other than sampling error (i.e.,
a “real” difference).
2. We can control the level of risk we are willing to take by using different alpha
values (e.g., .10 and .01), although it is generally set at .05 (or 5%) for the types
of studies we are most likely to undertake. By doing this we are saying, prior
to making even the first statistical calculation, we are willing to take a 5%
chance of making an error by rejecting a true null hypothesis.
3. The alpha value can help us create a range called a confidence interval. We can
use a confidence interval to predict a population parameter using a sample
statistic.
4. The alpha value can help us test a hypothesis about the population using data
drawn from a sample.

In short, because statistical decision making is based on sampling and probability,

we can never be 100% sure that any decision we make is going to be correct. We can,
however, use our alpha value to help us feel more comfortable in the decisions we do
make.

Type I and Type II Errors

Often alpha is referred to as the Type I error rate or the probability of making a
Type I error; they both mean the same thing. Because we are dealing with samples, we
are never 100% sure if the decisions we make are correct. If we reject a null hypoth-
esis when we should not have (i.e., due to chance), then we have made a Type I error.
Unfortunately, these things happen from time to time; along with that comes even
more good news and bad news. The good news is that we can control for Type I errors
using our alpha value, but changing the alpha value, as we will see shortly, may also
lead to other problems.
128 | STATISTICS TRANSLATED

The counterpart of a Type I error is called a Type II error and occurs when, because
of sampling error, we fail to reject a null hypothesis when we actually should reject
it (i.e., we say the values are not significantly different when they really are). In some
texts, you’ll see the Type II error rate referred to as beta; the two are synonymous. A
mistake that many people make is to assume that, if our alpha value is .05, then our
beta value must be .95—that is not true! Calculating beta is a very complicated process,
but, luckily for us, it is often computed as part of the output when using a statistical
software package.
Although beta is not often a topic for beginning statisticians, it does play an inte-
gral part in computing the statistical power of an inferential test. Statistical power,
defined as the probability of not making a Type II error (i.e., we will reject the null
hypothesis when it is false), is computed by subtracting beta from 1. This value, of
course, is affected by the alpha value we use, the actual difference between values we
are comparing and the sample size. Generally speaking, any power value over .80 is
considered acceptable. Again, this is something that consumers of statistics are not
frequently faced with, but we must be able to recognize and understand terms like this
when we see them!

Criticality of Type I and Type II Errors

Understanding the criticality and relationship of Type I and Type II errors is impera-
tive, so let’s use an analogy that is easy to understand. Suppose you have a null hypoth-
esis that reads:

„ The defendant is not guilty.

The research hypothesis would read:

„ The defendant is guilty.

In this case, a Type I error (i.e., rejecting the null hypothesis when you should not)
would mean that an innocent person would be sentenced to jail because you would be
supporting the research hypothesis. Failing to reject the null and not supporting the
research hypothesis due to sampling error (i.e., a Type II error) would mean a criminal
would be found not guilty and be allowed to go free. Which is worse, an innocent man
in jail or a criminal on the street? Of course, we could argue circumstances here, but
suffice it to say the first scenario is worse!
Having said all of that, Table 5.7 shows the relationship between null hypotheses
and the two types of errors.

TABLE 5.7.  Type I and Type II Errors

If we do not reject If we reject the
the null hypothesis null hypothesis
If null hypothesis is true Correct decision Type I error
If null hypothesis is false Type II error Correct decision
 Choosing the Right Statistical Test  |  129

Predicting a Population Parameter Based

on a Sample Statistic Using Confidence Intervals
In this section, we are only going to look at making estimations where one sample is
taken from a population. In doing so, we will learn to estimate a population param-
eter (e.g., a population mean) based on a sample statistic (e.g., a sample mean). We can
use the same tools to estimate other values (e.g., differences between two independent
population means, population proportions, and differences between two dependent
populations), but we are going to stick with the basics. The reasoning is simple; we
will focus on the most basic form of estimation because it is easy to understand, and
the underlying principles are the same for other types of estimations. That being said,
let’s move forward.
Let’s go back to our example using the students in the Management class. This
time, however, instead of hypothesizing that our group will be different from the
population, we want to use a sample statistic to predict a population parameter. In
this case, we are going to use our class average to estimate the average for all the other
Management students.
It would be impossible to ask every student in the university what they scored and
then compute the population average, so that option is out of the question. Instead,
we can use a confidence interval to predict a population parameter, in this case the
mean, based on a mean of the sample. A confidence interval is defined as

„ A range of numbers around a sample statistic within which the true value of
the population is likely to fall.

As we have said, any time we have a sample from a population, the sample statistic
may or may not be equal to the corresponding population parameter. Because of that,
we can never be 100% sure of any estimates made based on the sample data. We can
build a range around the sample statistic, called the confidence interval, within which
we predict the population parameter will fall.
To compute the confidence interval, you need two things: a sample from the pop-
ulation and an alpha value. We will stick with the standard alpha value of .05 (i.e., 5%).
This means that you want to create a range that, based on the data from the sample,
will have a 95% (i.e., 100% minus your alpha value of 5%) probability of including the
population average. Obviously, it also means there is a 5% probability that the popula-
tion mean will not fall into the confidence interval. Let’s use the following data to put
this all together:

1. The sample size is 50 (the number of students in your class).

2. Your average test score for your sample is 92.
3. You are setting alpha to .05—this means we are going to have a 95% confi-
dence interval.

Using this information you can use the following formula to compute the confidence
interval you need:
130 | STATISTICS TRANSLATED

 σ 
Confidence interval = x ± (za 2 ) 
 n 
The individual parts are easy to understand:

X x stands for the mean of the sample we are dealing with (i.e., 92).
X z is the z value for the alpha level we chose (i.e., 1.96). As you can see, we will
be dividing our z value by 2; we will talk more about that in a minute.
X a is our alpha value (i.e., .05).
X n is the square root of the sample size (i.e., 50) we are dealing with (i.e.,
7.07).
X s is our sigma value and represents the population standard deviation.

Here it seems we have a potential problem. This last step in the equation, shown
below, is the formula for the standard error of the mean, but we still do not know the
population standard deviation (i.e., s) in order to compute it:
σ

n
As we have said, if we do not know the population standard deviation we can
replace it with the sample standard deviation and get a very good estimate. In this
case, however, we’ll make it easy and use a population standard deviation of 5 so we
can move forward. That will give us everything we need for our formula:
 5 
Confidence interval = 92 ± (1.96) 
 7.07 
At first glance, everything here is pretty straightforward except for the whole idea
of the z score for alpha divided by 2. We do that because, if you are trying to create
a range that contains a population parameter, obviously you will have fluctuation
around that parameter due to sampling error. Since you know this is going to occur,
you have to be willing to make a mistake in either direction—you may estimate too
high, you may estimate too low, or you may estimate just right.
If we set alpha to .05 and then divide it by 2, we get a value of .025 or 2.5%. That
means you are willing to make a 2.5% mistake on either side of the mean. When you
subtract this value from the total percentage (50%) on either side, you are now saying
your range will include from 47.5% below the mean to 47.5% above the mean (remem-
ber 47.5% + 47.5% = 95%, the size of our confidence interval). When you find 47.5 (or
.4750) in our area under the normal curve table, you find the corresponding z value of
1.96. To verify this, here is the row in the table showing these values. As you can see
in Table 5.8, .4750 lies at the intersection of the 1.9 row and the .06 column; obviously,
1.9 and .06 equals 1.96. All of this is shown in Figure 5.9.

TABLE 5.8.  Percentage Value for z Score of 1.96

z score 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
1.9 .4713 .4719 .4726 .4732 .4738 .4744 .4750 .4756 .4761 .4767
 Choosing the Right Statistical Test  |  131

We can now compute the actual confidence interval using what we know up to
this point:

1. When alpha = .05, we are looking for a confidence interval that contains 95%
of the values (i.e., 1.00 – .05 or 100% – 5%).
2. Since we must establish a range on both sides of the value we are predicting,
it is necessary to divide the size of our confidence interval by 2. This leaves us
with 95%/2 or 47.5%.
3. The area under the normal curve table shows a z value of 1.96 for 47.5%.
4. The population standard deviation (5) divided by the square root of the sam-
ple size (n = 50) is .707.
5. Confidence interval = 92 ± (1.96)(.707)
6. Confidence interval = 92 ± 1.39
7. Confidence interval = 90.61 to 93.39

Pay Close Attention Here

Right about here, many beginning statis-
ticians (and if the truth were told, many
statisticians who should know better)
think a 95% confidence interval means
we are 95% sure the confidence interval
contains the population parameter of
interest. This is not true. The confidence
interval you compute around the sample
statistic either contains the population
parameter or it does not.
Instead, what it does say is, if we
FIGURE 5.9.  Relationship between the SEM and were to repeatedly take samples from
percentages of values in a distribution. the same population and calculate the
confidence interval, the intervals com-
puted would contain the population parameter 95% of the time. In this case, it means
that 95% of the time, the true population parameter would be somewhere between
90.61 and 93.39. While this may sound like semantics, it is a very important point.
In short, there is a 95% probability that the population mean is somewhere
between these two values. Again, we have a 5% chance of making an error, but a 95%
chance of being right is a good bet! This is shown in Figure 5.10.
Let’s look at one more case before we move on. Here let’s imagine our class aver-
age is 70, our population standard deviation is 10, and our sample size is 100:

1. When alpha = .05, we are looking for a confidence interval that contains 95%
of the values (i.e., 1.00 – .05 or 100% – 5%).
2. Since we must establish a range on both sides of the value we are predicting,
it is necessary to divide the size of our confidence interval by 2. This leaves us
with 95%/2 or 47.5%.
132 | STATISTICS TRANSLATED

3. The area under the normal curve table shows a z value of 1.96 or 47.5%.
4. The population standard deviation (10) divided by the square root of the
sample size (n = 100) gives us
10/10 or 1.
5. Confidence interval = 70 ±
(1.96)(1.00)
6. Confidence interval = 70 ± 1.96
7. Confidence interval = 68.04 to
71.96

Here, we are 95% confident that

our population average is somewhere
between 68.04 and 71.96. Again, we are
not saying that 95 out of 100 values fall
FIGURE 5.10.  95% confidence interval around a in this range; this is just the probability
mean of 92. that any given value does.

Confidence Intervals for Alpha = .01 and Alpha = .10

At this point, we run into a sticky situation. We just saw that we were able to find the
correct percentage for the 95% confidence interval because that meant we were look-
ing for a range of numbers that represented 47.5% above the mean and 47.5% below
the mean. We simply went into the body of the area under the normal curve table and
found that 47.5% represents a z score of 1.96.
Unfortunately, it is not quite so easy to determine the z scores for the other com-
mon alpha values (i.e., .01 and .10). For example, if our alpha value is .10, that means
we want to compute a 90% confidence interval; this means 45% on either side of the
mean. If we look at another section of the area under the normal curve table, shown
in Table 5.9, we can see this creates somewhat of a problem.

TABLE 5.9.  Percentage Value for z Score of 1.645

z score 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
1.6 .4452 .4463 .4474 .4484 .4495 .4505 .4515 .4525 .4535 .4545

In looking at the table, there is not an entry for 45% (i.e., .4500) but we do have
entries for 44.95% (i.e., .4495 in the table and a z score of 1.64) and 45.05% (i.e., .4505
in the table and a z score of 1.65). Since the value we are looking for, 45.00%, is in the
exact middle between 44.95% and 45.05%, we must compute a z score that represents
the average between the two. This process is as follows:

1. When alpha = .10, we are looking for a z score for 45% (i.e., 90%/2).
2. The table shows scores of 44.95% (z = 1.64) and 45.05% (z = 1.65).
3. 1.64 + 1.65 = 3.29
4. z = 3.29/2
 Choosing the Right Statistical Test  |  133

5. z = 1.645

Let’s use the same data as from the first example above; our mean is 92, our
population standard deviation is 5, and we have a sample size of 50. This leads us to a
confidence interval ranging from

1. Confidence interval = 92 ± (1.645)(.707)

2. Confidence interval = 92 ± 1.16
3. Confidence interval = 90.84 to 93.16

The same holds true when we have an alpha value of .01. In this case, we want
to compute a 99% confidence interval that represents 49.5% of the values above the
mean and 49.5% of the values below the mean. That means we are looking for a value
of .4950 in the area under the normal curve table; let’s use Table 5.10 to help us.

TABLE 5.10.  Percentage Value for z Score of 2.575

z score 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
2.5 .4938 .4940 .4941 .4943 .4945 .4946 .4948 .4949 .4951 .4952

Here we can see there is no entry for 49.5% (.4950), but there are entries for the
two values that surround it, .4949 and .4951. Because 49.50% is exactly in the middle
between 49.49% and 49.51%, we must again use the z score that represents the average
of the two (i.e., (2.57 + 2.58)/2 = 2.575). Again, this process is

1. When alpha = .01, we are looking for a z score for 49.5% (i.e., 99%/2).
2. The table shows scores of 49.49% (z = 2.57) and 49.51% (z = 2.58).
3. 2.57 + 2.58 = 5.15
4. z = 5.15/2
5. z = 2.575

Using this value, we go through the same process to compute our confidence
interval:

1. Confidence interval = 92 ± (2.575)(.707)

2. Confidence interval = 92 ± 1.82
3. Confidence interval = 90.18 to 93.82

Let me reiterate just one more time. This only means that we are 99% confident
that our population average lies somewhere between 90.18 and 93.82. Remember,
although it is close, 99% is not 100%; because of random error, we can never be abso-
lutely certain.
134 | STATISTICS TRANSLATED

Another Way to Think about z Scores in Confidence Intervals

It seems like we have gone to a lot of trouble to find the necessary z scores for these
three common confidence intervals, doesn’t it? If so, here is the good news. The z scores
we have just computed are obviously going to be constant every time we create a 90%,
a 95%, or a 99% confidence interval. Knowing that, I have created a reference chart
that is a lot easier to deal with. As you can see, the common confidence interval sizes,
along with their corresponding information, are shown in Table 5.11.

TABLE 5.11.  z Scores for Common Confidence Intervals

Confidence Area between
interval a a/2 mean and tails z a/2
.90 .10 .05 45% 1.645
.95 .05 .025 47.5% 1.96
.99 .01 .005 49.5% 2.575

Tying This All Together

Now that we have seen how to compute confidence intervals for different alpha val-
ues, let’s use one case to better understand how changing our alpha value can affect
our overall decision making. Suppose you are hired as a school psychologist and are
interested in finding the average IQ in your school. You do not have time to test all
500 students in your school and you want to work with a representative sample of 30.
How do you proceed?
First, let’s assume you have a sample size of 30 with a mean IQ of 110; assume fur-
ther you have a population standard deviation of 15 and an alpha value of .05. Using
these values, you can compute the confidence interval just like we did in the preceding
example:
First, let’s assume you have a sample size of 30 with a mean IQ of 110; assume fur-
ther you have a population standard deviation of 15 and an alpha value of .05. Using
these values you are able to compute the confidence interval just like we did in the
preceding example:
 15 
Confidence interval = 110 ± (1.96) 
 5.48 
1. First, 15 (i.e., the population standard deviation) divided by 5.48 (i.e., the
square root of the sample size) gives us 2.74.
2. When we multiply 2.74 by 1.96 (the z score of a/2), we get 5.37.
3. When we subtract 5.37 from our mean of 110, we get 104.63 (the lower limit
of our confidence interval).
4. When we add 5.37 to 110, we get 115.37 (the upper limit of our confidence
interval).

The results, shown in Figure 5.11, indicate there is a 95% probability that the aver-
age IQ in the school is between 104.63 and 115.37.
 Choosing the Right Statistical Test  |  135

Now, let’s lower our alpha level to

.01; this means we are going to create
a 99% confidence interval (i.e., 100% –
1% = 99%). The calculation is the same
except we need to substitute 2.575 for
1.96 (this is the z score for a/2 or .005%
[i.e., .5%]). Using our information, we
can compute our confidence interval
using the following formula:

 15 
Confidence interval = 110 ± (2.575) 
FIGURE 5.11.  95% confidence interval around a  5.48 
mean of 110.
In this case, there is a 99% probabil-
ity that the average IQ in the school is
between 102.94 and 117.06. By using the same data and decreasing the size of our
alpha value, we can see that we have widened the confidence interval (Figure 5.12).
In order to give ourselves the greatest opportunity to create a range that includes the
population mean, this means we have to use a smaller value for alpha.
Let’s do the same thing but increase our alpha value to .10. Again, our formula
would look the same except our z value now is 1.645.
 15 
Confidence interval = 110 ± (1.645) 
 5.48 
In this case, we have narrowed our confidence interval down to 105.50–114.50.
Here in Figure 5.13, we can see, by using a larger alpha value, we have decreased the
range of the confidence interval.

Be Careful When Changing Your Alpha Values

Although you have seen it is easy enough to change your alpha value
and compute the appropriate statistics, be careful when you do—you
might be creating a problem when you
analyze your data. For example, in the
prior three examples, we originally set
our alpha level at .05 and then saw what
the difference in the confidence inter-
val would be if we set alpha to .01 and
then to .10. The type of decision and
the needed precision of your results
are going to dictate the alpha level you
choose. This idea is shown in Table 5.12.
If you look closely at this table,
you can see that it conceptually makes
sense. The size of your alpha value is
inversely related to the width of your
FIGURE 5.12.  99% confidence interval around a confidence interval; larger alpha val-
mean of 110.
136 | STATISTICS TRANSLATED

ues lead to smaller confidence intervals

and vice versa. Of course, this creates
havoc when you are trying to estimate
a population parameter. Do you try to
“cast as large a net” as possible with a
small alpha value, or do you try to be as
accurate as possible with a large alpha
value? Unfortunately, this is the dilem-
ma that statisticians face when dealing
with uncertainty. How do we strike a
happy medium? Easy—set alpha equal to
.05. That is a standard that researchers
in the social sciences are familiar and
FIGURE 5.13.  90% confidence interval around a comfortable with. There is still some
mean of 110. degree of error, but it is small enough to
live with.

TABLE 5.12.  The Effect of Alpha on the Size of a Confidence Interval

Confidence Lower Mean Upper Width
interval (CI) Alpha z score end of CI of CI end of CI of CI
90% .10 1.645 105.50 110 114.50 9.0
95% .05 1.96 104.63 110 115.37 10.74
99% .01 2.575 102.94 110 117.06 14.12

Do We Understand Everything We Need to Know about Confidence Intervals?

As we did with the section on the central limit theorem, let’s go ahead and test our
comprehension of the chapter up to this point. In Table 5.13, you are supplied with the
sample mean, the alpha value, the number of values in the sample, and the population
standard deviations. Use these values to compute the lower and upper limits of the
confidence interval.

TABLE 5.13.  Learning to Compute the Width and Limits of a Confidence Interval
Lower limit Upper limit Width
Alpha n s of CI of CI of CI
100 .10 25  5
500 .05 50 25
 20 .01 15  3
 55 .05 20  7
 70 .01 22  6
220 .10 40 10
 Choosing the Right Statistical Test  |  137

Testing Hypotheses about a Population Parameter

Based on a Sample Statistic
In the last section, we learned to predict a population parameter based on data from
a sample. We saw there are many different types of predictions we could make, but
we used only the mean, as an example. At this point, we are again going to use one
example that will lay the groundwork for the various types of hypothesis testing we
will do throughout the remainder of the book. In this case, we are going to learn to
make decisions about a population parameter based on sample information. In order
to do that, let’s use the following example.
Let’s imagine you are the dean of the graduate school of psychology at ABC Uni-
versity. Unlike many other graduate schools in the United States, your students are
not required to take the Graduate Record Examination (GRE) as part of the admissions
process. Critics have suggested that this attracts weaker students and, because of that,
they will not score as high on the national board certifications as their peers in schools
where the GRE is required.
You admit to yourself that you have never really given it too much thought and,
to tell the truth, you are not sure how your students do in comparison to students in
other graduate schools. Interested in finding out, you visit one of the statistics profes-
sors in your department and tell her what is going on.
After telling her your problem, your professor immediately tells you, “In order to
conduct your study, you have to remember two things. First,” she says, “our student
body is assumed to be nothing more than a sample of the population of all graduate
psychology students throughout the United States. There are an infinite number of
other samples you could select from that same population.” She illustrates what she
was saying by drawing Figure 5.14 on the board.
She continues by saying, “What you want to know is if our students are real-
ly members of that population or, because of our admissions process, they are sig-
nificantly different from other samples
from that population. In other words,
you want to know if the average score
our students make on the national cer-
tification examination is significantly
different from the average score of the
other samples.” She also draws Figure
5.15 on the board.
“As you can see,” she says, “our crit-
ics are claiming that our students will
not do as well on the certification exams
as the other schools, and we do not
know if we are doing significantly bet-
ter or worse than they are. Given that,
you need to develop the hypothesis you
want to test and then collect the data
FIGURE 5.14.  The relationship between a popu-
lation and a sample from the population. you need.” After thinking about what
she said, you decide on the following
two-tailed research hypothesis:
138 | STATISTICS TRANSLATED

FIGURE 5.15.  Different versus significantly different.

„ There will be a significant difference in scores on the national certification

exam between graduate psychology students at ABC University and
graduate psychology students at other universities.

This means you will also have a corresponding null hypothesis:

„ There will be no significant difference in scores on the national certification

exam between graduate psychology students at ABC University and
graduate psychology students at other universities.

Can you tell me why we made the hypothesis two-tailed (i.e., nondirectional)?
That is easy—we do not know if our students are doing better or worse than other stu-
dents in the country. Because our scores could be greater than or less than the popula-
tion’s scores, we must make the hypothesis nondirectional.

Making a Decision about the Certification Examination Scores

Finally, after all this time, we are going to test our first hypothesis. Before we do, how-
ever, let’s summarize what we know about populations and samples up to this point.
First, a population represents all possible subjects about which we might want to col-
lect data—voters in the United States, every student in a given university, all managers
within a corporation who have a graduate degree, and so on. We have also said that,
when we take a sample, we are hoping it is representative of the population from
which it was taken—smaller samples tend to be less representative, larger samples more
so. Finally, we have shown that, when we plot values, the mean for example, from a
sample of 30 or more, the distribution is generally bell-shaped or, as we call it, a nor-
mal distribution. Having a sample that is normally distributed brings with it some
good news when it comes to hypothesis testing:
 Choosing the Right Statistical Test  |  139

1. We can test hypotheses to determine if one of the sample means is significant-

ly different from the population mean by using the z score we have already
learned to compute. We call this a one-sample z-test.
2. We saw earlier that the z score allowed us to report distances from the mean
in terms of the standard deviation. The one-sample z-test allows us to do the
same but, instead of using the standard deviation, we will use the standard
error of the mean.
3. The empirical rule still applies.
4. The principles underlying the z-test are fairly straightforward and are basi-
cally the same for most of the other statistical tests we will use.

In our case, let’s suppose we have 100 students in our graduate program. Knowing
that, as we just said, we can use the z statistic to test our hypothesis. This means using
the following values to compute our z value.

1. The mean score ( x ) we want to compute a z score for.

2. Either the population mean (m) or the mean of means ( m x ).
3. The population standard error of the mean ( σ x ).

Let’s start with the actual equation:

z=
(x − m)
σx
For the sake of our example, let’s assume the population mean is 800 (i.e., the
average score for graduate students at all universities), the SEM is 10, and the mean
value for our students is 815. We can substitute those into our formula:

(815 − 800)
z=
10
This would result in a z score of 1.5. From this point forward, this will be called
our computed value of z. In order to test our hypothesis, we have to compare this value
to a critical value of z. This is nothing more than a value from the area under the nor-
mal curve table we used earlier. The given value we choose will be based on the alpha
value we’re using and the type of hypothesis we’ve stated (i.e., directional or nondirec-
tional). Let’s use an example to better understand this.

Determining the Critical Value of z

We saw, earlier in the book, that a normal curve has a distribution of z scores with a
mean of zero. Fifty percent of all possible z scores fall at or below the mean of zero,
and 50% fall at or above the mean of zero. These z values represent all the possible
critical values of z. Our problem is to figure out which one to use.
In order to do this, we must decide on our alpha value, so we will use the tradi-
tional .05 (i.e., 5%). We then must subtract that from the range of all possible z values
140 | STATISTICS TRANSLATED

(i.e., 100%) to give us the range of possible z values we can use to test our hypothesis.
This means our possible range is 95% of all possible z values (i.e., 100% – 5% = 95%).
Next, we must use the type of hypothesis we have stated to determine the distribu-
tion of all the possible critical values of z. If we have a two-tailed hypothesis, we must
distribute the z scores equally under the curve; if we have a one-tailed hypothesis we
must distribute them according to the direction of the hypothesis. For now, let’s look
at the nondirectional (i.e., two-tailed) hypothesis; we will do the same for a one-tailed
hypothesis shortly.
If we already know that we want to build a range that includes 95% of all possible z
values and that range must be equally distributed on either side of the mean, it is easy
to determine what the range would be. We would equally divide the 95%, meaning
47.5% of the values would fall on one side of the mean and 47.5% of all values would
fall on the other. We would then have 2.5% of all possible values left over on either
side. We can see that in Figure 5.16.
At this point, things start coming
together. If we have 2.5% of all possible
values left over on either side of the
mean, it gives us a total of 5% remain-
ing. This is the same as our alpha value
(i.e., .05 or 5%). This tells us that, when
we have a two-tailed hypothesis, all we
must do is divide alpha by 2 to deter-
mine the percentage of all possible val-
ues on either side of the mean.
Now that we have identified an area
that contains 95% of all values of z, our
job is to determine exactly which value
we want to use as our critical value. In
FIGURE 5.16.  Equally dividing 95% of z values order to do that, we need to use the area
under the curve. under the normal curve table.
As you’ll remember, this table only
shows values for one side of the curve. That is fine. If we know values for one side of
the curve, it is easy to determine what the corresponding value would be on the other
side; it is just the numeric opposite. Although we have gone through all of this before
back when we talked about confidence intervals, let’s look at Table 5.14; this is a small
part of the critical value of the z table and will help us determine the critical z value
we need.

TABLE 5.14.  Percentage Value for a z Score of 1.96

z score 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
1.9 .4713 .4719 .4726 .4732 .4738 .4744 .4750 .4756 .4761 .4767

We can see that the value .475 (i.e., 47.5%) lies at the intersection of the row
labeled 1.9 and the column labeled .06. If you add these two values together, it gives
us a z value of 1.96. This means, for 47.5%, our critical value of z on the right side of
the mean is 1.96 and the critical z value on the left side of the mean is –1.96. Putting
 Choosing the Right Statistical Test  |  141

all of this together, we can see that a range of 95% (i.e., 47.5% on either side of the
mean) encompasses a range of critical z scores from –1.96 to +1.96. This is shown in
Figure 5.17.

We Are Finally Going to Test Our

Hypothesis!
We are now ready to test our hypothesis.
Let’s use the same values we did earlier:
the population mean is 800, the SEM
is 10, and our university’s mean score
is 815. That gave us a computed z score
of 1.5. All we must do to determine if
our group is significantly different is to
see if our computed z score falls within
the 95% range of all possible critical z
scores. In this case, it is easy to see that
it does; 1.5 is between –1.96 and +1.96.
This means we fail to reject our null
hypothesis; there is no significant differ-
FIGURE 5.17.  The relationship of z scores and
percentages. ence between our scores and the scores
of the other universities throughout the
nation.
Believe it or not, it is as easy as that! We will plot what we have done in just a min-
ute, but for now let’s look at another example using the national mean score of 800
again, but this time let’s use a mean score of 780 and a SEM of 10. What would be our
computed value of z? Putting those values into our formula we get:

z=
(780 − 800)
10
This gives us a z score of –2. Since this doesn’t fall between –1.96 and +1.96, we
will reject the null hypothesis. In this case, our score of 780 is significantly less than
the national average of 800. Given this, perhaps your critics are right; maybe it is time
to implement the GRE.
In Figure 5.18, I have used a population mean of 800 and SEM of 10 and plotted
a few more mean values for our students. This, along with the corresponding z scores,
will help solidify this idea.

Testing a One-Tailed Hypothesis

Much like the two-tailed hypothesis, the logic behind testing a one-tailed hypothesis
is quite straightforward. As its name implies, we will be using our alpha value on only
one side of the distribution, but that is the only difference. Let’s use the following case
to explain what I mean.
Imagine you are the dean of a school of education where the average profes-
sor’s annual salary is $73,000. One day, one of your more outspoken faculty members
comes into your office and asks, “How do you ever expect us to stay here? We work
142 | STATISTICS TRANSLATED

FIGURE 5.18.  Using the z score to reject or fail to reject a null hypothesis.

harder, teach more classes, and have a larger dissertation load than faculty at other
universities. Unless you can show us we are making significantly more than the aver-
age faculty salary at those other places, we are going on strike!” After saying this, the
professor leaves your office in a huff, slamming the door behind him. After careful
consideration, you decide he might have a point. He is good professor and he does a
lot of fine work; perhaps an investigation is called for.
What is the first thing you do? It is easy. Since he wants you to show that your
faculty is earning more than the national average, you state the following research
hypothesis:

„ Your faculty’s average annual salary will be significantly higher than the
national average salary of faculty members.

The corresponding null hypothesis would read:

„ There will be no significant difference between your faculty’s average annual

salary and the national average salary of faculty members.

In order to begin investigating this, you call the human resources department and
find that the average salary in your department is $74,000. After a little investigation,
they tell you the national average is $70,000 with a SEM of $2,000. This means your
faculty’s salary, $74,000, has a z score of +2:

1. z =
(74000 − 70000)
2000

2. z =
(4000)
2000

3. z = +2
 Choosing the Right Statistical Test  |  143

Is that large enough to consider the difference significant? Let’s compare the
computed z score to the critical z score and find out. Before we do, however, let’s take
a minute to consider where we are going conceptually.
In the prior example, when we were using a two-tailed hypothesis, we had to
divide our alpha value by 2 and mark off 2.5% on each end of the distribution. In
order to determine the critical z score, we had to take the remaining 47.5% (remem-
ber, 50% on either side of the distribution, minus 2.5%) and consult the area under
the normal curve table. In the table, we found that a z score of 1.96 corresponded to
47.5%, so we marked off a range from –1.96 to +1.96 on our curve. We then compared
our computed value of z to that range in order to test our hypothesis.
In this case, we must do something a bit different. Since we have a one-tailed
hypothesis but still have an alpha value of .05, we must mark the entire 5% off on one
end of the distribution or the other. Here we have a “greater than” hypothesis so we
must mark the 5% on the positive end of the distribution. You can see this in Figure
5.19.
Now, it is time to determine the critical value of z we are going to use. Using the
excerpt from the area under the normal curve table (Table 5.15), we can see that the
critical value for z for 45% of the area under the curve lies between 1.64 and 1.65; this
leaves us with a critical value of 1.645. That can be seen in Figure 5.20.
Now, let’s compare our computed z to the critical value of z that we have plotted. Is
our computed z (i.e., +2) greater than our critical value of z (i.e., 1.645)? Since it is, we
will reject the null hypothesis. Your faculty members are making, on average, $74,000
per year. This is significantly more than the national average of $70,000. We can see
this in Figure 5.21.

TABLE 5.15.  Percentage Value for a z Score of 1.645

z score 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
1.6 .4452 .4463 .4474 .4484 .4495 .4505 .4515 .4525 .4535 .4545

In Figure 5.22, I have shown a couple of other examples using different values
of z. When our computed value of z is equal to 1.5, we would not reject the null

FIGURE 5.19.  5% on the positive side of the FIGURE 5.20.  Critical z for 5% above the mean.
mean.
 144 | STATISTICS TRANSLATED

FIGURE 5.21.  Rejecting the null hypothesis with FIGURE 5.22.  Examples of rejecting and failing
a computed z of 2. to reject based on computed z scores.

hypothesis; there is no significant difference between our faculty’s salaries and the
national average. We do, however, reject the null when z is equal to 1.7. This means our
research hypothesis stating that our faculty members are already making more than
the national average is supported.

Testing a One-Tailed “Less Than” Hypothesis

In the preceding example, we had a one-tailed “greater than” hypothesis. Here, let’s
look at a “less than” case by using an average annual faculty salary of $66,500. Since
we already know the national average is $70,000 with a SEM of $2,000, we are con-
cerned that the faculty may have a justifiable complaint. Let’s begin to test their claims
with this research hypothesis:

„ Your faculty’s average annual salary will be significantly less than the
national average salary of faculty members.

The corresponding null hypothesis would read:

„ There will be no significant difference between your faculty’s average annual

salary and the national average salary of faculty members.

First, let’s compute our z score using our faculty member’s salary along with our
population mean and standard error:

z=
(66,500 − 70,000)
2000
Here the computed z score is –1.75. We can plot this on our distribution to see if
a significant difference exists but, before we do, let’s consider what our distribution
should look like.
 Choosing the Right Statistical Test  |  145

As we just saw, when we had a one-tailed hypothesis where we were investigating

a “greater than” condition, we had to put the entire 5% alpha value on the right side
(i.e., greater than zero) of the distribution. Now, since we are looking at a “less than”
relationship, we must put the 5% on the left (i.e., negative) side of the distribution
shown in Figure 5.23.
Since we are using 5%, the critical value of z (i.e., 1.645) will be the same as above;
we will just put a negative sign in front of it and use it on the opposite side of the dis-
tribution. We can see this in Figure 5.24.
If we compare our computed value of z (i.e., –1.75) to our critical value of z, it is
less than –1.645. Because of that, the professor seems to have a case; the faculty’s aver-
age annual salary is significantly lower than the national average. This can be seen in
Figure 5.25.
I have included two more examples in Figure 5.26. In the case where the comput-
ed z value is –3.0, you might need to consider giving your faculty a raise. In the other,
where z = –1, perhaps the administration should consider giving you a raise for having
to deal with employees like this.

Summarizing What We Just Said

Although we went into a lot of detail to get to this point, we can summarize determin-
ing our critical value of z using the following steps:

1. Determine the alpha value we want to use. In this example, let’s use a = .05.
2. If we have a two-tailed hypothesis (i.e., nondirectional), divide our alpha value
by 2; in this case that would leave us with .05/2 = .025. If we have a one-tailed
(i.e., directional) hypothesis, we will use the entire alpha value.
3. We then subtract the result from Step 2 from .50. In this case, if alpha is equal
to .05 and we have a two-tailed hypothesis, we would be left with .475 (i.e., .50

FIGURE 5.23.  5% on the negative side of the FIGURE 5.24.  Critical z for 5% below the mean.
mean.
 146 | STATISTICS TRANSLATED

FIGURE 5.25.  Rejecting the null hypothesis with FIGURE 5.26.  Examples of rejecting and failing
a computed z of –1.75. to reject based on computed z scores.

– .025 = .475). If we have a directional hypothesis, we will subtract our entire

alpha value from .50, leaving .450 (i.e., .50 – .05 = .45).
4. We then use either of those two values to determine our critical value of z
in the area under the normal curve table. In this case, our critical value for
a two-tailed hypothesis would be 1.96; our value for a one-tailed hypothesis
would be 1.645.
5. We have to determine our range of critical z scores:

a. If we have a two-tailed hypothesis, our range of z scores would be from

–1.96 to +1.96.
b. If we have a one-tailed “greater than” hypothesis, our range of critical z
scores would include everything to the left of zero and everything to the
right of zero up to and including +1.645.
c. If we have a one-tailed “less than” hypothesis, our range of critical z
scores would include everything to the right of zero and everything on
the left of zero down to and including –1.645.

6. At this point we compare our computed value of z to our critical value of z

and either reject or fail to reject our null hypothesis.

Be Careful When Changing Your Alpha Values

We saw this same heading earlier in this chapter—it almost makes you
think we need to pay close attention when we are changing alpha val-
ues, doesn’t it? While we do usually use an alpha value of .05, there are
instances where we might want to change. When we do, we have to think back to the
idea of Type I and Type II errors. Let’s include the same table we used back then but
add another column on the right showing the probability of a Type I or Type II error;
this is shown in Table 5.16.
 Choosing the Right Statistical Test  |  147

TABLE 5.16.  Tying Together the z Score, Alpha Value, Confidence Interval, and Error
Probability
Lower Upper
Confidence end of Mean end of Width Hypothesis error
interval Alpha z score CI of CI CI of CI probability
90% .10 1.645 105.50 110 114.50 9.0 Higher probability of
Type I error
Lower probability of
Type II error
95% .05 1.96 104.63 110 115.37 10.74 Acceptable probability
of Type I and Type II
errors
99% .01 2.575 102.94 110 117.06 14.12 Higher probability of
Type II error
Lower probability of
Type I error

Remember, using this table we demonstrated that higher alpha values lead to
smaller z scores and confidence intervals. Let’s tie those ideas together with the idea of
hypothesis testing so we can understand how these same values affect the probability
of Type I and Type II errors.
When we test a hypothesis and use a larger alpha value, we create a narrower con-
fidence interval. This means we are creating a smaller range of values that we would
consider not significantly different from the value we are interested in. In this case,
if we have a two-tailed hypothesis and use an alpha value of .10, we say that any value
from 105.50 to 114.50 is not significantly different from our average of 110 and we
would fail to reject the null hypothesis based on that value. Anything outside of that
range would be considered significantly different and would cause us to reject our
null hypothesis. It is easy to see, then, that a larger alpha gives us a far better chance
of rejecting the null hypothesis. Unfortunately, this also means that we are increasing
our Type I error rate; we may be rejecting a null hypothesis when we should not.
At the other end of the spectrum, if we decrease our alpha value to .01, we are
greatly widening the range of values we would consider not significantly different
from our mean score. In this case, we would consider anything from 102.94 to 117.06
to be different from 110 strictly due to chance. This, of course, lowers our probability
of rejecting the null hypothesis and greatly increases the probability of a Type II error;
we might fail to reject the null hypothesis when we should.
As we said earlier, at this point a good consumer of statistics knows two things.
First, we cannot count on being 100% right in our statistical decision making because
of random error. Second, because of that, we can control our level of risk by using an
appropriate alpha value. As we just saw, having too large or too small of an alpha value
creates problems; a good consumer of statistics will usually use an alpha value of .05
and its acceptable probability of making either a Type I or Type II error.

The Heart of Inferential Statistics

We have seen, in just a few pages, the very heart of inferential statistics. In this case,
using a one-sample z-test, we tested a hypothesis about a population mean using infor-
148 | STATISTICS TRANSLATED

mation from both the population and the sample. In upcoming cases we will com-
pare two samples, three samples, and so on. Although the different tests we will use
(t-tests, ANOVAs, etc.) have different formulas, the basic premise of hypothesis testing
remains the same.
As easy as this is, I am the first to admit that calculating all these values, using
the tables, and plotting the results can be somewhat tedious. Given that, let’s look at
an alternate way of making decisions about our hypotheses. I believe you’ll find this
way a lot easier.

Probability Values
Instead of working through all those calculations, statisticians use statistical software
to compute a p value. This p value is the probability that a particular outcome is due to
chance. In the case we just discussed, another way of describing the p value is to call it
the probability of our university’s salaries being equivalent to the salaries from other
universities. These p values range from 0.00 (no probability that the sample mean
came from the population being considered) to 1.00 (an absolute certainty that the
sample mean came from the population being considered). Obviously, if we are trying
to reject the null hypothesis, we want as small a p value as possible. This will help us
ensure that any differences we find are “real” and not due to chance.
Rather than testing a hypothesis by computing an observed value of z, finding a
critical value of z, and then comparing the two, statisticians can just compare their
computer-generated p value to their predetermined alpha value. For a particular case,
if the computed p value is less than the alpha value, a researcher will reject the null
hypothesis; this means the differences are significant and not attributable to chance.
If the p value is greater than or equal to the alpha value, the statistician will fail to
reject their null hypothesis. This means any differences found are not significant or
are attributable only to chance. Table 5.17 shows an example of these ideas.

TABLE 5.17.  Tying Together the z Score and the p Value

Computed Critical
value of z value of z Alpha Reject null hypothesis? p value
2.00 1.645 .05 Yes—the computed The p value has to be
value of z is greater less than alpha since we
than the critical value are rejecting the null
of z. hypothesis.
1.5 1.645 .05 No—the computed The p value has to be
value of z is less than greater than or equal
the critical value of z. alpha because we are
not rejecting the null
hypothesis.
1.645 1.645 .05 No—the computed The p value has to be
value of z is equal to greater than or equal
the critical value of z. alpha because we are
not rejecting the null
hypothesis.
 Choosing the Right Statistical Test  |  149

A Few More Examples

Look at the following hypotheses and their alpha and p values. Based on them, let’s
decide whether we need to reject or fail to reject our null hypothesis.

„ There will be a significant difference in achievement between students who

eat breakfast and students who do not eat breakfast.

The null hypothesis would read:

„ There will be no significant difference in achievement between students who

eat breakfast and those who do not eat breakfast.

Suppose we have collected achievement scores for a random sample of students

and then asked them to tell us whether they eat breakfast. After using a computer to
calculate the descriptive statistics, we have found the average score for students eating
breakfast is 82 while that of students skipping breakfast is 79. So far, it looks like we
might be supporting our research hypothesis, but we cannot be sure.
Using our standard alpha value of .05, let’s assume the output from that proce-
dure shows a p value of .05; what would we do? Looking at the table, we only reject
the null hypothesis if the p value is less than the alpha value. In this case, p is equal
to alpha, so we will fail to reject the null. Even though the mean scores are different,
apparently eating breakfast does not significantly affect achievement in school one
way or the other.
Suppose we had the following research hypothesis:

„ There will be a significant difference in scores on the Beck Depression

Inventory between patients receiving behavioral therapy and those receiving
cognitive therapy.

Our null hypothesis is:

„ There will be no significant difference in scores on the Beck Depression

Inventory between patients receiving behavioral therapy and those receiving
cognitive therapy.

At this point, we are going to make up a pretty far-fetched case study. Let’s sup-
pose we have a pool of depressed patients. We randomly select one group to receive
cognitive therapy and the other to receive behavioral therapy. After an appropriate
period, say 10 sessions with the psychologist, we ask each of the patients to complete
the Beck Depression Inventory. After computing the descriptive statistics, we see that
the cognitively treated patients have a mean score of 15 while the behaviorally treated
patients have a mean score of 30.
Things are looking pretty good for the cognitive therapists, aren’t they? Let’s
use our alpha value of .05, and, using the data we collected, our software computes
a p value of .01. Since this is less than our alpha value of .05, obviously the cognitive
150 | STATISTICS TRANSLATED

therapists know something the behaviorists do not; their clients’ scores are signifi-
cantly less than those of the behaviorists’ clients.
Let’s look at one last research hypothesis:

„ There will be a significant difference in elementary school achievement

between students attending public school and those being home-schooled.

Here is the corresponding null hypothesis:

„ There will be no significant difference in elementary school achievement

between students attending public school and those being home-schooled.

In order to compare levels of achievement, let’s use the results of a standardized

test that the students take at the end of each school year. The results show a mean
score of 56 for the public-school students and 62 for the home-schooled students. This
time, let’s set alpha equal to .10 and find that p is equal to .11—what would you do? In
this case, p is greater than our alpha value, so we fail to reject our null hypothesis. It
is apparent that it makes no difference where students attend elementary school—they
all do equally as poorly on the standardized tests!
Everything we have talked about concerning the p value is summarized in Table
5.18.

TABLE 5.18.  The Relationship between the p Value and a Decision Made about
a Hypothesis
What happens to the What happens to the
Situation null hypothesis? research hypothesis?
Computed p is greater than alpha. We fail to reject the null We do not support the
  or hypothesis. Any observed research hypothesis. Any
Computed p is equal to alpha. differences may be due to observed differences may
random sampling error. be due to sampling error.

Computed p is less than alpha. We reject the null We support the research
hypothesis. The observed hypothesis. The observed
differences are probably differences are probably
not due to sampling error. not due to sampling error.

Great News—We Will Always Use Software to Compute Our p Value

Although it is important to understand why and how we perform the manual calcula-
tions we have just gone through, we should always rely on a statistical software pack-
age to compute our p value. Not only is p difficult to compute by hand but using the
software will eliminate the probability of us making errors in our math. Starting in
the next chapter, you’ll see how easy it is to use p when testing our hypotheses. We are
about to finally learn to identify the appropriate statistical test we need in order to
test a given hypothesis but, before we do, let’s check our knowledge of this last section.
Look at Tables 5.19, 5.20, and 5.21. In each, you will see a null and research
hypothesis, data values for the population and sample, an alpha value, and the stan-
 Choosing the Right Statistical Test  |  151

dard error of the mean. Use these values to compute the appropriate z score, obtain
the critical z score, and then determine whether you should reject the null hypothesis.
Finally, based on everything, determine if the p value would be less than .05. Again,
you do not have to compute a p value; just state whether it would be less than based on
whether you would reject the null hypothesis based on the other computed values. In
the last line of the table, support your decision about the p value. I have supplied the
answers at the end of the book so you can check your work.

TABLE 5.19.  Intelligence—Is There a Significant Difference?

Null hypothesis: There will be no significant difference in the intelligence
level between our class and the overall university level.
Research hypothesis: The average intelligence level of our class will be
significantly higher than the average intelligence level of
all other classes in the university.
Population average: 100
Class average: 105
SEM: 3
Alpha: .05
Computed z value:
Critical z value:
Reject null hypothesis?
How do you know?
Is p less than .05?
How do you know?

TABLE 5.20.  Rainfall—Is There a Significant Difference?

Null hypothesis: There will be no significant difference in the yearly
amount of rainfall between Texas and Florida.
Research hypothesis: There will be a significant difference in the yearly
amount of rainfall between Texas and Florida.
Texas average: 70 inches
Florida average: 62 inches
SEM: 10 inches
Alpha: .05
Computed z value:
Critical z value:
Reject null hypothesis?
How do you know?
Is p less than .05?
How do you know?
152 | STATISTICS TRANSLATED

TABLE 5.21.  Truancy—Is There a Significant Difference?

Null hypothesis: There will be no significant difference in the number
of truancy cases at our school and the district average
number of truancy cases.
Research hypothesis: There will be a significant difference in the number
of truancy cases in our school and the district average
number of truancy cases.
District average: 12 cases
Our school’s average: 17 cases
SEM: 2 cases
Alpha: .05
Computed z value:
Critical z value:
Reject null hypothesis?
How do you know?
Is p less than .05?
How do you know?

S TEP 5 Choose the Right Statistical Test

It is finally time to learn how to choose the right statistical test
for a given situation. Although we are going to go about it in
an easily understood manner, it has been my experience that
this is one of the most difficult things for the beginning stat-
istician to do. As I said before, one might choose from literally hundreds of different
parametric and nonparametric procedures. We will only look at the small number of
tests you’ll probably use.

You Already Know a Few Things

First, you already know a few things about choosing the right statistical procedure
for the data you have collected. For instance, we talked about how, if the data you
collected for your dependent variable are either nominal or ordinal, we will use non-
parametric procedures. If the data are quantitative (interval or ratio level), most of
the time we will use a parametric test. In rare instances, if we have quantitative data
that are not normally distributed, we will use a nonparametric procedure. We will talk
more about this later in the book.
After we have determined the type of data we are collecting, we can use the
number of independent variables, the number of levels of each of the independent
variables, and the number of dependent variables to select the appropriate statistical
test from Table 5.22. (This table is also found in Appendix G.)
 Choosing the Right Statistical Test  |  153

TABLE 5.22.  Choosing the Right Statistical Test

Number of
Number of levels in the Number of Type of data the
independent independent dependent dependent variable Alternate
variables variable variables represents Statistical test to use statistical test**
N/A* N/A* 1 Quantitative One-sample z-test or N/A
one-sample t-test
1 2 1 Quantitative Dependent-sample Wilcoxon
t-test or independent- t-test or Mann–
sample t-test Whitney U test
1 3 or more 1 Quantitative Analysis of variance Kruskal–Wallis
(ANOVA) H test
2 2 or more 1 Quantitative Factorial ANOVA N/A
1 or more 2 or more 2 or more Quantitative Multivariate ANOVA N/A
(MANOVA)
0*** 0 0*** Quantitative Pearson correlation N/A
0*** 0 0*** Ordinal Spearman N/A
correlation
1 or more 2 or more 1 or more Nominal Chi-square N/A
* Not all parametric tests have a nonparametric equivalent, or the test mentioned is already a nonparametric
test.
** When you’re working with one-sample z-tests or one-sample t-tests, it is really not logical to point out the
independent variable or its levels. As we have just seen with the one-sample z-test, we are just comparing a
sample statistic to a population parameter. We will see the same holds true, albeit with smaller sample sizes,
when we visit the one-sample t-test in the next chapter.
***In correlational procedures, you simply want to look at the relationship between two variables; these are
called predictor and criterion variables.

A Couple of Notes about the Table

Table 5.22 is very straightforward; it takes information from each of the steps and
shows us the appropriate statistical test for a given situation. You will notice, however,
that there are three places marked with asterisks. We need to discuss these before
moving forward.
In the second row from the top, you can see that the one-sample t-test and the
one-sample z-test are marked with one asterisk; this indicates that it is not logical to
consider either the number of independent variables or their levels. You are simply
comparing a sample statistic to a population parameter.
The rightmost column of the table, labeled “Alternate Statistical Test,” is marked
with two asterisks. As you can see, alternate statistical tests are listed in only a few
of the rows; this is because not every parametric test has a corresponding nonpara-
metric test, and, when one does exist, they are used only when a statistician has col-
lected quantitative data that are not normally distributed or in instances where rank-
level data were collected. These two things are generally exceptions to the rule. Most
consumers of statistics will seldom be asked to use these nonparametric tests. I will
explain more about this when we get to the specific tests.
You can see that the third place marked with asterisks is in the row that shows
the Pearson and Spearman correlation procedures. You do not have independent and
dependent variables in correlations; instead you have predictor variables and criterion
154 | STATISTICS TRANSLATED

variables. While these names are different, they are used in a manner somewhat like
their independent and dependent cousins. We will discuss this in detail when we get
to that chapter.

Summary of Step 5: Choose the Right Statistical Test

The table we can use to select the statistical test we need is very straightforward and
includes the tests you’ll need to use, most of the time. Notice there is no mention of
tests such as the analysis of covariance (ANCOVA), multiple regression, structural
equation modeling or any of the other types of statistical tests that you, the consumer
of statistics, will hardly ever use.
Now we need to move forward and discuss the use of the statistical tests identi-
fied in the table. We will start each chapter with a brief overview of the statistical tool
being considered and then use case studies to help work though each of the six steps
in our model.

Do You Understand These Key Words and Phrases?

alpha value a priori
beta central limit theorem
confidence interval level of significance
lower confidence limit mean of means
p value sampling distribution of the means
sample standard error of the mean significantly different
statistic standard error of the mean (SEM)
Type I error Type II error
z score

Do You Understand These Formulas and Symbols?

Confidence interval:
 σ 
x ± (za 2 ) 
 n 

Mean of the sampling distribution of means:

mx

Population standard error of the mean:

σ
σx =
n
Sample standard error of the mean:
s
sx =
n
 Choosing the Right Statistical Test  |  155

z score for sampling distribution of the means when the population mean is
known:

z=
(x − m)
σx
Remember: In the z score formula, both the population mean and the standard
error of the mean can be replaced by the mean of means or the sample standard
error of the mean if needed.

Quiz Time!

Using the decision table, identify the statistical test that should be used to test each of the fol-
lowing hypotheses and then explain why you chose that test.

1. There will be no significant difference in the number of fish caught in lakes,

streams, or rivers.

2. Interest in science professions will be significantly higher in young females after

they’ve attended an after-school program than before they attended the after-school
program.

3. Employees assigned to well-lighted offices will have higher levels of productivity

than employees assigned to dimly lighted offices.

4. There will be a significant difference in health problems between citizens of

Canada, citizens of the United States, and citizens of Mexico.

5. There will be a significant difference in height and weight between persons raised
in an urban setting and persons raised in a rural setting.

6. The miles per gallon demonstrated by foreign-manufactured cars is significantly

different from 19.

7. There will be a significant difference in the number of children between families

labeled as lower, middle, and high socioeconomic status.

8. Audience appreciation for shows appearing on Broadway will be significantly lower

than the audience appreciation of the touring version of the same show.

9. There will be a significant correlation between the number of siblings and the
annual income of their parents.

10. There is a significant difference in government apathy between voters labeling

themselves as Republicans, voters labeling themselves as Democrats, and voters
labeling themselves as Liberals.

11. There will be no significant difference in the percentage of males and females in
our office building when compared to the national percentage.

12. There will be a significant difference in authors’ procrastination levels before being
called by their editor and after they have been called by their editor.
CHAPTER 6

The One-Sample t-Test

Introduction
We are finally up to the point where we can choose our statistical test, use SPSS to
analyze our data, and interpret our hypothesis based on the SPSS report. From here
through the end of the book, we will dedicate a chapter to each of the tests in our
table. We’ll work through a lot of cases, going through the entire six-step process, to
ensure we know when and how to use each test.
In the last chapter, we learned to use a statistic from a large sample of data to test
a hypothesis about a population parameter. In our case, using a z-test, we tested a
hypothesis about a population mean using the mean of one sample drawn from that
population. We agreed that by “large sample” we mean any sample that has 30 or
more data values. What happens, however, when we have a sample with less than
30 items? We cannot use the one-sample z-test, so what can we do? The answer
is, we can use its counterpart, the one-sample t-test. The logic underlying the one-
sample t-test is very similar to that of the one-sample z-test and is easily understood.
You’ll find that this material is critical to understanding the next two chapters, on
the independent-sample t-test and the dependent-sample t-test. For now, let’s take
a trip.

Welcome to the Guinness Breweries

All the sudden we’ve taken a dramatic turn, haven’t we? We have gone from discussing
statistics to an apparent trip to a brewery! I hate to burst your bubble, but we are only
going to visit the brewery to meet a man named William Gossett.
Gossett, a chemist, worked for Guinness in the early 1900s. His job involved test-
ing samples of ale to ensure quality. He decided it was too costly and time consuming
156
 The One-Sample t-Test | 157

to use large samples (i.e., 30 or more), so he was determined to develop a statistical

tool that would allow him to use smaller samples and still be able to test hypotheses
about a population parameter (i.e., the batch of ale). We can start understanding his
work by looking at the t distribution.

The t Distribution
We’ve seen, in prior chapters, when we had large (i.e., greater than 30), mound-shaped
quantitative data distributions, the empirical rule showed us how to understand the
distribution of the data. We were able to locate the relative position of a data point
in terms of the standard deviation by computing z scores using the sample mean, the
population mean, and the standard error of the mean. If we did not know the stan-
dard deviation of the population and therefore could not compute the standard error
of the mean, we also saw that we could use the standard deviation of the sample to
compute the sample standard error of the mean. This is a very close approximation
of the same population parameter and can be used in our formula to compute the z
score.
Gossett, while experimenting with possible solutions, noticed something interest-
ing about the mound-shaped distribution of data values when the sample size was
less than 30. Each time he decreased the sample size by one and plotted the means
of repeated samples of the same size, the shape of the distribution flattened out. You
can see the plots for 30 sample means, 25 sample means, and 20 sample means in
Figure 6.1.
By looking at the picture we can
see that, when the means of individual
samples with 30 sample means are plot-
ted, the distribution looks like a normal
curve. When we use only 25 sample
means, the sampling distribution gets
flatter and more spread out; this is even
more evident when we have only 20
sample means. In other words, the fewer
the number of data values you have,
the more spread out on both ends (i.e.,
platykurtotic) our distribution is.
FIGURE 6.1.  Changes in distribution shape based Using this idea, Gossett further dis-
on number of samples plotted. covered that the empirical rule applied
to this distribution and, if you compen-
sate for the number of data values less than 30, you can test a hypothesis by compar-
ing a computed value of t to a critical value of t. This is exactly what we did when we
compared a computed value of z to a critical value of z in the z-test. This was the key, he
determined, to testing a population parameter using only a relatively few data values.

Putting This Together

As I just said, when we have samples with less than 30 members, we can test a hypoth-
esis using a computed value of t just like we used the computed value of z. In order to
158 | STATISTICS TRANSLATED

better understand this, let’s state a hypothesis and then use a dataset with 15 values
to test it:

„ Persons living in urban environments have anxiety levels significantly

different from the national average.

Let’s use a test, designed to measure anxiety, with scores ranging from zero (no
anxiety) to 80 (high anxiety). The 15 scores in Table 6.1 serve as an example so we can
continue our discussion. Since we already know how to use SPSS to compute descrip-
tive statistics so we can move past that part; Figure 6.2 shows the output.

TABLE 6.1.  Anxiety Scores

34 32 28 32 35
32 28 30 39 32
29 33 34 32 30

Descriptive Statistics

N Mean Std. Deviation Variance

Statistic Statistic Std. Error Statistic Statistic

Anxiety 15 32.0000 .74322 2.87849 8.286
Valid N (listwise) 15
FIGURE 6.2.  Descriptive statistics.

The formula for the computed value of t is like the formula for the computed
value of z; we just need to include different values:
x −µ
t=
sx
As always, x represents the mean of the dataset—in this case 32. From the descrip-
tive statistics in Figure 6.2, we see the standard error of the sample (i.e.,  µ x) is .74322.
In this case, let’s assume the mean of our population (i.e., μ) is 30; this is the value
that we want to compare our sample mean against. We can insert these values into our
equation and calculate a t value of 2.691 using the following three steps:

32 − 30
1. t =
.74322
2
2. t =
.74322

3. t = 2.691
 The One-Sample t-Test | 159

Let’s use SPSS to verify our calculations. We have entered our data in Figure 6.3,
selected Analyze, Compare Means, and One-Sample T Test. We then select the Anxi-
ety value and set the Test Value to 30, shown in Figure 6.4; this represents the mean
value of our population. Clicking on OK causes SPSS to run the one-sample t-test,
comparing the mean score of our data to the population mean of 30, and will produce
Figure 6.5.

FIGURE 6.3.  Selecting the Compare Means and One-Sample T-Test command.

We are familiar with these terms, so let’s not go into detail just yet; we will refer
to our p value, shown as Sig. 2-tailed by SPSS, and the mean difference later. For now,
the value that we are most interested in is 2.691, our computed value of t. This is the
value we will compare to a critical value of t in order to test our hypothesis.

Determining the Critical Value of t

Just as we did with the z score, we will determine our critical value of t using a table
developed specifically for that purpose. The table showing the area under the curve
for t values (i.e., a t table) will look somewhat different from that for z values since it
must compensate for the fewer number of items in the sample. We compensate for the
smaller sample sizes by using a value called degrees of freedom.

Degrees of Freedom
Throughout the remainder of the book, most of the inferential statistics we compute
will include degrees of freedom; this is the number of values in the dataset we’ve col-
160 | STATISTICS TRANSLATED

FIGURE 6.4.  Identifying the Test Variable and Test Value for the One Sample T-Test.

lected that are “free to vary” when you try to estimate any given value. For example,
suppose we have a one-sample t-test where we know the population mean is 10. I might
ask you to tell me five numbers that, when summed, would give me an average of 10.
You might start out by saying “6, 8, 10, 12 . . .” but then I would have to interrupt you
by saying, “That is enough, if I know the first 4 numbers, I can calculate the fifth.”
Here’s how I know. In order to compute a mean of 10 with a dataset of five values,
the sum of the values in the dataset would have to equal 50 (i.e., 50/5 = 10). We could
set up an equation to show the relationship between the values we know and the sum
we need; x is used in the equation to represent the value we do not know:

6 + 8 + 10 + 12 + x = 50

A little simple math would show the sum of the four values we know is 36:

6 + 8 + 10 + 12 = 36

One-Sample Test

Test Value = 30

95% Confidence Interval of the

Difference
t df Sig. (2-tailed) Mean Difference Lower Upper

Anxiety 2.691 14 .018 2.00000 .4059 3.5941

FIGURE 6.5.  Inferential statistics from a One-Sample Test.

 The One-Sample t-Test | 161

We can subtract that value from 50 and see the value we need to complete our
equation is 14:

50 – 36 = 14

In order to double-check our work, we can then put these numbers together and
do the required math:

6 + 8 + 10 + 12 +14 = 50
And then:

50/5 = 10

In short, we’ve shown that, if we know the mean and all the values in the dataset
except one, we can easily determine the missing value. In this case, four of the values
in the dataset can vary, but the fifth one cannot. By definition then, our degrees of
freedom are 4, that is, (n – 1).

Be Careful Computing Degrees of Freedom

Although it is just that easy to compute the degrees of freedom for a
one-sample t-test, you must be careful. As we will see in later chapters,
different statistical tests compute degrees of freedom in different ways;
it depends on the number of items in your dataset and the number of levels of the
independent variable. For now, let’s use our degrees of freedom to look at the critical
values of t table (Table 6.2). An example of a complete critical values of t is shown in
Appendix B.

TABLE 6.2.  Critical Values of t

df a = .05 a = .025 df a = .05 a = .025 df a = .05 a = .025
1 6.314 12.70 11 1.796 2.201 21 1.721 2.080
2 2.920 4.303 12 1.782 2.179 22 1.717 2.074
3 2.353 3.182 13 1.771 2.160 23 1.714 2.069
4 2.132 2.776 14 1.761 2.145 24 1.711 2.064
5 2.015 2.571 15 1.753 2.131 25 1.708 2.060
6 1.943 2.447 16 1.746 2.120 26 1.706 2.056
7 1.895 2.365 17 1.740 2.110 27 1.703 2.052
8 1.860 2.306 18 1.734 2.101 28 1.701 2.048
9 1.833 2.262 19 1.729 2.093 29 1.699 2.045
10 1.812 2.228 20 1.725 2.086 30 1.697 2.042

Although the t table looks a lot like the z table, there are a few differences. First,
you can see degrees of freedom abbreviated as df. Next, there are three highlighted
columns. The leftmost highlighted column shows degrees of freedom from 1 through
10, the middle, highlighted column shows degrees of freedom from 11 through 20,
162 | STATISTICS TRANSLATED

and the rightmost column shows degrees of freedom from 21 through 30. To the right
of each of these values, you can see the critical value of t for that given degree of free-
dom; this value is shown for both alpha = .05 and alpha = .025.
Before we move forward, one thing you may have noticed is there is no column
for α = .01; I did that on purpose simply to get the idea across. This is just an example
of the critical values of the t table. You’ll find that others have all the values of t in one
column; others include different alpha values, and so on. For now, we’ll just use what
we have.
In order to use the t table, let us suppose we have a scenario with 12 degrees of
freedom and want the critical value of t for alpha = .05. All we need to do is find 12
in the table and look directly to the right in the column labeled alpha = .05. As you
can see, the critical value of t is 1.782. In another case, we might have 29 degrees of
freedom and want the critical value of t when alpha is .025. Again, we would find the
appropriate row and see that the critical value of t is 2.045.

Let’s Get Back to Our Anxiety Hypothesis

Remember we are dealing with a two-tailed (i.e., nondirectional) hypothesis. Just as we
did with the z-test, this means we must divide our alpha value by 2 (i.e., .05/2 = .025).
Our degrees of freedom are 14 (i.e., our sample size of 15 minus 1). Using these values,
the table shows we need to use a critical value of t equal to 2.145; let’s plot that value
on our curve and test our hypothesis.

Plotting Our Critical Value of t

In our earlier discussion, we agreed that the fewer the number of items in a dataset,
the flatter the curve will be. In Figure 6.6, we have a fairly flat curve, and have marked
the critical values of t on both ends.
Since our computed value of t (i.e., 2.691) is not within the range of the critical
values, we can reject the null hypothesis. By looking at our descriptive statistics, it
appears that the anxiety levels of people living in urban areas may be significantly
higher than the national average. Let’s check that, however, with our output from
SPSS that was shown in Figure 6.5.
We can see here that our p value, labeled “Sig. (2-tailed)” is .018; this is less than
our alpha value of .025. This verifies the decision we made when we compared the
critical value of t to the computed value of t; we must reject our null hypothesis and
support our research hypothesis.

The Statistical Effect Size of Our

Example
Although we’ve rejected the null hypoth-
esis, some statisticians feel this isn’t
enough. At this point, we know the
groups being compared are significant-
FIGURE 6.6.  The critical value of t for a two- ly different, but we know nothing about
tailed alpha value of .05. the degree to which they are different.
 The One-Sample t-Test | 163

In order to address this issue, statisticians have developed tools, called effect size indi-
ces, which help us better understand the magnitude of any significant difference we’ve
uncovered. An easy way to think about the effect size is that it helps us better under-
stand the extent to which the independent variable affects the dependent variable;
this is sometimes called the practical significance. In this case, the effect size for the
one-sample t-test, called Cohen’s delta (i.e., d), is computed by subtracting the mean
of the population from the mean of our sample; we then divide the remainder by the
standard deviation of our sample.

x −µ
d=
s
Let’s replace the symbols in our formula using the values from above:

32 − 30
1. d =
2.87849
2
2. d =
2.87849

3. d = .695

This leaves us with an effect size of .695, but what does that mean?
First, it’s plain to see that the computed effect size is nothing more than the per-
centage of the standard deviation that the difference in the mean scores represents.
In discussing this difference, Cohen defined effect sizes as being “small” (i.e., .2 or
smaller), “medium” (i.e., between .2 and .5), or “large” (i.e., greater than .5). In this
case, we have a large effect size; the groups are significantly different, and the levels
of the independent variable had a dramatic effect on the dependent variable. In short,
there’s a strong relationship between where a person lives and their level of anxiety.
As you will see in the following chapters, different inferential tests require effect
sizes to be computed using different values and different formulas. Even so, inter-
preting them will be about the same. Because they contribute so much to our goal
of becoming a good consumer of statistics, we will make computing the effect size a
normal part of our descriptive statistics from now on.

Let’s Look at a Directional Hypothesis

Let’s move forward with our discussion by looking at the following one-tailed hypoth-
esis:

„ Men who use anabolic steroids will have significantly shorter life
expectancies than men who do not.

In this case, suppose we know the average life expectancy for men is 70 but we
only have access to information on 12 men who used steroids of this type. Their age
164 | STATISTICS TRANSLATED

of death is shown in Table 6.3. Using this data, SPSS would compute the descriptive
statistics shown in Figure 6.7. Since SPSS does not compute the effect size for a one-
sample t-test, we can use these statistics, along with the formula, and compute an
effect size of –.535:

68 − 70
1. d =
3.742
−2
2. d =
3.742

3. d = .535

TABLE 6.3.  Steroid Users’ Ages at Death

64 66 77 67
66 66 70 66
68 72 64 70

We are interested only in the absolute value, so we drop the negative sign and
wind up with an effect size of .535. SPSS would then compute the inferential statistics
in Figure 6.8.

Descriptive Statistics

N Mean Std. Deviation

Statistic Statistic Std. Error Statistic

Age 12 68.0000 1.08012 3.74166

Valid N (listwise) 12
FIGURE 6.7.  Descriptive statistics from a One-Sample Test.

One-Sample Test

Test Value = 70

95% Confidence Interval of the

Difference
t df Sig. (2-tailed) Mean Difference Lower Upper

Age -1.852 11 .091 -2.00000 -4.3773 .3773

FIGURE 6.8.  Inferential statistics from a One-Sample Test.

Here our computed value of t is –1.852, and our critical t value is 1.796 for alpha
= .05. Remember, though, the table doesn’t show negative values, so, if we are going
to test a one-tailed “less than” hypothesis, we need to put a negative sign in front of
our value; this means we wind up with –1.796. Since we have a one-tailed hypothesis,
 The One-Sample t-Test | 165

as seen in Figure 6.9, we would then plot our critical value on the “less than” side of
the distribution.
Our computed t value, –1.852, is
less than the critical value of t (i.e.,
–1.796). Because of that, we reject
our null hypothesis and support our
research hypothesis. This is supported
by a large effect size of .535; the inde-
pendent variable does have a large effect
on the dependent variable. Apparently,
men who use anabolic steroids live sig-
FIGURE 6.9.  Comparing the computed and nificantly fewer years than their peers
critical values of t for a one-tailed alpha value
of .05.
who do not.
Just as was the case with the z-test,
this is very easy to understand. To make
things even better, just as we did with the large samples, we can use our p value to help
us test our null hypothesis.

Using the p Value

As we saw earlier, the p value for our anxiety hypothesis was .018, thereby supporting
our decision to reject the null hypothesis and support the research hypothesis. Since
we were dealing with a nondirectional hypothesis, that was as far as we needed to take
it. In this case, we are dealing with a directional hypothesis; because of that, we need
to stop and talk about issues related to p values and one-tailed tests.
We know that when we have a two-tailed hypothesis, we must allow for error
on both sides of the mean; in order to do that, we must divide our alpha value by 2.
When we have a one-tailed (i.e., directional) hypothesis, we use the entire alpha value
on one side of the distribution or the other, depending on the direction stated in the
hypothesis.
A directional hypothesis also affects how we interpret our p value. SPSS only com-
putes the value for a nondirectional hypothesis. That means our Sig. (2-tailed) value
of .091 in Figure 6.8 represents the p value for a two-tailed test. Since we are testing a
directional hypothesis, we must divide our p value by 2, leaving us with a one-tailed p
value of .0455. This means we now have a p value less than our alpha value of .05 and
supports our decision to reject the null hypothesis. It is interesting to note, however,
that had we stated a nondirectional hypothesis and used the corresponding two-tailed
value (i.e., .091), we wouldn’t have rejected the null hypothesis since it is greater than
our alpha value of .05. Be aware, however, that the way the p value is presented differs
between software packages; some compute only the two-tailed p value, some only the
one-tailed value, and others compute both. In order to accurately test your hypoth-
eses, make sure you know which value you are dealing with.

Check Your Mean Scores!

When we reject a null hypothesis, we always need to ensure the mean scores are in
the order hypothesized. In this case, the average life span of males using steroids was
166 | STATISTICS TRANSLATED

68, two years less than the average life span of 70 and in the order we hypothesized.
If, however, the average life span of males using steroids was 72, we would still have a
difference of two years and we would reject the null hypothesis. In that case, however,
since the values are not in the order hypothesized, we would not support the research
hypothesis. We can see this in Table 6.4.

TABLE 6.4.  Using the p Value to Test the Steroid Use Hypothesis
Average
lifespan Steroids Difference p Decision
70 68 –2 .0455 Reject null hypothesis.
Support research hypothesis.
70 72 +2 .0455 Reject null hypothesis.
Fail to support research hypothesis.

One More Time

In the last chapter, when we talked about using the z score to test hypotheses, we
compared certification exam scores from nursing students at ABC University to the
national average of 800. In that case, we were concerned whether our students’ scores
were significantly less than or significantly greater than the national average. Given
that, we used a two-tailed, or nondirectional, hypothesis. Let’s use the same scenario
but imagine we are making the argument that the scores of our students are signifi-
cantly greater than the national average of 800. Because of that, we would use a one-
tailed “greater than” hypothesis:

„ Nursing students at ABC University will have significantly higher scores on

the national certification exam than nursing students at other universities.

Let’s use the data in Table 6.5 to test our hypothesis. Figure 6.10 shows the descrip-
tive statistics from SPSS. Based on these descriptive statistics, we would compute a
very large effect size of 1.4 and a t value of 7.426. To test our hypothesis, all we need to
do is compare our computed value of t, shown in Figure 6.11, to the critical value of t.

TABLE 6.5.  Certification Exam Scores

810 800 810 810 810
810 815 815 815 815
810 815 820 830 830
840 835 840 835 850
848 842 844 811 865

Unlike the nondirectional hypothesis, where we had to divide our alpha value
by 2, here we have a one-tailed hypothesis, so we will determine our critical value of
t using the entire alpha value of .05. We can determine our degrees of freedom by
subtracting 1 from the sample size and then, by looking back at Table 6.2, see that our
 The One-Sample t-Test | 167

Descriptive Statistics

N Mean Std. Deviation

Statistic Statistic Std. Error Statistic

Certification 25 825.0000 3.36650 16.83251

Valid N (listwise) 25
FIGURE 6.10.  Descriptive statistics from a One-Sample Test.

One-Sample Test

Test Value = 800

95% Confidence Interval of the
Difference
t df Sig. (2-tailed) Mean Difference Lower Upper
Exam Score 7.426 24 .000 25.00000 18.0519 31.9481

FIGURE 6.11.  Inferential statistics from a One-Sample Test.

critical value of t is 1.711. Since our critical value of t is much less than our computed
value of t (i.e., 7.426), we must reject our null hypothesis and support our research
hypothesis. Based on this, it is apparent that our nursing students do score signifi-
cantly higher on the national certification exams than those students at other univer-
sities. This, again, is supported by the large effect size. The relationship between the
computed and critical values of t is shown in Figure 6.12.
Just as was the case earlier, we need
to ensure our mean scores are in the
direction hypothesized. In this case, the
average score of our nursing students
is 825, a difference from the mean of
+25. Let’s compare our current results
against the results had our students
averaged 775 (Table 6.6).

FIGURE 6.12.  Comparing the computed and Important Note about Software
critical values of t for a one-tailed alpha value Packages
of .05.
As we saw in our example output, SPSS
only gives us the two-tailed p value.
When testing a directional hypothesis, that means we had to divide our p value by 2
prior to comparing it to our alpha value. Remember, this is not the case with all soft-
ware packages; some will supply both the one-tailed and two-tailed values. These ideas
are summarized in Table 6.7.
168 | STATISTICS TRANSLATED

TABLE 6.6.  Using the p Value to Test the Certification Exam Score Hypothesis
All
universities ABC Difference p Decision
800 825 +25 .0000 Reject null hypothesis.
Support research hypothesis.
800 775 –25 .0000 Reject null hypothesis.
Fail to support research hypothesis.

TABLE 6.7.  The Relationship between Hypothesis Type, Alpha, and p

Type of hypothesis Alpha p
Nondirectional (two-tailed) Divide by 2. Use entire value.
Directional (one-tailed) Use entire value. Divide by 2.

Let’s Use the Six-Step Model!

Now that we know how to test a hypothesis by both comparing a computed value of t to
a critical value of t and comparing a computed p value to a pre-established alpha value,
let’s put this all together and use our six-step model. We’ll look at three separate cases
to make sure we have the hang of it; let’s get started with ambulance drivers.

The Case of Slow Response Time

Residents in a lower socioeconomic section of town recently complained to local health-
care officials that ambulance response times to their neighborhood were slower than
the average response time throughout town. This, of course, worried the officials, and
they decided to investigate. Their records showed the average response time through-
out the city to be 90 seconds after an emergency was phoned in. They decided to moni-
tor the next 20 response times to the concerned neighborhood to help determine if
the complaint had any merit.

S TEP 1 Identify the Problem

In this case, we can clearly see a problem; ambulance response
times are perceived to be slower in the lower socioeconomic sec-
tions of town. This could lead to undesirable outcomes while
waiting for medics at the scene of an accident or illness.
Let’s see if this meets our criteria for a good problem:

1. Yes—city officials are interested in this problem.

2. Yes—the scope of the problem is manageable by city officials; they would sim-
ply need to collect response times for each section of town.
 The One-Sample t-Test | 169

3. Yes—city officials have the knowledge, time, and resources needed to investi-
gate the problem.
4. Yes—the problem can be researched through the collection and analysis of
numeric data.
5. Yes—investigating the problem has theoretical or practical significance.
6. Yes—it is ethical to investigate the problem.

Given that, we’ll use this problem statement as we move forward:

„ This study will investigate whether ambulance response times vary between
different sections of the city.

Our problem statement is clear and concise, it includes all the variables we want
to investigate, and we have not interjected personal bias.

S TEP 2 State a Hypothesis

„ Ambulance response times to the lower socioeconomic

community are significantly greater than 90 seconds.
Does this seem odd to you? Were you expecting to say that
response times to the lower socioeconomic sections of town were greater than those
to the more affluent parts of town? If you were, you are essentially right. In this case,
however, we know the exact population value, 90 seconds, we want to compare; there-
fore the null hypothesis would read:

„ Ambulance response times to the lower socioeconomic community are not

significantly greater than 90 seconds.

S TEP 3 Identify the Independent Variable

In this case we are interested in looking at times for 20 ambu-
lance calls to the affected neighborhood (i.e., our sample) and
compare them to a known population parameter (i.e., the aver-
age time for the total number of calls in the town). Because of
that, our independent variable is “calls for an ambulance,” but
we only have one level of that variable: calls to the specific part of town. While it seems
that calls to the town overall are another variable, that is not the case. They represent
the overall population and cannot be considered as a level of an independent variable;
this will become clearer as we move through the next steps.
170 | STATISTICS TRANSLATED

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is the average of the 20 ambulance
response times in the sample. Just as was the case with the z-test,
we will compare it to the population average (i.e., the overall
average response time for the community) of 20.
In order to find the average response time for the 20 calls
the officials are going to monitor, let’s use the data in Table 6.8. Figure 6.13 shows the
descriptive statistics for the sample.

TABLE 6.8.  Response Time Data

90 90 101 106 80
82 90 95 95 94
80 89 100 99 99
97 104 102 99 88

One-Sample Statistics

N Mean Std. Deviation Std. Error Mean

Response Time 20 94.0000 7.69142 1.71985

FIGURE 6.13.  Descriptive statistics from a One-Sample Test.

Just as the residents believed, the average response time for calls in their neigh-
borhood is greater than response times for the city in general. Let’s move on to the
next step to see if their complaints are justified by answering this question, “Is the
response time to their neighborhood higher or is it significantly higher?”

S TEP 5 Choose the Right Statistical Test

We have already established, in Step 2, that we are comparing
the mean of a small sample (i.e., less than 30 values) to that of
the population. Given that, we have one independent variable
with one level; we also have one dependent variable represent-
ing quantitative data. As a result, we are going to use the one-
sample t-test.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Figure 6.14 shows us the output we’ll use to test our hypothesis.
Our one-tailed p value of .031 is much smaller than our alpha
value of .05, so we reject the null hypothesis. This significant
difference is supported by a somewhat large effect size of .520.
 The One-Sample t-Test | 171

One-Sample Test

Test Value = 90

95% Confidence Interval of the

Mean Difference
T Df Sig. (2-tailed) Difference Lower Upper

Response Time 2.326 19 .031 4.00000 .4003 7.5997

FIGURE 6.14.  Inferential statistics from a One-Sample Test.

This is further substantiated by noting that our computed value of t, 2.326, is much
larger than our critical value of t, 1.729.

Interpreting Our Research Hypothesis

Again, be careful when you interpret the results of these tests. We rejected the null
hypothesis, but can we support the research hypothesis? In the case, the mean of our
sample is 94 and the population mean is 90. Because of that, the citizens have a legiti-
mate complaint. Remember, however, that if the difference of 4 seconds had been in
the other direction (i.e., the mean score of the sample was 86), the same p value would
have been computed. Like we have said before, the computer does not know our
hypothesis, it only processes the data. It is up to us to make sense of what it tells us.

The Case of Stopping Sneezing

Physicians at a premiere hospital are constantly working to lower the amount of time
it takes a patient to stop sneezing after being exposed to an allergen. Their experience
has shown, on average, that it takes about 2 minutes for standard drugs to be effective.
Today, however, they’re listening to a representative from a pharmaceutical company
who is introducing a drug the company claims will end sneezing in significantly less
time. The physicians skeptically agreed to try the drug and decided to use it on a
sample of 10 patients. The results of that sample, they believe, will help them decide
whether to use the new “miracle drug.”

S TEP 1 Identify the Problem

The problem here is very clear; patients are exposed to aller-
gens and want to stop sneezing as quickly as possible. While
there are drugs that work, the physicians are interested in the
problem and are always looking for something better for their
patients. Physicians do have the time and expertise and can col-
172 | STATISTICS TRANSLATED

lect numeric data showing the time it takes patients to stop sneezing. Finally, inves-
tigating the problem is ethical and very practically significant (at least to those of us
who suffer from allergies!).

„ Persons suffering from allergies desire to be treated with medications that

relieve their symptoms as quickly as possible. Physicians will evaluate
the new medication to determine if it works as well as claimed by the
pharmaceutical representative.

S TEP 2 State a Hypothesis

We have a one-tailed “less than” directional hypothesis:

„ Sneezing treated with the new medication will end in

significantly less than 2 minutes.

Our null hypothesis will read:

„ Sneezing treated with the new medication will not end in significantly less
than 2 minutes.

S TEP 3 Identify the Independent Variable

Here our independent variable is “sneezing,” but we only have
one level: sneezing in the sample we are interested in. We will
be comparing the average time it takes to end sneezing from
our sample (i.e., the sample statistic) to the average time for
ending sneezing in the overall population (i.e., the population
parameter). Because of that, we have an independent variable with only one level.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is the time it takes sneezing to stop. We
know the population average is 2 minutes (i.e., 120 seconds),
and we are going to compare it to the average of our 10 cases.
We will use data in Table 6.9 with each time shown in seconds;
the following Figure 6.15 shows the descriptive statistics for the
sample. The pharmaceutical salesperson may be on to something: the new drug does
stop sneezing in less time. What we want to know, however, is whether 117 seconds is
significantly less than 120 seconds.

TABLE 6.9.  Sneezing Data

95 123 120 110 125 113 114 122 122 126
 The One-Sample t-Test | 173

One-Sample Statistics

N Mean Std. Deviation Std. Error Mean

Time Until Sneezing Stops 10 117.00 9.416 2.978

FIGURE 6.15.  Descriptive statistics from a One-Sample Test.

S TEP 5 Choose the Right Statistical Test

Since we have one independent variable with one level and one
dependent variable representing quantitative data, obviously
we are going to use the one-sample t-test. Just like before, we
are comparing the mean of our small sample (i.e., 10 cases treat-
ed with the new drug) to the average time the physicians have
experienced with other treatments.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Figure 6.16 shows what the SPSS output for the one-sample
t-test. It seems the salesperson may be out of luck; the mean
time for the new drug was lower, but it was not significantly
lower. Our one-tailed p value is greater than our alpha value,
so we fail to reject the null hypothesis. We could also compute
an effect size of .319 to verify that our independent variable had only a moderate
effect on the dependent variable. Obviously, our computed value of t, –1.007, is much
smaller than our critical value of t, –1.833.

One-Sample Test

Test Value = 120

95% Confidence Interval of the

Mean Difference
t df Sig. (2-tailed) Difference Lower Upper
Time Until Sneezing Stops -1.007 9 .340 -3.00000 -9.7360 3.7360

FIGURE 6.16.  Inferential statistics from a One-Sample Test.

There’s an interesting issue here, however, that comes up from time to time. In
this case, the new drug has an average response time of 117 seconds, while the estab-
lished drug’s average time is 120 seconds. Even though the difference is not signifi-
cant, the new drug’s time is lower. The question to the physicians then remains, “Do
we use the new drug and save 3 seconds or, is that so insignificant in the overall scope
of things that we continue with what we are used to?”
174 | STATISTICS TRANSLATED

The Case of Growing Tomatoes

I have lived in big cities most of my life and recently decided to move out into the
country. One of the pleasures I have discovered is the ability to grow your own garden
and “live off the fruit of the land.” In talking to some of the other folks around here,
I have discovered that many people consider growing tomatoes a science somewhat
akin to nuclear physics. They talk about the amount of rain needed, the type of fertil-
izer to use, how far plants should be spaced apart, and other exciting topics. One thing
I found interesting is that they consider a 10-ounce tomato to be about average; most
folks cannot grow them any bigger than that. After hearing that, I figured a good way
for a city boy to fit in would be to grow the biggest, juiciest tomatoes in town!
In order to set my plan in action, I went to the local garden supply store and
inquired as to which fertilizer I should use. The guy running the place said he had
basically two types. The cheaper variety worked well enough, but he told me, by spend-
ing a few extra bucks, I could grow tomatoes I would be proud of! Given that, I bought
the costlier of the two, drove home, and put my plan into action.

S TEP 1 Identify the Problem

Without going into all the detail, it’s apparent I have an interest-
ing, manageable problem: I want to grow the biggest tomatoes
I can. I have the time and expertise, the weight of my tomatoes
is the numeric data I need, and growing tomatoes is certainly
practical and very ethical. Here is my problem statement:

„ Steve wants to demonstrate to his neighbors that he can grow tomatoes that
are equal to or larger than the community’s historical average.

S TEP 2 State a Hypothesis

In this case I have a one-tailed “greater than” directional
hypothesis:

„ Steve’s tomatoes will weigh significantly more than

10 ounces.

My null hypothesis is

„ Steve’s tomatoes will not weigh significantly more than 10 ounces.

 The One-Sample t-Test | 175

S TEP 3 Identify the Independent Variable

In this case, the independent variable is “tomato,” but I only
have one level; those that I have grown. I will be comparing the
weight of my tomatoes (i.e., the sample statistic) to the average
weight of 10 ounces (i.e., the population parameter).

S TEP 4 Identify and Describe the Dependent Variable

I show the weights of each tomato in Table 6.10. In this case, the
dependent variable is the weight of the tomatoes I have grown.
The population average is 10 ounces; the weights of the first
20 tomatoes I picked and the statistics that describe them are
shown in Figure 6.17.

TABLE 6.10.  Tomato Data

10 11 12 11 10
11 11 12 12 11
8 9 7 8 8
8.5 9 8 7.5 9

One-Sample Statistics

N Mean Std. Deviation Std. Error Mean

Tomato Weight 20 9.6500 1.63916 .36653

FIGURE 6.17.  Descriptive statistics from a One-Sample Test.

Things are not looking too good, are they? I was determined to grow bigger toma-
toes, but mine are a little smaller than the average weight reported by my neighbors.
Maybe I can salvage some of my dignity, though; I can still hold my head up high if
they are not significantly smaller.

S TEP 5 Choose the Right Statistical Test

As we know, I have one independent variable with one level and
my dependent variable, the weight of my tomatoes, is quantita-
tive. Because of that, I’ll use a one-sample t-test.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Based on our data, Figure 6.18 shows what SPSS will compute
for us. Things might not be as bad as I had feared. My p value is
.352 but remember, I am dealing with a one-tailed test, so I must
divide that by 2. Even when I do that, my p value of .176 is still
176 | STATISTICS TRANSLATED

One-Sample Test

Test Value = 10

95% Confidence Interval of the

Difference
t df Sig. (2-tailed) Mean Difference Lower Upper

Tomato Weight -.955 19 .352 -.35000 -1.1172 .4172

FIGURE 6.18.  Inferential statistics from a One-Sample Test.

much larger than the alpha value; this is reflected by a moderate effect size of .213.
Because of that, I fail to reject my null hypothesis; this means the research hypothesis
is not supported. What does this mean for me? Easy—I get to sit around talking to my
neighbors and tell them that my tomatoes are just about average.

Summary
At the beginning of the chapter I told you that, because of the similarities to the one
sample z-test, it wouldn’t take long to discuss the one-sample t-test. The key thing to
remember from this chapter is the t distribution; it is created when multiple samples,
each with less than 30 values, are collected and their means plotted. We then used
the concept of degrees of freedom to help understand the additional random error
introduced by smaller sample sizes. Again, just like the z-test, hypotheses can be
tested by computing a t value and comparing it to a critical value of t or by using a
computed p value and comparing it to alpha. Regardless of whether we reject or fail
to reject the null hypothesis, we can look at the strength of the relationship between
our independent and dependent variables using a computed effect size. In the next
two chapters we will look at two other tests that use the t distribution; the indepen-
dent and dependent-sample t-tests. While many of the underlying features are the
same, we will see that these tests are used to specifically control for the relationship
between levels of the independent variable.
Before we move forward with our quiz and into the next chapter, let me tell you
a short story about Mr. Gossett. When he was working for Guinness, Gossett was
not allowed to publish his findings because the brewery considered his work their
intellectual property. After all, they reasoned, the ideas were developed by their
employee while he was working in their brewery. Gossett, realizing the importance
of his work, felt it was imperative that he publish his findings. Rather than give
his employer a reason to fire him, he published the work under the pseudonym
“A. ­Student.” Because of that, many of the older textbooks refer to this distribution
as “Student’s t distribution” and to the test as “Student’s t-test.” While you may not
think this is important, just think how much money you could win on the television
show Jeopardy if one of the categories was “Famous Statisticians”!
 The One-Sample t-Test | 177

Do You Understand These Key Words and Phrases?

computed value of t critical value of t
critical value of t table degrees of freedom
effect size t distribution

Do You Know These Formulas?

Computed value of t:
x −µ
t=
Effect size:
sx
x −µ
d=
s

Quiz Time!

Using the SPSS output provided, what decision would you make for each of the hypotheses
below? Manually compute the effect size for each hypothesis; what does that tell you in rela-
tionship to the decision made?

1. The average number of cavities for elementary school children in areas without
fluoride in the water will be significantly higher than the national average of 2.0
(use Figures 6.19 and 6.20 to test this hypothesis).

One-Sample Statistics

N Mean Std. Deviation Std. Error Mean

Number of Cavities 15 2.33 1.589 .410

FIGURE 6.19.  Descriptive statistics from a One-Sample Test.

One-Sample Test

Test Value = 2

95% Confidence Interval of the

Difference
t Df Sig. (2-tailed) Mean Difference Lower Upper

Number of Cavities .813 14 .430 .333 -.55 1.21

FIGURE 6.20.  Inferential statistics from a One-Sample Test.

178 | STATISTICS TRANSLATED

2. Over the course of their careers, physicians trained in Ivy League schools will have
significantly fewer malpractice suits filed against them than the national average of
11 (use Figures 6.21 and 6.22 to test this hypothesis).

One-Sample Statistics

N Mean Std. Deviation Std. Error Mean

Malpractice Suits 15 5.20 2.426 .626

FIGURE 6.21.  Descriptive statistics from a One-Sample Test.

One-Sample Test

Test Value = 11

95% Confidence Interval of the

Difference
T df Sig. (2-tailed) Mean Difference Lower Upper

Malpractice Suits -9.259 14 .000 -5.800 -7.14 -4.46

FIGURE 6.22.  Inferential statistics from a One-Sample Test.

3. The graduation rate of universities that require students to spend their first two
years at a community college will be significantly greater than 82% (use Figures
6.23 and 6.24 to test this hypothesis).

One-Sample Statistics

N Mean Std. Deviation Std. Error Mean

Graduation Rate 15 79.60 4.188 1.081

FIGURE 6.23.  Descriptive statistics from a One-Sample Test.

One-Sample Test

Test Value = 82

95% Confidence Interval of the

Difference
T df Sig. (2-tailed) Mean Difference Lower Upper

Graduation Rate -2.219 14 .043 -2.400 -4.72 -.08

FIGURE 6.24.  Inferential statistics from a One-Sample Test.

4. Adult turkeys in the wild will weigh significantly less than 12 pounds (use Figures
6.25 and 6.26 to test this hypothesis).
 The One-Sample t-Test | 179

One-Sample Statistics

N Mean Std. Deviation Std. Error Mean

Weight of Turkey 15 10.53 2.066 .533

FIGURE 6.25.  Descriptive statistics from a One-Sample Test.

One-Sample Test

Test Value = 12

95% Confidence Interval of the

Difference
T Df Sig. (2-tailed) Mean Difference Lower Upper

Weight of Turkey -2.750 14 .016 -1.467 -2.61 -.32

FIGURE 6.26.  Inferential statistics from a One-Sample Test.

5. Residents in lower socio-economic areas of town are concerned that the time it
takes to restore electricity after a storm-related power outage (i.e., 5.15) is signifi-
cantly longer than the town’s average restoration time (i.e., 4.0; use Figures 6.27
and 6.28 to test this hypothesis).

FIGURE 6.27.  Descriptive statistics from a One-Sample Test.

FIGURE 6.28.  Inferential statistics from a One-Sample Test.

6. A local golf-pro has noticed that the number of holes-in-one made on her course
during the current season seems to be significantly higher than in years before. By
looking at the number from prior years, she can see that some years are greater
than the current year, and some less. She decided to compare her current years’
results to those in the past 20 years, to see if there is a significant difference. If she
has seen 25 holes-in-one during the current year and using Figures 6.29 and 6.30,
what could she determine?
180 | STATISTICS TRANSLATED

FIGURE 6.29.  Descriptive statistics from a One-Sample Test.

FIGURE 6.30.  Inferential statistics from a One-Sample Test.

CHAPTER 7

The Independent-Sample t-Test

Introduction
We just learned how Gossett successfully used the t distribution, with a one-sample
t-test, to compare a sample of ale to the batch (i.e., the population) of ale it came
from. Gossett soon realized, however, that a problem arose if you have small sam-
ples from two different populations and need to determine if they are significantly
different from one another.
For example, suppose we are interested in looking at the difference in the number of
disciplinary referrals between children from single-parent homes and children who live
with both parents. Our independent variable would be “Home Environment,” and there
would be two levels, “Single Parent” and “Both Parents.” Our dependent variable
would be the number of times children were disciplined. In another case, we might be
interested in comparing the time it takes pigeons to learn via operant or classical con-
ditioning. Again, we would have one independent variable, type of reinforcement, with
two levels—operant conditioning and classical conditioning. Our dependent variable
would represent the amount of time it took pigeons to learn to do a particular task.
Both scenarios meet all the criteria for an independent or dependent sample
t-test. Both have one independent variable with two levels, one dependent variable,
and the data being collected is quantitative. Our only problem is, how do we decide
whether to use the independent-sample t-test or the dependent-sample t-test?
We can answer that question by looking closely at the relationship between the
two levels of each of the independent variables. Do you notice anything remarkable?
In the first case, you are measuring the number of referrals of two groups of stu-
dents—those from single-parent homes and those from homes where there are two
parents. These two groups are independent of one another; a child who is in one group
can’t be in the other. The same goes for the groups of pigeons—one group is being

181
182 | STATISTICS TRANSLATED

trained with operant conditioning, and the other with classical conditioning. Being in
one group excludes the possibility of being in the other.
In both instances the researcher would collect data from two groups, with each
group representing a unique population. In Chapter 8, we will see that samples are
not always independent, so we will use the dependent-sample t-test. In the mean-
time, in Figure 7.1 we can use our example of students from different home environ-
ments to help explain what we’ve covered up to this point.

If We Have Samples from Two Independent Populations, How Do We

Know If They Are Significantly Different from One Another?
With the one-sample t-test, we were able to test hypotheses using computed t values,
critical T-scores, and p values; we were also able to use Cohen’s delta to allow us to
measure the effect of the independent variable on the dependent variable. When we
have samples from two unrelated populations, we can do the same thing by using the
independent-sample t-test. The general idea of hypothesis testing is the same, but we
need to use a slightly different sampling distribution. Instead of using the sampling
distribution of the means, which we talked about earlier, we will use the sampling distri-
bution of mean differences. The decision as to which t-test to use is shown in Figure 7.1.

The Sampling Distribution of Mean Differences

Before we get into detail about this distribution, let’s conceptually understand where
we are going. First, back when we were comparing means from a sample to the popula-

FIGURE 7.1.  Comparing samples from two populations.

 The Independent-Sample t-Test | 183

tion mean, it made sense to take single random samples from a population, calculate
the mean of the sample and then plot it. By doing this, we created the sampling distri-
bution of the means. Given both large enough sample sizes and number of samples, the
sample mean was equal to the population mean, and the main measure of dispersion
was called the standard error of the mean. We were then able to compare our sample
mean to the sampling distribution of the means to determine any significant difference.
In this case, however, we are not interested in determining if a single sample
is different from the population from which it was drawn. Instead we are interest-
ed in determining if the mean difference between samples taken from two like, but
independent, populations are significantly different from the mean difference of any
other two samples drawn from the same populations. This can be done by creating a
sampling distribution of mean differences (SDMD) using the following steps.

1. Take repeated samples from each of the two populations we are dealing with.
2. Calculate the mean of each sample.
3. Compute the difference between the two means.
4. Plot the difference between the two means on a histogram.

If we continued doing this and plotted each difference on a histogram, we would

wind up with a bell-shaped distribution; the larger the number of differences plotted,
the more normal the distribution would become. Unlike the sampling distribution of
the means, where the mean of the new distribution equals the population mean, the
mean of the new distribution will be zero. This is shown in Figure 7.2.

FIGURE 7.2.  Creating a sampling distribution of mean differences.

184 | STATISTICS TRANSLATED

To help us understand this, keep in mind that our hypothesis concerning disci-
plinary referrals involves samples from two populations—students from single-parent
homes and students who live with both parents. As we saw with the central limit theo-
rem, most of the values in each sample will cluster around the mean—a mean that
we’ve hypothesized is the same for both populations. When we take samples from
each population and then subtract the mean of one from the mean of the other, it
naturally follows that the majority of the means of the repeated samplings are eventu-
ally going to cancel one another out, leaving a distribution mean of zero. Those means
that do not cancel out will leave negative and positive values that create the tails of
the distribution.
Let’s use the hypothesis about the number of parents and disciplinary referrals
to demonstrate this idea. Table 7.1 shows data for 10 samples representing the aver-
age number of referrals for each level of the independent variable (i.e., single parent
or both parents). After that, I’ve subtracted the mean number of referrals from two-
parent homes from the mean number of referrals from a single-parent home and have
included that value in the fourth column:

TABLE 7.1.  Referral Data

Single- Both Difference
Sample parent home parents home (single – both)
 1 3 2  1
 2 2 6 –4
 3 2 1  1
 4 6 5  1
 5 1 2 –1
 6 0 1 –1
 7 3 6 –3
 8 5 7 –2
 9 5 2  3
10 3 8 –5
Average 3 4 –1

If we use a histogram to plot the difference values from the rightmost column,
SPSS will create Figure 7.3. If we continued calculating an infinite number of these
mean differences and plotted them on a histogram, we would wind up with the same
qualities of the t distribution discussed in Chapter 6.

1. The shape of the distribution will be different for different sample sizes. It
will generally be bell-shaped but will be “flatter” with smaller sample sizes.
2. As the number of samples plotted grows, the distribution will be symmetrical
around the mean.
3. The empirical rule applies.
4. The measure of dispersion, the standard error of mean differences (SEMD),
is conceptually the same as the standard error of the mean.
 The Independent-Sample t-Test | 185

5. The percentage of values under

the curve depends on the alpha
value and the degrees of free-
dom (df).
6. The t distribution table is used
to determine critical t values
for hypothesis testing.

Calculating the t Value for the

Independent-Sample t-Test
Looking back, our data show that the
students from single-parent homes aver-
FIGURE 7.3.  Bar chart of frequency distribution. age three trips per year to see the prin-
cipal. Their classmates with two parents
at home have a higher average: four dreaded trips down the hallway! Although there’s
a difference between these two groups, is this difference significant or is it due to
chance? We can state our question in the form of a null hypothesis and then use the
computed t value and our table showing the critical values of t to help us answer the
question.

„ There will be no significant difference in the number of referrals between

students from single-parent homes and students who live with both parents.

Obviously, our research hypothesis would read:

„ There will be a significant difference in the number of referrals between

students from single-parent homes and students who live with both parents.

Pay Attention Here

Just as I have for the other values throughout the book, I want to show you how to
compute the t value for the independent-sample t-test. While you’ll probably never be
asked to compute a t value “by hand,” conceptually understanding the process will
make you a better consumer of statistics. At first glance, this formula may look compli-
cated, but remember what I said way back in the Introduction; we are only using basic
math—addition, subtraction, multiplication, and division.
x1 − x2
t=
 SS1 + SS 2   1 1
 n + n − 2   n + n 
1 2 1 2
186 | STATISTICS TRANSLATED

Some of this may look like Greek (no pun intended), so let’s use Table 7.2 to
refresh our memory of the symbols we’ve already covered as well as add a few new
symbols to our gray matter.

TABLE 7.2.  Symbols Related to the Independent-Sample t-Test

Symbol Value
x1 The individual values in Group 1.
x2 The individual values in Group 2.

x12 The squared values in Group 1.

x22 The squared values in Group 2.

Sx1 The sum of values in Group 1.

S x12 The sum of the squared values in Group 1.

Sx2 The sum of the values in Group 2.

2 The sum of the squared values in Group 2.

S x2

x1 The mean of Group 1.

x2 The mean of Group 2.

n1 The number of values in Group 1.

n2 The number of values in Group 2.
SS Sum of squares.

Now, let’s start putting these together to solve the equation. The first thing we see
in the denominator (the bottom of the fraction) is that we need to compute the sum of
squares for both groups. We have already talked about the idea of the sum of squares
earlier, but let’s compute it again using the following formulas:

1. SS1 = ∑ x12 −
( ∑ x1 )
2

n1

2. SS2 = ∑ x 2 − ( ∑ x2 )
2

2
n2

First, we will modify Table 7.1 by squaring each of the values in our dataset; this is
shown in the second and fourth columns of Table 7.3 and is labeled Step 1. In Step 2,
we then total the values for each column. Finally, in Step 3, we will compute the mean
for our original two sets of values, shown in the bottom row of Table 7.3.
 The Independent-Sample t-Test | 187

TABLE 7.3.  Computing the Sum of Squares

Original data Original data
from Group 1 Step 1 from Group 2 Step 1
x1   x22 x2   x22
3  9 2  4
2  4 6 36
2  4 1  1
6 36 5 25
1  1 2  4
0  0 1  1
3  9 6 36
5 25 7 49
5 25 2  4
3  9 8 64

Step 2 Sx1 = 30 2
Sx2 = 122 Sx2 = 40 2
S x2 = 224
Step 3 x1 = 3 x2 = 4

We can then compute the sum of squares by using the data from the table.

1. SS1 = 122 − (30)

2
(40)2
SS2 = 224 −
10    10

2. SS1 = 122 −  900   1600 

SS2 = 224 − 
 10   10 
  
3. SS1 = 122 – 90     SS2 = 224 – 160
4. SS1 = 32         SS2 = 64

We can now include our SS1 and SS2 values, along with the values we already
knew, into our equation and compute our actual t value:

3−4
1. t =
 32 + 64   1 1
   + 
10 + 10 − 2 10 10
2. t = −1
 96   2 
   
18 10
−1
3. t =
(5.33)(.2)
188 | STATISTICS TRANSLATED

−1
4. t =
1.066
−1
5. t =
1.0325

6. t = –.968

In this example, using SPSS requires us to do something a bit differently. On the

data screen in Figure 7.4, you see two columns, parents and referrals. In the parents
field, I’ve included a value of 1 if the student has one parent at home and a value of
2 if the student has two parents living at home. These two values will represent the
two levels, that is, the number of parents, and the independent variable parents. The
referrals field, ranging from zero to 8, is the actual number of referrals for that child:

FIGURE 7.4.  Data View spreadsheet including number of parents and referrals.

To move forward, as seen in Figure 7.5, we have to select Analyze, Compare

Means, and Independent-Samples T test. The labels used by SPSS diverge from those
we normally use. In Figure 7.6 we’re asked to identify the Test Variable (i.e., the depen-
dent variable is referrals), the Grouping Variable (i.e., the independent variable is
parents), and the Groups (i.e., the levels of the independent variable are one parent or
two parents). Clicking on OK will give us the descriptive statistics shown in Figure 7.7.
Following that, SPSS provides us with the appropriate inferential statistics needed
to test our hypothesis. Notice, in the output, that there are several values we’ve not yet
talked about. We will get to them shortly, but for now let’s focus on what we need in
 The Independent-Sample t-Test | 189

FIGURE 7.5.  Using the Compare Means and Independent-Samples T Test commands.

FIGURE 7.6.  Defining the Test and Grouping Variables for the Independent-Samples T Test.
190 | STATISTICS TRANSLATED

Group Statistics

Parents N Mean Std. Deviation Std. Error Mean

Referrals One Parent 10 3.0000 1.88562 .59628

Both Parents 10 4.0000 2.66667 .84327

FIGURE 7.7.  Descriptive statistics from the Independent-Samples T Test.

Figure 7.8. The column to the far right is labeled “Referrals” and represents the infer-
ential statistics for our dependent variable. Beneath that, is a column titled “Equal
Variances Assumed.” If you look down that column to the row labeled “T,” you’ll see
the value –.968, exactly what we computed by hand. We can now use this information
to help us make a decision about the hypothesis we stated.

Independent Samples Test

Referrals

Equal variances Equal variances

assumed not assumed

Levene's Test for Equality of F 4.787

Variances Sig. .042
t-test for Equality of Means T -.968 -.968

Df 18 16.200

Sig. (2-tailed) .346 .347

Mean Difference -1.00000 -1.00000
Std. Error Difference 1.03280 1.03280

95% Confidence Interval of Lower -3.16982 -3.18723

the Difference Upper 1.16982 1.18723

FIGURE 7.8.  Inferential statistics from the Independent-Samples T Test.

Testing Our Hypothesis

As we saw, students from one-parent homes average three referrals, while their class-
mates average four. Again, they are different, but are they significantly different? We
can determine that by using essentially the same hypothesis-testing procedures we
used in the last two chapters. We’ve computed our t value, and we’re going to compare
it against the critical value of t from the table. In order to determine the critical value
of t, we have to go through a couple of steps.
First, we need to determine which alpha value to use. Let’s stick with .05 for now
but take a minute to think back. Based on our hypothesis, are we going to use the
entire alpha value to determine our critical value? Of course, the answer is “No”; since
we have a two-tailed hypothesis, we’re going to divide our alpha value by 2, giving us
an alpha value of .025.
Next we need to determine the necessary degrees of freedom. We do that by using
 The Independent-Sample t-Test | 191

the following formula; in it you can see we add the two sample sizes and then subtract
the number of levels in our independent variable, in this case 2:

df = (n1 + n2) – 2

Since both of our samples sizes are 10, our degrees of freedom value is 18 (i.e.,
10 + 10 – 2 = 18). We can check ourselves by looking at Figure 7.8, shown above.
With an alpha value of .025 and 18 degrees of freedom, we can use Table 7.4, an
exact copy of Table 6.2, and find our critical value of t is 2.101.

TABLE 7.4.  Critical Values of t

df α = .05 α = .025 df α = .05 α = .025 df α = .05 α = .025
 1 6.314 12.706 11 1.796 2.201 21 1.721 2.080
 2 2.920 4.303 12 1.782 2.179 22 1.717 2.074
 3 2.353 3.182 13 1.771 2.160 23 1.714 2.069
 4 2.132 2.776 14 1.761 2.145 24 1.711 2.064
 5 2.015 2.571 15 1.753 2.131 25 1.708 2.060
 6 1.943 2.447 16 1.746 2.120 26 1.706 2.056
 7 1.895 2.365 17 1.740 2.110 27 1.703 2.052
 8 1.860 2.306 18 1.734 2.101 28 1.701 2.048
 9 1.833 2.262 19 1.729 2.093 29 1.699 2.045
10 1.812 2.228 20 1.725 2.086 30 1.697 2.042

We can then plot both the critical and computed value and t and test our hypoth-
esis. Remember, in Figure 7.9, we must plot 2.101 as both a positive and a negative
value since we’re testing a two-tailed hypothesis.
Should we reject or fail to reject our
null hypothesis? In this case, we would
fail to reject the null since our computed
value of t (i.e., –.968) is within the range
created by our critical value of t. That
means, in this case, although the stu-
dents with both parents at home have a
higher number of referrals, the differ-
ence isn’t significant.

The p Value
We can also look at our p value to help
us decide. Since we are testing a two-
FIGURE 7.9.  Plotting the critical and computed
values of t. tailed hypothesis, we need to use the
two-tailed p value. If we look in the col-
umn labeled “Equal Variances Assumed” up in Figure 7.8, we can see that p (i.e., Sig.
2-tailed) of .346 is much greater than our alpha value, and we support our decision to
not reject the null hypothesis.
192 | STATISTICS TRANSLATED

Note on Variance and the t-Test

Although there are columns for both “Equal Variances Assumed” and “Equal Vari-
ances Not Assumed,” we’re only going to use the “Equal Variances Assumed” at this
point. We will go into detail regarding the use of these values after we’ve worked
through a couple of examples and have a thorough understanding of the basic ideas
of the independent-sample t-test.

The Statistical Effect Size of Our Example

As was the case with the one-sample t-test, we need to compute an effect size to better
understand the strength of the relationship between the mean scores from each of the
levels of our independent variable. In this case, we will again use Cohen’s d, but we will
compute it in a slightly different manner:
x −x
d= 1 2
s pooled
You can see, this time, that we are subtracting the mean of one group from the
mean of the other and then dividing that value by the pooled standard deviation (i.e.,
Spooled). The pooled standard deviation is calculated by:

1. Adding the variance of the two groups.

2. Dividing that value by 2.
3. Computing the square root of that value.

This is shown in the following formula:

2 2
s1 + s2
s pooled = 2
Using the values from above, we can compute the effect size in three steps. Notice,
SPSS does not include the variance, so we had to square our standard deviation values
to include them in the formula:

3.56 + 7.11
1. s pooled = 2

10.67
2. S pooled =
2

3. S pooled = 5.34
4. Spooled = 2.31

Using this value, we can then compute Cohen’s delta:

3−4
d=
2.31
 The Independent-Sample t-Test | 193

This gives us an effect size of –.433. As we said earlier, we always use the absolute
value of an effect size; that means we just need to drop the negative sign and wind up
with an effect size of .433. According to Cohen, this is a medium effect size and, as
might be expected given the relatively large p value, indicates that the number of par-
ents in a home does not have a large effect on the number of referrals a given student
receives.

Let’s Try Another Example

In this case, we’re interested in determining if there is a significant difference in the
frequency of nightmares between men and women. In order to do so, we will test the
following null hypothesis:

„ There will be no significant difference in the number of nightmares between

males and females.

Our two-tailed research hypothesis would read:

„ There will be a significant difference in the number of nightmares between

males and females.

In order to test this hypothesis, we could ask a representative sample of males and
females to report how often they have dreams of this type in a month. Our indepen-
dent variable is gender and has two levels, male and female. We have one dependent
variable, frequency of nightmares, and the data collected is quantitative. Since a given
participant can fall into only one category, male or female, we will use an indepen-
dent-sample t-test.
Using the data in Table 7.5, we might see that males averaged 12 nightmares
and females averaged 7 nightmares (Figure 7.10). The mean difference between these
groups is 5, meaning that males have an average of 5 more nightmares per month
than females. We can use this information in Table 7.6 to compute the square of the
number of dreams for each student.

TABLE 7.5.  Nightmare Data

Males Females
12  2
13  3
17  4
10  9
10  8
10  9
 9 10
12  8
15  8
12  9
194 | STATISTICS TRANSLATED

Group Statistics

Gender N Mean Std. Deviation Std. Error Mean

Nightmares Male 10 12.00 2.494 .789

Female 10 7.00 2.867 .907

FIGURE 7.10.  Descriptive statistics from the Independent-Samples T Test.

TABLE 7.6.  Computing the Sum of Squares

x1 x22 x2 x22
12 144  2   4
13 169  3   9
17 289  4  16
10 100  9  81
10 100  8  64
10 100  9  81
 9  81 10 100
12 144  8  64
15 225  8  64
12 144  9  81

Sx1 = 120 2
Sx2 = 1496 Sx2 = 70 2
S x2 = 564

x 1 = 12 x1 = 7

Let’s compute the t value by using the values in Table 7.6; I’ve already computed
the sum of squares for you.
12 − 7
t=
 56 + 74   1 1
   + 
18 10 10

If we worked on through the formula, we would end up with a computed t value of

4.16. In Figure 7.11, we can see this exactly matches what SPSS would generate for us.
Since we have a two-tailed hypothesis, we must divide our alpha value of .05 by
2; we can then use the new alpha value of .025, along with 18 degrees of freedom, to
determine that our critical value of t is 2.101. Since our computed value of t is much
greater than our critical value of t, we will reject our null hypothesis (Figure 7.12).
On the SPSS output, our decision is supported, in the column labeled “Equal
Variances Assumed,” by a two-tailed p value of .001. In short, males do have a signifi-
cantly greater number of nightmares than females.
 The Independent-Sample t-Test | 195

Independent Samples Test

Nightmares

Equal variances Equal variances

assumed not assumed
Levene's Test for Equality of F .810
Variances Sig. .380
t-test for Equality of Means T 4.160 4.160
Df 18 17.661

Sig. (2-tailed) .001 .001

Mean Difference 5.000 5.000

Std. Error Difference 1.202 1.202
95% Confidence Interval of Lower 2.475 2.472
the Difference Upper 7.525 7.528

FIGURE 7.11.  Inferential statistics from the Independent-Samples T Test.

Remember the Effect Size

Let’s not forget about our effect size. We
know we have a significant difference,
but what effect did the levels of the inde-
pendent variable have on the dependent
variable? Again, we could use our effect
size formula. Let’s compute the pooled
variance first,

FIGURE 7.12.  Plotting the critical and computed

6.22 + 8.22
values of t for a two-tailed test. 1. Spooled =
2

2. Spooled = 14.44
2
3. Spooled = 7.22
4. Spooled = 2.69

We can include this into our effect size formula along with the values of the two
means resulting in an effect size of 1.86. Obviously, the independent variable has a
large effect on the dependent variable.

12 − 7
1. d =
2.69

2. d = 1.86

How Does This Work for a Directional Hypothesis?

Using this same example let’s suppose we felt, from the outset, that men would have
more nightmares than women. That means we would test the same null hypothesis:
196 | STATISTICS TRANSLATED

„ There will be no significant difference in the number of nightmares between

males and females.

In this case, we would have to change our research hypothesis to be one-tailed,

so it would read:

„ Males will have significantly more nightmares than females.

This means, of course, that we’re going to have the same descriptive statistics
and we will compute the same value for t. We do need a different critical value of t,
however, because we are testing a directional hypothesis. Because we’re dealing with a
one-tailed hypothesis, however, we need to refresh our memories on using our alpha
value to determine the correct critical value of t we need to use.
As you know, when you state a directional hypothesis, you’re saying one of the
mean values you are comparing is going to be significantly higher OR significantly
lower than the other. Since you’re not looking at the probability of it being either, this
means you are going to have to use your entire alpha value on one end of the curve
or the other. Here, we’re hypothesizing that the mean of one group is going to be sig-
nificantly higher than the other. Because of that, we wouldn’t divide alpha by 2 as we
did with the two-tailed test; instead we would use the t table to look at the one-tailed t
value for an alpha level of .05 and 18 degrees of freedom. We would find this gives us
a critical value of 1.734; let’s plot this in Figure 7.13 and see what happens.
In this case, our computed value
of t (i.e., 4.16) is greater than our criti-
cal value of t (i.e., 1.734). There seems
to be a significant difference in the val-
ues, and we can use our p value to help
support our decision. We have to be
careful here, however. Remember, we
are looking at a one-tailed hypothesis,
and we used our entire alpha value to
determine the critical value of t. Having
changed those things, our p value is also
affected; it has dropped to .0005. When
we compare it to our alpha value of .05,
FIGURE 7.13.  Plotting the critical and computed we do support the decision we made
values of t for a one-tailed test. when we compared the computed and
critical values of t. In Table 7.7, we can
see the logical pattern of when to divide either alpha or p by 2 depending on the type
of hypothesis.

TABLE 7.7.  The Relationship between the Direction of the Hypothesis,

Alpha, and p
Type of hypothesis Alpha p
Directional (one-tailed) Use entire alpha value Divide p by 2
Nondirectional (two-tailed) Divide alpha by 2 Use entire p value
 The Independent-Sample t-Test | 197

Reminder—Always Pay Attention to the Direction of the Means!

We can see that our computed value of t is greater than our critical value of t, and our
p value is less than our alpha value, but we must always be careful. SPSS doesn’t know
our hypothesis; it simply works with the data we supply. In cases where the p value
is less than alpha, you must make your mean values match the direction you hypoth-
esized before you decide to support your research hypothesis.
For example, here we hypothesized that males would have a significantly high-
er number of nightmares than females. Our mean scores of 12 for males and 7 for
females enable us to reject the null hypothesis. Obviously, that means we can support
the research hypothesis. Suppose, however, the mean values had been reversed. We
may have collected our data and found a mean score of 7 for males and 12 for females.
If that’s the case, SPSS would compute the same p value. This means you would reject
the null hypothesis, but you wouldn’t support the research hypothesis because the
mean difference is not in the order you hypothesized. Be very, very careful with this.

Putting the Independent-Sample t-Test to Work

Having said all of this and having gained a good understanding of how the indepen-
dent-sample t-test works, it is now time to put it to work for us. Just as we’ve done in
the prior chapters, we will use case studies to help us. After reading the case we will
again use the six-step process to help us understand what we need to do. Let’s get
started with a case about an age-old complaint, “I could do better if I only had better
surroundings, equipment, and so forth . . .”

The Case of the Cavernous Lab

A university lab manager has computers set up in two locations on campus. One of
these locations, the Old Library, is a dark, cavernous old building; the other is the new
brightly lighted, well-decorated Arts and Sciences building. The manager has noted
that the equipment layout is about the same in each building and both labs receive the
same amount of use.
Each term approximately 50 students use these labs for an “Introduction to Com-
puters” course, with about half in each lab. The faculty member who teaches the class-
es has noticed that the two groups of students are about the same in that they both
have the same demographic and achievement characteristics. Despite this, the teacher
has noticed that the students using the lab in the Old Library do not seem to perform
as well as students working in the new laboratory. The only thing the professor can
figure is that “students always perform better in better surroundings.” The teacher has
asked the lab manager to investigate and report the findings.
The manager decides the best thing to do is to collect final examination scores
from all 50 students and compare the results of the students in the Library to those
in the Arts and Sciences building. After collecting the data, the manager realizes he
will probably need to use a statistical test to analyze the data but isn’t sure which one.
198 | STATISTICS TRANSLATED

Because of that, the manager asks us to get involved; let’s see what we can do to help
him.

S TEP 1 Identify the Problem

The problem here is very straightforward. The university lab
manager has installed new computer labs, but their use seems
to have affected student achievement. Not wanting his labs to
play a part in students’ grades, he decides to investigate. He is
certainly interested in a very manageable problem. Since he
installed the lab and is working with the teachers, he has the knowledge, time, and
resources necessary. He can easily collect numeric data to help ethically address a
problem with important practical significance. That allows us to make a very simple
problem statement:

„ Despite similar demographic and achievement characteristics, students in

the computer lab demonstrate higher performance than students in the Old
Library.

S TEP 2 State a Hypothesis

The first thing we tell the manager is that we need to determine
if differences in achievement, if they do exist, are significant.
Knowing that, the first thing we do is state the hypothesis we
will be investigating:

„ Students taking Introduction to Computers in the new Arts and Sciences

laboratory will score significantly higher on their final examination than
students taking the same course in the Old Library.

As we can see, this is a directional hypothesis since we believe one group will
perform significantly better than the other group. We stated it that way because that’s
the scenario the manager wants to investigate. We could have just as easily stated a
nondirectional hypothesis by saying there would be a significant difference between
the groups with no direction stated. That isn’t, however, what the scenario calls for.
The null hypothesis would be:

„ There will be no significant difference in levels of achievement between

students taking Introduction to Computers in the new Arts and Sciences lab
and students taking the same course in the Old Library.

S TEP 3 Identify the Independent Variable

Now we need to look at the cause and effect we are trying to
investigate. The teacher is suggesting that the location of the
laboratory (cause) affects how students perform (effect) in the
Introduction to Computers class. This means our independent
 The Independent-Sample t-Test | 199

variable is location and there are two levels: students in the Old Library and students
in the new Arts and Sciences building.

S TEP 4 Identify and Describe the Dependent Variable

The manager has decided to collect final examination scores
from all the students to help make a decision; this will be our
dependent variable. Although it isn’t explicitly stated, we can
assume the data will be quantitative, with scores ranging from
0 to 100. For our example, let’s use the data in Table 7.8.

TABLE 7.8.  Achievement Data

Students in the
Arts and Sciences building Students in the Old Library
70 73 73 76 76 65 69 69 72 72
72 79 79 79 79 71 75 75 75 75
82 82 86 82 82 77 78 78 78 78
85 85 85 85 88 81 81 81 82 84
88 88 91 91 94 84 84 87 88 90

The descriptive statistics are shown in Figure 7.14.

Group Statistics

Building N Mean Std. Deviation Std. Error Mean

Achievement Arts & Sciences 25 82.0000 6.39010 1.27802

Old Library 25 77.9600 6.34087 1.26817

FIGURE 7.14.  Descriptive statistics from the Independent-Samples T Test.

We can see that the students in the Arts and Science building do have a higher
average score than their counterparts in the Old Library. While it is apparent that
the independent variable does have some effect on the dependent variable, that’s not
enough. We must determine if the difference between the two groups is significant or
if it is due to chance.

S TEP 5 Choose the Right Statistical Test

In this case, we have one independent variable we have named
“Building” with two levels—students in the Old Library and stu-
dents in the new Arts and Sciences building. We are collecting
quantitative data, and our data are fairly normally distributed.
If we consider all this information, it appears that the inde-
pendent-sample t-test is exactly what we need to use to test our
hypothesis.
200 | STATISTICS TRANSLATED

S TEP 6 Use Data Analysis Software to Test the Hypothesis

By looking at the prior table, we’ve already seen that the mean
score of the students in the Arts and Sciences building is higher
than the mean score of the students in the Old Library, but we
also know the difference might be due to chance. We can use
the statistics in Figure 7.15 to help us make a decision about our
hypothesis.

Independent Samples Test

Achievement

Equal variances Equal variances

assumed not assumed
Levene's Test for Equality of F .012
Variances Sig. .913
t-test for Equality of Means T 2.244 2.244
df 48 47.997

Sig. (2-tailed) .029 .029

Mean Difference 4.04000 4.04000

Std. Error Difference 1.80044 1.80044
95% Confidence Interval of Lower .41996 .41996
the Difference Upper 7.66004 7.66004

FIGURE 7.15.  Inferential statistics from the Independent-Samples T Test.

Up to this point, we’ve used only the “Equal Variances Assumed” column to test
our hypotheses. As I said earlier, now that we have a good feeling for how the inde-
pendent-sample t-test works, it is important to understand the meaning of these two
values.

Testing Equal Variance in the Groups

Although it rarely affects beginning statisticians, the amount of variance within each
group affects the way the t value is computed. Again, although SPSS doesn’t show the
variance as part of the t-test results, we could easily square the standard deviation to
determine the variance of each group. In this case, these values are very close: 40.833
and 40.207. SPSS uses these values to test the null hypothesis comparing the variance
between the groups:

„ There will be no significant difference between the variance of the Arts and
Sciences group and the variance of the Old Library group.
 The Independent-Sample t-Test | 201

Notice this is a two-tailed hypothesis and it is always stated in this manner. To test
it, SPSS computes an analysis of variance (this is often called ANOVA, but more on that
later). The results of this ANOVA are shown in the top two rows of Figure 7.15 labeled
“Levene’s Test for Equality of Variance.”
To interpret the results, we first need to set our alpha level at .05. We then com-
pare our alpha value to the computed value of p for the Levene test. Since, in this
case, the computed value of p is .913 and our alpha value is .05, we fail to reject the
null hypothesis (be careful here, we are still talking about the values in the rows for
the Levene’s test, not the p values farther down the table). This means there’s no sig-
nificant difference in the variance between the two groups. Given that, we will use
the column reading “Equal Variances Assumed” for our independent-sample t-test. If
the computed p value for the Levene test was less than .05, we would reject the null
hypothesis and then have to use the column reading “Equal Variances Not Assumed.”
This idea is summarized in Table 7.9.

TABLE 7.9.  Interpreting the Levene Test

Levene test p value Resultant t-test decision
p value less than .05 Equal variances not assumed
p value greater than or equal to .05 Equal variances assumed

Testing Our Hypothesis

We are finally ready to test our hypothesis. We have a one-tailed hypothesis; this
means we must divide our Sig. (2-tailed) value of .029 by 2; this leaves us with a one-
tailed p value of .0145. This is less than our alpha value, .05, and means we reject our
null hypothesis. This decision is supported by a fairly large effect size of .635.

Determining the Critical Value of t When You Have More Than 30 Degrees
of Freedom
We could also check ourselves by comparing the computed value of t to the critical
value of t. In this case, in order to do that, we need to look at an issue that arises when
our degrees of freedom value is greater than 30.
Up to this point, we’ve been dealing with cases where we have had 30 or fewer
degrees of freedom and each of those values has been listed in the t distribution table.
In this case, we have 48 degrees of freedom, but if you look at the complete t distri-
bution table in the back of the book, you’ll see the values skip from 30 to 40, 40 to
60, 60 to 120, and then 120 to infinity (that’s what the symbol ∞ means). You might
ask, “What happened to all of the other values?” The short answer is, once we get to
a certain number of degrees of freedom, the table can be abbreviated because the
area under the curve for t changes so little between the differing degrees of freedom,
there’s no need to have a value that is 100% correct.
For example, if we have a problem with 40 degrees of freedom and our alpha
value is .05, our critical value for t would be 1.684; if we had 60 degrees of freedom, it
would be 1.671. Since the difference between these two values is so small (i.e., .013) the
table does not break the critical values down any further. Instead, when we are manu-
202 | STATISTICS TRANSLATED

ally computing statistics, we use the next highest df value in the table that’s greater
than our actual df. In this case, we have 48 degrees of freedom so we would use the
value for df = 60 (or 1.671). Since our computed value of t is 2.211, we again know we
should reject our null hypothesis.

Make Sure You Check the Direction of the Mean Scores

As we said earlier, if you’ve stated a directional hypothesis and you have a one-tailed p
value less than .05, you have to ensure your mean values match the manner in which
you hypothesized in order to support your research hypothesis. For example, here we
hypothesized the students in the Arts and Sciences building would do better than the
students in the Old Library. Our mean scores of 82 for the Arts and Sciences students
and 77.96 for the Old Library students, combined with a p value of .0145, enabled us
to reject the null hypothesis. Be careful, though; we might have collected the data
and found a mean score of 82 for students in the Old Library and a score of 77.96
for students in the Arts and Sciences building. In that case, even though SPSS would
have computed the same p value and we would have rejected the null hypothesis, we
still couldn’t support the research hypothesis since the means are not in the order
hypothesized. Our software doesn’t know the hypothesis you’re testing, it just works
with the data you entered. Because of that, you must be careful to interpret the output
correctly.

The Case of the Report Cards

Let’s continue by looking at a situation in which instructors at a high school are interest-
ed in determining if the frequency of report cards has any effect on students’ intrinsic
motivation. After researching the literature, they’re amazed to find many studies have
shown mixed results. Some authors contend that immediate feedback contributes to
higher levels of motivation, while others believe students might consider it controlling
and rebel against it. To investigate this, the teachers agree to give half of their classes
weekly report cards and the other half will receive report cards on the normal 9-week
schedule. After doing so they plan to analyze their data and come to a conclusion. In
order to help them, let’s go through the six-step process.

S TEP 1 Identify the Problem

In this case it seems all of the criteria are met, but let’s take a
look at the significance of the problem; is it practically or theo-
retically important? Actually, one could argue both are true.
The report cards are being used to address a practical problem;
higher achievement is linked to higher levels of motivation. At
the same time, the researchers are addressing a problem for which the literature has
shown mixed results. Given that, one could argue that the study also addresses cur-
 The Independent-Sample t-Test | 203

rent motivational theory and contributes back to the literature about what’s known
about feedback and intrinsic motivation.

„ Research has shown that the frequency of report cards may affect student
motivation. In order to be able to provide feedback to students in the most
meaningful way possible, teachers will investigate the effect on motivation of
weekly report cards versus report cards given once per 9 weeks.

S TEP 2 State a Hypothesis

„ Students receiving report cards each week will have

significantly different levels of intrinsic motivation
than students receiving report cards every 9 weeks.

In this case, the research hypothesis is nondirectional because the researchers

are stating a difference will exist, but they’re not sure which group will be higher. It is
based on prior observations and research, the results of which have been mixed.
The null hypothesis would read:

„ There will be no significant difference in intrinsic motivation between

students receiving weekly report cards and students receiving report cards
every 9 weeks.

S TEP 3 Identify the Independent Variable

The independent variable is “Report Card Frequency,” and
there are two levels; one group of students will receive report
cards every 9 weeks, and the other group will receive them each
week. The two levels are independent of one another. A student
in one group cannot be in the other group.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable in this case is the level of a student’s
intrinsic motivation; we will assume our instrument has 20
questions with a range of 1 to 5 on each question. That means,
of course, that scores can range from 20 (a low level of intrinsic
motivation) to 100 (a high level of intrinsic motivation). Know-
ing that, let’s use the data in Table 7.10 for 10 students in each group. SPSS would
create the descriptive statistics shown in Figure 7.16.
204 | STATISTICS TRANSLATED

TABLE 7.10.  Report Card Data

Weekly report cards 9-week report cards
75 80 69 68
68 82 68 64
87 75 69 68
80 75 72 67
65 55 69 68

Group Statistics

Report N Mean Std. Deviation Std. Error Mean

Motivation Weekly 10 74.2000 9.34285 2.95447

Nine-week 10 68.2000 1.98886 .62893

FIGURE 7.16.  Descriptive statistics from the Independent-Samples T Test.

We can see the weekly group has a mean score (i.e., 74.2) considerably larger
than do those students getting the report cards every 9 weeks (i.e., 68.2). Remember,
though, the teachers are interested in a significant difference, not just a difference.
Knowing that, they need to look at the second table. Our job now is to select the
appropriate statistic to help us interpret our data.

S TEP 5 Choose the Right Statistical Test

In this case we again have one independent variable with two
groups. The groups are independent of one another, and we
have one dependent variable that represents quantitative data.
Based on the kurtosis and skewness values, the data appear to
be fairly normally distributed so we are going to use the inde-
pendent-sample t-test.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Using Figure 7.17, what should the teachers conclude regarding
their null hypothesis? First, the teachers must use Levene’s test
to decide whether there’s a significant difference in the variance
between the two groups. In this case, since the p value is well
below .05, they must use the column labeled “Equal Variances
Not Assumed” to test their hypothesis. Notice, when this line is used, the degrees of
freedom field isn’t calculated in the same manner it is when we use the line that says
“Equal Variances Assumed.” This is done to control for error introduced by the differ-
ences in variance; SPSS will always calculate this for us.
 The Independent-Sample t-Test | 205

Independent Samples Test

Motivation
Equal variances Equal variances
assumed not assumed

Levene's Test for Equality of F 8.881

Variances Sig. .008
t-test for Equality of Means T 1.986 1.986

Df 18 9.814

Sig. (2-tailed) .062 .076

Mean Difference 6.00000 6.00000
Std. Error Difference 3.02067 3.02067

95% Confidence Interval of Lower -.34619 -.74780

the Difference Upper 12.34619 12.74780

FIGURE 7.17.  Inferential statistics from the Independent-Samples T Test.

As we said earlier, since the teachers are looking only for a significant differ-
ence in their hypothesis and no direction is implied, the hypothesis is two-tailed; that
means they should use the Sig. (2-tailed) value shown in the table. In this case, the p
value is .076, larger than the alpha value of .05. Because of this, the null hypothesis
is not rejected; there’s not a significant difference in levels of motivation between
students who receive report cards every week and those who receive them only every
9 weeks. We could verify that by checking the computed value of t (1.986) against our
critical value of t (2.101); again, this would show that the teachers would not reject the
null hypothesis.

Notice the Effect Size

In this case, although we are failing to reject the null hypothesis, we would compute an
effect size of .889. Even though this is very large, remember that the p value is a func-
tion of both the effect size and the sample size. In this case, although the independent
variable had a large effect on the dependent variable, the p value was somewhat higher
than you might expect because of the small sample size.

A Point of Interest
We have already agreed that a good hypothesis is consistent with prior research or
observations. In this case, based on the literature, the teachers hypothesized there
would be a significant difference in motivation. They didn’t say significantly higher
or significantly lower, they just said “different.” Interestingly, had they opted for a
one-tailed hypothesis such as “Students receiving weekly report cards will have signifi-
cantly higher levels of intrinsic motivation than those receiving report cards each 9
weeks,” they would have rejected the null hypothesis. We know that because, had we
divided our two-tailed p value by 2, the resultant one-tailed p value of .038 would be
less than our alpha value. Again, based on the literature, the hypothesis is two-tailed
206 | STATISTICS TRANSLATED

so the teachers have to live with it the way it is. Let’s look at another scenario and see
what happens.

The Case of the Anxious Athletes

Researchers at a major university are interested in determining if participation in
organized sports has any relationship to levels of anxiety. Their initial feelings are
that persons involved in physical activity are less likely to report levels of anxiety so
high they affect their quality of life. Wanting to test this, the researchers decide to use
a well-known anxiety inventory to measure levels of anxiety between two groups—uni-
versity students involved in organized sports and those not involved in some type of
sport. In order to collect their data, professors propose to identify 15 students in each
group, ask them to complete the inventory, and then calculate the results.

S TEP 1 Identify the Problem

Here, again, we have a problem that is worth investigating.
Without going into all of the details, since higher levels of anxi-
ety may lead to lower grades, it is in the university’s best interest
to determine if this is true. Again, much like the problem inves-
tigating motivation, this could be considered both a practical
and a theoretical problem.

„ High levels of anxiety have been shown to be detrimental to students’

well-being and quality of life. Since research has shown that involvement
in physical activity may lead to lower levels of anxiety, researchers at the
university investigate that by comparing anxiety levels of students involved
in organized sports versus those who are not.

S TEP 2 State a Hypothesis

In this case, the researchers feel athletes will have signifi-
cantly lower levels of anxiety than their less active classmates.
Given that, they decide to use the following hypothesis:

„ University students involved in organized sports programs will have

significantly lower levels of anxiety than university students not involved in
organized sports programs.

The null hypothesis would read:

 The Independent-Sample t-Test | 207

„ There will be no significant difference in levels of anxiety between university

students involved in organized sports programs and university students not
involved in organized sports programs.

S TEP 3 Identify the Independent Variable

The independent variable in this case is, for lack of a better
term, “university student.” The two levels are students involved
in organized sports and students not involved in organized
sports.

S TEP 4 Identify and Describe the Dependent Variable

Again, the professors are going to measure anxiety using an
anxiety inventory. The scores reported by the instrument are
quantitative in nature and range from 20 (a low level of anxiety)
to 80 (a high level of anxiety). Let’s use the data in Table 7.11 to
represent 15 students in each group. SPSS would generate the
descriptive statistics shown in Figure 7.18.

TABLE 7.11.  Anxiety Data

Students in organized sports Students not in organized sports
35 32 31 24 42 31
33 41 35 25 41 32
25 31 41 31 31 33
40 29 61 31 24 20
44 31 25 41 25 25

Group Statistics

Sports N Mean Std. Deviation Std. Error Mean

Anxiety Plays Sports 15 35.6000 9.04591 2.33565

Does Not Play Sports 15 30.4000 6.80126 1.75608

FIGURE 7.18.  Descriptive statistics from the Independent-Samples T Test.

So far it seems we are doing fine. We have two groups of quantitative data that
are fairly normally distributed. The means are 5.2 points apart, and we have a rather
large effect size of .650 so we may be on to something. Let’s move on to the next step
and see what happens.
208 | STATISTICS TRANSLATED

S TEP 5 Choose the Right Statistical Test

Here again we have one independent variable with two levels—
one group that participates in sports and one that doesn’t par-
ticipate in sports. The dependent variable, anxiety scores, shows
two groups with nice bell-shaped distributions. This means we
should use an independent-sample t-test.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

In Figure 7.19, the p value for Levene’s test is greater than .05,
so we will use the scores in the “Equal Variances Assumed” col-
umn. Thinking back to our hypothesis, it was one-tailed, so we
need to divide the Sig. (2-tailed) value by 2; this leaves us with
.043. This means we reject the null hypothesis; there are sig-
nificant differences in anxiety between students involved in sports and students not
involved in sports.

Independent Samples Test

Anxiety

Equal variances Equal variances

assumed not assumed

Levene's Test for Equality of F .466

Variances Sig. .501
t-test for Equality of Means T 1.780 1.780

Df 28 25.995

Sig. (2-tailed) .086 .087

Mean Difference 5.20000 5.20000

Std. Error Difference 2.92216 2.92216

95% Confidence Interval of Lower -.78578 -.80665

the Difference Upper 11.18578 11.20665

FIGURE 7.19.  Inferential statistics from the Independent-Samples T Test.

Wait a minute; are we jumping the gun here? We are right; we do have a p value
less than our alpha value, but look at the mean scores from the two groups. The stu-
dents involved in sports have a higher mean score (35.6) than their less active class-
mates (30.4). In this case, the researchers would technically be able to reject the null
hypothesis, but since the results weren’t in the hypothesized direction, there is no
support for the research hypothesis. For us, this shows we need to pay attention to
the relationship between our printout and the hypothesis we’ve stated. Many good
students have been led astray by not paying strict attention to what they were doing!
 The Independent-Sample t-Test | 209

Just in Case—A Nonparametric Alternative

As we said earlier, the use of the t-test mandates we have quantitative data that are nor-
mally (or nearly normally) distributed. Although the t-test is robust enough to handle
minor variances in the normality of the distribution, if the problem is bad enough,
we are forced to use the nonparametric Mann–Whitney U test. Unlike the independent
sample t-test, the U test uses the median of the dataset, rather than the mean, as the
midpoint. Because of this, the U test can be used for both quantitative data that’s not
normally distributed and in cases where the researcher collects ordinal (i.e., rank)
level data. Generally speaking, the software packages create output for the U test that
is very similar to the t-test; it is also just as easy to interpret.

Summary
Using the independent-sample t-test is very straightforward. Once we state our hy-
pothesis and identify the statistical tool we need, interpreting the results of a sta-
tistical software package is easy. Remember, the key things to look for are Levene’s
variance test and the computed value of p for the data you’ve entered. Always keep
in mind, however, that if you reject the null hypothesis, ensure the means are in the
order you hypothesized prior to supporting the research hypothesis. Table 7.12 can
act as a guide to help you understand and use the t-test.

TABLE 7.12.  Decision Making Using the t-Test

Step Decisions to be made
Decide if a t-test is needed One independent variable with two levels?
One dependent variable with quantitative data that is
  normally distributed?

Decide which t-test If levels of independent variable are related, use the
is needed   dependent sample t-test.
If levels of independent variable are not related, use
  the independent sample t-test.

Once the test is run, check If p is less than .05, equal variances are not assumed.
Levene’s p value If p is greater than or equals .05, equal variances are
 assumed.

Determine if one-tailed or If the hypothesis is directional (one-tailed), use the

two-tailed p value is needed  one-tailed p value. If you only have the two-tailed
  value, divide it by 2.
If the hypothesis is nondirectional (two-tailed), use the
 two-tailed p value.

Use appropriate p value to If p is less than alpha, reject null hypothesis. Be sure
compare to alpha value   to check if the mean values are in the order you’ve
  hypothesized prior to making any decisions.
If p is greater than or equals alpha, fail to reject null
 hypothesis.
210 | STATISTICS TRANSLATED

Do You Understand These Key Words and Phrases?

computed value of t effect size
equal variances assumed equal variances not assumed
independent-sample t-test Levene’s test
Mann–Whitney U test pooled standard deviation
sampling distribution of mean differences standard error of mean difference
sum of squares

Do You Understand These Formulas?

Degrees of freedom:
df = n1 + n2 – 2

Effect size:
x1 − x2
d=
s pooled
Pooled standard deviation:
2 2
s1 + s2
s pooled = 2

Sum of squares for group 1:

SS1 = ∑ x12 −
( ∑ x1 )
2

n1

Sum of squares for group 2:

SS2 = ∑ x22 −
( ∑ x2 )
2

n2

t-score:
x1 − x2
t=
 SS1 + SS 2   1 1
 n + n − 2   n + n 
1 2 1 2
 The Independent-Sample t-Test | 211

Quiz Time!

The Case of the Homesick Blues

Psychologists are interested in determining whether males or females make more phone calls
home during military basic training.

1. What is the hypothesis being tested?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 7.20 and 7.21, what decision should they make?
Group Statistics

Freshman Gender N Mean Std. Deviation Std. Error Mean

Phone Calls Males 15 93.2667 9.80136 2.53070

dimension1

Females 15 69.8000 22.68165 5.85638

FIGURE 7.20.  Descriptive statistics from the Independent-Samples T Test.

Independent Samples Test

Phone Calls

Equal variances Equal variances

assumed not assumed
Levene's Test for Equality of F 10.936
Variances Sig. .003
t-test for Equality of Means T 3.678 3.678
Df 28 19.052

Sig. (2-tailed) .001 .002

Mean Difference 23.46667 23.46667

Std. Error Difference 6.37978 6.37978
95% Confidence Interval of Lower 10.39828 10.11612
the Difference Upper 36.53505 36.81721

FIGURE 7.21.  Inferential statistics from the Independent-Samples T Test.

The Case of the Cold Call

Employees of a telemarketing firm traditionally worked in cubicles where it was easy to hear
their coworkers around them. After a recent expansion, employees started complaining that
the increased noise level in the room was beginning to interfere with their productivity. Fear-
212 | STATISTICS TRANSLATED

ing this could lead to fewer sales, the owners of the company decided to put half of their
employees into private offices to see if it would lead to more calls per hour.

1. What is the hypothesis being tested?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 7.22 and 7.23, what decision should they make?
Group Statistics

Office Location N Mean Std. Deviation Std. Error Mean

Marketing Calls Cubicle 8 15.8750 3.35676 1.18679

Enclosed Office 8 7.5000 3.66450 1.29560

FIGURE 7.22.  Descriptive statistics from the Independent-Samples T Test.

Independent Samples Test

Marketing Calls
Equal variances Equal variances
assumed not assumed

Levene's Test for Equality of F .122

Variances Sig. .732
t-test for Equality of Means T 4.767 4.767

Df 14 13.894
Sig. (2-tailed) .000 .000

Mean Difference 8.37500 8.37500

Std. Error Difference 1.75700 1.75700

95% Confidence Interval of Lower 4.60661 4.60390

the Difference Upper 12.14339 12.14610

FIGURE 7.23.  Inferential statistics from the Independent-Samples T Test.

The Case of the Prima Donnas

Female actresses on Broadway are convinced that putting their names on the marquee in front
of their male counterparts leads to a larger audience. The owners of a theater are interested
in finding out if this is true, so they track tickets sales while changing the order of the names
on the marquee periodically.

1. What is the hypothesis being tested?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 7.24 and 7.25, what decision should they make?
 The Independent-Sample t-Test | 213

Group Statistics

Gender N Mean Std. Deviation Std. Error Mean

Ticket Sales Prima-donna 15 955.0667 266.98434 68.93506

Males 15 879.8667 257.97145 66.60794

FIGURE 7.24.  Descriptive statistics from the Independent-Samples T Test.

Independent Samples Test

Ticket Sales

Equal variances Equal variances

assumed not assumed

Levene's Test for Equality F .006

of Variances Sig. .939
t-test for Equality of Means T .784 .784

Df 28 27.967

Sig. (2-tailed) .439 .439

Mean Difference 75.20000 75.20000
Std. Error Difference 95.85750 95.85750

95% Confidence Interval of Lower -121.15519 -121.16561

the Difference Upper 271.55519 271.56561

FIGURE 7.25.  Inferential statistics from the Independent-Samples T Test.

The Case of the Wrong Side of the Road

Some believe that gasoline stations going into a city charge less per gallon than stations do
on roads leading out of a city. Their logic is that drivers need to fill up as they’re leaving on
a trip; since that’s the case, the owners of gasoline stations feel they can charge more and get
away with it. I’m not sure if I believe this, so I’m going out and collecting gasoline prices from
stations going in both directions.

1. What is the hypothesis being tested?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 7.26 and 7.27, what decision should I make?
Group Statistics

Gas Station Location N Mean Std. Deviation Std. Error Mean

Price per Gallon Into Town 10 3.0180 .16430 .05196

dimension1

Out of Town 10 3.0480 .20286 .06415

FIGURE 7.26.  Descriptive statistics from the Independent-Samples T Test.

214 | STATISTICS TRANSLATED

Independent Samples Test

Price per Gallon

Equal variances Equal variances

assumed not assumed

Levene's Test for Equality of F 1.818

Variances Sig. .194
t-test for Equality of Means T -.363 -.363
Df 18 17.255
Sig. (2-tailed) .721 .721
Mean Difference -.03000 -.03000

Std. Error Difference .08255 .08255

95% Confidence Interval of Lower -.20343 -.20397
the Difference Upper .14343 .14397

FIGURE 7.27.  Inferential statistics from the Independent-Samples T Test.

The Case of Workplace Satisfaction

The human resources manager of a large corporation has noted a higher resignation rate by
female employees than their male coworkers. In order to understand the situation better, the
manager asks all employees to complete a workplace satisfaction survey. She will then compare
the result by gender. The scores on the survey range from zero (low satisfaction) to 20 (high
satisfaction). The manager feels that one of the factors contributing to higher female attrition
is they are less satisfied with their workplace than their male coworkers.

1. What is the hypothesis being tested?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 7.28 and 7.29, what has the manager learned?

FIGURE 7.28.  Descriptive statistics from the Independent-Samples T Test.

 The Independent-Sample t-Test | 215

FIGURE 7.29.  Inferential statistics from the Independent Samples T Test.

The Case of the Flower Show

Gardeners who grow flowers specifically for competition in flower shows have two differing
opinions about the quality of major sources of fertilizer—commercially made or that coming
from nature. In order to try to use the fertilizer that would best bolster her flowers’ chances
of winning, one grower decided to grow two plots of roses—one using natural fertilizer and
one using commercially produced. After the roses bloomed, the gardener invited local florists
to come to her nursery and judge the flowers without knowing the type of fertilizer used for
a given flower; she asked that they be evaluated on a scale from 1 (least beautiful) to 5 (most
beautiful).

1. What is the hypothesis being tested?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 7.30 and 7.31, what decision should the grower make?

FIGURE 7.30.  Descriptive statistics for Independent-Samples T Test.

216 | STATISTICS TRANSLATED

FIGURE 7.31.  Inferential statistics for Independent-Samples T Test.

CHAPTER 8

The Dependent-Sample t-Test

Introduction
In the last chapter, we saw that when we have one independent variable with two
levels that are independent of one another, we used the independent-sample t-test.
What happens, however, when the levels of the independent variable represent two
different measurements of the same thing? For example, imagine comparing pretest
and posttest scores for one group of students. If we did, we would have two mea-
surements of the same group. The independent variable would be “Student Test
Scores” and there would be two levels, “Pretest Scores” and “Posttest Scores.” In
this case, a given student’s posttest score would be directly related to his or her
pretest score. Because of that, we would say the levels are dependent upon, or influ-
ence, one another.
As another example, suppose we are interested in determining if a particular drug
has an effect on blood pressure. We would measure a person’s blood pressure prior
to the study, have the patient take medicine for a period of time, and then measure
the blood pressure again. In this case, for a given set of measurements, the “Before
Blood Pressure” and the “After Blood Pressure” are the two levels of the independent
variable “Blood Pressure.” Again, these two levels are related to one another, so we
would use the dependent-sample t-test (sometimes called the paired-samples t-test)
to check for significant differences.

That’s Great, but How Do We Test Our Hypotheses?

Just like in the prior chapter, here we have an independent variable with two levels
and one dependent variable where we have collected quantitative data. Given that,
217
218 | STATISTICS TRANSLATED

you might be asking, “Why not just use the independent-sample t-test?” While that is
a logical question, unfortunately, we cannot do that. As we’ll soon see, the relation-
ship between the levels of the independent variable creates a problem if we try to do
it that way.

Independence versus Dependence

When we talked about the independent-sample t-test, the key word was “indepen-
dent”; scores from one group did not influence scores in the second group. That is not
the case when we have two related groups; let’s use an example to help us understand
what we are getting at.
Using the idea of pretests and posttests we alluded to earlier, let’s use the data in
Table 8.1 to help us better understand where we are going with this idea.

TABLE 8.1.  Student Test Data

Student Pretest Posttest
1 52 89
2 48 77
3 51 81
4 45 69
5 50 80
6 60 90
Average 51 81

In this case, the key to understanding what is meant by dependence is based on

two things. First, obviously each individual pretest score was made by one given stu-
dent. Given that, when a student takes the posttest, his score will be “related” to his
pretest score. Second, the fact that two sets of scores were collected from one set of stu-
dents makes them “dependent” on each other; the posttest score should generally be
higher than the pretest score for a given student. This might sound somewhat confus-
ing, but we will see more examples that make this concept very clear. The bottom line,
for now, is that we have two levels: student pretest scores and student posttest scores.

Computing the t Value for a Dependent-Sample t-Test

Because of the relationship between the two levels of the independent variable, we
cannot compute the t value using the same formula as the independent-sample t-test;
if we did, we would dramatically increase the probability of making a Type I error (i.e.,
rejecting the null hypothesis when we should not). In order to avoid that, we have a
different formula to compute the t value:

D
t=
 (∑ D)2 
∑D −
2

 n 
 n ( n − 1) 
 
 The Dependent-Sample t-Test | 219

Testing a One-Tailed “Greater Than” Hypothesis

To help understand this formula, let’s look at my wife’s job. She’s a media specialist
and firmly believes that it doesn’t matter what children read (obviously within limits);
it is getting them to read anything that makes them better readers.
In order to test her theory, she could start by asking her students to tell her how
many hours each week they spend reading outside of school. She could then introduce
them to the Harry Potter series, books that appeal to children at that age level. After a
few weeks she would again ask the children how many hours they read per week. Obvi-
ously, she hopes the second number is greater than the first. Her research hypothesis
would be:

„ The average number of hours spent reading per week will be significantly
greater after allowing students to read books that appeal to them.

In Table 8.2, we have data for five students with the number of hours weekly
they read before the new books were introduced, as well as the number of books they
read after the new books were introduced. In order to compute the t value, we need
to include two extra columns to create Table 8.3. The first new column, labeled D,
shows us the difference between the Before and After scores. The second new column,
labeled D2, shows the squared value of D. At the bottom of each of those columns, you
can see we have summed these values.

TABLE 8.2.  Number of Books Read

Student Before After
1 4 6
2 4 6
3 6 6
4 7 9
5 8 12
Average 5.8 7.8

TABLE 8.3.  Computing the Difference Values for Books Read

Student Before After D D2
1 4  6 +2  4
2 4  6 +2  4
3 6  6 0  0
4 7  9 +2  4
5 8 12 +4 16
Average 5.8 7.8 SD = 10 SD2 = 28

Looking back at the equation, you can see the first value we need to compute is
D, the average of the difference between scores in each dataset. We can see the sum
of the differences (i.e., ∑D2) is 10; to compute the average, we need to divide that by
220 | STATISTICS TRANSLATED

5, the number of values in our dataset. This leaves us with an average of 2; we can put
that into our formula before moving on to the next step.
2
t=
 (∑ D)2 
∑D −
2

 n 
 n ( n − 1) 
 

We can now deal with the denominator of the equation. Again, let’s go through
step by step. First, in the third column of our table we have calculated the difference
between both values and summed them (i.e., 10). In the rightmost column, we have
taken each of the difference values and squared it. Adding these values gives us the
sum of the differences squared (i.e., ∑D2). We can insert these values, along with n
(i.e., 5) into the equation. We can finish computing the equation using the following
steps.

1. t = 2
 100 
28 −
 5 
 n ( n − 1) 
 

2
2. t =
 100 
28 −
 5 
 n ( n − 1) 
 

2
3. t =
 28 − 20 
 5 ( 5 − 1) 

2
4. t =
 8
 
20
2
5. t =
(.4)
2
6. t =
.6325
7. t = 3.16

This leaves us with a computed t value of 3.16. Before we can test the hypothesis,
we must determine the critical value of t from the same table we used with the inde-
 The Dependent-Sample t-Test | 221

pendent-sample t-test. This time, however, we will compute our degrees of freedom
by subtracting one from the total pairs of data; when we subtract 1 from 5, we are left
with 4 degrees of freedom. Using the traditional alpha value of .05, we would refer to
our table and find that the critical value of t is 2.132.
We can then plot that on our t distribution shown in Figure 8.1; remember, we
have a one-tailed hypothesis, so the entire t value goes on one end. Here our com-
puted value of t is greater than our critical value of t. Obviously we have rejected the
null hypothesis and supported my wife’s research hypothesis: children do read more
when they are interested in what they are reading.
In order to check our work using
SPSS, we need to set up our spread-
sheet, shown in Figure 8.2, to include
two variables, Before and After; we
would then include the data for each.
Following that, in Figure 8.3, we select
Analyze, Compare Means, and Paired
Samples T Test. As shown in Figure 8.4,
we would then identify the pairs of data
we want to compare and click on OK.
FIGURE 8.1.  Comparing the computed and criti- SPSS would first provide us with Figure
cal values of t. 8.5; it verifies what we computed earlier
in Table 8.3.

FIGURE 8.2.  Before and after data in the Data View spreadsheet.

As we can see in Figure 8.6, the t value of 3.16 is exactly what we computed earlier,
and our p value is less than .05. This means we can support the research hypothesis;
kids reading books they enjoy actually did spend significantly more time reading per
week than they did when they had no choice in their reading material. What is the
bottom line? My wife is happy!
222 | STATISTICS TRANSLATED

FIGURE 8.3.  Using the Compare Means and Paired-Samples T Test command.

The Effect Size for a Dependent-Sample t-Test

Just as was the case with the independent-sample t-test, we can compute an effect size
for the dependent-sample t-test. Interpreting it is the same as in our earlier example,
but the formula is a lot easier to compute; all you do is divide the average mean differ-
ence by the standard deviation of the difference.
xdifference
d=
s difference

FIGURE 8.4.  Creating the pairs of data to be analyzed using the Paired-Samples T Test.
 The Dependent-Sample t-Test | 223

Mean N Std. Deviation Std. Error Mean

Pair 1 Before 5.8000 5 1.78885 .80000
After 7.8000 5 2.68328 1.20000

FIGURE 8.5.  Descriptive statistics from the Paired-Samples test.

Paired Samples Test

Pair 1
After - Before

Paired Differences Mean 2.00000

Std. Deviation 1.41421

Std. Error Mean .63246

95% Confidence Interval of Lower .24402

the Difference Upper 3.75598
T 3.162
Df 4
Sig. (2-tailed) .034
FIGURE 8.6.  Inferential statistics from the Paired-Samples test.

Using the values from above and the following two steps, we can compute an
effect size of 1.42:

2
1. d =
1.41

2. d = 1.42

According to Cohen’s standards, this is very large and indicates that the treatment
had quite an effect on the dependent variable. Just as we did in the preceding chapter,
we will always include the effect size as part of our descriptive statistics.

Testing a One-Tailed “Less Than” Hypothesis

We have seen how well this works with a “greater than” one-tailed hypothesis; now let’s
look at an example where we are hypothesizing that one average will be significantly
less than another. Let’s suppose we are working with a track coach at our local high
school who is trying to improve his team’s times in the 400-meter run. The coach has
heard that a diet high in protein leads to more muscle mass and figures this should
contribute to his athletes running faster. He decides to test the new diet for 6 weeks
and measure the results:
224 | STATISTICS TRANSLATED

„ A track athlete’s time in the 400-meter run will be significantly less after
following a high-protein diet for 6 weeks.

During the diet regimen, the coach collected data shown in Table 8.4; each of the
times is measured in seconds.

TABLE 8.4.  Race Time Data

Athlete Before After
Al 105  98
Brandon 110 105
Charlie 107 100
De’Andre 112 107
Eduardo 101  91
Frank 108 103

Just by looking, it seems that the “after” scores are lower, but let’s go ahead and
compute our t value. First, since we are looking at a one-tailed “less than” hypothesis,
we need to subtract the “Before” from the “After” value. This can be seen in Table 8.5.

TABLE 8.5.  Computing the Difference Values for Race Times

Athlete Before After D D2
Al 105  98  –7  49
Brandon 110 105  –5  25
Charlie 107 100  –7  49
De’Andre 112 107  –5  25
Eduardo 101  91 –10 100
Frank 108 103  –5  25

Sum SD2 = 39 SD2 = 273

To compute x , we are again going to divide the sum of the differences (–39) by
the number of values in the dataset (6); this gives us an average difference of –6.5.
Let’s go ahead and enter that into our equation:
−6.5
t=
 (∑ D)2 
∑D −
2

 n 
 n ( n − 1) 
 

We continue by inserting the sum of the differences squared (i.e., 273), the sum
of the differences (i.e., –39), and n (i.e., 6), into our equation:
 The Dependent-Sample t-Test | 225

−6.5
t=
 ( −39)2 
 273 − 6 
 
 6 (6 − 1) 
 

If we simplify that, we come up with the following formulas:

−6.5
1. t =
 1521 
273 −
 6 
 30 
 

−6.5
2. t =
 19.5 
 
30 

3. t = −6.5
.065

−6.5
4. t =
.806

5. t = –8.06

We can now help the coach test his hypothesis. First, using our alpha value of
.05 and our degrees of freedom of 5, we find that we have a critical t value of 2.015.
In order to plot this, keep in mind that we have a one-tailed hypothesis, so the entire
critical t value goes on one end of the distribution as shown in Figure 8.7. In this case,
since we have a one-tailed “less than” hypothesis, it needs to go on the left tail of the
distribution. Since our computed value of t is also negative, it needs to go on the left
side of the distribution as well.
We can clearly see our computed
value of t of –8.06 is far less than our
critical value of t; this means we must
reject the null hypothesis. It appears
that athletes really do run faster if they
follow a high-protein diet.
In Figures 8.8 and 8.9, we can con-
firm what we have done using SPSS. We
FIGURE 8.7.  Using the computed and critical are further assured of our decision since
values of t to test the hypothesis. the p value of .000 is less than our alpha
value; we can also compute an effect size
by dividing the mean difference of our
scores by the standard deviation.
226 | STATISTICS TRANSLATED

Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 Before the Diet 107.1667 6 3.86868 1.57938

After the Diet 100.6667 6 5.75036 2.34758

FIGURE 8.8.  Descriptive statistics from the Paired-Samples test.

Paired Samples Test

Pair 1

After the Diet -

Before the Diet
Paired Differences Mean -6.50000

Std. Deviation 1.97484

Std. Error Mean .80623

95% Confidence Interval of Lower -8.57247

the Difference Upper -4.42753
T -8.062
Df 5
Sig. (2-tailed) .000
FIGURE 8.9.  Inferential statistics from the Paired-Samples test.

−6.5
d=
1.97
This yields an effect size of –3.30, but as we did with the independent-sample
t-test, we must drop the negative sign and use the absolute value. When we do, we see
that our effect size means our intervention had a definite effect on our dependent
variable.

Testing a Two-Tailed Hypothesis

In order to help us thoroughly understand what we are doing, let’s consider a case
where we have a two-tailed hypothesis. In this case, we are investigating a drug
designed to stop migraine headaches. While we have every indication that it should
work, we have discovered that the drug may negatively affect a person’s systolic (i.e.,
the upper number) blood pressure. In some instances, the drug might cause the per-
son’s blood pressure to rise; in other cases, it might drop significantly.
Let’s create a dataset by taking a patient’s blood pressure at the start of our study,
administer the migraine drug for 2 weeks and then measure the blood pressure again.
After we have completed our study, we wind up with the data in Table 8.6; notice that
I’ve already computed the difference and the squared difference for you.
 The Dependent-Sample t-Test | 227

TABLE 8.6.  Computing the Difference Values for Blood Pressure

Patient Prior After D D2
 1 120 135 15 225
 2 117 118  1   1
 3 119 131 12 144
 4 130 128 –2   4
 5 121 121  0   0
 6 105 115 10 100
 7 128 124 –4  16
 8 114 111 –3   9
 9 109 117  8  64
10 120 120  0   0
SD = 37 SD2 = 563

Just like before, we compute D by taking the average of the difference scores (i.e.,
37/10 = 3.7) and include it in our formula:
3.7
t=
 (∑ D)2 
∑D −
2

 n 
 n ( n − 1) 
 

We can insert the values for SD2, SD, and n into our equation and use the follow-
ing steps to compute t:
3.7
1. t =
 1369 
563 −
 10 
 10 (10 − 1) 
 
3.7
2. t =
 563 − 136.9 
 
90
3.7
3. t =
 426.1
 
90 
3.7
4. t =
(4.73)

5. t = 3.7
2.18
Finally, we are left with a t value of 1.70; this matches Figures 8.10 and 8.11 that
would be created by SPSS.
228 | STATISTICS TRANSLATED

Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 Before the Medication 118.3000 10 7.66014 2.42235

After the Medication 122.0000 10 7.49815 2.37112

FIGURE 8.10.  Descriptive statistics from the Paired-Samples test.

Paired Samples Test

Pair 1

After the
Medication -
Before the
Medication
Paired Differences Mean 3.70000
Std. Deviation 6.88073
Std. Error Mean 2.17588

95% Confidence Interval of Lower -1.22218

the Difference Upper 8.62218
T 1.700
Df 9
Sig. (2-tailed) .123

FIGURE 8.11.  Inferential statistics from the Paired-Samples test.

You can see from the descriptive statistics that the average blood pressure before
the medication is 118.3, with a slightly higher blood pressure of 122 after the medica-
tion has been taken. If we compute our effect size of .538 based on these values, we
can see that the intervention had a moderate influence on our dependent variable.
We can now plot our computed t value and the appropriate critical value. Remem-
ber, since we are dealing with a two-tailed hypothesis, it is necessary to divide alpha by
2 and find the critical value for alpha = .025. Using the table, along with 9 degrees of
freedom, our critical value of t is 2.262. As shown in Figure 8.12, we would then mark
that off on both ends of our normal curve.
Since our computed value is within
the range of the positive and negative
critical values, we do not reject the null
hypothesis; even though the average
“After” blood pressure rose slightly, it
wasn’t a significant difference. This is
verified by the two-tailed p value of .123.
In other words, after all of this, we have
shown that the drug manufacturers
have nothing to worry about. Apparent-
FIGURE 8.12.  Using the computed and critical ly, their new migraine medicine doesn’t
values of t to test the hypothesis.
significantly affect the average systolic
blood pressure.
 The Dependent-Sample t-Test | 229

Let’s Move Forward and Use Our Six-Step Model

Since we are now proficient in doing the calculations by hand, let’s use our six-step
approach to work with a middle school math teacher. In this case, the teacher has
noticed that many of her students are very apathetic toward math. This, she knows,
can contribute to low levels of achievement. The teacher, realizing that something must
be done, begins to investigate ways to get her students more interested in their studies.
During her investigation, the teacher finds a new software package that seems
to be exactly what she needs. The particular package starts by asking students about
their personal interests and activities, and then it creates word problems based on
that information. By tailoring the math lesson to each individual student, the teacher
hopes to foster more interest in her subject. She hopes, of course, this will lead to less
apathy and higher achievement.
This really excites the teacher and a plan is immediately put into effect. The teach-
er plans to measure student interest in math and then use the software for 10 weeks.
At the end of the 10 weeks, she plans on using the same instrument to see if there has
been any change in their feelings toward math.

Identify the Problem

S TEP 1
There is definitely a straightforward, important problem in the
teacher’s room, and she has the expertise, resources, and time
to investigate it. It would be easy for her to collect numeric data
to determine levels of apathy toward math, and there appear to
be no problems with the ethicality of her investigation. Here is
her problem statement:

„ Math teachers have found many students are apathetic toward their subject
matter; this, they feel, may lead to lower achievement. This study will
investigate using a software package that tailors the text of word problems to
each student’s given interests and activities. This, they believe, will lead to
lower apathy and higher achievement.

State a Hypothesis
S TEP 2
From the text of the scenario it is easy to develop the hypothesis
the teacher will be investigating:

„ Levels of student apathy will be significantly lower

after using the new software system for 10 weeks than
they were prior to using the software.

As we can see, we have a directional hypothesis because we have stated that stu-
dents will have lower levels of apathy after the intervention than they did before it.
Again, we stated it in this manner because it reflects the situation that the teacher
wants to investigate. The corresponding null hypothesis would read:
230 | STATISTICS TRANSLATED

„ There will be no significant difference in levels of student apathy before

using the software and levels of student apathy after using the software.

Identify the Independent Variable

S TEP 3
We have two groups we can easily identify—students before
using the software and the same students after using the soft-
ware. Given that, we can call the independent variable “Stu-
dent” and use the two groups as our levels. Since we are mea-
suring the same group at two different points in time (i.e., their
pretest scores and their posttest scores), we will use a depen-
dent-sample t-test.

Identify and Describe the Dependent Variable

S TEP 4
Our dependent variable is student apathy, and we can see the
scores below in Table 8.7. The first column, “Student ID Num-
ber,” is the unique number identified for each student. The sec-
ond column shows the student’s apathy score at the start of the
10 weeks, and the third column shows the same student’s apathy
score at the end of the 10 weeks.

TABLE 8.7.  Student Apathy Scores

Student Starting Ending
ID number apathy score apathy score
1 62 50
2 67 55
3 67 55
4 60 68
5 69 60
6 66 60
7 70 65
8 72 65
9 68 70

SPSS would produce Figure 8.13, our descriptive statistics. In this case, we used
nine pairs of data representing “Starting Apathy” and “Ending Apathy” scores. In
the Mean column, we see two values. The Starting Apathy scores show that students
had an average score of 66.78 on the questionnaire administered at the start of the 10
weeks. The Ending Apathy scores, collected at the end of the 10 weeks, show an aver-
age of 60.89. For each of the mean values, we also see the standard deviation and the
standard error of the mean.
 The Dependent-Sample t-Test | 231

Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 Ending Apathy 60.8889 9 6.67915 2.22638

Starting Apathy 66.7778 9 3.76755 1.25585

FIGURE 8.13.  Descriptive statistics from the Paired-Samples test.

Choose the Right Statistical Test

S TEP 5
In this case, we have one independent variable with two levels
that are related; we say the levels are related because the quan-
titative data collected for one level is related to the quantitative
data in the other level. It is clear that we need to use a depen-
dent-sample t-test.

Use Data Analysis Software to Test the Hypothesis

S TEP 6
Our software would compute the results shown in Figure 8.14.
We can see a mean difference of –5.889, which would allow
us to compute an effect size of .856. While these numbers look
promising, do not let them take your attention away from test-
ing our hypothesis using the p value. In this case, we have to
divide the (Sig. 2-tailed) value of .033 by 2; this gives us a one-tailed p value of .0165
so we will reject the null hypothesis; the ending apathy score is significantly lower
than apathy at the outset of the study; hopefully this will lead to higher achievement
in math!

Paired Samples Test

Pair 1

Ending Apathy -
Starting Apathy
Paired Differences Mean -5.88889
Std. Deviation 6.88194

Std. Error Mean 2.29398

95% Confidence Interval of Lower -11.17882

the Difference Upper -.59896
T -2.567
Df 8
Sig. (2-tailed) .033

FIGURE 8.14.  Inferential statistics from the Paired-Samples test.

232 | STATISTICS TRANSLATED

The Case of the Unexcused Students

In this case, let’s try to help one of our friends: the principal of a local high school.
It seems, with every year that passes, that students miss more and more days due to
unexcused absences. The principal, being a believer in the tenets of behavioral psy-
chology, decides to try motivating the students with a reward. The students are told
that, if they can significantly decrease the number of times they are absent during the
term, they will be rewarded with a party. Let’s use our six-step approach to see if we
can help him investigate whether or not his plan will work.

Identify the Problem

S TEP 1
The principal is faced with somewhat of a two-pronged prob-
lem. Students are missing more and more days of school; since
absences are directly related to lower achievement, it would be
in the best interest of the students to be at school as often as
possible. At the same time, many states base school funding
partially on student attendance rates; it would be in the principal’s best interests (and
career aspirations) to get students to be in school as often as possible. This is clearly a
problem that can be investigated using inferential statistics.

„ The school in question is experiencing a problem with low student

attendance. Since both achievement and school funding are related to
student attendance, it is imperative that action be taken to address the
problem. The principal will investigate whether an extrinsic reward will
help increase attendance.

State a Hypothesis
S TEP 2
Here is the research hypothesis that corresponds to the princi-
pal’s plan:

„ Students will have significantly fewer absences after

the party program than they did before the party
program.

We can see this is a directional hypothesis in that we are suggesting the total num-
ber of absences will be less than they were before starting the study. We can state the
null hypothesis in the following manner:

„ There will be no significant difference in the number of absences prior to the

party program and after the party program is announced.
 The Dependent-Sample t-Test | 233

Identify the Independent Variable

S TEP 3
We could easily identify the independent variable and its lev-
els by noting our interest in one set of students at two distinct
points in time: before the promise of the party and after the
party plan was announced.

Identify and Describe the Dependent Variable

S TEP 4
We know we must collect the number of absences for each stu-
dent, both in the semester before the plan and at the end of the
semester in which the plan was implemented. In Table 8.8, we
can see we have a row that contains the number of absences for
students A through H before the start of the study and another
row that shows the number of absences for the same students at the end of the term.
SPSS would calculate the descriptive statistics shown in Figure 8.15.

TABLE 8.8.  Absence Data

A B C D E F G H
Before the plan 8 7 6 5 6 8 4 3
After the plan 7 6 7 4 6 7 4 3

Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 Ending Absences 5.50 8 1.604 .567
Starting Absences 5.88 8 1.808 .639

FIGURE 8.15.  Descriptive statistics from the Paired-Samples test.

Things are looking good; the average number of absences after the implementa-
tion of the party program is less than the average number before starting the pro-
gram. We have to be careful, though; the values are very close. We should run the
appropriate statistical test and see what the output tells us.

Choose the Right Statistical Test

S TEP 5
Here we have the perfect scenario for a dependent-sample t-test
because there is one independent variable with two related lev-
els and one dependent variable representing quantitative data.

Use Data Analysis Software to Test the Hypothesis

S TEP 6
We can use SPSS to easily compute the inferential statistics we
would need to test our hypothesis; these statistics are shown in
Figure 8.16.
We have a very small standard deviation of the difference
234 | STATISTICS TRANSLATED

Paired Samples Test

Pair 1

Ending
Absences -
Starting
Absences

Paired Differences Mean -.375

Std. Deviation .744
Std. Error Mean .263

95% Confidence Interval of Lower -.997

the Difference Upper .247
T -1.426
Df 7
Sig. (2-tailed) .197

FIGURE 8.16.  Inferential statistics from the Paired-Samples test.

(i.e., .744) and an even smaller mean difference score (i.e., –.375); this gives us an effect
size of .504; things are not looking good for the principal but let’s move forward.
Knowing that we have a one-tailed hypothesis, we would divide our Sig. (2-tailed)
value and arrive at a one-tailed p value of .0985; this means we cannot reject our null
hypothesis. We could verify this by comparing our computed value of t, –1.416, to our
critical t value, 2.365. Although the principal tried, the offer of a party at the end of
the term just was not enough of a stimulus to get students to come to school.

The Case of Never Saying Never

Never one to give up, the principal of the high school decides to try again. Thinking
back over the original plan, the principal decides the offer of a party just was not
enough of a reward to get students to change their behavior. The principal brain-
storms with some teachers, and they develop what they feel is a better plan. The litera-
ture suggests that the traditional school day, because it starts so early in the morning,
is not conducive to getting teenagers to come to school. Knowing that, the principal
suggests starting school at noon and running it until 6:00 P.M. This, the principal
feels, will allow the students to sleep in and still have time in the evening for all of their
activities. Upon announcing the plan to the public, the principal is inundated with
criticism as many feel forcing children to stay later in the day will cause even more
absenteeism. The principal, vowing at least to test the plan, concedes they may be right
but continues to plan for setting up the new schedule.
 The Dependent-Sample t-Test | 235

Identify the Problem

S TEP 1
Obviously, the problem is exactly the same as what the principal
addressed before. His plan worked to a minor degree, but as we
have talked about, that difference may just be caused by error
due to the exact students he worked with, the type of year the
study was conducted, and so on. Given that, it’s time to move
forward and try something else. Remember, his students’ future, not to mention his
job, may hang in the balance! While the problem is the same, the problem statement
will change slightly:

„ The school in question is experiencing a problem with low student

attendance. Since both achievement and school funding are related to
student attendance, it is imperative that action be taken to address the
problem. The principal will investigate whether a change in the school’s
schedule will help increase attendance.

State a Hypothesis
S TEP 2
Based on this story, we can clearly see a hypothesis has formed.
Since the principal is not sure if attendance will go up or down,
the hypothesis must be stated as two-tailed (Step 1):

„ Students attending high school under the new

schedule will have a significantly different number
of absences than they did when they were under the
old schedule.

The null hypothesis would read:

„ There will be no significant difference in absences between the new schedule

and the old schedule.

Identify the Independent Variable

S TEP 3
Just as in the prior example, the independent variable is the
student body. Remember, however, we are measuring them at
two different points in time (i.e., before the new schedule and
after the new schedule).

Identify and Describe the Dependent Variable

S TEP 4
The dependent variable is the number of times the students
were absent; let’s use the data in Table 8.9 to compute the
descriptive statistics shown in Figure 8.17.
236 | STATISTICS TRANSLATED

TABLE 8.9.  Absence Data

A B C D E F G H
Old schedule 4 5 6 5 6 7 6 7
New schedule 5 6 7 7 6 8 6 8

Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 Old Schedule 5.75 8 1.035 .366

New Schedule 6.63 8 1.061 .375

FIGURE 8.17.  Descriptive statistics from the Paired-Samples test.

Choose the Right Statistical Test

S TEP 5
This scenario calls for a dependent sample t-test since the two
levels of the independent variable are related, and we use the
dependent variable to collect parametric data.

Use Data Analysis Software to Test the Hypothesis

S TEP 6
Using the results shown in Figure 8.18, we could compute a fairly
large effect size (i.e., 1.37) but that might be the only good news.
First, we can clearly see the mean number of absences went up
during the trial period. This bad news is compounded when we
look at the two-tailed p value, .006, and see that it is clearly less

Paired Samples Test

Pair 1
Old Schedule -
New Schedule

Paired Differences Mean -.875

Std. Deviation .641

Std. Error Mean .227

95% Confidence Interval of Lower -1.411
the Difference Upper -.339
T -3.862
Df 7
Sig. (2-tailed) .006
FIGURE 8.18.  Inferential statistics from the Paired-Samples test.
 The Dependent-Sample t-Test | 237

than our alpha value of .05. We have to reject our null hypothesis; there was a signifi-
cant difference in the number of absences. Unfortunately for the principal, this means
the new schedule caused a significant increase in the number of absences!

Just in Case—A Nonparametric Alternative

Just as was the case with the independent-sample t-test, there are instances where
either the data distribution is not conducive to using parametric statistics or we have
collected data that are ordinal (ranked) in nature. In cases where this happens and
the levels of the independent variable are dependent on one another, we have to use
the nonparametric Wilcoxon t-test. Setting up, running, and interpreting the output
of the Wilcoxon test is very similar to that of the dependent-sample t-test. Again, this
is not something that happens often, and, like the Mann–Whitney U test, this test is
something you need to keep in the back of your mind for those rare instances.

Summary
The dependent-sample t-test, much like its independent counterpart, is easy to un-
derstand, both conceptually and from an applied perspective. The key to using both
of these inferential techniques is to keep in mind that they can be used only when
you have one independent variable with two levels and when the dependent vari-
able measures quantitative data that is fairly normally distributed. Again, the labels
“independent” and “dependent” describe the relationship between the two levels of
the independent variable that are being compared.
As I said earlier, these two inferential tests are used widely in educational re-
search. What happens, though, when you have more than one independent variable,
more than two levels of an independent variable, or even multiple dependent vari-
ables? These questions, and more, will be answered in the following chapters.

Do You Understand These Key Words and Phrases?

dependent-sample t-test Wilcoxon t-test

Do You Understand These Formulas?

Effect size for a dependent-sample t-test:

xdifference
d=
s difference
238 | STATISTICS TRANSLATED

t score for the dependent-sample t-test:

D
t=
2 − ∑
 ( D)2 
∑ D
n 

 n ( n − 1) 
 

Quiz Time!

As usual, before we wind up this chapter, let’s take a look at a couple of case studies. Read
through these and answer the questions that follow. If you need to check your work, the
answers are at the end of the chapter.

The Case of Technology and Achievement

Proponents of technology in the classroom have suggested that supplying children with laptop
computers will significantly increase their grades.

1. What is the hypothesis being tested?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 8.19 and 8.20, what decision should they make?
Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 Laptops Used 85.2667 15 3.80726 .98303

No Laptops 86.5333 15 4.79385 1.23777

FIGURE 8.19.  Descriptive statistics from the Paired-Samples test.

 The Dependent-Sample t-Test | 239

Paired Samples Test

Pair 1
Laptops Used -
No Laptops

Paired Differences Mean -1.26667

Std. Deviation 6.08824

Std. Error Mean 1.57198

95% Confidence Interval of Lower -4.63822

the Difference Upper 2.10489
T -.806
Df 14
Sig. (2-tailed) .434

FIGURE 8.20.  Inferential statistics from the Paired-Samples test.

The Case of Worrying about Our Neighbors

Citizens are concerned that the annexation of property adjoining their town will decrease the
value of their homes. The city government insists there will be no change.

1. What is the hypothesis being investigated?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 8.21 and 8.22, are the citizens’ concerns warranted?
Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 Before Annex $116,027.73 15 $10,229.535 $2,641.255

After Annex $106,555.87 15 $8,521.641 $2,200.278

FIGURE 8.21.  Descriptive statistics from the Paired-Samples test.

240 | STATISTICS TRANSLATED

Paired Samples Test

Pair 1

After Annex -
Before Annex
Paired Differences Mean $-9,471.867

Std. Deviation $14,517.440

Std. Error Mean $3,748.387

95% Confidence Interval of Lower $-17,511.357

the Difference Upper $-1,432.377
t -2.527
df 14
Sig. (2-tailed) .024

FIGURE 8.22.  Inferential statistics from the Paired-Samples test.

The Case of SPAM

In order to attract new customers, a local entrepreneur is advertising an Internet service that
guarantees customers will receive fewer unsolicited emails per month than they would using
other services. After enrolling for the service, several customers complained that, contrary to
the advertisements, they were actually getting more SPAM!

1. What is the hypothesis the entrepreneur is stating?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 8.23 and 8.24, should the customers complain?
Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 SPAM Before 46.00 16 5.465 1.366

SPAM After 65.88 16 5.976 1.494

FIGURE 8.23.  Descriptive statistics from the Paired-Samples test.

 The Dependent-Sample t-Test | 241

Paired Samples Test

Pair 1

SPAM After -
SPAM Before
Paired Differences Mean 19.875

Std. Deviation 9.444

Std. Error Mean 2.361

95% Confidence Interval of Lower 14.843

the Difference Upper 24.907
T 8.418
Df 15
Sig. (2-tailed) .000

FIGURE 8.24.  Inferential statistics from the Paired-Samples test.

The Case of “We Can’t Get No Satisfaction”

The board of directors at a large corporation recently hired a new President, and, after a few
weeks, the corporation’s administration felt they should survey the management staff to see if
they were satisfied with their choice. Based on the results, managers seemed happy, leaving the
administration feeling they had found exactly the right person for the job. After three months,
however, the complaints starting piling in—something was going seriously wrong, management
seemed very upset! Based on that, the members of the management team were asked to com-
plete the same satisfaction survey again. To the board of director’s dismay, the management’s
satisfaction was lower; the question is, however, was it significantly lower?

1. What is the hypothesis the board of directors is stating?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 8.25 and 8.26, what should the board of directors do?
Paired Samples Statistics

Mean N Std. Deviation Std. Error Mean

Pair 1 Satisfaction Before 60.1333 15 6.63181 1.71233

Satisfaction After 42.2667 15 5.57375 1.43914

FIGURE 8.25.  Descriptive statistics from the Paired-Samples test.

242 | STATISTICS TRANSLATED

Paired Samples Test

Pair 1
Satisfaction
Before -
Satisfaction
After
Paired Differences Mean 17.86667

Std. Deviation 7.70776

Std. Error Mean 1.99013

95% Confidence Interval of Lower 13.59825
the Difference Upper 22.13508
T 8.978
df 14
Sig. (2-tailed) .000

FIGURE 8.26.  Inferential statistics from the Paired-Samples test.

The Case of “Winning at the Lottery”

I’ve often told friends that playing the lottery is a pass-time for people who don’t understand
probability. To investigate this idea, let’s imagine that I can identify a group of 20 people who
will tell me how much they spend on lottery tickets each week. I would then ask them to watch
a brief video on probability as it relates to gambling. My thought is that this would lead to them
to spending less money the following week.

1. What is the hypothesis I am testing?

2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Based on Figures 8.27 and 8.28, what will I learn?

FIGURE 8.27.  Descriptive Statistics from the Paired-Samples Test.

 The Dependent-Sample t-Test | 243

FIGURE 8.28.  Inferential statistics from the Paired-Samples Test.

CHAPTER 9

Analysis of Variance
and Multivariate Analysis of Variance

Introduction
In the last two chapters we dealt with situations in which we have one independent
variable with two levels. In many instances, however, we may have an independent
variable with three or more levels. For example, suppose we are interested in deter-
mining if there is a significant difference in the per capita income between residents
of Canada, the United States, and Mexico. We could easily state a hypothesis to
investigate this question.
There will be a significant difference in the average per capita income between
citizens of Canada, the United States, and Mexico.
Our independent variable is “Country” and the three different countries represent
three levels. The dependent variable is the average income for each country’s citi-
zens. Given all of this, which statistical test can we use?
At about this point, a lot of beginning statisticians say, “This is easy; we will use
three separate independent-sample t-tests. We will compare Canada to the United
States, Canada to Mexico, and then the United States to Mexico. That covers all the
bases.” They are right. Those three comparisons are inclusive of all the comparisons
that could be made between the three countries but using three different t-tests cre-
ates a problem. Let me show you what I mean.
Let’s suppose we actually decided to use the three separate t-tests; by doing so,
we would have three separate computed values of t, up to three separate critical
values of t (depending on the number in each group), and we would be using alpha in
three separate tests. The first two of these things are not so important, but the three

244
 Analysis of Variance and Multivariate Analysis of Variance  |  245

alpha values create quite a problem. In order to understand what might go wrong,
think back to a very early chapter where we talked about Type I and Type II errors.
We agreed that Type I errors occur when we reject a null hypothesis when we
should not; the larger the alpha value, the greater the probability of making an error
of this type. We also said that, if we decrease alpha, we increase the probability of a
Type II error where we fail to reject a null hypothesis when we should. We settled on
a traditional alpha value of .05 because it is a good balance between the two types
of error.
Keeping that in mind, think what might happen if we used three independent
sample t-tests to test our hypothesis. In essence, we would be using an alpha value
of .05 three different times in order to test one overall hypothesis. Obviously, this is
going to greatly inflate our possibility of making a Type I error. In order to address this
problem, an English agronomist, Ronald Fisher, developed the analysis of variance
or, as it is most always abbreviated, ANOVA.
Fisher, originally an astronomer, began working in 1919 as a biologist for an ag-
ricultural station in England. During his tenure there, Fisher contributed greatly to
the fields of genetics and statistics. Like his friend Gossett and his t-test, Fisher was
interested in working with small data samples and eventually developed the analysis
of variance to help him in his studies. His list of accomplishments in statistics and
genetics is so great and his awards are so impressive that he was even knighted by
Queen Elizabeth in 1952!

Understanding the ANOVA

The key to understanding when and how to use the ANOVA is much the same as it is
for any statistical test. First, you need to know the number of independent variables
and the number of levels within each variable. You also need to know how many
dependent variables you have, and the type of data collected; let’s use an example we
can relate to:

„ There will be a significant difference in achievement between students

studying math in a lecture-based classroom, students studying math in a
technology-based classroom, and students studying math in a classroom
where both lecture and technology are used.

It is easy to see, in this hypothesis, that the independent variable is “instructional

type” and there are three levels—lecture, technology, and a combination of both; the
dependent variable is achievement. Notice, also, that we have stated a nondirectional
hypothesis since it would be difficult to state one null hypothesis that would cover all
the different combinations we could possibly check. Given that, when you are dealing
with a situation that calls for an analysis of variance, you always state a nondirectional
hypothesis.
246 | STATISTICS TRANSLATED

Fisher felt that the only way to accurately investigate significant differences
between groups was to look at the variance of scores within each of the levels of the
independent variable as well as the variance between the different groups. We will
discuss the underlying logic and formulas in a bit, but for now let’s look at the differ-
ent types of ANOVAs.

The Different Types of ANOVAs

Like the t-test, there are several types of ANOVA; the one we use in a particular case
is based on the number of independent variables, the number of dependent variables,
and the type of data collected. From the outset, we need to be clear on one thing. If
your dependent variables represent nonparametric data, you will be using one of the
nonparametric ANOVAs, techniques that are rarely used by beginning statisticians.
Because of that, they are beyond the scope of this book. Suffice it to say that non-
parametric ANOVAs exist; we just will not be talking about them. Having said that,
let’s focus on the types of ANOVAs that can be used to analyze a single quantitative
dependent variable.

One-Way ANOVA
The most elementary ANOVA is called the one-way, or simple, ANOVA. When we
say something is “one way,” we mean there is only one independent variable and one
dependent variable. The independent variable can have three or more levels, and the
dependent variable represents quantitative data.

Factorial ANOVA
When we say an ANOVA is “factorial,” we simply mean there is more than one inde-
pendent variable and only one dependent variable. One of the most common ways
we refer to factorial ANOVAs is by stating the number of independent variables. For
example, if we have a three-way ANOVA, we are saying we have a factorial ANOVA
with three independent variables. Sometimes we take it a step further and tell the
reader the number of levels within each of the independent variables. For example,
let’s say we had a two-way ANOVA where the independent variables were “college class”
and “gender” and we are interested in using it to determine if there is a significant dif-
ference in achievement. As we know, “college class” has four levels: freshman, sopho-
more, junior, and senior; obviously, gender has two levels, female and male. In order
to be as clear as possible, someone performing research using this ANOVA might call
it a 4 by 2 (sometimes you see it abbreviated as 4 × 2) ANOVA rather than just a facto-
rial ANOVA. This tells the reader two things.
First, there are two independent variables. The first independent variable is repre-
sented by the “4,” which also shows it has four levels (i.e., freshman, sophomore, junior,
and senior). The second independent variable is represented by the “2,” which means
it has two levels (i.e., male and female). The great thing about using this annotation is
we can tell exactly how many measurements we are going to make by multiplying the
 Analysis of Variance and Multivariate Analysis of Variance  |  247

two numbers together. In this case, we would be collecting eight measurements (4 × 2

= 8). This is shown in Table 9.1.
In another case, we might have a three-way ANOVA where we are collecting data
about (1) the gender of a child, (2) whether the child went to kindergarten, and (3) the
number of parents in the home. Using these variables, we are interested in determin-
ing if there is a significant difference in achievement in the first grade. Here, we
would have a three-way ANOVA because we have three independent variables. (This
three-way ANOVA could also be called a 2 × 2 × 2 ANOVA because gender has two
levels, male or female; kindergarten attendance has two levels, yes or no; and number
of parents in the home has two levels, in this case one or two parents.) We would then
have to collect data about eight different groups (2 × 2 × 2 = 8). Again, we can graphi-
cally represent this in Table 9.2.

TABLE 9.1.  Cells for Gender and Class for Factorial ANOVA
Achievement of Achievement of Achievement of Achievement of
male freshmen male sophomores male juniors male seniors
Achievement of Achievement of Achievement of Achievement of
female freshmen female sophomores female juniors female seniors

TABLE 9.2.  Cells for Gender and Number of Parents for Factorial ANOVA
Males attending Females attending Males not attending Females not
kindergarten— kindergarten— kindergarten— attending
one parent at home one parent at home one parent at home kindergarten—one
parent at home
Males attending Females attending Males not attending Females not
kindergarten— kindergarten— kindergarten— attending
two parents at two parents at two parents at kindergarten—two
home home home parents at home

The factorial ANOVA’s underlying logic, and particularly the computations, are
very complicated. We will touch briefly on its use at the end of the chapter, but we will
spend most of our effort focusing on the one-way ANOVA.

Multivariate ANOVA (MANOVA)

There is one last type of ANOVA that we need to discuss, one where we have more than
one dependent variable, regardless of the number of independent variables. For exam-
ple, suppose we have two groups of soldiers and we want to determine if there is a signifi-
cant difference in levels of accuracy and self-confidence between soldiers trained to fire a
rifle on a live firing range, and soldiers trained to fire a rifle using a virtual training system.
In this case we would have one independent variable, training type (i.e., live or virtual) and
two dependent variables (i.e., accuracy and self-confidence). The MANOVA allows us to
test the effect of training type on a combination of accuracy and self-confidence. This
is done by creating a matrix showing the relationship between each set of values (i.e., an
individual soldier’s accuracy and self-confidence). In cases where we find a significant
difference for the combination of dependent variables, the MANOVA also allows us to
look at each of the dependent variables separately (i.e., the effect of training type on
accuracy and the effect on training type on self-confidence).
248 | STATISTICS TRANSLATED

The multivariate analysis of variance, because of many interactions that multiple

independent and dependent variables can create, is often a complicated test to com-
pute, analyze, and interpret. Much as we did with the factorial ANOVA, we will learn
to develop multivariate hypotheses to address problem statements calling for them
and then use SPSS to analyze data necessary for testing them. Before we move any
further, however, let’s discuss the assumptions we make when using any ANOVA.

Assumptions of the ANOVA

Unfortunately, even when you have the right number of independent and dependent
variables and you are collecting the right type of data, sometimes the ANOVA is
not the appropriate test to work with. The ANOVA, in fact, works under four major
assumptions:

Random Samples
The first of these assumptions is that, if your dependent variable scores represent
a sample of a larger population, then it should be a random sample. Other types of
samples, especially where the researcher picks subjects because of ease or availability,
can make the results of the ANOVA not generalizable to other samples or situations.

Independence of Scores
The second assumption is that the values collected are independent of one another.
For example, if you are collecting data from school children, this means the score that
one child gets is not influenced by, or dependent on, the score received by another
child. This is the same idea as was the case with the independent-sample t-test.

Normal Distribution of Data

The third assumption is that each sample or population studied needs to be somewhat
normally distributed. As we saw in the t-test, this does not mean a perfect bell-shaped
distribution is required, but there must be a certain degree of “normality” for the
underlying calculations to work correctly.

Homogeneity of Variance
The fourth assumption requires that there be “homogeneity of variance.” This means
that the degree of variance within each of the samples should be about the same.
As we will see later, SPSS will provide us with this information, thereby making our
decision-making process far easier.
The first two assumptions do not have anything to do with the actual statistical
process; instead they are methodological concerns and should be addressed when the
data are being collected. For the third assumption, we have already seen that it is very
straightforward to check if data are normally distributed using various numeric and
graphical descriptive statistics. The last assumption, homogeneity of variance, is some-
 Analysis of Variance and Multivariate Analysis of Variance  |  249

thing we address with secondary statistical procedures appropriate for each test. We
saw the same thing when we used the Levene test when interpreting the results of the
independent-sample t-test; we must do the same thing when we are using the ANOVA.
In other words, if we are using the ANOVA and we find the p value of the Levene sta-
tistic is less than .05, we have decisions to make. For now, however, it is important to
discuss the math and logic underlying the one-way ANOVA.

Calculating the ANOVA

The computations for the ANOVA are very straightforward. Simply put, Fisher dem-
onstrated that, in order to test hypotheses using a one-way analysis of variance, one
needed to understand four things:

1. The total variance in the data. This is the sum of the within-group variance and
the between-groups variance.
2. The total degrees of freedom. This is the sum of the within-group degrees of freedom
and the between-groups degrees of freedom.
3. The mean square values for the within-group and between-groups variance.
4. The F value that can be calculated using the data from the first three steps.

We can better understand these values by using the data in Table 9.3. In this case,
we are interested in looking at the effect of different types of medication on headache
relief. Here we have three levels: Pill 1, Pill 2, and Pill 3. The dependent variable,
which we will call “time to relief,” shows the number of minutes that elapse between
the time the pill is taken and the time the headache symptoms are relieved. This will
allow us to state the following hypothesis:

„ There will be no significant difference in the amount of time needed to

relieve headache symptoms between Pill 1, Pill 2, and Pill 3.

TABLE 9.3.  Amount of Time for Pill Type

Pill 1 Pill 2 Pill 3
8 6 8
8 6 9
7 5 2
3 4 8
5 2 2

Descriptive Statistics
We can use the actual ANOVA procedure to calculate the descriptive statistics but,
before we do that, let’s look at how we would set up the data in SPSS. In Figure 9.1, you
250 | STATISTICS TRANSLATED

FIGURE 9.1.  Data for pill and time in the Data View spreadsheet.

can see a variable labeled Pill; there are three values representing the three types of
pills. There is also a variable labeled Time; this is the amount of time before relief.
Once we have entered our data, we will then tell SPSS to run the ANOVA for
us. In Figure 9.2, we begin by selecting Compare Means and then One-Way ANOVA.
At this point, as shown in Figure 9.3, we will be asked to do several things. First, we

FIGURE 9.2.  Using the Compare Means and One-Way ANOVA commands.
 Analysis of Variance and Multivariate Analysis of Variance  |  251

FIGURE 9.3.  Identifying the independent and dependent variables and the statistics to be
computed.

include our variable Time in the Dependent List (i.e., the dependent variable); we fol-
low that by including Pill in the Factor Type (i.e., the independent variable). In this
case, unlike the Independent Sample t-Test, we do not need to include the exact values
for the levels; SPSS simply uses each value entered under Pill as a level. We have also
selected Options and asked for Descriptives and the Homogeneity of variance test.
This will generate the output shown in Figure 9.4.
By looking at the mean scores, it appears that the time it takes Pill 2 to work may
be significantly less than the other two pills, but in order to make sure, we need to
continue by computing the component parts of the F statistic.

The Total Variance

Fisher began by using these data to calculate three values; the total sum of squares, the
between sum of squares, and the within sum of squares. We have seen this idea of sum of
squares earlier, so it will not be a lot of trouble. We start by squaring each value in

Descriptives
Time

Pill One Pill Two Pill Three Total

N 5 5 5 15
Mean 6.2000 4.6000 5.8000 5.5333
Std. Deviation 2.16795 1.67332 3.49285 2.47463
Std. Error .96954 .74833 1.56205 .63895
95% Confidence Interval for Lower Bound 3.5081 2.5223 1.4631 4.1629
Mean Upper Bound 8.8919 6.6777 10.1369 6.9037
Minimum 3.00 2.00 2.00 2.00
Maximum 8.00 6.00 9.00 9.00

FIGURE 9.4.  Descriptive statistics from the ANOVA.

252 | STATISTICS TRANSLATED

Table 9.4; I have substituted x1 for Pill 1, x2 for Pill 2, and x3 for Pill 3 and have gone
ahead and totaled each column.

TABLE 9.4.  Computation of Sums of Squares

x1 x32 x2 x32 x3 x32

8 64 6 36 8 64
8 64 6 36 9 81
7 49 5 25 2  4
3  9 4 16 8 64
5 25 2  4 2  4
2
Sx1 = 31 S x32 = 211 Sx2 = 23 S x32 = 117 Sx3 = 29 S x3 = 217

The Total Sum of Squares

At this point, we have all of the information necessary to compute the total sum of
squares using the following formula; I have abbreviated “sum of squares” as SS:

2
SSTotal = ∑ x −
(∑ x )2
N
First we need to determine the value Sx2. Since it represents each value in the
dataset squared, we sum the second, fourth, and sixth columns:

1. Sx2 = 211 + 117 + 217

2. Sx2 = 545
( x ) ), we follow these steps:
To calculate the second part of the equation (i.e., ∑
2

N
1. Compute Sx by adding the totals of the first, third, and fifth columns (i.e.,
31 + 23 + 29 = 83).
2. Compute (Sx) 2 by squaring Sx (i.e., 83 * 83 = 6889).
3. Divide (Sx) 2 by the total number of data values collected (i.e., 6889/15 =
459.27).

After we have completed these steps, we now have everything we need to include
in the equation:

1. SSTotal = ∑ x 2 − (∑ )2
x
N
2. SSTotal = 545 — 459.27
3. SSTotal = 85.73
 Analysis of Variance and Multivariate Analysis of Variance  |  253

The Between Sum of Squares

The Between Sum of Squares also has a rather lengthy formula but, again, it is just basic
math:
( ∑ x1)2 + (∑ x 2) + ( ∑ x 3)2 − ( ∑ x )2
2
SS Between =
n1 n2 n3 N
First, we square the totals for columns 1, 3, and 5 and then divide each total by the
number of data items in the column. Let’s break this down into four steps:

1. SS Between =
(31)2 + (23)2 + (29)2 − ( ∑ x )2
5 5 5 N

961 529 841 ( ∑ x )

2
2. SS Between = + + −
5 5 5 N

3. SS Between = 192.2 + 105.8 + 168.2 −

( ∑ x )2
N

4. SS Between = 466.2 −
( ∑ x )2
N
Now, in order to determine the value that we must subtract from 466.2, all we do
is complete the following three steps. Notice, this is exactly what we did when we were
computing the total sum of squares:

1. Add the totals for the first, third, and fifth columns (i.e., 31 + 23 + 29 = 83).
2. Square the value from Step 1 (i.e., 83 * 83 = 6889).
3. Divide the value from Step 2 by the total number of data values collected (i.e.,
6889/15 = 459.27).
4. SSBetween = 466.2 – 459.27.

We can now enter that value into our equation:

SSBetween = 466.2 – 459.27

This leaves us with

SSBetween = 6.93

The Within Sum of Squares

Just as we did with the others, we will break the formula for the Within Sum of Squares
down into its component parts:

SSWithin = SS1 + SS2 + SS3

254 | STATISTICS TRANSLATED

Computing the Within Sum of Squares requires that we first compute the Sum
of Squares for each of the levels of the independent variable; they each take the same
general form:
(∑ x1)2
SS1 = ∑ X 12 −
N
To complete this equation, we go through the following three general steps:

1. We square the sum of all of the values from the first level of the independent
variable shown in column one (i.e., 31 * 31 = 961).
2. We divide the result of Step 1 by the number of data values in that column by
N (i.e., 5).
3. We then subtract that value from the sum of the squared values shown in
column 2 of the table (i.e., 211).

We can then include our values in the equation and compute SS1 using the fol-
lowing four steps:

(31)2
1. SS1 = 211 −
5
961
2. SS1 = 211 −
5
3. SS1 = 211 – 192.2
4. SS1 = 18.8

We can then use the same formula to compute SS2 (i.e., 11.2) and SS3 (i.e., 48.8)
and then insert all three values into our original equation:

SSWithin = 18.8 + 11.2 + 48.8

This leaves us with

SSWithin = 78.8

A SHORTCUT

There is an easier way to compute the Total Sum of Squares; all you need to do is add
the Between Sum of Squares to the Within Sum of Squares. This is shown in the fol-
lowing formula:

SSTotal = SSBetween + SSWithin

That means if we know only two of the values, we can easily compute the third.
For example, if we know the Total Sum of Squares and the Sum of Squares Between,
we can insert them into the following formula:
 Analysis of Variance and Multivariate Analysis of Variance  |  255

85.73 = 6.93 + SSWithin

We then subtract 6.93 from both sides and arrive at a Total Sum of Square value
of 78.8:

85.73 – 6.93 = 6.93 + SSWithin – 6.93

If you manually computing these values, you might find that computing the
Between Sum of Squares is the most tedious. Because of that, it is much easier to
compute the Total Sum of Squares and the Within Sum of Squares and use those to
compute the Between Sum of Squares value.

Computing the Degrees of Freedom

As has been the case throughout the book, computing the degrees of freedom is
very straightforward. For the ANOVA, however, you have to compute three different
degrees of freedom values - one for Between Groups, one for Within Groups and a
Total value:

1. The Between Groups degrees of freedom are equal to the number of levels of
the independent variable, minus one. In this case, we have three levels, so we
have two degrees of freedom.
2. The Within Group degrees of freedom are equal to the total number of data
items minus the number of levels in the independent variable. In this case, we
have 12 degrees of freedom (i.e., 15 - 3 = 12).
3. The Total degrees of freedom are equal to the Within Group degrees of free-
dom plus the Between Group degrees of freedom (i.e., 12 + 2 = 14 degrees of
freedom). You can also compute this value by subtracting one from the total
number of data items (i.e., 15 – 1 = 14).

Computing the Mean Square

As we said earlier, the mean square is one of primary components needed to compute
the F value. The mean square is the sum of the Between Group Mean Square and the
Within Group Mean Square. Both are easy to compute by dividing the Sum of Squares
for each group by the number of degrees of freedom for that group. In this case, the
Mean Square values are:

1. Between Groups Mean Square = 3.47 (i.e., 6.93 / 2)

2. Within Group Mean Square = 6.57 (i.e., 78.8 / 12)

Computing the F Value

Finally, we can compute the F value, obviously named for Fisher, by dividing the
between-groups mean square by the within-group mean square. This gives us an F
value of .528 (i.e., 3.47/6.57).
256 | STATISTICS TRANSLATED

We Have the Four Components—Now What?

Before we go any further, look at the output from SPSS, shown in Figure 9.5, it match-
es our computations exactly.

ANOVA
Pill
Sum of Squares df Mean Square F Sig.

Between Groups 6.933 2 3.467 .528 .603

Within Groups 78.800 12 6.567
Total 85.733 14

FIGURE 9.5.  Inferential statistics from the ANOVA.

The F Distribution
At this point, we have our computed value of F. Now we need an F distribution so that
we can determine the critical value needed to test our hypothesis. This process is simi-
lar to how t distributions are created, in
that an infinite number of F values are
computed from random samples of data
and then plotted. This process is shown
in Figure 9.6.
This distribution is certainly a lot
different from the normal curves we
have been using up to this point. First,
there are no plotted values less than
zero; this is explained by the fact that
we are using the variance which, as you
know, can never be less than zero. As for
the strange shape of the distribution,
that takes a little more explaining.
When Fisher first developed the
F distribution, he discovered that the
shape of the distribution depends on
the “within” and “between” degrees of
freedom. In the following example, you
can see a plot of three distributions
with different degrees of freedom. As
you can see in Figure 9.7, the larger the
between-groups degrees of freedom,
shown on the horizontal axis, the more
the distribution is skewed to the right;
the more degrees of freedom you have
for within groups, shown on the vertical
axis, the more peaked the distribution.
FIGURE 9.6.  Creating the F distribution.
 Analysis of Variance and Multivariate Analysis of Variance  |  257

Determining the Area under the

Curve for F Distributions
Like the z (i.e., normal) distribution and
the t distribution, we get the critical
value of F from a table related to its dis-
tribution. An example is shown in Table
9.5.
Since it is so small, obviously this
FIGURE 9.7.  Changes in the shape of the F distri-
isn’t a complete F table; in fact, it is only
bution based on sample sizes.
one part of the F table for alpha = .05.
Since, like the other tests, the size of the
critical region changes depending on the alpha value, there are completely different F
tables for alpha = .10 and alpha = .01. This is done because having two separate degrees
of freedom requires entries on the top and left sides of the table. In order to avoid
confusion, it is far easier to limit the entries on a given table to a specific alpha value.
Examples of tables for all alpha values are shown in Appendices C, D, and E.

TABLE 9.5.  Critical Values of F

Within-group Between-groups degrees of freedom
degrees of
freedom 1 2 3 4 5 6 7 8 9 10
 1 1.61 2.00 2.16 2.25 2.30 2.34 2.37 2.39 2.41 2.42
 2 18.5 19.0 19.2 19.3 19.3 19.4 19.4 19.4 19.4 19.4
 3 10.1 9.55 9.28 9.12 9.01 8.94 8.89 8.85 8.81 8.79
 4 7.71 6.94 6.59 6.39 6.26 6.16 6.09 6.04 6.00 5.96
 5 6.61 5.79 5.41 5.19 5.05 4.95 4.88 4.82 4.77 4.74
 6 5.99 5.14 4.76 4.53 4.39 4.28 4.21 4.15 4.10 4.06
 7 5.59 4.74 4.35 4.12 3.97 3.87 3.79 3.73 3.68 3.64
 8 5.32 4.46 4.07 3.84 3.69 3.58 3.5 3.44 3.39 3.35
 9 5.12 4.26 3.86 3.63 3.48 3.37 3.29 3.23 3.18 3.14
10 4.96 4.10 3.71 3.48 3.33 3.22 3.14 3.07 3.02 2.98
11 4.84 3.98 3.59 3.36 3.20 3.09 3.01 2.95 2.90 2.85
12 4.75 3.89 3.49 3.26 3.11 3.00 2.91 2.85 2.80 2.75

Having said that, if you are like me, upon first seeing the F table, you are probably
thinking, “This isn’t so bad, it looks a lot like our t and z tables.” If so, you are right;
there is only one small difference. First, you can see the two degrees of freedom val-
ues we have talked about. The horizontal row across the top shows you the “between
groups degrees of freedom,” and the vertical column on the left shows you the “with-
in-group degrees of freedom.”
Now we are going to use this table, along with our data about the different head-
ache remedies, to test this hypothesis:

„ There will be no significant difference in the amount of time needed to

relieve headache symptoms between Pill 1, Pill 2, and Pill 3.
258 | STATISTICS TRANSLATED

First, we must locate the critical value of F for situations where we have 2 between-
groups degrees of freedom and 12 within-group degrees of freedom. To do this, we go
across the top row until we get to the column showing 2 degrees of freedom; we then
go down the left column until we get to 12 degrees of freedom. The point where that
row and that column intersect shows us an F value of 3.89; we can mark that on our F
distribution shown in Figure 9.8. In Figure 9.9, we then plot our computed value of F.
In order to test our hypothesis, we
compare these two values. If our com-
puted value of F falls to the right of the
critical value, we reject the null hypoth-
esis. If our computed value of F is equal
to or to the left of the critical value of F,
then we do not reject the null hypoth-
esis. This, of course, is the same logic
we used with the z and t-tests. In this
case it means we cannot reject the null
hypothesis; obviously, there is no signifi-
cant difference in headache relief time
between the three types of pills.

FIGURE 9.8.  The critical value of F equals 3.89. The p Value for an ANOVA
The decision to fail to reject the null
hypothesis is supported by a p value of
.603, shown as Sig., back in Figure 9.5.
This simply means there are no signifi-
cant differences between any of the com-
binations of pills: Pill 1 compared to Pill
2, Pill 1 compared to Pill 3 or Pill 2 com-
pared to Pill 3. You will see that when we
reject a null hypothesis, we must take it
a step further. We’ll use tools called post-
hoc tests to help us determine where the
difference lies.

Effect Size for the ANOVA

FIGURE 9.9.  Comparing a computed value of F
to a critical value of F. You can also compute an effect size for
the analysis of variance. Again, all we
are doing is investigating the effect of the independent variable on the dependent
variable. Calculating the effect size, called eta-squared, is easy:

h2 = between sum of squares/total sum of squares

Substituting our values into the formula, we will compute an effect size of .08.
 Analysis of Variance and Multivariate Analysis of Variance  |  259

1. h2 = 6.933/85.733
2. h2 = .08

This supports our decision not to reject the null hypothesis. Apparently, the type
of pill you take does not make a lot of difference. At the same time, a person with a
really bad headache might look at the mean scores and choose Pill 2; while it is not
significantly faster, it is somewhat faster!

Testing a Hypothesis Using the ANOVA

We are going to use our six-step model to discuss another case and, again, we will do
all the calculations manually.

The Case of Multiple Means of Math Mastery

A schoolteacher has a large group of children in her math class. The class is equally dis-
tributed in terms of gender, ethnic breakdown, and academic achievement levels. The
teacher is a progressive graduate student and has learned in one of her classes that
computer-assisted instruction (CAI) may help children realize higher achievement in
mathematics.
The teacher decides to divide her class into three equal groups; the first group
will receive a lecture, the second will use computer-assisted instruction, and the last
group will receive a combination of both. The teacher is interested in finding out,
after a semester of instruction, if there is any difference in achievement. She can then
use the six-step model in her investigation.

S TEP 1 Identify the Problem

In this case, a problem or opportunity for research clearly
exists. The teacher wants to know if using technology will help
increase achievement in her class, but we need to ensure that it
meets the criteria for a good problem:

1. Yes—she is interested in this problem.

2. Yes—the scope of the problem is manageable; she would simply need to col-
lect grades from each of her students.
3. Yes—the teacher has the knowledge, time, and resources needed to investi-
gate the problem.
260 | STATISTICS TRANSLATED

4. Yes—she can collect numeric data representing her students’ grades.

5. Yes—investigating the problem has theoretical or practical significance.
6. Yes—it is ethical to investigate the problem.

It seems that our problem meets all criteria; remember, as we write it, it must be
clear and concise, contain all variables to be considered, and not interject the bias of
the researcher:

„ This study will investigate the effect of technology on the achievement, both
alone and in combination with lecture, of students in a math class.

S TEP 2 State a Hypothesis

The hypothesis the teacher wants to investigate is clearly formu-
lated. Remember, because of the shape of the distribution, we
will only state nondirectional hypotheses.

„ There will be a significant difference in levels of

achievement between students taking math in a
lecture group, students taking math in a CAI group,
and students getting a combination of CAI and
lecture.

Our null hypothesis would be:

„ There will be no significant difference in levels of achievement between

students taking math in a lecture group, students taking math in a CAI
group, and students getting a combination of CAI and lecture.

S TEP 3 Identify the Independent Variable

The independent variable is “type of instruction” or “instruc-
tional method,” and there are three levels: students in the lec-
ture group, students in the CAI group, and students using a
combination of lecture and CAI.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is achievement and she will be using
a final grade based on lecture, CAI, or a combination; this is
shown in Table 9.6. Figure 9.10 shows our descriptive statistics.
 Analysis of Variance and Multivariate Analysis of Variance  |  261

Descriptives
Grade

Lecture CAI Combination Total

N 10 10 10 30
Mean 78.0000 86.4000 93.6000 86.0000
Std. Deviation 5.73488 3.09839 2.95146 7.61124
Std. Error 1.81353 .97980 .93333 1.38962
95% Confidence Interval for Lower Bound 73.8975 84.1835 91.4887 83.1579
Mean Upper Bound 82.1025 88.6165 95.7113 88.8421
Minimum 68.00 82.00 90.00 68.00
Maximum 88.00 92.00 100.00 100.00

FIGURE 9.10.  Descriptive statistics from the ANOVA.

TABLE 9.6.  Final Grade Data for Instructional Type

Lecture CAI CAI/lecture
80 84 90
82 86 92
78 84 94
76 82 90
80 88 100
72 90 94
82 92 94
88 88 96
74 86 92
68 84 94

In this case, the mean scores between the three groups are different, but the same
question must be asked, “Are they significantly different?”

S TEP 5 Choose the Right Statistical Test

Here we have one independent variable with three levels and one
dependent variable where the data are quantitative. In this case,
it appears we need to use an analysis of variance (ANOVA).

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Before we commit ourselves to an ANOVA, however, think back
to the assumption regarding the homogeneity of variance. As
we said, we need to have about the same amount of variance
within each of the samples. It is important that we consider
this because, if the degree of variance is significantly different
between the groups, the calculations underlying the ANOVA could be thrown off and
262 | STATISTICS TRANSLATED

would affect the validity of our results. Because manually calculating these values is
something we would never do, SPSS creates the table shown in Figure 9.11.

Test of Homogeneity of Variances

Grade

Levene Statistic df1 df2 Sig.

2.557 2 27 .096
FIGURE 9.11.  Homogeneity of variance statistics from the ANOVA.

We will use this table to test the null hypothesis: “There is no significant differ-
ence in the variance within the three sets of data.” Just like with the t-test, all we must
do is compare our p value (i.e., Sig.) to an alpha value of .05. Obviously, since our
p value is larger than alpha, there is no significant difference in the degree of variance
within the three sets of data. If the p value was less than .05, we might need to revert
to the ANOVA’s nonparametric equivalent; the Kruskal–Wallis H test. Keep in mind,
however, that the difference in the variances between the groups must be very, very
large for that to ever happen. Given that, we can go ahead and use the SPSS output
shown in Figure 9.12.

ANOVA
Grade

Sum of Squares Df Mean Square F Sig.

Between Groups 1219.200 2 609.600 35.719 .000

Within Groups 460.800 27 17.067
Total 1680.000 29

FIGURE 9.12.  Inferential statistics from the ANOVA.

We finally have everything we need to test our hypothesis:

„ There will be no significant difference in levels of achievement between

students taking math in a lecture group, students taking math in a CAI
group, and students getting a combination of CAI and lecture.

To begin with, since this is our first case with an ANOVA, let’s plot both our com-
puted value of F and our critical F value. Remember, we determine the critical value
of F by using both the within and between-groups degrees of freedom in combination
with our desired alpha value. In Figure 9.13, we would find that these values represent
a critical value of 3.35.
Our computed value of F (i.e., 35.719) falls very far to the right of our critical value
 Analysis of Variance and Multivariate Analysis of Variance  |  263

of F (i.e., 3.35); there is a significant dif-

ference between the three groups. SPSS
also computed a p value of .000, and
we could manually compute an effect
size of .726. Both values support our
decision. Here we have somewhat of a
conundrum; we have rejected the null
hypothesis and can feel assured that a
significant difference exists between the
groups. The questions that arise, howev-
er, are “Which groups are different?” Is
the CAI group different from Lecture?
Is the Lecture group different from
FIGURE 9.13.  Comparing critical and computed the Combination group? Is the Com-
values of F to test a hypothesis.
bination group different from the CAI
group? How can this be determined?”

The Post-Hoc Comparisons

As we said earlier, in order to determine which groups are significantly different, we
could use separate t-tests to compare the lecture and CAI groups, the lecture and
combination groups, and the CAI and combination groups. Remember, however, that
approach is not statistically sound. This would mean that you would have a separate
alpha value for each of the t-tests, which, of course, would greatly inflate the Type I
error rate. Instead of doing that, we can use a set of tools called multiple comparison
tests to help us make a sound decision.

Multiple-Comparison Tests
The multiple-comparison tests do exactly what their name implies. They make com-
parisons between the levels of an ANOVA and tell us which are significantly different
from one another. We won’t get into the manual calculations; suffice it to say that they
are a modified version of the independent-sample t-test that controls for the inflated
Type I error rate. We will always use computer software to compute the right value
for us.
There are several multiple-comparison tests we can choose from, and most of the
time they will provide us with the same results. There are situations that require a
specific multiple-comparison test, but these do not arise that often. For our purposes,
one of the most commonly used multiple-comparison tests is the Bonferroni test. The
procedure is easy to understand, and it gives the researcher a good estimate of the
probability that any two groups are different while, at the same time, controlling for
Type I and Type II errors. Typical output for the Bonferroni procedure is shown in
Figure 9.14.
You can see that there are four rows in the table; the bottom three rows represent
the three levels of our independent variable—the Lecture group, the CAI group, and
the Combination group. If we look at the lecture row, for example, we can see we are
given statistics representing it and its relationship to the other levels of the indepen-
264 | STATISTICS TRANSLATED

Multiple Comparisons
Grade
Bonferroni
(I) Method (J) Method Mean 95% Confidence Interval
Difference (I-J) Std. Error Sig. Lower Bound Upper Bound
*
Lecture CAI -8.40000 1.84752 .000 -13.1157 -3.6843
*
Combination -15.60000 1.84752 .000 -20.3157 -10.8843
*
CAI Lecture 8.40000 1.84752 .000 3.6843 13.1157
*
Combination -7.20000 1.84752 .002 -11.9157 -2.4843
*
Combination Lecture 15.60000 1.84752 .000 10.8843 20.3157
*
CAI 7.20000 1.84752 .002 2.4843 11.9157

*. The mean difference is significant at the 0.05 level.

FIGURE 9.14.  Multiple-comparison tests following the ANOVA.

dent variable. There is one quirk, however: because of the total number of possible
comparisons, part of the table is redundant.
For example, in the third row where you compare the CAI group to the Combina-
tion group, the mean difference is –7.2 points. In row 4, you compare the same two
groups; only, this time, you subtract the average CAI score from the average combina-
tion score. This gives you a difference, obviously, of +7.2 points. The order in which
you make the subtraction doesn’t matter; the point here is that there is a 7.2-point
difference between the groups.
We can also see an asterisk next to each of the mean difference values; this indi-
cates that the mean difference is significant when alpha equals .05. For example, we
see that the mean difference between the Lecture and the CAI group is –8.4; this
difference is statistically significant. In this case, we can see asterisks next to all our
mean difference values; each of the groups is significantly different from one another.
We can verify our results by looking at the p value in the fourth column (i.e., Sig.).
Here, we can see the p value between each of these; this means the two groups in that
given comparison are significantly different from one another. Again, the p value
supports the significant difference shown by the asterisks next to all mean difference
comparisons.

Always Observe the Means!

At this point, we have significant differences in scores between each of the three
groups. The average score of the combination group is significantly higher than the
average score of the CAI group, and it, in turn, is significantly higher than the lecture
group. As was the case with the t-tests, it is imperative to remember that we must look
at the mean values from the descriptive statistics before we make any decisions. In this
case, educators might want to adopt the Combination approach, but remember, the
same low p values would exist if the mean differences were in the opposite direction
(e.g., the Combination group was significantly lower than the Lecture group).
 Analysis of Variance and Multivariate Analysis of Variance  |  265

So, what do we know from all of this? Apparently, the use of a combination of
lecture and CAI results in achievement scores significantly higher than either lecture
alone or CAI alone. At the same time, the use of CAI is significantly more effective
than the use of lecture alone. The teacher should use this information to plan future
class sessions accordingly.

The Case of Seniors Skipping School

A principal (are we back to this guy again?) has heard a rumor that seniors, because
they are getting close to the end of their high school careers, are starting to skip class
regularly which, if bad grades follow, could affect their ability to graduate. Obviously,
the principal doesn’t think this is a very good idea, and so he initiates an investigation.

S TEP 1 Identify the Problem

Yes, the principal is still interested in this problem and has the
authority, time, resources, and knowledge to address it. This
problem is practically significant and well within the scope of
the principal’s authority. It would be unethical if the principal
didn’t investigate this problem by simply counting the number
of absences of each senior. We could state it in the following manner:

„ The administration and faculty in the high school have noticed what they
believe is an inordinate number of absences by students in the senior class
when compared to students in other grades. Afraid that higher levels of
absenteeism could lead to lower grades and fewer graduates, personnel at
the school will attempt to determine if seniors are skipping more classes than
students in other grades.

S TEP 2 State a Hypothesis

For the principal to decide if a significant difference exists in
the number of absences by class in the high school, he could use
this hypothesis:

„ There will be a significant difference in the number

of absences between freshmen, sophomores, juniors,
and seniors.

The corresponding null hypothesis would state that no significant difference

would exist.
266 | STATISTICS TRANSLATED

S TEP 3 Identify the Independent Variable

Our independent variable is the class a given student is in and
there are four levels: freshman, sophomore, junior, and senior.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is the number of days absent; we’ll use
Table 9.7 to show data the principal collected from the school’s
registrar. Figure 9.15 shows the descriptive statistics that would
be computed by SPSS. Here, the means for the freshman, soph-
omore, and junior classes are all approximately equal; the aver-
age for seniors is much larger. Our question, as always is, “Are the number of absences
different, or are they significantly different?” Let’s pick the appropriate statistical test
and find out.

Descriptives
Absences

Freshman Sophomore Junior Senior Total

N 18 18 18 18 72
Mean 2.6667 2.7222 2.7778 5.1667 3.3333
Std. Deviation 1.45521 1.56452 1.06027 1.24853 1.69506
Std. Error .34300 .36876 .24991 .29428 .19977
95% Confidence Interval for Lower Bound 1.9430 1.9442 2.2505 4.5458 2.9350
Mean Upper Bound 3.3903 3.5002 3.3050 5.7875 3.7317
Minimum .00 .00 1.00 3.00 .00
Maximum 5.00 5.00 4.00 7.00 7.00

FIGURE 9.15.  Descriptive statistics from the ANOVA.

TABLE 9.7.  Number of Absences for Each Class

Freshman Sophomore Junior Senior
4 1 5 1 3 2 5 6
3 2 4 1 4 4 5 7
2 1 3 2 2 2 5 6
4 5 4 3 3 4 7 7
5 3 5 2 4 4 5 4
2 4 1 3 3 2 4 3
1 4 1 5 1 4 4 5
0 3 2 4 1 2 6 3
2 2 0 3 2 3 6 5
 Analysis of Variance and Multivariate Analysis of Variance  |  267

S TEP 5 Choose the Right Statistical Test

Since we have one independent variable with four levels and our
dependent variable is quantitative data, we are going to use the
one-way analysis of variance.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

In Figure 9.16, our Levene p value is much greater than .05, so
we have no problems with the homogeneity of variance between
the different datasets. This means we can continue with the
ANOVA table shown in Figure 9.17. Since our p value is less
than our alpha value of .05, we will reject our null hypothesis
and support our research hypothesis. We could compute a moderate effect size of .396
that supports our decision. After looking at the descriptive statistics, we are fairly sure
where the significant difference lies, but we can look at our multiple comparison tests
in Figure 9.18 to make sure.

Test of Homogeneity of Variances

Absences
Levene Statistic df1 df2 Sig.

1.318 3 68 .276
FIGURE 9.16.  Homogeneity of variance statistics from the ANOVA.

ANOVA
Absences
Sum of Squares df Mean Square F Sig.
Between Groups 80.778 3 26.926 14.859 .000
Within Groups 123.222 68 1.812
Total 204.000 71

FIGURE 9.17.  Inferential statistics from the ANOVA.

This table verifies what we suspected. Since there are asterisks beside each of the
mean difference values between the seniors and every other class, it is apparent they
have a significantly higher number of absences than those in the other classes. There
is no significant difference in the mean scores between any of the other classes. As
always, be careful; the same statistics would have been computed had the seniors had
significantly fewer absences.
268 | STATISTICS TRANSLATED

Multiple Comparisons
Absences
Bonferroni

(I) Class (J) Class Mean 95% Confidence Interval

Difference (I-J) Std. Error Sig. Lower Bound Upper Bound
Freshman Sophomore -.05556 .44871 1.000 -1.2750 1.1639

Junior -.11111 .44871 1.000 -1.3305 1.1083

*
Senior -2.50000 .44871 .000 -3.7194 -1.2806
Sophomore Freshman .05556 .44871 1.000 -1.1639 1.2750
Junior -.05556 .44871 1.000 -1.2750 1.1639
*
Senior -2.44444 .44871 .000 -3.6639 -1.2250
Junior Freshman .11111 .44871 1.000 -1.1083 1.3305
Sophomore .05556 .44871 1.000 -1.1639 1.2750
*
Senior -2.38889 .44871 .000 -3.6083 -1.1695
*
Senior Freshman 2.50000 .44871 .000 1.2806 3.7194
*
Sophomore 2.44444 .44871 .000 1.2250 3.6639
*
Junior 2.38889 .44871 .000 1.1695 3.6083

*. The mean difference is significant at the 0.05 level.

FIGURE 9.18.  Multiple-comparison tests following the ANOVA.

The Case of Quality Time

Many social workers are concerned that, as their number of children grows, parents
are spending less and less quality time with each newborn child. They are concerned
because they believe this lack of parental involvement leads to higher levels of anger,
hostility, and instances of antisocial behavior as the child grows older. In order to
investigate the social workers’ concerns, a group of researchers decides to identify a
population, determine the birth order of each person in the population, and then ask
each person the number of times they have been arrested. They begin with a clear,
well-defined problem statement.

S TEP 1 Identify the Problem

This problem clearly meets all six criteria for a good problem
statement:

„ Social workers are concerned that children with

older siblings will receive less parental attention as
they grow older. This leads to higher levels of anger,
hostility, and antisocial behavior later in life.
 Analysis of Variance and Multivariate Analysis of Variance  |  269

S TEP 2 State a Hypothesis

„ There will be a significant difference in the number

of arrests and the relative birth order of a child.

S TEP 3 Identify the Independent Variable

In our hypothesis, relative birth order is our independent vari-
able. For our example, we are going to assume there are no
more than five levels. A child was either born first, second,
third, fourth, or fifth.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is the number of arrests; Table 9.8
shows data for 50 people. Figure 9.19 shows our descriptive sta-
tistics.

Descriptives
Arrests

Born First Born Second Born Third Born Fourth Born Fifth Total
N 10 10 10 10 10 50
Mean 3.0000 2.7000 3.4000 3.9000 1.1000 2.8200
Std. Deviation 1.24722 .82327 1.42984 1.10050 .73786 1.42414
Std. Error .39441 .26034 .45216 .34801 .23333 .20140
95% Confidence Interval for Lower Bound 2.1078 2.1111 2.3772 3.1127 .5722 2.4153
Mean Upper Bound 3.8922 3.2889 4.4228 4.6873 1.6278 3.2247
Minimum 1.00 1.00 1.00 3.00 .00 .00
Maximum 4.00 4.00 5.00 6.00 2.00 6.00

FIGURE 9.19.  Descriptive statistics from the ANOVA.

TABLE 9.8.  Number of Arrests by Birth Order

Born first Born second Born third Born fourth Born fifth
4 3 3 4 1
2 2 4 3 2
3 3 2 3 1
3 3 5 5 0
4 4 5 4 2
1 3 3 6 1
4 3 2 3 1
4 2 1 3 0
4 1 5 5 2
1 3 4 3 1
270 | STATISTICS TRANSLATED

S TEP 5 Choose the Right Statistical Test

Here again we have met the criteria for a one-way ANOVA; one
independent variable with five levels, and our dependent vari-
able is quantitative.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Again, we assume we are using the one-way ANOVA, but we
must check the Levene test in Figure 9.20 to make sure.
Our p value is fairly low but still above our alpha level of .05;
this means we do not have to worry about our assumption that
the variance is homogeneous between the five groups. Based on
that, Figure 9.21 shows us the ANOVA table.

Test of Homogeneity of Variances

Arrests

Levene Statistic df1 df2 Sig.

2.218 4 45 .082
FIGURE 9.20.  Homogeneity of variance statistics from the ANOVA.

ANOVA
Arrests

Sum of Squares df Mean Square F Sig.

Between Groups 45.080 4 11.270 9.340 .000

Within Groups 54.300 45 1.207
Total 99.380 49

FIGURE 9.21.  Inferential statistics from the ANOVA.

Here we have a p value less than .05, so we will reject our null hypothesis. This is
supported by a moderate effect size of .454. Apparently, the social workers are right;
there is a significant difference in the number of arrests when you look at birth order.
We do need to look at our post-hoc tests in Figure 9.22 to see where those differences
lie. If they are right, the people born last will have a significantly higher number of
arrests.
What does this tell us? The first thing it tells us is that, for the social workers,
things can go from looking pretty good to looking pretty bad in just a matter of sec-
onds. First, if they had paid attention to the descriptive statistics in Step 3, they would
have already known that children born last have fewer arrests; the fact that they are
significantly less is supported by the p value. The birth order of children born first
through fourth makes no difference in the number of arrests, but children born fifth
have significantly fewer arrests than their siblings.
 Analysis of Variance and Multivariate Analysis of Variance  |  271

Multiple Comparisons
Arrests
Bonferroni
(I) Birth Order (J) Birth Order Mean 95% Confidence Interval
Difference (I-J) Std. Error Sig. Lower Bound Upper Bound

First Second .30000 .49126 1.000 -1.1502 1.7502

Third -.40000 .49126 1.000 -1.8502 1.0502

dimension3

Fourth -.90000 .49126 .736 -2.3502 .5502

*
Fifth 1.90000 .49126 .004 .4498 3.3502
Second First -.30000 .49126 1.000 -1.7502 1.1502

Third -.70000 .49126 1.000 -2.1502 .7502

dimension3

Fourth -1.20000 .49126 .186 -2.6502 .2502

*
Fifth 1.60000 .49126 .021 .1498 3.0502
Third First .40000 .49126 1.000 -1.0502 1.8502
Second .70000 .49126 1.000 -.7502 2.1502
dimension2 dimension3

Fourth -.50000 .49126 1.000 -1.9502 .9502

*
Fifth 2.30000 .49126 .000 .8498 3.7502

Fourth First .90000 .49126 .736 -.5502 2.3502

Second 1.20000 .49126 .186 -.2502 2.6502
dimension3

Third .50000 .49126 1.000 -.9502 1.9502

*
Fifth 2.80000 .49126 .000 1.3498 4.2502
*
Fifth First -1.90000 .49126 .004 -3.3502 -.4498
*
Second -1.60000 .49126 .021 -3.0502 -.1498
*
dimension3

Third -2.30000 .49126 .000 -3.7502 -.8498

*
Fourth -2.80000 .49126 .000 -4.2502 -1.3498

*. The mean difference is significant at the 0.05 level.

FIGURE 9.22.  Multiple-comparison tests following the ANOVA.

The Case of Regional Discrepancies

The president of a manufacturing company with factories in four locations through-
out the United States has recently been informed that one of the factories may be lag-
ging the others in production. This bothers the president since each factory supports a
specific region of the country. Lower production means fewer products for customers.
When stores run out of a specific product, buyers might just turn to the competitor’s
brand; it is time to act!
272 | STATISTICS TRANSLATED

S TEP 1 Identify the Problem

The company’s president has reason to be alarmed; factories
lagging in production could certainly affect the company’s bot-
tom line. This, of course, might cause the Board of Directors to
consider the president’s bottom line. Since he has three kids in
college, it behooves him to investigate whether there is actually
a problem!

„ Different areas of the country depend on one of four regional manufacturing

factories for needed products. If a given factory falls behind in production, it
is possible that consumers will turn to products from another manufacturer
to meet their needs.

S TEP 2 State a Hypothesis

The company’s president would begin with the following
research hypothesis:

„ There will be a significant difference in production

between the four plants.

The corresponding null hypothesis would simply state that no significant differ-
ence would exist.

S TEP 3 Identify the Independent Variable

If you violate any of the assumptions that underlie using an
analysis of variance, the best you can do is to run a Kruskal–
Wallis H test for each of the independent variables. This test is
the nonparametric alternative for the one-way ANOVA. This, of
course, inflates the probability of making a Type I error but, if
you are ambitious, there is a test called Friedman’s ANOVA that can be modified to
work in this scenario.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is production. We’ll use the data values
in Table 9.9 to represent the number of units manufactured by
each plant during the past 12 months. Figure 9.23 shows our
descriptive statistics.
 Analysis of Variance and Multivariate Analysis of Variance  |  273

Descriptives
Units Produced

95% Confidence Interval for Mean

N Mean Std. Deviation Std. Error Lower Bound Upper Bound Minimum Maximum

Plant A 12 1003.2500 32.15057 9.28107 982.8225 1023.6775 954.00 1050.00

Plant B 12 1000.7500 31.53678 9.10388 980.7125 1020.7875 953.00 1032.00
Plant C 12 1003.1667 27.65315 7.98278 985.5967 1020.7366 953.00 1035.00
Plant D 12 1001.9167 27.62561 7.97483 984.3642 1019.4691 957.00 1044.00
Total 48 1002.2708 28.86781 4.16671 993.8885 1010.6532 953.00 1050.00

FIGURE 9.23.  Descriptive statistics from the ANOVA.

TABLE 9.9.  Number of Units Manufactured by Each Plant

Plant A Plant B Plant C Plant D
1037  968  983  969
1002  953  953 1013
1003  959 1024 1015
1013 1026  984 1019
1050  968  965 1014
 954 1027 1033 1029
1034 1015 1019  970
1036 1032 1019 1001
 998 1024 1006  957
 980 1031 1028 1044
 975 1023  989 1018
 957  983 1035  974

We can see, right away, that there is probably not a significant difference in pro-
duction rates between the four plants since the mean scores seem to be very, very
close. Using the correct statistical test will help us be sure.

S TEP 5 Choose the Right Statistical Test

Here we have one independent variable with four levels: Plant
A, Plant B, Plant C, and Plant D, and our dependent variable
is quantitative in nature; it appears we will use an ANOVA to
analyze our data.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

There is no problem with the homogeneity of variance. Given
that, we’ll use the analysis of variance; we’ll start by looking at
Figure 9.24 to check our homogeneity of variance and then use
the ANOVA table in Figure 9.25 to test our hypothesis.
274 | STATISTICS TRANSLATED

Test of Homogeneity of Variances

Units Produced

Levene Statistic df1 df2 Sig.

.428 3 44 .734
FIGURE 9.24.  Homogeneity of variance statistics from the ANOVA.

ANOVA
Units Produced

Sum of Squares Df Mean Square F Sig.

Between Groups 50.396 3 16.799 .019 .996

Within Groups 39117.083 44 889.025
Total 39167.479 47

FIGURE 9.25.  Inferential statistics from the ANOVA.

Our p value, .996, is very large; obviously, we will not reject our null hypothesis.
We could also compute a very small effect size of .0012 that supports our decision.
Although we would not have to, we can check our multiple comparison tests in Figure
9.26 just to see how they look when no significant difference is detected.
In this case, there are no asterisks marking any of the mean differences; this veri-
fies what we saw with the large p value and small effect size. Apparently, the rumor
regarding low production was only that. Now all the president must worry about is if
they are all running up to capacity!

The Factorial ANOVA

At the beginning of this chapter, we learned that a factorial ANOVA is used when a
researcher has more than one independent variable but only one dependent variable
that represents quantitative data. Most entry-level statistics books do not cover this
topic, for three reasons.

1. Although it may appear to be quite clear cut, you will find when you get into
scenarios where you have three or more independent variables that things
can get quite complicated in a hurry. For example, if you have a 4 × 2 × 3 facto-
rial ANOVA, you are dealing with 24 different comparisons.
2. Second, the math underlying this test is far beyond what we want to cover in
this book. Because of that, we will not manually calculate it.
3. If you violate any of the assumptions that underlie using an ANOVA, the best
solution is to run a Kruskal–Wallis H test, the nonparametric alternative for
the one-way ANOVA, for each of the independent variables. This, of course,
inflates the probability of making a Type I error. If you are ambitious, there
 Analysis of Variance and Multivariate Analysis of Variance  |  275

Multiple Comparisons
Units Produced
Bonferroni
(I) (J) Plant Mean Difference (I-J) Std. Error Sig. 95% Confidence Interval
Plant Lower Bound Upper Bound

Plant A ensi on3

Plant B 2.500 12.173 1.000 -31.13 36.13

Plant .083 12.173 1.000 -33.55 33.71

C
Plant 1.333 12.173 1.000 -32.30 34.96
D
Plant B dimensio
Plant A -2.500 12.173 1.000 -36.13 31.13
n3
Plant -2.417 12.173 1.000 -36.05 31.21
C
Plant -1.167 12.173 1.000 -34.80 32.46
D
Plant C dimensio
Plant A -.083 12.173 1.000 -33.71 33.55
n3
Plant B 2.417 12.173 1.000 -31.21 36.05
Plant 1.250 12.173 1.000 -32.38 34.88
D
Plant D dimensio
Plant A -1.333 12.173 1.000 -34.96 32.30
n3 Plant B 1.167 12.173 1.000 -32.46 34.80

Plant -1.250 12.173 1.000 -34.88 32.38

C

FIGURE 9.26.  Multiple-comparison tests following the ANOVA.

is a test called Friedman’s ANOVA that can be modified to work in this sce-
nario.

Having said all of that, we can look at the basics.

The Case of Age Affecting Ability

Recently, a university’s accrediting agency mandated that professors, in all disciplines,
should use technology more frequently in their classrooms. One of the young assistant
professors noted, however, that it was not as easy as that. What, the professor won-
dered, would happen if technology proved to have a negative effect in the classroom?
To investigate this issue, the young professor arranged to have two classes in
statistics taught in different ways. In the first class, students would use hand-held
calculators to work through their problems. In the second class, students would use
276 | STATISTICS TRANSLATED

personal computers to help in their work. The professor followed this plan through-
out the term and, after collecting final grades, decided to test whether the grades in
the two groups were equivalent. Noting there was one independent variable (class)
with two levels (the calculator section and the computer section) and one dependent
variable (course grade), the professor analyzed the data using an independent-sample
t-test. Upon running the analysis, the results showed a p value greater than the pre-
determined alpha value of .05. Given that, rejecting the null hypothesis was not fea-
sible—there was apparently no significant difference between the two groups.
As the young professor was presenting the results in a faculty meeting, one of
his colleagues noted that a lot of variables affect whether technology adoption is suc-
cessful. The young professor decided to investigate this and, after some additional
research, found that a student’s age is an important predictor of success with technol-
ogy. In fact, he read, students that are older than 30 appear to have more problems
when trying to use technology in a classroom. “That is true.” the young professor
thought. “I did notice some of the older students looking puzzled when we started
using the computers. Maybe I should look at this.”

S TEP 1 Identify the Problem

Here the university has a problem it is interested in addressing;
in this case, the faculty pointed out the problem. Given that, it
is certainly ethical and practically significant; numeric data can
be collected by which a hypothesis can be tested. The scope of
the problem is well within the authority of the university, and
they had better make sure they have the knowledge, time, and resources needed to
address it. They begin with a clear statement of the problem:

„ An accrediting agency has mandated that technology be used at all levels of

university instruction. Faculty members are concerned that the achievement
of older students may be affected because of this requirement.

S TEP 2 State a Hypothesis

Before we state a hypothesis, we need to take a closer look at
what we are doing. First, it is obvious we are going to be hypoth-
esizing using two independent variables, age and technology
type used (computer or calculator). If we were using a one-way
ANOVA to look at each of these variables alone, we would state
the following two null hypotheses:

„ There will be no significant difference in achievement between students

who use computers in a statistics class and those who use calculators in
a statistics class. There will be no significant difference in achievement
between students in a statistics class who are older than 30 and those who
are 30 or younger.
 Analysis of Variance and Multivariate Analysis of Variance  |  277

If we think about it, however, this is not really what we are interested in; there are
four groups that we want to look at. To make this clear, look at Table 9.10.

TABLE 9.10.  Cells for Age and Technology Type for Factorial ANOVA
People 30 or younger People older than 30
Computers Group 1 Group 2
Calculators Group 3 Group 4

You might be saying to yourself, “What is the big deal? We have four groups; we
still have the two independent variables and have already stated hypotheses about
them—what is the problem?” The problem is that we cannot just look at the effects
of each of the independent variables, called the main effects; we also must look at the
interaction between the two variables. In other words, it is important to know if these
two variables interact to cause something unexpected to happen. For example, the
literature suggests that younger people will perform better than older people when
computers are involved; this is called an interaction of technology type and age. It is
imperative that we look at the interaction of these variables to see if achievement is
different based on the way these independent variables affect one another. Given that,
in addition to our two main effect hypotheses that we stated earlier, we also must state
an interaction effect hypothesis:

„ There will be no significant interaction between technology use and age and
their effect on achievement in a statistics class.

Obviously, this is radically different from any hypotheses we have stated up to this
point, but it will become clear as we work through the example.

S TEP 3 Identify the Independent Variable

Here our independent variables are age and technology type.
Age has two levels: over 30 and 30 and below. Technology type
also has two levels: computers or calculators.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is achievement in the statistics class.
We’ll use the data in Table 9.11 and move forward with our
analysis.
278 | STATISTICS TRANSLATED

TABLE 9.11.  Score Data for Technology and Age Groups

Technology Age group Scores
68 69
67 70
30 or younger 66 71
73 73
75 74
Computer
79 75
80 75
Over 30 76 72
71 71
72 74
73 77
73 78
30 or younger 77 69
77 70
75 71
Calculator
70 69
74 73
Over 30 75 75
72 76
72 72

We have two groups of 20 students: those who use calculators and those who use
computers. Each of these groups includes 10 students age 30 and younger; the remain-
ing 10 students are older than 30. In order to input this into the SPSS spreadsheet,
shown in Figure 9.27, we would identify the students using computers as 1 in the Tech-
nology column; those with calculators would be identified as 2. Students 30 or younger
would be identified as 1; students older than 30 would be identified with a 2.
Our software would generate the descriptive statistics shown in Figure 9.28. As
you can see, the average scores for the computer group (72.55) and the calculator
group (73.40) are about the same; so are the average scores for the 30 or younger
group (72.30) and the students older than 30 (73.65). We can take this a step further
and break the scores down by age group within the technology group; our software
would give us the following results:
Again, just by looking at the data, our young professor might be right; there
doesn’t appear to be any great difference between the age of a student and whether
they are using computers or calculators. Our analysis will help us make sure we have
made the right decision.
 Analysis of Variance and Multivariate Analysis of Variance  |  279

FIGURE 9.27.  Technology type, age, and score entered into the Data View spreadsheet.

Descriptive Statistics
Dependent Variable:Scores

Technology Age Mean Std. Deviation N

d
Computer 30 or Younger 70.6000 3.09839 10
i Over 30 74.5000 3.17105 10
m

Total 72.5500 3.64872 20

e

Calculator 30 or Younger 74.0000 3.26599 10

Over 30 72.8000 2.25093 10

n

Total 73.4000 2.79850 20

i

Total 30 or Younger 72.3000 3.55557 20

o

n
Over 30 73.6500 2.81490 20

1
Total 72.9750 3.23829 40
FIGURE 9.28.  Descriptive statistics for the factorial ANOVA.
280 | STATISTICS TRANSLATED

S TEP 5 Choose the Right Statistical Test

Here we have two independent variables and one dependent
variable that is quantitative in nature. Given that, we are going
to use a factorial analysis of variance. Remember, also, we can
call this a two-way analysis of variance since we have two inde-
pendent variables or we can call it a “2 × 2” analysis of variance
because there are two levels in the first independent variable (age group) and two
levels in the second independent variable (type of technology).

S TEP 6 Use Data Analysis Software to Test the Hypothesis

When we look at the averages for all students using calculators,
they seem to be fairly equal (i.e., 74.0 and 72.8), regardless of
a given student’s age. There is a larger difference between stu-
dents using computers in that persons 30 and younger have an
average score of 70.6 while their older classmates have an aver-
age of 74.5. Our question, however, is the same as it always is, “Are these differences
due to chance, or are the averages from the two groups significantly different?” We
would use our computer to compute both the Levene test and the factorial ANOVA.
In order to do that, we would select Analyze, General Linear Model, and Univariate,
shown in Figure 9.29. This would be followed by identifying the Dependent Variable
and Fixed Factors (Figure 9.30).
Figure 9.31 shows that the p value from the Levene test is .462, much greater than
our alpha value of .05. This means we can proceed with testing our hypotheses with
the ANOVA table shown in Figure 9.32.

FIGURE 9.29.  Selecting the General Linear Model and Univariate to run a factorial ANOVA.
 Analysis of Variance and Multivariate Analysis of Variance  |  281

FIGURE 9.30.  Identifying the independent variables, dependent variable, and descriptive
statistics for the factorial ANOVA.

Levene's Test of Equality of Error Variances

Dependent Variable:Scores

F df1 df2 Sig.

.878 3 36 .462
FIGURE 9.31.  Levene’s test for the factorial ANOVA.

Tests of Between-Subjects Effects

Dependent Variable:Scores

Type III Sum of Partial Eta

Source Squares df Mean Square F Sig. Squared
a
Corrected Model 90.475 3 30.158 3.409 .028 .221
Intercept 213014.025 1 213014.025 24076.938 .000 .999
technology 7.225 1 7.225 .817 .372 .022
age 18.225 1 18.225 2.060 .160 .054
technology * age 65.025 1 65.025 7.350 .010 .170
Error 318.500 36 8.847
Total 213423.000 40
Corrected Total 408.975 39

a. R Squared = .221 (Adjusted R Squared = .156)

FIGURE 9.32.  Inferential statistics from a factorial ANOVA.

282 | STATISTICS TRANSLATED

The table in Figure 9.32 is very similar to the ANOVA table we used earlier, but
there are a few differences. First, there is a lot of technical information given in the
first two rows, so move down to the third and fourth rows, labeled “Technology” and
“Age.” These lines, called the Main Effects, show what the results of a one-way ANOVA
would be if it were calculated using only one independent variable at a time. In this
case, if we separated the groups by either Age or Technology used, the differences are
not significant; our p value (i.e., Sig.) is greater than our alpha value of .05. This, of
course, validates the observation we made earlier when we noticed that the means for
each of these groups were fairly close to one another.

Interpreting the Interaction p Value

After we have looked at the results of analyzing the independent variables separately,
we then need to look at the p value of the interaction between Technology and Age,
shown in the fifth row. In this case, we can see the p value is equal to .010, far less than
our alpha value of .05. This means we have a significant interaction effect and can
reject our null hypothesis.

„ There will be no significant interaction between technology and age and

their effect on achievement in a statistics class.

We have rejected the null hypothesis, but what does it mean? How do we interpret
these results so that they are meaningful to us? To do that, we will rely on that adage
“one picture is worth a thousand words”; many people find it easier to help interpret
the results of a factorial ANOVA by plotting the interactions using the average scores
from each subgroup. We would do this by selecting the Plots subcommand from the
earlier screen and identifying the x-axis and y-axis variables as shown in Figure 9.33.
SPSS would then create Figure 9.34, an interaction plot.
To interpret this chart, first look at the legend at the top right of the chart. It tells
us a solid line represents the scores for students 30 and younger and a dashed line rep-
resents the scores of students older than 30. When we look at the actual graph, we can
see that the bottom left of the graph represents those students in the computer group
and the bottom right side shows those students in the calculator group.
Given this, you can see that the dashed line on the left side of the graph (com-
puter group) starts at the point on that line which represents the average score of the
over-30 group (74.5). It then goes down to the point on the right side of the graph
(the calculator side) that is equal to the average score of that group (72.8). The solid
line for the 30 and younger group starts on the computer side at 70.6 and goes up to
the level of the calculator group (74.00). Based on this, it seems the young professor’s
colleague was right: age as an independent variable does interact with the type of
technology a student uses. We can look at this in even more detail.
When looking at the right side of the chart, the first thing we notice is that there
is not a big difference in achievement, by age, for those students using calculators. We
already knew this because, using the descriptive statistics, we saw students under 30
who used a calculator had an average score of 74.0 while students 30 and older using
calculators had an average of 72.8. The more striking observation we can make is
 Analysis of Variance and Multivariate Analysis of Variance  |  283

FIGURE 9.33.  Selection of the dependent variable and fixed factors in a factorial ANOVA.

that students 30 and younger have somewhat lower achievement (i.e., 70.6) than their
older classmates when using computers (i.e., 74.5). While we knew this already from
looking at the data, plotting them on this graph makes the differences quite clear.
At this point, we can safely state that a significant interaction does exist (remember,
however, we already knew that because of the low p value). We can also see that when
considered by the type of technology used, the average scores change when we move
between levels of age. Before we move forward, we need to talk about the types of
interactions that might exist.

Ordinal and Disordinal Interactions

As we saw in our graph, the two lines,
when plotted, crossed one another. We
call this a disordinal interaction, and
it exists when one level of an indepen-
dent variable has an effect dramatical-
ly opposite that of another level of the
same variable. For example, here we saw
that the age of a student affects achieve-
ment differently depending on whether
the student is using a computer or a cal-
culator.
In other cases, we might have an
ordinal interaction. In an ordinal inter-
FIGURE 9.34.  Disordinal interaction. action, the lines, when plotted, do not
284 | STATISTICS TRANSLATED

cross one another. For example, if our plotted data had looked like the following
graph, we would have had an ordinal interaction. If you look at Figure 9.35, the rea-
soning behind this is clear; the independent variable called Technology Type only
affects one level of the independent vari-
able called Age Group. As can be seen,
students whose ages are less than or
equal to 30 have a higher average than
their older classmates when calculators
are being used. The two age groups are
equal, however, when they are all using
computers.

Always Use Your p Value!

Remember, although it is easy to inter-
pret this graph, you cannot use it alone
to decide if your interaction is signifi-
cant; for that you must rely on the p
FIGURE 9.35.  Ordinal interaction. value. The graphs are only used to help
demonstrate the degree of the interac-
tion.

The Case of the Coach

A high school coach is interested in improving the endurance of his football play-
ers. He’s heard there are two things that greatly contribute to endurance: the type of
conditioning exercise players are involved in and their diet. In order to investigate
this issue, he sets up a training regimen where half of his students will work out by
running; others will go to an aerobics class. He also assigns half of each group of
student athletes to a carbohydrate-rich diet, while the others will eat as much protein
as possible. The coach is interested in determining if either of these approaches, or a
combination of both, will lead to higher levels of endurance.

S TEP 1 Identify the Problem

This case doesn’t specifically state a problem, but one is certain-
ly implied. The coach is interested in winning as many games
as possible, so he’s looking for a way to increase endurance.
I’m sure there is some way the coach can numerically measure
endurance. All the other criteria for a good problem are met:
 Analysis of Variance and Multivariate Analysis of Variance  |  285

„ This study will investigate the effect of combining different diets and exercise
regimens on the endurance of high school athletes.

S TEP 2 State a Hypothesis

The coach will be investigating the following two main-effect
hypotheses:

„ There will be no significant difference in the degree

of endurance of high school athletes following a diet
rich in protein and athletes eating a diet rich in
carbohydrates. There will be no significant difference
in the degree of endurance between high school
athletes who run and athletes who go to aerobics
classes.

He will also be investigating a hypothesis for the interaction effect:

There will be no significant interaction between exercise regimen and


diet for high school athletes.

S TEP 3 Identify the Independent Variable

There are two independent variables, diet and exercise regi-
men. Diet has two levels: carbohydrate rich and protein rich.
Exercise regimen also has two levels: running and aerobics.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is endurance and, for the purposes of
this example, it will be measured by a physical skill test the
coach has used in the past. Results of the test are quantitative
and range from 0 to 100.
At this point, it is obvious that we are going to use the fac-
torial ANOVA since we have two independent variables, with one dependent variable
used to collect quantitative data. Rather than create an entire dataset, we can use the
results shown in Figure 9.36 to test the coach’s hypothesis.
Using the descriptive statistics, we can see something interesting is happening.
The protein group has a much higher mean (i.e., 87.1) than the carbohydrate group
(i.e., 73.0), but the overall mean scores of the aerobic group (i.e., 79.7) and the running
group (i.e., 80.4) are nearly identical.
When we look at the interaction between the variables, it appears to make no dif-
ference if the protein group runs (i.e., 87.4) or takes aerobics (i.e., 86.8). It does seem
that athletes eating carbohydrates and taking aerobics have a slightly lower average
(i.e., 72.6) than others in the aerobics class who are eating protein (i.e., 73.4). In this
case the values are different but are they significantly different?
286 | STATISTICS TRANSLATED

Descriptive Statistics
Dependent Variable:Endurance

Food Exercise Mean Std. Deviation N

Protein Aerobic 86.8000 .83666 5

Running 87.4000 1.81659 5

Total 87.1000 1.37032 10

Carbohydrate Aerobic 72.6000 1.14018 5
Running 73.4000 2.40832 5
Total 73.0000 1.82574 10
Total Aerobic 79.7000 7.54321 10

Running 80.4000 7.64780 10

Total 80.0500 7.40181 20

FIGURE 9.36.  Factorial ANOVA descriptive statistics.

S TEP 5 Choose the Right Statistical Test

Since we have two independent variables and our dependent
variable is quantitative, we are going to use the factorial ANOVA.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

First, we need to check the Levene statistic in Figure 9.37 to
make sure we are not violating the homogeneity of variance
assumption.
Our p value is exactly equal to our alpha of .05. Remember,
we only reject the null hypothesis when p is less than .05. There
are no problems here, so we can use the following ANOVA table in Figure 9.38 to test
our hypothesis.
Immediately we can tell, from our computed p value (i.e., .895), that a significant
interaction between diet and type of exercise does not exist. Even though the interac-
tion is not significant, we can still look at the interaction plot in Figure 9.39.
Here we have an ordinal interaction; the lines do not cross one another. In fact,
the lines are nearly parallel. Since the interaction is not significant, it is necessary to
look at the p values of the main effects.

Levene's Test of Equality of Error Variances

Dependent Variable:Endurance

F df1 df2 Sig.

3.234 3 16 .050
FIGURE 9.37.  Factorial ANOVA Test of Equality.
 Analysis of Variance and Multivariate Analysis of Variance  |  287

Tests of Between-Subjects Effects

Dependent Variable:Endurance

Source Type III Sum of Partial Eta

Squares df Mean Square F Sig. Squared
a
Corrected Model 996.550 3 332.183 119.706 .000 .957
Intercept 128160.050 1 128160.050 46183.802 .000 1.000
Food 994.050 1 994.050 358.216 .000 .957
Exercise 2.450 1 2.450 .883 .361 .052
Food * Exercise .050 1 .050 .018 .895 .001
Error 44.400 16 2.775
Total 129201.000 20
Corrected Total 1040.950 19
a. R Squared = .957 (Adjusted R Squared = .949)

FIGURE 9.38.  Inferential statistics for the factorial ANOVA.

In this case, we can see there is not a significant main effect for the type of
exercise that a person performs because the p value is equal to .361. Of course, we
suspected that before examining the p
value because our mean scores were so
close to one another for the two groups.
A significant main effect does exist for
the type of diet an athlete follows. In
this case, the p value of .000 shows that
athletes eating a diet high in protein can
develop higher levels of endurance than
athletes eating a diet high in carbohy-
drates. We can verify this by looking at
the interaction plot and noting that the
scores on the left side (protein) are dra-
matically higher than those on the right
side (carbohydrates). Given all of this,
the only null hypothesis we can reject is
our first one; it seems that a diet rich in
FIGURE 9.39.  Ordinal interaction of food and protein is better for developing superior
exercise. levels of endurance.

The Multivariate ANOVA (MANOVA)

As we discussed earlier, the primary reason we use a multivariate ANOVA is to test
hypotheses with two or more dependent variables, and one or more independent vari-
ables. We are only going to look at scenarios with one independent variable, however,
due to several reasons:

1. Like the factorial ANOVA, when you have a scenario with more than one
dependent variable and two or more independent variables, your analysis can
get very complicated.
288 | STATISTICS TRANSLATED

2. As was the case with the factorial ANOVA, the math underlying the M
­ ANOVA
is far beyond what we want to cover in this book. Because of that, we will rely
on our software to do the calculations for us.
3. Just as was the case with the ANOVA, if you violate any of the assumptions
that underlie using a MANOVA, you run the risk of making a Type I error.
Each additional independent or dependent variable worsens this problem.

Assumptions of the MANOVA

Like the ANOVA, there are assumptions we must keep in mind when running the
MANOVA. We talked about several of these earlier in the chapter—the need for ran-
dom samples, independence of scores, a normal distribution of the data, and homo-
geneity of variance. We still must ensure those assumptions are met but, because we
have multiple dependent variables, we also must concern ourselves with homogeneity of
covariance matrices and multivariate normality.
First, the homogeneity of the covariance matrices is an expansion of the homo-
geneity of variance computed and used by the ANOVA. As we saw then, we had to
ensure that the variance between each of the groups was approximately equal. Failing
to do that could possibly lead to a Type I error or call for the use of a non-parametric
alternative. We have the same type of concern with the MANOVA but here we must
ensure that the correlation between any of the groups being evaluated is equal, this is
called homogeneity of covariance matrices.
Following that, the assurance of multivariate normality is an extension of the need
for a normal distribution when we used the ANOVA. In this case, however, instead
of normality between the data representing the levels of one independent variable,
with multivariate normality we assume that the dependent variables, collectively, have
normality within groups.
To completely understand these final two assumptions requires math skills far
beyond what we need to know in order to effectively use the MANOVA. Because of
that, we’ll discuss how violations of any of these assumptions are dealt with at a more
general level. Knowing that, we can move forward to discussing the MANOVA in
general.

Using the MANOVA

We can begin looking at the mechanics underlying the MANOVA using an example
where a marketing manager is trying to determine if there is a specific region, urban
or rural, where the company’s new brand of coffee might be best introduced. If so,
possible marketing strategies could be developed to increase sales. In order to do this,
the sales manager proposes trying to determine if there is a relationship between
region, the consumers’ intention to buy coffee, and their taste preferences in coffee.
In this case, we would have one independent variable, region, with two levels, urban
and rural. There would be two dependent variables: intention to buy coffee, measured
on a scale from 1 (Not Likely) to 5 (Very Likely), and taste preference, mild to strong,
measured on a scale from 1 (Very Mild) to 5 (Very Strong).
 Analysis of Variance and Multivariate Analysis of Variance  |  289

Looking at these, you might ask, “Why not just use two ANOVAs, one for each of
the dependent variables?”

„ There will be a significant difference in intention to buy between regions.

„ There will be a significant difference in taste preference between regions.

We could certainly do that, but it would create two problems. First, as we said ear-
lier, the use of multiple ANOVAs would increase our Type I error rate. Additionally,
we would not be able to accurately determine if there is a relationship between each of
the regions and a combination of the consumer’s taste preference and their intention
to buy coffee. Knowing if the combination does show a relationship could strongly
affect decisions on where and how to market the new coffee.
Keeping in mind that the purpose of the MANOVA is to test both dependent vari-
ables simultaneously, it creates a third variable that contains a linear combination of
the dependent variables. This results in the ability to compute an overall mean and, by
using it, we can determine if there are significant group differences, by region, on the
set of dependent variables. This will help increase the statistical power of our analysis.
In short, we can determine if there is a significant interaction between customers’
intention to buy and their taste preference by region. Knowing that, we can examine
all variables simultaneously with the following interaction hypothesis:

„ There will be a significant difference in taste preference and intention to buy

coffee based on the geographic region of the consumer.

There are also the two main effect hypotheses we discussed earlier:

„ There will be a significant difference in intention to buy between regions.

„ There will be a significant difference in taste preference between regions.

An example of the data we could use to test this hypothesis is shown in Table 9.12.

TABLE 9.12.  Data for Coffee Drinker Region, Intention, and Taste
Region Intention Taste
Urban 2 3
Urban 3 4
Urban 4 2
Urban 4 3
Urban 3 5
Rural 1 5
Rural 1 5
Rural 2 5
Rural 3 4
Rural 2 4
290 | STATISTICS TRANSLATED

The left-most column shows our two regions, Urban and Rural. The middle col-
umn shows a customer’s intention to buy coffee, ranked from 1 (Not Likely) to 5 (Very
Likely) and the right-most column shows Taste preference, ranked from 1 (Very Mild)
to 5 (Very Strong). Note that I’m only showing an example of 10 potential customers
here, the actual dataset that we would use for our SPSS analysis would be much larger.
We begin our analysis by entering the larger dataset into our SPSS spreadsheet
(Figure 9.40). We follow that by selecting Analyze, General Linear Model, and Multi-
variate, shown in Figure 9.41. This would be followed by identifying the Dependent
Variable and Fixed Factors, as well as choosing to compute the Descriptive Statistics
and Estimate of Effect Size (Figure 9.42).

FIGURE 9.40.  Data for multivariate analysis of variance (MANOVA). Reprint Courtesy of Interna-
tional Business Machines Corporation, © International Business Machines Corporation.

FIGURE 9.41.  Selecting the General Linear Model and Multivariate to run a MANOVA. Reprint
Courtesy of International Business Machines Corporation, © International Business Machines
Corporation.
 Analysis of Variance and Multivariate Analysis of Variance  |  291

FIGURE 9.42.  Selecting the Independent and Dependent Variables and Options Needed for the
MANOVA. Reprint Courtesy of International Business Machines Corporation, © International
Business Machines Corporation.

Figure 9.43 shows the descriptive statistics, and we can see that the mean scores
for Intention to Buy are very close for Urban and Rural customers (i.e., 3.3 and 3.1);
the same is true for Taste Preference (i.e., 2.8 and 2.7). Box’s Test of Equality of Covari-
ance Matrices (i.e., Box’s M), shown in Figure 9.44, tests the assumption of homogene-
ity of the covariance matrices. In this case, our Sig. (i.e., p) is not less than an alpha
of .05, so we are supporting the hypothesis that the variance in the groups is equal.
Because there is equal variance between groups, we can use the results shown in
Figure 9.45 to test our multivariate hypothesis. Therein we want to focus on the row
labeled “Region” and we can see that there are Sig. (i.e., p) values for four tests, Pillai’s
Trace, Wilk’s Lambda, Hoteling’s Trace, and Roy’s Largest Root. As we have seen in
other places throughout this book, there are specific times, and researchers’ opinions,
as to which of these tests should be used for a given data set. It is agreed, however,
that Pillai’s Trace is the most robust of these tests, so we will use it in our journey to
become good producers and consumers of research.

FIGURE 9.43.  Descriptive Statistics for the MANOVA.

292 | STATISTICS TRANSLATED

FIGURE 9.44.  Box’s M for the MANOVA.

FIGURE 9.45.  Multivariate Test results for the MANOVA.

In the section labeled Region, Pillai’s trace is not significant (i.e., F (2, 37) =
151.182, p = .895). This means we fail to reject the interaction null hypothesis; there
is no significant difference between urban consumers and rural consumers when con-
sidering intent to buy and taste preference simultaneously. At this point, we would
not go further with our analysis by testing the main effect hypotheses focusing on
each dependent variable individually (i.e., a one-way ANOVA). As we’ll see in the next
example, when we do see significant multivariate differences, we’ll take our analysis a
step further in trying to determine the exact cause of the difference.

The Case of Balancing Time

Many universities are introducing first-year seminars wherein new students are intro-
duced to all aspects of campus life—the library, financial aid, career services, academic
advising, and health services. Most universities require this course in the first term
of a student’s academic career. While many students really like the experience, there
are complaints about the number of class sessions per week and the length of class
 Analysis of Variance and Multivariate Analysis of Variance  |  293

sessions. In order to investigate ways of addressing these student concerns, one uni-
versity’s administrators want to determine if factors such as course satisfaction and
perceived value are directly related to whether a student’s class meets once per week,
or if they are required to be in class two days per week.

S TEP 1 Identify the Problem

Many universities offer required first-year seminars where stu-
dents are introduced to many aspects of the college experience.
Administrators are concerned that student complaints about
time spent in class could negatively affect student satisfaction
and the perceived value of the class.

S TEP 2 State a Hypothesis

There will be a significant difference between levels of student
satisfaction and perceived value, considered conjointly, between
students taking a weekly 90-minute class, and students taking
two 45-minute classes in a week.

S TEP 3 Identify the Independent Variable

Here we can name the independent variable “days per week”
and there are two levels, one day per week and two days per
week.

S TEP 4 Identify and Describe the Dependent Variable

There are two dependent variables, student satisfaction and
perceived value, both will be measured on a scale from 1 (High-
ly Dissatisfied) to 5 (Highly Satisfied). Descriptive statistics for
each are shown in Figure 9.46. The p value in Box’s M, shown
in Figure 9.47, is greater than .05 indicating that the variance in
the groups is equal.

FIGURE 9.46.  Descriptive Statistics the MANOVA.

294 | STATISTICS TRANSLATED

FIGURE 9.47.  Box’s M for the MANOVA.

S TEP 5 Choose the Right Statistical Test

Because there is one independent variable and two dependent
variables measuring quantitative data, we will use the MANO-
VA to test our hypothesis.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

We can begin by looking at Days under the Effects column in
Figure 9.48. Therein, the Pillai’s Trace Sig. (i.e., p value) is .029
indicating that we will reject our null hypothesis. The question
then becomes, “Which of the dependent variables is significant-
ly different between the two levels of the independent variable?”
We can determine this by looking at the output in Figure 9.49.

FIGURE 9.48.  Multivariate test results for the MANOVA.

Before we do, however, we need to think back to the problem with power had we
chosen to use two separate ANOVAs rather than one MANOVA. The same problem
still exists with the MANOVA but to a much lesser extent. In order to control for
potential issues, we will adjust our alpha value using the Bonferroni correction. Our
new alpha value is computed by dividing our traditional alpha value of .05 by the num-
ber of hypotheses we have stated. In this case our corrected alpha is .025 (i.e., .05/2).
We will then use this new value in our comparisons to our completed p value when
testing our univariate hypotheses.
In Figure 9.49, our Sig. (i.e., p) values for Satisfaction and Value are .105 and
 Analysis of Variance and Multivariate Analysis of Variance  |  295

FIGURE 9.49.  Between Subject Effects for the MANOVA.

.019 respectively. This means we would fail to reject the null hypothesis for Satisfac-
tion (i.e., p of .105 > alpha of .025) but reject the null hypothesis for Value (i.e., p of
.019 < alpha of .025). This means, if we look back at our descriptive statistics, we can
determine that the mean Satisfaction level of students taking the class once per week
(i.e., 4.2308) is significantly greater than the mean value of students taking the class
twice per week (i.e., 3.3077). In seeing this, faculty and administrators might want
to reconsider the twice weekly option or find new or better ways to increase student
satisfaction.

Summary
Although we have covered quite a lot of material in this chapter, the analysis of vari-
ance is straightforward. We use the simple (one-way) ANOVA when we have one
independent variable that has three or more levels and one dependent variable that is
represented by quantitative data. When we have more than one independent variable
and only one dependent variable, we will use a factorial ANOVA and a multivariate
ANOVA when we have two or more dependent variables. As I mentioned, there are
other types of ANOVAs but, for now, we have talked about as much as we need to; it
is time to move forward!
296 | STATISTICS TRANSLATED

Do You Understand These Key Words and Phrases?

analysis of variance assumptions
between groups variance Bonferroni correction
cell degrees of freedom disordinal interaction
F value factorial analysis of variance
homogeneity of covariance matrices interaction effect
main effect mean square multiple-comparison test
multivariate ANOVA (MANOVA) multivariate normality
one-way ANOVA ordinal interaction Pillai’s Trace
post-hoc tests within-group variance

Do You Understand These Formulas?

Between sum of squares:

( ∑ x1)2 + (∑ x 2) + ( ∑ x 3)2 − ( ∑ x )2
2
SS Between =
n1 n2 n3 N

Total sum of squares:

SSTotal = ∑ x
2
−
( ∑ x)
2

Within sum of squares:

SSW = SS1 + SS2 + SS3

Quiz Time!

To test mastery of this topic, work through these examples and check your work at the end of
the book.

The Case of Degree Completion

A university is interested in knowing if age is a determining factor in the number of years it
takes to complete a doctoral degree. To aid in their investigation, they’ve placed each of their
graduates into one of four age categories: 20–29, 30–39, 40–49, and 50 and above.

1. What are the null and research hypotheses the university is investigating?

2. What is the independent variable?

 Analysis of Variance and Multivariate Analysis of Variance  |  297

3. What are the levels of the independent variable?

4. What is the dependent variable?
5. Based on the information in Figures 9.50, 9.51, and 9.52, what could the university
conclude?

Descriptives
Time to Finish

95% Confidence Interval for Mean

N Mean Std. Deviation Std. Error Lower Bound Upper Bound Minimum Maximum

20 thru 29 5 4.4000 1.08397 .48477 3.0541 5.7459 2.50 5.00

30 thru 39 5 3.7000 .83666 .37417 2.6611 4.7389 3.00 5.00
40 thru 49 5 3.6000 .89443 .40000 2.4894 4.7106 3.00 5.00
50 and older 5 4.8000 .83666 .37417 3.7611 5.8389 4.00 6.00
Total 20 4.1250 .98509 .22027 3.6640 4.5860 2.50 6.00

FIGURE 9.50.  ANOVA Descriptive Statistics.

Test of Homogeneity of Variances

Time to Finish
Levene Statistic df1 df2 Sig.
.073 3 16 .973
FIGURE 9.51.  ANOVA tests of homogeneity of variance.

ANOVA
Time to Finish

Sum of Squares df Mean Square F Sig.

Between Groups 4.937 3 1.646 1.951 .162

Within Groups 13.500 16 .844
Total 18.438 19

FIGURE 9.52.  Inferential statistics for the ANOVA.

The Case of Seasonal Depression

Psychologists are interested in determining whether there is a difference in the number of
anxiety attacks during the different seasons of the year. In order to investigate, they tracked
44 people for 12 months and asked them to record the number of anxiety attacks they had
during each season.

1. What are the null and research hypotheses the psychologists are investigating?

2. What is the independent variable?

3. What are the levels of the independent variable?
298 | STATISTICS TRANSLATED

4. What is the dependent variable?

5. Using Figures 9.53, 9.54, 9.55 and 9.56, what could the psychologists conclude?
Descriptives
Anxiety Attacks

95% Confidence Interval for Mean

N Mean Std. Deviation Std. Error Lower Bound Upper Bound Minimum Maximum

Winter 6 6.1667 1.47196 .60093 4.6219 7.7114 4.00 8.00

Spring 6 3.8333 .75277 .30732 3.0433 4.6233 3.00 5.00
Summer 6 3.8333 .75277 .30732 3.0433 4.6233 3.00 5.00
Fall 6 3.8333 .75277 .30732 3.0433 4.6233 3.00 5.00
Total 24 4.4167 1.38051 .28179 3.8337 4.9996 3.00 8.00

FIGURE 9.53.  ANOVA Descriptive statistics.

Test of Homogeneity of Variances

Anxiety Attacks

Levene Statistic df1 df2 Sig.

1.997 3 20 .147
FIGURE 9.54.  ANOVA tests of homogeneity of variance.

ANOVA
Anxiety Attacks

Sum of Squares df Mean Square F Sig.

Between Groups 24.500 3 8.167 8.448 .001
Within Groups 19.333 20 .967
Total 43.833 23

FIGURE 9.55.  Inferential statistics for the ANOVA.

Multiple Comparisons
Anxiety Attacks
Bonferroni
(I) Season (J) Season Mean 95% Confidence Interval
Difference (I-J) Std. Error Sig. Lower Bound Upper Bound
*
Winter dime
Spring 2.33333 .56765 .003 .6718 3.9949
*
nsion
Summer 2.33333 .56765 .003 .6718 3.9949
*
d

3
Fall 2.33333 .56765 .003 .6718 3.9949
i

*
Spring dime
Winter -2.33333 .56765 .003 -3.9949 -.6718
m

e
nsion
Summer .00000 .56765 1.000 -1.6616 1.6616

n
3
Fall .00000 .56765 1.000 -1.6616 1.6616
*
s Summer dime
Winter -2.33333 .56765 .003 -3.9949 -.6718
i

nsion
Spring .00000 .56765 1.000 -1.6616 1.6616
o

3
Fall .00000 .56765 1.000 -1.6616 1.6616
*
Fall Winter -2.33333 .56765 .003 -3.9949 -.6718
n

dime

nsion
Spring .00000 .56765 1.000 -1.6616 1.6616
3
Summer .00000 .56765 1.000 -1.6616 1.6616

*. The mean difference is significant at the 0.05 level.

FIGURE 9.56.  Multiple Comparisons for the ANOVA.

 Analysis of Variance and Multivariate Analysis of Variance  |  299

The Case of Driving Away

Teachers, worried over the rising number of tardy students, noticed that students who have
their own car are late more often than their peers who only have access to their parents’ car.
Both groups, they believe, are tardy more often than students who have no access to an auto-
mobile for their personal use. In order to investigate their theory, the teachers identified 10
students in each of these groups and tracked the number of times they were late for school
during the month.

1. What are the null and research hypotheses the teachers are investigating?

2. What is the independent variable?

3. What are the levels of the independent variable?
4. What is the dependent variable?
5. By looking at Figures 9.57, 9.58, 9.59 and 9.60, what could the teachers conclude?

Descriptives
Times Late

95% Confidence Interval for Mean

N Mean Std. Deviation Std. Error Lower Bound Upper Bound Minimum Maximum

Owns Car 10 3.8000 1.22927 .38873 2.9206 4.6794 2.00 6.00

Has Access to Car 10 5.3000 .94868 .30000 4.6214 5.9786 4.00 7.00
No Access to Car 10 6.8000 1.13529 .35901 5.9879 7.6121 5.00 9.00
Total 30 5.3000 1.64317 .30000 4.6864 5.9136 2.00 9.00

FIGURE 9.57.  Descriptive statistics for the ANOVA.

Test of Homogeneity of Variances

Times Late

Levene Statistic df1 df2 Sig.

.385 2 27 .684
FIGURE 9.58.  Homogeneity of Variance for the ANOVA.

ANOVA
Times Late

Sum of Squares df Mean Square F Sig.

Between Groups 45.000 2 22.500 18.243 .000

Within Groups 33.300 27 1.233
Total 78.300 29

FIGURE 9.59.  Inferential statistics for the ANOVA.

300 | STATISTICS TRANSLATED

Multiple Comparisons
Times Late
Bonferroni

(I) Access (J) Access Mean 95% Confidence Interval

Difference (I-J) Std. Error Sig. Lower Bound Upper Bound
*
Owns Car Has Access to Car -1.50000 .49666 .016 -2.7677 -.2323
*
No Access to Car -3.00000 .49666 .000 -4.2677 -1.7323
*
Has Access to Car Owns Car 1.50000 .49666 .016 .2323 2.7677
*
No Access to Car -1.50000 .49666 .016 -2.7677 -.2323
*
No Access to Car Owns Car 3.00000 .49666 .000 1.7323 4.2677
*
Has Access to Car 1.50000 .49666 .016 .2323 2.7677
*. The mean difference is significant at the 0.05 level.

FIGURE 9.60.  Multiple comparisons for the ANOVA.

The Case of Climbing

Mount Rainier is the largest glaciated mountain in the United States, rising some 14,400 feet
above sea level. While thousands attempt to summit this massive mountain every year, not all
of them make it and turn back short of their goal. A group of climbers was discussing this one
day and argued that success depended on the route the climbers took, as well as on whether
they were with a professional guide. To try to investigate this, they asked the next 64 people
coming down from the mountain three things: the route they took, the altitude they reached
(remember, 14,400 is the summit), and whether they used a guide service.

1. What are the three main-effect null hypotheses the climbers are investigating?
2. What are the three research hypotheses the climbers are investigating?
3. What are the independent variables?
4. What are the levels of the independent variables?
5. What is the dependent variable?
6. Using the information in Figures 9.61 and 9.62, what conclusion would the climbers
come to?
 Analysis of Variance and Multivariate Analysis of Variance  |  301

FIGURE 9.61.  Descriptive statistics for the ANOVA.

FIGURE 9.62.  Between-Subject Effects for the ANOVA.

The Case of Employee Productivity

Throughout the United States, the number of companies hiring employees to work from home
is growing. Many administrators who support these arrangements believe this will lead to
302 | STATISTICS TRANSLATED

higher productivity than for that of employees working in the traditional office environment.
In order to investigate this, administrators from one company collected data from 60 employ-
ees working as telephone technical support personnel, 30 working from home, and 30 from a
traditional workplace. The data collected measured employee satisfaction, on a scale from 1 to
5, and the average number of customers an employee was able to work with over time.

1. What is the interaction hypothesis the managers are investigating?

2. What are the two main-effect research hypotheses the managers are investigating?
3. What is the independent variable?
4. What are the levels of the independent variable?
5. What are the dependent variables?
6. Using the information in Figures 9.63 through 9.66, what conclusion would the
administrators come to?

FIGURE 9.63.  Descriptive statistics for the MANOVA.

FIGURE 9.64.  Box’s M for the MANOVA.

 Analysis of Variance and Multivariate Analysis of Variance  |  303

FIGURE 9.65.  Multivariate statistics for the MANOVA.

FIGURE 9.66.  Tests of Between-Subjects Effects for the MANOVA.

CHAPTER 10

The Chi-Square Tests

Introduction
Up to this point we have dealt with quantitative data and, because of that, we have
been using parametric statistical tools. As I said in the very beginning, this is normal
for introductory statistics students. Other than the topic of this chapter, the chi-square
test, it is unusual for beginning students to use other nonparametric statistical tests.
In order to begin understanding the purpose of the chi-square, as well as how it
works, think back to our chapter on the analysis of variance. In it, we saw the one-
way analysis of variance, where we investigated an independent variable with three
or more levels, and the factorial analysis of variance (e.g., a two-way ANOVA) where
we compared more than one independent variable, regardless of the number of lev-
els within each of the variables. The chi-square tests work using basically the same
principles but instead of each level representing a quantitative value (e.g., test scores),
the levels of the chi-square test contain counts of values (i.e., nominal data). Let’s look
at an example of both a “one-way” and “factorial” chi-square to help us understand
what we are getting into.

The One-Way Chi-Square Test

The one-way chi-square test, or the chi-square goodness of fit as it is more commonly
called, helps us determine if a distribution of data values we observe matches how we
expect the data to be distributed. The following example will explain what I mean.
Suppose we are in a situation where we have posed this question to a random
group of 300 people walking down the street: “What do you think of allowing 16-year-
old children to vote?” Suppose, when you looked at your data, you found that 95 peo-
304
 The Chi-Square Tests  |  305

ple supported the idea, 105 didn’t like the idea, and 100 didn’t care; these are called
the observed values in each of the cells. This is shown in Table 10.1.

TABLE 10.1.  Observed Values of Vote Data

It is all right with me. I do not like the idea. I do not care.
Observed 95 105 100

Once we have collected our observed values, the question then becomes, “We
just defined goodness of fit as the comparison of a set of observed values to a set of
expected values; how do we determine the expected values?” That’s easy; since we are
testing goodness of fit, we are interested in comparing the observed values to values
we expect. For example, if we are interested in comparing our results to national aver-
ages showing that 40% of people like the idea and the rest are evenly split (i.e., 30%
each) between not liking the idea or not caring, this results in Table 10.2.

TABLE 10.2.  Observed and Expected Values of Vote Data

It is all right with me. I do not like the idea. I do not care.
Observed 95 105 100
Expected 120 (i.e., 40%) 90 (i.e., 30%) 90 (i.e., 30%)

In order to test your hypothesis, you would compare the proportion of data values
in the table to determine if the differences between what you observed and what you
expected are significant.
Before we move forward, it is important to know that the chi-square test is not
designed to be used with a small number of expected values. Because of that, it is
important to ensure you expect at least five occurrences of each of the cell values; try-
ing to do any calculations with less than that will adversely affect your results.

The Factorial Chi-Square Test (the Chi-Square Test of Independence)

The factorial chi-square, or as it is more often called, the chi-square test of indepen-
dence, is very akin to the factorial ANOVA in that you are collecting data where you
have more than one independent variable. For example, imagine you are the ultimate
authority charged with deciding whether sex education should be taught in your dis-
trict’s elementary schools. Before you make your decision, you decide that you need
to take into consideration the will of the public. In order to do that, you randomly
select 60 male and 60 female parents within the community (i.e., gender is the first
independent variable). You ask them to tell you their opinion, yes or no, whether they
agree with the proposed new curriculum (i.e., their opinion about the subject is the
second independent variable).
We have just seen something that might look counterintuitive. In this case, it
seems natural that “opinion” would be a dependent variable in that we would be inter-
ested in measuring how a person feels about a subject. Remember, however, in this
example we are not interested in how any one person feels about the subject; what we
306 | STATISTICS TRANSLATED

really want to know is how many people with a given opinion fall into each category.
Given that, “opinion” is the independent variable, and the count of each opinion is the
dependent variable where we are collecting nominal data.
In testing your hypothesis, your feeling is that there will be an equal distribution
of “yes” and “no” answers (i.e., the expected values) between the genders. For the sake
of discussion, I have gone ahead and entered a set of equal expected values, along with
a set of observed values, in Table 10.3.

TABLE 10.3.  Observed and Expected Values of Vote Data by Gender

Yes No
Observed Expected Observed Expected
Females 35 30 25 30
Males 25 30 35 30

Now, in order to test our hypothesis, we will need to compare our observed values
to our expected values. Before we can do that, we first need to have a better under-
standing of the concepts underlying the chi-square.

Computing the Chi-Square Statistic

In the tests we have examined up to this point, we have compared a statistic generated
from data we collected to a value from a table appropriate to a given statistical test
(e.g., the z table, the t table, or F table). By doing so, we have been able to determine if
a significant difference existed between the statistic we computed and a specific point
on a given data distribution.
At first glance, the chi-square procedure works in much the same manner except,
when using it, we are not interested in comparing a single statistic computed from a
specific data value to a table value. Instead we are trying to determine if the distribu-
tion of the variable in question is similar to the distribution we are expecting. We do
this by comparing a computed value of chi-square, based on the entire observed dis-
tribution, to a critical value of chi-square based on the expected distribution. When
we do, we are able to tell if a distribution of data we have observed is similar to or
significantly different from the distribution we are expecting.
Because it is very easy to understand, let’s go back and use our example of the
goodness-of-fit test where we asked for opinions about persons as young as 16 being
allowed to vote. Table 10.4 is the same one we used as part of that discussion.

TABLE 10.4.  Observed and Expected Values of Vote Data

It is all right with me. I do not like the idea. I do not care.
Observed  95 105 100
Expected 120  90  90

Our null hypothesis would read:

„ There will not be a significant difference in the number of responses expected

in each cell and the number of responses observed in each cell.
 The Chi-Square Tests  |  307

Obviously the research hypothesis would read:

„ There will be a significant difference in the number of responses expected in

each cell and the number of responses observed in each cell.

In short, I have hypothesized that the observed distribution of responses will not
be different from the expected distribution of responses. Testing this hypothesis will
involve comparing a chi-square value computed from the observed responses to a
critical chi-square value based on a given degrees of freedom value. As we are used to
doing, we will reject our null hypothesis if the computed value of chi-square is greater
than the critical value of chi-square. If we are saying nothing more than the distribu-
tion we observed is significantly different from the distribution we expected. The
formula for computing the chi-square statistic is shown below:
S (O − E )
2
c2 =
E
First, c, the Greek letter “chi”, indicates that we will be computing its squared
value. As you can see in Table 10.5, for each row we are going to:

1. Subtract the expected value (i.e., E) in each row from the observed value
(i.e., O) in each row. This is shown in the second column.
2. Square the value from Step 1. This is shown in the third column.
3. Divide the value from Step 2 by the expected value (i.e., E) for that row; this
gives us the chi-square value for that row. This is shown in the fourth column.
4. Add all of the chi-square values from each row (Step 3) to get the comput-
ed chi-square value. This is shown at the bottom of the fourth column of
Table 10.5.

TABLE 10.5.  Computing the Chi-Square Statistic

Observed – Observed – Chi-square =
expected expected squared Column 2 ÷ number expected
It is all right with me. 95 – 120 = –25 625 625/120 = 5.21
I do not like the idea. 105 – 90 = 15 225 225/90 = 2.5
I do not care. 100 – 90 = 10 100 100/90 = 1.11

Computed chi-square = 8.82

Now all we must do is compare our computed value to the table (i.e., critical)
value. To do that, we need to know what the distribution looks like.

The Chi-Square Distribution

As has been the case throughout this book, we will use a formula to create the chi-
square distribution. Like the F distribution, we will base this distribution on squared
values (i.e., chi-square), so the distribution starts at zero and all values in the distribu-
308 | STATISTICS TRANSLATED

tion are positive. This distribution is also based on the number of degrees of freedom
from the data we have collected, but how we compute the degrees of freedom is going
to depend on the particular chi-square test we are using. The fewer the degrees of
freedom, the more peaked the distribution; a larger number of degrees of freedom
will result in a flatter distribution that is more skewed to the right. This idea is shown
in Figure 10.1.
As I said earlier, we will compute
degrees of freedom differently for each
of the chi-square tests but let’s focus on
the goodness-of-fit test.

df = (number of entries) – 1

The number of entries in the for-

mula represents the number of levels of
the independent variable; in this case
we have three (e.g., “It is all right with
me”) so we have 2 degrees of freedom
(i.e., 3 – 1 = 2). We will use this value,
with our chi-square table, to determine
the critical chi-square value. Just so you
are aware: many chi-square tables show
critical values for a range of different
FIGURE 10.1.  Shape of chi-square distribution
for different degrees of freedom. alpha values. Since we have primarily
used the traditional alpha value of .05
throughout the book, I have only shown
that part of Table 10.6; there is a com-
plete table in Appendix F.

TABLE 10.6.  Critical Values of Chi-Square

Degrees of freedom Alpha = .05
 1  3.84
 2  5.99
 3  7.82
 4  9.49
 5 11.07
 6 12.59
 7 14.07
 8 15.51
 9 16.92
10 18.31

Here I have only shown critical values of chi-square up to 10 degrees of freedom;

I did that for two reasons. First, since computing the degrees of freedom is a function
of the number of cells and rows, you will rarely wind up with more than 10 degrees of
 The Chi-Square Tests  |  309

freedom. Second, as you can see, the critical value of chi-square with only 10 degrees
of freedom is 18.31; that’s a really large value. In order to compute a chi-square value
larger than that, you would need observed and expected values that were totally out
of proportion. Needless to say, that does not happen very often.
Knowing that, let’s get back to testing our null hypothesis:

„ There will not be a significant difference in the number of responses expected

in each cell and the number of responses observed in each cell.

Looking at the table, since we have 2 degrees of freedom, our critical chi-square
value is 5.99. Since our computed value of 8.82 is greater than our critical value, we
reject our null hypothesis; there appears to be a significant difference in the observed
number of responses for each answer and the expected number of responses for each
value. We can see this in Figure 10.2.
In order to use SPSS, let’s first enter
our data. As you can see, we have one
variable, Vote. For the chi-square test to
work, we would need to enter an appro-
priate value for each of our 300 voters.
This means we would enter, for exam-
ple, 95 values of 1 for people who sup-
port the idea, 105 values of 2 for those
who do not like the idea, and 100 values
of 3 for people who do not care (Figure
10.3).
In order to run the actual test, we
would select Analyze, Nonparametric
Tests, Legacy Dialogs, and Chi-Square;
FIGURE 10.2.  Testing a hypothesis using the this is seen in Figure 10.4.
computed and critical values of chi-square. We would then identify Vote as
our Test Variable and include Expected
Values of 120, 90, and 90. We have also
asked for descriptive statistics to be computed as shown in Figure 10.5.
SPSS would generate two tables for us. In Figure 10.6, we can see the number of
votes we observed in each category, the number of votes we expected in each category,
and the difference (i.e., the Residual) between the two. In Figure 10.7, we can see a
chi-square value of 8.82, just as we calculated, and a p value (i.e., Asymp. Sig.) of .012,
which also supports our decision to reject the null hypothesis.

What about the Post-Hoc Test?

At this point, you would expect us to run a post-hoc test to determine where the sig-
nificant difference lies. Unfortunately, running post-hoc tests with a chi-square test is
very difficult to both explain and calculate; SPSS does not do that for us. Because of
310 | STATISTICS TRANSLATED

FIGURE 10.3.  Vote data in the Data View spreadsheet.

FIGURE 10.4.  Selecting the chi-square test.

 The Chi-Square Tests  |  311

FIGURE 10.5.  Identifying the Test Variable and the Expected Values.

that we are happy knowing that a significant difference does exist; where that differ-
ence lies is obvious.

Working with an Even Number of Expected Values

In this case, we knew the expected values for our formula. At other times, if you
have an even distribution of expected values, you can have SPSS calculate them. For
example, if we were expecting an equal number of responses for each question, SPSS
would have computed the expected value for each cell by dividing the total number
of observed values by the number of cells. In this case, we would wind up with an
expected frequency of 100 for each response, 300 observed values divided by 3 pos-
sible answers. We can see how we would set that up in Figure 10.8; instead of input-
ting the values we wanted, we have simply told SPSS that all categories are equal. The
resulting observed and expected values are shown in Table 10.7.

Vote

Observed N Expected N Residual

It is all right with me 95 120.0 -25.0
I do not like the idea 105 90.0 15.0
I do not care 100 90.0 10.0
Total 300

FIGURE 10.6.  Observed and expected values for vote data.

312 | STATISTICS TRANSLATED

Test Statistics As you can see in Figures 10.9 and

Vote 10.10, since our observed values are
Chi-square 8.819
not much different from the expected
Df 2
values, the computed value of p (i.e.,
Asymp. Sig. .012
.779) would be larger than our tradi-
tional alpha value of .05. This means, of
FIGURE 10.7.  Inferential statistics for the chi-
square goodness-of-fit test. course, that we would not reject our null
hypothesis if we expected an equal num-
ber of responses for each of the answers.

TABLE 10.7.  Observed and Expected Values of Vote Data with All
Categories Equal
It is all right with me. I do not like the idea. I do not care.
Observed  95 105 100
Expected 100 100 100

FIGURE 10.8.  Identifying the Test Variable and the Expected Values when all categories are
equal.

vote

Observed N Expected N Residual

It is all right with me 95 100.0 -5.0
I do not like the idea 105 100.0 5.0
I do not care 100 100.0 .0
Total 300

FIGURE 10.9.  Observed and expected values for the vote data.
 The Chi-Square Tests  |  313

Test Statistics

vote

Chi-square .500
df 2
Asymp. Sig. .779

FIGURE 10.10.  Inferential statistics for the chi-square goodness-of-fit test.

The Case of the Belligerent Bus Drivers

We will better learn how to interpret the results of the chi-square in just a few pages,
but for now let’s go ahead and put the six-step model to work for us. Let’s suppose a
statistician was recently hired to act as an expert witness in a court case. The case
involved a local school system that was suing the bus driver’s union, charging that the
bus drivers were malingering. This, the school system stated, was evidenced by the
fact that far too many bus drivers called in sick on Mondays and Fridays, effectively
extending a normal weekend into a three-day holiday. The school system presented
Table 10.8 showing total absences for given workdays throughout the school year.

TABLE 10.8.  Observed Values of Absence by Day of Week

Monday 25
Tuesday 10
Wednesday 16
Thursday 19
Friday 30

The school system pointed out to the driver’s union that they felt the number of
absences should be about equally distributed over the course of the week (e.g., 20%
of the absences should occur on Monday, 20% on Tuesday, etc.). The driver’s union
agreed that the numbers seemed higher on Mondays and Fridays but felt it was just a
coincidence. They dismissed the concerns of the school system and told the adminis-
trators they would have to support their allegations before they would take any action.
At this point, the school system had no resort other than to attempt to make its point.

S TEP 1 Identify the Problem

In this case, only the school board is interested in this problem
from a practical perspective; the bus driver’s union is happy
with the status quo. The school system has already collected the
numeric data it needs since it had the time, and so on, necessary
to do so; there was nothing unethical about it collecting the
314 | STATISTICS TRANSLATED

data. Again, although the drivers may disagree, this is something the schools should
investigate.

„ Providing school bus transportation is integral to student attendance and

ultimately, achievement. Administrators are attempting to determine if the
drivers are manipulating the system to allow more time away from the job.

S TEP 2 State a Hypothesis

By reading this scenario, the null hypothesis the school board
wants to investigate is apparent:

„ There will be no significant difference in the number

of absences expected on a given day and the actual
number of absences observed on a given day.

The research hypothesis would be

„ There will be a significant difference in the number of absences expected on

a given day and the actual number of absences observed on a given day.

S TEP 3 Identify the Independent Variable

Having stated this, it is easy to see that our independent vari-
able is “day of week.” We have five levels: Monday, Tuesday,
Wednesday, Thursday, and Friday.

S TEP 4 Identify and Describe the Dependent Variable

The dependent variable is the number of absences for each day.
Notice, in Figure 10.11, the total number of observed absences
is 100. This means our expected value for each day of the week
is 20 since the school district believes absences should be evenly
distributed.

Day of Week

Observed N Expected N Residual

Monday 25 20.0 5.0

Tuesday 10 20.0 -10.0
Wednesday 16 20.0 -4.0
Thursday 19 20.0 -1.0
Friday 30 20.0 10.0
Total 100

FIGURE 10.11.  Observed and expected values for day of the week data.
 The Chi-Square Tests  |  315

Since we are dealing with nominal data, it does not make sense to compute most
of the descriptive statistics; the most important value, the mode, was easily seen back
in Table 10.8—the most absences occurred on Fridays. We could, however, use SPSS
to create the bar chart in Figure 10.12 if we wanted; it verifies what we already know.

FIGURE 10.12.  Bar chart for day of the week data.

S TEP 5 Choose the Right Statistical Test

In this case, we have one independent variable with five lev-
els; we have one dependent variable representing nominal data.
Because of this, we will use the chi-square goodness-of-fit test
(i.e., the one-way chi-square test).

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Just by looking at the graphical output and the descriptive sta-
tistics, we had a fairly good idea that a problem existed. By look-
ing at Figure 10.13, we can see there is a significant difference
between the number of absences the
school district expects each day of the
Test Statistics week and the actual number of absences
Day of Week
the drivers take; this decision is support-
Chi-square 12.100
a ed by a p value (i.e., .017) less than an
Df 4
alpha value of .05. Based on this, even
Asymp. Sig. .017
though we do not have a post-hoc test,
the results are obvious; calling in sick 25
FIGURE 10.13.  Inferential statistics for the chi-
times on Monday and 30 times on Fri-
square goodness-of-fit test.
day, the drivers are ruining their case.
316 | STATISTICS TRANSLATED

The Case of the Irate Parents

In this case, we have been hired to investigate the contention of a group of parents
that there is gender discrimination in a local elementary school’s gifted program. Our
job is to try to determine if the number of boys and girls in the program represents
what we could expect in the general population. Let’s use our six-step model:

S TEP 1 Identify the Problem

This is certainly a significant, practical problem and one that
is within the ability and purview of the school to investigate;
to not do so would be highly unethical. Investigating it would
simply involve counting the number of each gender in the pro-
gram and comparing it to the percentage of each gender in the
population. Our problem statement could read:

„ Because of perceived inequalities, this study will investigate whether the

number of males and females in the gifted program is proportionate to the
gender distribution throughout the rest of the school.

S TEP 2 State a Hypothesis

Our null hypothesis is easy enough:

„ There will not be a significant difference in the

percentage of boys and girls in the gifted program
and the percentage of boys and girls in the general
population.

It would lead to the following research hypothesis:

„ There will be a significant difference in the percentage of boys and girls

in the gifted program and the percentage of boys and girls in the general
population.

S TEP 3 Identify the Independent Variable

Our independent variable is gender, and the two levels are boys
and girls.
 The Chi-Square Tests  |  317

S TEP 4 Identify and Describe the Dependent Variable

Our dependent variable is the number of each gender in the
gifted program. Let’s assume there are 48 boys and 52 girls.
Obviously, the mode is “girls”; it doesn’t make a lot of sense to
compute any other descriptive statistics.

S TEP 5 Choose the Right Statistical Test

In this case, we have an independent variable with two levels:
boys and girls. We also have a dependent variable that repre-
sents nominal data. We will use the computer to compute the
chi-square goodness-of-fit test but, first, let’s decide on our
expected values. Remember, we could either expect an equal
number of boys and girls or we could base our expected values on some other criteria.
Suppose we weren’t sure so we did a little research and found that approximately 45%
of the elementary school students in the United States are males (i.e., 45 out of 100
students would be expected to be males); that would give us the following observed
and expected values shown in Figure 10.14.

Gender

Observed N Expected N Residual

Male 48 45.0 3.0

Female 52 55.0 -3.0
Total 100

FIGURE 10.14.  Observed and expected values for gender.

There doesn’t appear to be a lot of difference between our observed and expected
values but, as always, let’s use SPSS to check.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

As can be seen in Figure 10.15, our p value (i.e., .546) is much
greater than an alpha value of .05; this means we fail to reject
the null hypothesis. Because of this, parents of neither gender
should complain; the proportion of boys and girls in the gifted
classes is not significantly different from the national average.
318 | STATISTICS TRANSLATED

Test Statistics

Gender
Chi-square .364
df 1
Asymp. Sig. .546

FIGURE 10.15.  Inferential statistics for the chi-square goodness-of-fit test.

The Chi-Square Test of Independence

At the start of the chapter I said that the chi-square test of independence is much like
the factorial ANOVA in that we are interested in looking at data based on two or more
independent variables. Again, the only difference is that with the chi-square test we
are using nominal rather than quantitative data.
Suppose, for example, we are back to the group we just visited in the last case.
This time they are complaining about an unequal distribution of gender within ability
groups in the local school system. They point out that if one looks at the number of
females in each of the ability groups, it seems there is a disproportionate number of
males in the highest group. They use Table 10.9 to show their point.

TABLE 10.9.  Observed Values of Gender within Ability Group

Male Female Total
Low-ability group 40 24 64
Middle-ability group 35 24 59
High-ability group 56 21 77

Total 131 69 200

By looking at the chart, we can see where the parents might be concerned. There
are more males than females in the low- and middle-ability groups but there are nearly
three times as many males as females in the high-ability group. The parents may really
have something to complain about this time!
Since we have used two independent variables to identify nominal-level data, we
will use a two-way chi-square test, often called the chi-square test of independence. This
test is used in situations where we want to decide if values represented by one indepen-
dent variable are not influenced, or are independent of, the same values represented
by another independent variable. In this case, we want to see if the total number of
students, when we look at it from a gender perspective, is independent of the same
total number of students when we look at it from an ability group perspective. We can
investigate this question using the following research hypothesis:

„ Gender has a significant effect on membership in an ability group (i.e., they

are not independent of one another).
 The Chi-Square Tests  |  319

Our null hypothesis is:

„ Gender has no significant effect on membership in an ability group (i.e.,

they are independent of one another).

By looking at these hypotheses you can see we have two independent variables:
ability group and gender. Ability group has three levels: low, middle, and high; and
gender has two: male and female. Just like with the factorial ANOVA, this means we
are dealing with a 3 × 2 chi-square table.
Our dependent variable is the actual number of occurrences of each of these val-
ues. As we just saw, there is a total of 200 students—131 males and 69 females broken
into the three separate ability groups. There are 40 males and 24 females in the low-
ability group, 35 males and 24 females in the middle-ability group, and 56 males and
21 females in the high-ability group.

Computing Chi-Square for the Test of Independence

Since we have more than one variable, we have to compute chi-square a bit differently.
First, we will look at how we determine our expected values; after that we will look
at computing the cell and overall chi-square values, and then we will worry about the
degrees of freedom.

Computing Expected Values for the Test of Independence

We will use our table of observed values to compute our expected values using the
following formula:
(Row Total)(Column Total)
E=
Sample Size
Using this equation, we can compute the expected value for the first cell (i.e., low-
ability group males) by inserting our row total of 64, our column total of 131, and then
our sample size of 200 into the formula. We then go through the following three steps.

1. E = (64)(131)
200

2. E = 8384
200

3. E = 41.9

The resulting table (i.e., Table 10.10) is sometimes called a contingency table since
we want to decide if placement in one group is contingent on placement in a second
group. If the expected proportions in each cell are not significantly different from the
320 | STATISTICS TRANSLATED

actual proportions, then the groups are independent; in this case, it would mean that
gender does not have an effect on ability group. If the differences are significant, then
we will say the two variables are dependent on one another. That would mean gender
does influence ability group.

TABLE 10.10.  Contingency Table for the Chi-Square Test of Independence

Male Female Total
Low-ability group Observed = 40 Observed = 24 Observed = 64
Expected = 41.9 Expected = 22.1 Expected = 64.0

Middle-ability group Observed = 35 Observed = 24 Observed = 59

Expected = 38.6 Expected = 20.4 Expected = 59

High-ability group Observed = 56 Observed = 21 Observed = 77

Expected = 50.4 Expected = 26.6 Expected = 77

Total Observed = 131 Observed = 69 Observed = 200

Expected = 131 Expected = 69 Expected = 200

Computing the Chi-Square Value for the Test of Independence

Since all the observed values are close to the expected values, things might not be as
bad as some parents think. In order to be sure, we will compute the chi-square values
for each of the cells and then add those together to get the overall chi-square value.
We will use the same formula we used with the goodness-of-fit test:
S (O − E )
2
c2 =
E
Unlike the goodness-of-fit test, now we will focus on each individual cell rather
than an entire row. These steps are shown in Table 10.11.

1. Subtract the expected value in each cell from the observed value in each cell
(column A in the table).
2. Square that value (column B).
3. Divide that value by the expected value for that cell; this gives us the chi-
square value for that cell (column C).
4. Add the chi-square values for each cell to get the overall chi-square value (the
bottom of column C).
 The Chi-Square Tests  |  321

TABLE 10.11.  Computing the Chi-Square Statistic

Observed – Observed – Chi-square =
expected expected squared Column 2 ÷ number expected
Low-ability males 40 – 41.9 = –1.9 3.61 3.61/41.9 = .08
Low-ability females 24 – 22.1 = 1.9 3.61 3.61/22.1 = .16
Middle-ability males 35 – 38.6 = –3.6 12.96 12.96/38.6 = .36
Middle-ability 24 – 20.4 = 3.6 12.96 12.96/20.4 = .64
 females
High-ability males 56 – 50.4 = 5.6 31.36 31.36/50.4 = .62
High-ability females 21 – 26.6 = –5.6 31.36 31.36/26.6 =1.18
Computed chi-square = 3.02

Determining the Degrees of Freedom for the Test of Independence

Here is the formula for the degrees of freedom for the test of independence:

df = (R – 1)(C – 1)

In this formula, C represents the number of columns in the contingency table; in

this case we have two, one for female and one for male. R represents the number of
rows; we have three, one for each ability group. When we subtract one from each of
these values and then multiply them together, we wind up with 2 degrees of freedom.

We Are Finally Going to Test Our Hypothesis

As usual, to test our hypothesis, we have to compare our computed value of chi-square
to the appropriate critical value. In the table, the critical value for 2 degrees of free-
dom is 5.99, greater than our computed value of 3.02. This means our p value would
be greater than or equal .05. As a result, we fail to reject the null hypothesis. Appar-
ently, ability group and gender are independent of one another.
You might be thinking, “This is great, but what does it mean? We have failed to
reject the null, but what’s the bottom line? What does this mean to the complaining
parents?” The answer is very straightforward; all we are saying is that gender and their
ability level are independent of one another. There is about the same proportion of
males and females in each of the ability groupings. To get even more specific, let’s look
at the percentages of each group in Table 10.12.

TABLE 10.12.  Percentage of Each Gender within Ability Group

Male Female Total
Low-ability group 62.5% 37.5% 100%
Middle-ability group 59.3% 40.7% 100%
High-ability group 72.7% 27.3% 100%

Total 65.5% 34.5% 100%

322 | STATISTICS TRANSLATED

Here, the actual number of boys and girls in each ability group has been replaced
by the percentage of the group they represent. For example, out of all of the students in
the low-ability group, 62.5% of them are males; obviously the rest are females. In the
high-ability group, 72.7% are males while only 27.3% are females. When you first look
at this, it does seem like there is some inequality. You cannot, however, go by just the
cell values; you also have to compare your cell percentages to the overall percentages.
For example, our sample of 200 kids is 65.5% male; that means you would expect
about 65.5% of the students in each group to be male. As you can see, the percentage
of males in each group is close to 65.5%. At the same time, the overall percentage
of females in the study is 34.5%; again, the cell percentages are close to this value.
Because of this, we can say that the two variables, gender and ability level, are inde-
pendent of one another and any differences we see are purely due to chance. In other
words, if we know something about a person’s gender, it doesn’t tell us a thing about
her ability level.
Suppose, however, the p value had been less than .05 and we were able to reject
the null hypothesis, thereby saying there was some dependence between the two vari-
ables. Have we seen a situation like that before? Sure we have. In essence, what we
would be stating is that the two variables interact with one another. This is the same
type of phenomenon we saw with the factorial ANOVA, but this time we are using
nominal data.

Using SPSS to Check What We Just Computed

In order to use SPSS to compute the chi-square test of independence, we first set up
our data as shown in Figure 10.16. In this case we used a value of 1 for male and 2 for

FIGURE 10.16.  Selecting the Crosstabs command on the Data View spreadsheet.
 The Chi-Square Tests  |  323

female and used values from 1 to 3 to represent low-, middle-, and high-ability groups.
At that point, we select Analyze, Descriptive Statistics, and Crosstabs.
Following that, as shown in Figure 10.17, we select Gender as our Row and Ability
Group as our Column; we also ask for Descriptive Statistics as shown in Figures 10.18
and 10.19.

FIGURE 10.17.  Identifying the rows and columns for the Crosstabs command.

Gender * Ability Group Crosstabulation

Count

Ability Group

Low Ability Middle Ability High Ability Total

Gender Male 40 35 56 131

Female 24 24 21 69
Total 64 59 77 200
FIGURE 10.18.  Counts of gender within ability group.

Because of rounding issues, the chi-square value here is slightly different from
that we computed manually, but the p value of .220, when divided by 2 (i.e., .110), still
shows that we fail to reject our null hypothesis. This means, as we said earlier, that
gender and ability level are independent of one another; any differences we see are
purely due to chance.
324 | STATISTICS TRANSLATED

Chi-Square Tests

Asymp. Sig. (2-

Value df sided)
a
Pearson Chi-Square 3.031 2 .220
Likelihood Ratio 3.073 2 .215
N of Valid Cases 200

a. 0 cells (.0%) have expected count less than 5. The minimum

expected count is 20.36.

FIGURE 10.19.  Output of the Crosstabs command used for the chi-square test of independence.

The Case of Corporal Punishment

The principal of a high school in a small town has recently started using corporal
punishment in his school. Many of the residents are opposed to this plan because they
feel that minority children are at the highest risk for being punished in this manner.
Because many minority students are already at risk for failure or dropout, concerned
community members have asked him to investigate their assertion. The principal
assures them there is no problem but agrees to investigate.

S TEP 1 Identify the Problem

Obviously, this is a problem that should be investigated, and it’s
certainly ethical to collect numeric data representing the num-
ber of cases of corporal punishment within each ethnic group.
It’s within the principal’s authority to investigate this issue, and
it is in the best interest of the students. Given that, let’s move
forward by stating the problem:

„ Corporal punishment has been shown to negatively affect achievement,

motivation, and persistence in school. To exacerbate the problem, studies
have shown that it is disproportionately applied to minority students, many
of whom are already at risk for failure. It is felt that the relationship between
the frequency of corporal punishment and minority groups be investigated at
the high school in question.

S TEP 2 State a Hypothesis

In order to investigate our hypothesis, we will ask 20 students
from three different ethnic groups whether they have been cor-
porally punished. Our null hypothesis will be:
 The Chi-Square Tests  |  325

„ Membership in an ethnic group has no significant effect on being corporally

punished (i.e., they are independent of one another).

Our research hypothesis would read:

„ Membership in an ethnic group has a significant effect on being corporally

punished (i.e., they are not independent of one another).

S TEP 3 Identify the Independent Variable

Here we have two independent variables. In the first, our inde-
pendent variable is ethnic group; for the sake of our discussion
we’ll call them A, B, and C. The second independent variable is
whether or not the student has been punished.

S TEP 4 Identify and Describe the Dependent Variable

We can call our dependent variable “Punishment” and, again,
there are two possible responses—yes or no. We will use the data
in Table 10.13 to represent our participants’ responses.

TABLE 10.13.  Punishment within Ethnic Group

Yes No Total
Group A Count  5 15 20
Group B Count 14  6 20
Group C Count  6 14 20

Total Count 25 35 60

If we wanted to, we could use a histogram to display our data; in this case, we can
use the histogram in Figure 10.20 to help us see the relationship between the ethnic
groups and their level of support.

S TEP 5 Choose the Right Statistical Test

Herein we have two independent variables, ethnic group and
punishment. Because the dependent variable represents the
number of occurrences of each level within the variables, we
will use the chi-square test of independence (i.e., the factorial
chi-square).

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Figure 10.21 provides us with the contingency table showing the
observed counts, while Figure 10.22 shows us the results of the
chi-square test.
326 | STATISTICS TRANSLATED

FIGURE 10.20.  Bar chart showing counts of punishment within ethnic group.

Our p value of .007 is well below our alpha value of .05, so we must reject our
null hypothesis; the use of corporal punishment is not independent of ethnic group.
This just means that corporal punishment is used inordinately among different ethnic
groups.

Count

punishment
Punished Not Punished Total

Ethnic Group A 5 15 20

Group B 14 6 20
Group C 6 14 20
Total 25 35 60

FIGURE 10.21.  Counts of punishment within ethnic group.

Chi-Square Tests

Asymp. Sig. (2-

Value df sided)

Pearson Chi-Square 10.011 2 .007

Likelihood Ratio 10.141 2 .006
Linear-by-Linear Association .101 1 .750
N of Valid Cases 60

FIGURE 10.22.  Inferential statistics for the chi-square test of independence.

 The Chi-Square Tests  |  327

Post-Hoc Tests Following the Chi-Square

As we said earlier, explaining and running the post-hoc tests for the chi-square tests is
somewhat difficult. Some statistical software packages will do it for us by running tests
that determine the nonproportional cells by using techniques called simultaneous
confidence-interval procedures, while others will use a modification of the Bonferroni
technique (the same tool we talked about when we discussed the ANOVA). If you are
lucky enough to have software that does this, you do not have a problem.
In most instances, however, the software will not do this, so you have to compare,
using the chi-square test, each cell to each of the others to find the nonproportional
scales. In our last example, this would have involved making six separate comparisons.
Not only is this time consuming, but it also increases the probability of a Type I error
(i.e., rejecting the null hypothesis when we should not) since, as we said when we dis-
cussed the ANOVA, the alpha values of .05 are cumulative—the more tests we do, the
larger alpha gets.
There is some good news, however. We can look at the table and begin to see
where the significant differences lie. For example, we can see that groups A and C
have far lower levels of punishment than does Group B. Obviously, that means that
groups A and C have a much larger number of students who say they have not been
corporally punished. While not as exact as a post-hoc test, these observations will go a
long way toward helping the researcher decide how and why the p value was so small.

The Case of Type of Instruction and Learning Style

Throughout the world, more and more colleges and universities are introducing dis-
tance learning programs as a viable alternative for students who need educational
programs that are time and place independent. Some people are critical of these
programs because they feel they may not equally support all learning styles and there-
fore may lead to failure in such a learning environment. In order to investigate their
concerns, let’s use learning style and attrition information from a sample of 80 stu-
dents. By analyzing the interaction of learning style and attrition, we will be able to
determine if the concerns about distance education are warranted.

S TEP 1 Identify the Problem

This is definitely a critical problem since I work primarily in a
distance learning environment. I am interested in this, I have
the ability, time, resources, and so on, to thoroughly investigate
it, and it is certainly ethical. All criteria seem to be met, so let’s
state our problem and see if I get to keep my job!

„ In an effort to meet the needs of students who cannot be time and place
dependent, distance education programs are offered at most institutions of
328 | STATISTICS TRANSLATED

higher education throughout the United States. Many critics believe that a
student’s chance of failure from these programs may be higher because of a
conflict between their learning style and the manner by which the course is
delivered.

S TEP 2 State a Hypothesis

Let’s use the following null hypothesis:

„ Learning style has no significant effect on

performance in different learning environments (i.e.,
they are independent of one another).

The research hypothesis is

„ Learning style has a significant effect on performance in different learning

environments (i.e., they are not independent of one another).

S TEP 3 Identify the Independent Variable

Let’s use learning style as our first independent variable. For the
purpose of this example, we will assume we are using a popular
learning style inventory where students answer a series of ques-
tions and are then classified into one of four learning styles:
Convergers, Divergers, Assimilators, and Accommodators. We
won’t get into the meaning of these four styles; we will just consider them to be the
four levels of our independent variable. Our second independent variable, status, has
two levels: success or failure.

S TEP 4 Identify and Describe the Dependent Variable

In this case, we will define success in the program as having
graduated. Let’s assume we have already created the dataset
and we have created the following contingency table shown in
Figure 10.23 and the bar chart shown in Figure 10.24.

Learning Style
Accomodator Assimilator Diverger Converger Total

Graduate Graduate 9 16 12 10 47

Non-graduate 5 12 6 10 33
Total 14 28 18 20 80

FIGURE 10.23.  Counts of graduation status within learning style.

 The Chi-Square Tests  |  329

FIGURE 10.24.  Bar chart showing counts of graduation within learning style.

Although we cannot be sure, it looks like there might be a difference in success

rates. Since their success and failure rates are equal, it seems the Convergers have pro-
portionately a larger number of students failing to finish the program than students
with other learning styles. Let’s move forward and see if our suspicions are right.

S TEP 5 Choose the Right Statistical Test

In this case, our dependent variable represents nominal data:
the count of each student for a given learning style and their
graduation status. Since we have two independent variables, we
should use the chi-square test of independence.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Despite our observations, because of our relatively large p value,
shown in Figure 10.25, we cannot reject our null hypothesis;
success in the distance education program appears to be inde-
pendent of a student’s learning style. Interestingly, this is data

Chi-Square Tests

Asymp. Sig. (2-

Value df sided)
Pearson Chi-Square 1.304 3 .728
Likelihood Ratio 1.308 3 .727
Linear-by-Linear Association .377 1 .539
N of Valid Cases 80
FIGURE 10.25.  Inferential statistics for chi-square test of independence.
330 | STATISTICS TRANSLATED

I have collected from distance education students over the past 15 years. Believe it or
not, and contrary to what we might think, it seems that learning style doesn’t make a
difference in distance education.

Summary
The chi-square test, more than likely, is the only nonparametric test that most begin-
ning statisticians will ever use. The key points to remember are that we are analyzing
nominal (i.e., categorical) data and, unlike the earlier tests we have looked at, we are
not going to use the chi-square test to compare means or variances. Instead, we are
going to compare the frequency of values in a data distribution; we will either compare
this observed frequency to an expected frequency that we specify or an expected
frequency that has been predetermined. As was the case with the analysis of variance,
we can use the chi-square test for either one independent variable (i.e., the goodness-
of-fit test) or multiple independent variables (i.e., the test of independence).

Do You Understand These Key Words and Phrases?

chi-square goodness-of-fit test chi-square test of independence contingency
table
goodness of fit expected value
one-sample chi-square test observed value

Do You Understand These Formulas?

Computed value of chi-square:

S (O − E )
2
c2 =
E

Degrees of freedom for goodness of fit:

df = (number of entries) – 1

Degrees of freedom for chi-square test of independence:

df = (R – 1)(C – 1)

Expected values for chi-square test of independence:

(Row Total)(Column Total)
E=
Sample Size
 The Chi-Square Tests  |  331

Quiz Time!

As you’ve just seen, working with the chi-square test is very easy. Before we move forward, how-
ever, take a look at each of the following scenarios. After you’ve read through these cases and
answered the questions, you can check your work at the end of the book.

The Case of Prerequisites and Performance

A psychology professor in a graduate school was concerned that the grade distribution for
his class was different for students whose undergraduate degree was in psychology and for
students whose major was in other areas. He collected and analyzed data that he had collected
from his classes and wound up with the results shown in Figures 10.26 and 10.27.

1. What are the null and research hypotheses the professor is investigating?
2. What are the independent variables and their levels?
3. What is the dependent variable?
4. Which chi-square test would he need to test the hypothesis?
5. What does this tell the professor about undergraduate major and performance in a
graduate school psychology class?

Grade
A B C D F Total

College Major Psychology 24 10 7 6 7 54

Other 10 15 13 11 12 61
Total 34 25 20 17 19 115

FIGURE 10.26.  Counts of grades within major.

Chi-Square Tests

Asymp. Sig. (2-

Value df sided)

Pearson Chi-Square 10.966 4 .027

Likelihood Ratio 11.172 4 .025
Linear-by-Linear Association 6.697 1 .010
N of Valid Cases 115

FIGURE 10.27.  Inferential statistics for chi-square test of independence.

332 | STATISTICS TRANSLATED

The Case of Getting What You Asked For

I once met a professor who only used a three-letter grading system. Everyone with an average
numeric grade greater than 1 standard deviation above the mean makes an A; averages fall-
ing between the mean ± 1 standard deviation are assigned a B, and anything lower than that
receives a C. Using the empirical rules, this means that about 68% of students should receive a
B, 16% should receive an A, and 16% should receive a C. One day I heard his class complaining
about what they perceived as an inordinate number of B grades. The professor responded by
showing the students the output presented in Figures 10.28 and 10.29.

1. What are the null and research hypotheses the professor is testing?
2. What is the independent variable and its levels?
3. What is the dependent variable?
4. Which chi-square test would he need to test the hypothesis?
5. What does this tell his class about the distribution of grades according to his grad-
ing scheme?

Grade

Observed N Expected N Residual

A 10 16.0 -6.0
B 67 68.0 -1.0
C 23 16.0 7.0
Total 100

FIGURE 10.28.  Counts of grades.

Test Statistics

Grade
Chi-square 5.327
df 2
Asymp. Sig. .070

FIGURE 10.29.  Inferential statistics for chi-square goodness of fit.

The Case of Money Meaning Nothing

In this case, imagine we are working with a group that wants to build a clinic to work with
drug-addicted clients. After picking the property and going to the city for approval they find
that many local citizens have complained that they do not want such a facility in their neigh-
 The Chi-Square Tests  |  333

borhood. “After all,” they said, “that means we will have a lot of indigent people wandering
around; they are the ones with drug problems!” In order to investigate their concerns, let’s
imagine we have collected relevant data and have produced Figures 10.30 and 10.31.

1. What are the null and research hypotheses we are investigating?

2. What are the independent variables and their levels?
3. What is the dependent variable?
4. Which chi-square test would we need to test this hypothesis?
5. What would we tell the persons concerned with building the clinic in their neigh-
borhood?

Income
$20,000 to $40,000 to $60,000 to $80,000 and
< $20,000 $39,999 $59,999 $79,999 higher Total

Drug Use Yes 22 27 33 21 24 127

No 5 6 7 9 7 34
Total 27 33 40 30 31 161

FIGURE 10.30.  Counts of drug use within income group.

Chi-Square Tests

Asymp. Sig. (2-

Value Df sided)

Pearson Chi-Square 2.055 4 .726

Likelihood Ratio 1.960 4 .743
Linear-by-Linear Association .715 1 .398
N of Valid Cases 161

FIGURE 10.31.  Inferential statistics for chi-square test of independence.

The Case of Equal Opportunity

A clothing manufacturer, always aware that it is both federally mandated as well as morally
the right thing to do, wants to ensure that gender distribution among employees is about the
same as the national average. With a little research, he finds that females represent 45% of the
population but only 41% of workers in the company. Rather than just push it aside as a fluke,
the manager decides to look at the SPSS output shown in Figures 10.32 and 10.33.

1. What are the null and research hypotheses the manager is investigating?
2. What is the independent variable and its levels?
3. What is the dependent variable?
334 | STATISTICS TRANSLATED

4. Which chi-square test would the manager use to test this hypothesis?
5. What would the manager learn about gender representation in the company?
Employee Gender

Observed N Expected N Residual

Female 41 45.0 -4.0
Male 59 55.0 4.0
Total 100

FIGURE 10.32.  Observed and expected values of gender.

Test Statistics

Employee
Gender
Chi-square .646
df 1
Asymp. Sig. .421

FIGURE 10.33.  Inferential statistics for chi-square goodness-of-fit test.

CHAPTER 11

The Correlational Procedures

Introduction
When we discussed descriptive statistics, we talked about the idea of using scatter-
plots to look at the relationship between two sets of quantitative data. We saw cases
where two values, votes and campaign spending for example, might have a positive
relationship—higher levels of spending lead to a greater number of votes. In another
case, we might collect data about the amount of time people exercise daily and their
blood pressure. In this case, we might find a negative relationship: when one value
goes up (i.e., the amount of exercise) the other value goes down (i.e., blood pressure).
In this chapter we are going to expand on that idea by learning how to numerically
describe the degree of relationship by computing a correlation coefficient. We will
continue by showing how to use scatterplots and correlation coefficients together
to investigate hypotheses. The operative word here is “investigate”; that word will be
important in just a few minutes.

Understanding the Idea of Correlations

As we saw when we discussed scatterplots, when we look at the relationship between
two sets of data, one of three things can happen. First, as the values in one dataset get
larger, so can the values in the other dataset. Second, as the values in one dataset get
larger, the values in the other dataset can get smaller. Third, as the values in one data-
set change, there might be no apparent pattern in the change of values in the other
dataset. This idea, called the correlation between two variables, is shown in Table 11.1.
In order to better understand correlation, let’s look at a case we used when we
discussed scatterplots. In this case, we investigated whether the amount of tuition

335
336 | STATISTICS TRANSLATED

reimbursement spent by a company, in any given year, was correlated with the number
of resignations during that same time period. As a reminder, the management of the
company was concerned that more and more employees were taking advantage of that
benefit, becoming better educated and moving to another company for a higher sal-
ary. Using the following data, the company’s manager decided to investigate whether
such a relationship really existed. In Table 11.2, each row represents a given year, the
amount of tuition for that year, and the number of resignations for that same year.

TABLE 11.1.  Understanding Correlation

One value goes up, the
Both values go up other goes down There is no clear pattern
Variable 1 Variable 2 Variable 1 Variable 2 Variable 1 Variable 2
1 5 1 9 1 3
2 6 2 8 2 5
3 7 3 7 3 2
4 8 4 6 4 4
5 9 5 5 5 1

TABLE 11.2.  Tuition and Resignations by Year

Year Tuition (thousands) Resignations
1 $32 20
2 $38 25
3 $41 28
4 $42 25
5 $50 30
6 $60 32
7 $57 40
8 $70 45
9 $80 45
10 $100 90

You can see that there is no easily identified dependent variable. Since the correla-
tion is a descriptive tool, we are interested in using our numeric statistics to look at the
relationship between two variables, tuition and resignations. As we just said, manage-
ment is worried that the more money they spend, the greater the number of resigna-
tions. As we said, though, that is only one of three possibilities. First, if they are right,
as the amount of tuition reimbursement goes up, so will the number of resignations.
A second possibility is that, when the amount of tuition money goes up, the number
of resignations goes down. The final option is that there will be no obvious pattern
between the number of resignations and the amount of tuition money.
In situations like this, we can investigate these relationships with any of several
correlational statistics. The key in determining which correlation tool to use depends
on the type of data we have collected. In this case, both resignations and tuition spent
are quantitative (i.e., interval or ratio data). Because of that, we will use the Pearson
 The Correlational Procedures  |  337

product–moment correlation coefficient (most people call it Pearson’s r). Karl Pearson was
interested in relationships of this type and developed a formula that allows us to calcu-
late a correlation coefficient that numerically describes the relationship between two
variables. Here is his formula; we will go through it step by step and soon be using it
with no problem.

r=
(∑ xy) − (∑ x * ∑ y)
n
n x2 −
(∑ x )  n∑ y2 − (∑ y) 
2 2
 ∑
First, let’s look at the data We will use in Table 11.3, specifically columns 2 and 3.

TABLE 11.3.  Tuition and Resignations

Tuition (thousands) Resignations
$32 20
$38 25
$41 28
$42 25
$50 30
$60 32
$57 40
$70 45
$80 45
$100 90

If you closely examine these data, you can see that, generally speaking, as tuition
goes up, so do resignations. In order to continue with our calculations, we will modify
Table 11.3 and create Table 11.4 and refer to tuition as x and resignations as y; we will
label these as columns A and C.

TABLE 11.4.  Values Needed for Computing the Pearson Correlation Coefficient
x y
32 20
38 25
41 28
42 25
50 30
60 32
57 40
70 45
80 45
100 90
(A) (C)
338 | STATISTICS TRANSLATED

Let’s make another modification and create Table 11.5 by adding two columns
where we can enter the values when we square each of the x and y values and one
column where we can put the product when we multiply each value of x by each value
of y. We are also going to put a row in at the bottom where we can sum everything in
each of the columns; these new columns will be labeled B, D, and E.

TABLE 11.5.  Values Needed for Computing the Pearson Correlation Coefficient
x c2 y y2 xy
32 1024 20 400 640
38 1444 25 625 950
41 1681 28 784 1148
42 1764 25 625 1050
50 2500 30 900 1500
60 3600 32 1024 1920
57 3249 40 1600 2280
70 4900 45 2025 3150
80 6400 45 2025 3600
100 10000 90 8100 9000
Sx = 570 Sx2 36562 3 Sy = 380 Sy2 = 18108 Sxy = 25238
(A) (B) (C) (D) (E)

In this case, remember that n is the number of pairs of data in the equation, in
this case 10. Knowing that, we now have everything we need to include in the formula,
and we can compute Pearson’s r using the following steps. Notice we are just taking the
values from the bottom row of Table 11.5 and inserting them into the formula; again,
we have these columns labeled A–E.

10( E) − ( A * C )
1. r =
[10 * B − ( A) 2][10 * D − (C ) 2]

10(25238) − (570 * 380)

2. r =
[10 * 36562 − (570) 2][10 *18108 − (380)2]
252380 − 216600
3. r =
[365620 − 324900][181080 − 144400]

(252380 − 216600)
4. r =
40720 * 36680
35780
5. r =
1493609600
35780
6. r =
38647.25

7. r = .926
 The Correlational Procedures  |  339

We could, of course, use SPSS to analyze the same data; we would start by input-
ting our data and selecting Analyze, Correlate, and Bivariate (i.e., two variables) as
shown in Figure 11.1.

FIGURE 11.1.  Using the Correlate and Bivariate options on the Data View spreadsheet.

In Figure 11.2, we identify the two fields we want to correlate—tuition and res-
ignations. Obviously, since we are dealing with two quantitative variables, we select
Pearson’s correlation.
These commands would produce the table in Figure 11.3. As you can see in the
figure, Pearson’s r is exactly what we calculated by hand. You can also see the software
has given us a p value (i.e., Sig. 2-tailed) of zero; this indicates a significant correlation.
We will discuss that shortly, but for now let’s focus on the r value itself.

Interpreting Pearson’s r
In this case our r value is .926 but it could have ranged from –1 to +1. When you are
computing a correlation, three things can happen:

1. If both values go up, or if both values go down, it is a positive correlation. The

greater the positive correlation, the closer r is to +1.
2. If one value goes up and the other goes down, they are negatively correlated.
The greater the negative correlation, the closer r moves to –1.
340 | STATISTICS TRANSLATED

FIGURE 11.2.  Selecting the variables to be correlated on the Data View spreadsheet.

Correlations

Tuition (in
thousands of
dollars) Resignations
**
Tuition (in thousands of Pearson Correlation 1 .926
dollars) Sig. (2-tailed) .000

N 10 10
**
Resignations Pearson Correlation .926 1

Sig. (2-tailed) .000

N 10 20
**. Correlation is significant at the 0.01 level (2-tailed).

FIGURE 11.3.  Positive Pearson correlation coefficients.

3. If r is close to zero, it is indicative of little or no correlation; there is not a clear

relationship between the values.

A Word of Caution
A positive correlation exists when the values of two variables both go
up or if both values go down. Only when one variable goes up and the
other goes down do we say there is a negative correlation. Do not make
the mistake of labeling a case in which both variables go down as a negative correla-
tion. The three possibilities are shown below in Table 11.6.
 The Correlational Procedures  |  341

TABLE 11.6.  Relationship between Positive and Negative Correlations

↓↑ ↓↓ ↑↑
Negative correlation Positive correlation Positive correlation

One value goes up, the Both values go down. Both values go up.
other value goes down.

Pearson’s r is between Pearson’s r is between Pearson’s r is between

zero and –1. zero and +1. zero and +1.

In the preceding case, we have a very high r value of .926, but what does that
really mean? How high does the r value have to be for us to consider the relationship
between the variables to be meaningful? The answer to this question is that the inter-
pretation of the r value depends on what the researcher is looking for.
Any time we analyze data, an r value of .90 or greater would be excellent for a pos-
itive correlation and a value less than –.90 would be great for a negative correlation.
In both instances we could clearly see that a relationship exists between the variables
being considered. At other times, we might consider an r value in the .60s and .70s (or
–.60 to –.70) to be enough; it would just depend on the type of relationship we were
looking for. Obviously, if our correlation coefficient is smaller than this, especially
anything between zero and .50 for a positive correlation or between –.50 and zero for
a negative correlation, it is apparent that a strong relationship apparently does not
exist between the two variables.

An Even More Important Word of Caution!

We just saw that a strong relationship existed between the amount of
tuition money spent and the number of resignations. Many beginning
statisticians make the mistake of thinking that a positive correlation
means one thing caused the other to happen. While you might be able to make an
inference, you cannot be sure of a cause-and-effect relationship; it might be due to
chance.
For example, did you know there is a large correlation between the number of
churches in a city and the number of homicides in the same city? While you may be
thinking “blasphemy,” it’s true—a large number of churches is highly correlated with
a large number of homicides. Think carefully about what I just said, though. Should
people insist that churches be torn down in order to lower the homicide rate? Of
course not! While the values are correlated, they certainly are not causal. The real
reason underlying the correlation is the size of the town or city. Larger cities have a
greater number of both churches and homicides; smaller towns have a smaller num-
ber of each. They are related to each other but not in any causal way.
In another case that we can all relate to, during the current Covid-19 pandemic,
many people are reporting high levels of stress and anxiety. While these feelings are
certainly understandable, it could be that the news reports that we read and hear
may contribute to this by reporting somewhat misleading results about the effect of
Covid-19 in a given location. For example, as I’m writing this, California, Florida,
Texas, New York, and Georgia are reporting the highest numbers of cases within the
342 | STATISTICS TRANSLATED

United States. While this is obviously a cause for great concern, and it is very impor-
tant information, at the same time it could be very misleading. Why? Doesn’t it stand
to reason that Florida has more cases than South Dakota, and New York has more
cases than Vermont? Of course, it does, the number of cases of Covid-19 is positively
correlated to the number of residents of the state! States with more residents have
larger numbers of residents testing positive for the disease than do states with smaller
populations. Wouldn’t it be more meaningful if we read or heard information such as
the percentage of residents affected or the number of cases per 100,000 residents in
a given location? Yes, this would be more accurate, but it doesn’t take away from the
message—regardless of where you are and the number of people who are ill, always
protect yourself and those around you!
These are perfect examples of being a good consumer of statistics. You can see
quite a few of these noncausal correlations being presented in various newspapers,
magazines, and other sources; in a lot of cases, people are trying to use them to bol-
ster their point or get their way about something. What is the lesson? Simple, always
look at the big picture.
Now, let’s use the same dataset but create Table 11.7 by changing the resignation
numbers just a bit.

TABLE 11.7.  Tuition and Resignation Data

Tuition (thousands) Resignations
$32 40
$38 35
$41 28
$42 25
$50 20
$60 22
$57 15
$70 25
$80 10
$100 30

Again, just by looking, we can see that an apparent relationship exists. In this
case, it appears that the number of resignations goes down as the amount of tuition
money spent goes up. Let’s check that using the information given in Figure 11.4.
Here SPSS computed an r value of –.419 and, since it is between –.5 and zero, it
is a rather small negative correlation. While there is a correlation between the money
spent and the resignations, it is not very strong. We can also see that the r value is
negative. This indicates an inverse relationship: when one of the values goes up, the
other value goes down. In this case we can see that, as the amount of tuition money
spent gets larger, there are fewer resignations. Remember, although a correlation
exists here, there may or may not be a causal relationship; be careful how you inter-
pret these values.
When you look at Table 11.8; do you see a pattern emerging? Does there seem to
be a logical correlation between the two sets of values? If you look closely, you will see
 The Correlational Procedures  |  343

Correlations

Tuition Resignations
Tuition Pearson Correlation 1 -.419

Sig. (2-tailed) .228

N 10 10
Resignations Pearson Correlation -.419 1
Sig. (2-tailed) .228
N 10 10
FIGURE 11.4.  Negative Pearson correlation coefficients.

there does not appear to be. Sometimes a value in the left column will have a much
greater value in the right column (e.g., 41 and 58), and sometimes the value in the
right column will be much lower (e.g., 32 and 20). As you can see in the SPSS output
in Figure 11.5, because of the lack of a pattern, Pearson’s r is only .013; this supports
our observation that there does not seem to be a meaningful correlation between the
two sets of values. Remember, though, this might be purely coincidental; you cannot
use these results to infer cause and effect.

TABLE 11.8.  Tuition and Resignation Data

Tuition (thousands) Resignations
$32 20
$38 35
$41 58
$42 25
$50 22
$60 72
$57 4
$70 15
$80 90
$100 5

Correlations

Tuition Resignations

Tuition Pearson Correlation 1 .013

Sig. (2-tailed) .971

N 10 10
Resignations Pearson Correlation .013 1
Sig. (2-tailed) .971

N 10 10

FIGURE 11.5.  Pearson correlation coefficients showing a weak positive correlation.

344 | STATISTICS TRANSLATED

A Nonparametric Correlational P rocedure

It stands to reason that, if we can compute a correlation coefficient for quantitative
data, we should be able to do the same for nonparametric data. For beginning statisti-
cians, the only nonparametric correlation we generally use is the Spearman rank differ-
ence correlation (most often called Spearman’s rho). Given its name, it only works with
rank-level (i.e., ordinal) data. Let’s use the data in Table 11.9 as an example.

TABLE 11.9.  Employee and Management Rank of Issues

Issue Employee rank Management rank
A 3 3
B 4 7
C 7 9
D 8 1
E 1 5
F 5 2
G 2 6
H 6 4
I 10 10
J 9 8

In this case, let’s imagine we are members of the board of directors of a company
where several issues have arisen that seem to be affecting employee morale. We have
decided that our first step is to ask both the management team and employees to rank
these issues in terms of importance. We then want to determine if there is agreement
between the two groups in terms of the importance of the issues labeled A–J.
Here we have three columns: the first column represents the 10 issues facing the
company (i.e., A–J) that employees and management are asked to rank. The second
column shows how the issues were ranked by the employees. The third column shows
how the issues were ranked by the management. We can use the Spearman rank-dif-
ference correlation formula to help us determine if the employees and management
ranked them in a similar order:

rs = 1−
6 (∑ d 2 )
n(n2 − 1)
This formula is more straightforward than the Pearson formula, but there are
components we have not used up to this point. First, rs is the symbol for Spearman’s
rho; that is the correlation coefficient we are interested in computing. Notice, we are
using the same symbol as Pearson’s r except we are adding a lower-case s to differenti-
ate it. Second, the number 6 is exactly that: a constant value of 6. Next, the lower-case
n is the number of objects we are going to rank; in this case we have 10.
Computing the square value of all the rankings (i.e., d2) involves first determining
the difference between the employee and management rankings; this is shown in the
column labeled d in Table 11.10. For example, the employees’ ranking for issue D is
an 8; management’s ranking for that same issue is 1; the difference between the two
 The Correlational Procedures  |  345

rankings is 7. The d2 value for that issue is 49 (i.e., 7 * 7); this is nothing more than
the difference value squared.

TABLE 11.10.  Necessary Values for Spearman’s Rho

Employee Management
Issue rank rank d d2
A 3 3 0 0
B 4 7 –3 9
C 7 9 –2 4
D 8 1 7 49
E 1 5 –4 16
F 5 2 3 9
G 2 6 –4 16
H 6 4 2 4
I 10 10 0 0
J 9 8 1 1
Sd2 = 108

We can compute our Spearman’s rho coefficient using the following steps:

6(108)
1. r s = 1 −
10(100 − 1)
648
2. r s = 1 −
990

3. rs = 1 – .655
4. rs = .345

Interpreting Spearman’s rho is the same as interpreting Pearson’s r. The value of

rho can range from –1 to +1, with values of rho approaching +1 indicating a strong
agreement in rankings; values approaching –1 indicate an inverse agreement in rank-
ings. In Figure 11.6 we have a rho value of .345 indicating there is very little agreement
in the rankings of the issues between employees and management. As usual, we can
use SPSS to verify our manual computations:

The p Value of a Correlation

Pay attention also to the p value shown in the table. Sometimes this confuses the
beginning statistician because he wants to use it to reject or fail to reject the hypoth-
esis stated in the second step of the six-step process; that is not what it is for. Remem-
ber, the research hypothesis stated in Step 2 for a correlation is simply looking for a
meaningful relationship between the two variables under consideration; we use our
subjective interpretation of the r value and scatterplot to help us make that decision.
346 | STATISTICS TRANSLATED

Correlations

Employees Management

Spearman's rho Employees Correlation Coefficient 1.000 .345

Sig. (2-tailed) . .328

N 10 10
Management Correlation Coefficient .345 1.000
Sig. (2-tailed) .328 .

N 10 10

FIGURE 11.6.  Spearman’s rho correlation coefficient showing a moderate positive correlation.

The p value shown as part of the output for a correlation tests the hypothesis, “the
correlation coefficient in the population is significantly different from zero.” By using
this hypothesis, statisticians can make decisions about a population r value by using
sample data. This is very, very rarely done by beginning statisticians, so we won’t go
into any further detail. Knowing that, let’s move forward to our first case study.

The Case of the Absent Students

It seems we have talked a lot about students missing class, haven’t we? One would
guess I think going to class is important! Perhaps, if I think that way, then I would
enjoy looking at the correlation between the number of times students are absent dur-
ing a semester and their score on the final exam. In order to do that, I could use our
six-step procedure.

S TEP 1 Identify the Problem

In this case, it’s apparent that I’m concerned about the num-
ber of absences and their relationship to the final grade. Since
that meets all the criteria for a problem I might investigate (e.g.,
within my control, ethical, etc.), we can state the following prob-
lem.

„ Research has shown a relationship between a student’s number of absences

and class achievement. This study will investigate whether there is such a
relationship between grades and the number of absences in my classes.

Before we move forward, look at what I just wrote. Is this really a problem? No,
it may not be; since the correlation measures relationships that are not necessarily
causal, we are simply trying to determine if a problem may exist.
 The Correlational Procedures  |  347

S TEP 2 State a Hypothesis

Our hypothesis will be that a negative relationship exists
between the number of absences and the grade on the final
exam. In other words, the fewer days a student misses, the
higher their final exam grade will be. Given that, our research
hypothesis would read:

„ There will be a negative correlation between the number of times a student is

absent and their grade on the final exam.

Right off the bat, you can see something different about this hypothesis in that
it does not include the word “significant.” This is because of three things. First, the
correlational procedures are descriptive statistics. When we learned to calculate the
correlations earlier, we saw that they are used to tell us whether a linear relationship
exists between two variables. Second, despite the fact they are descriptive in nature,
many people will state a hypothesis and then use the correlation to investigate rather
than test it. Third, because we are not involved in testing a null hypothesis, we are
going to subjectively evaluate our research hypothesis. Because of these things, the
wording and processes of our six-step model are going to change slightly.

S TEP 3 Identify the Independent Variable

Here is one of those changes I just mentioned. By reading the
case, did you notice anything else different about the hypoth-
esis we are investigating? If not, consider the fact that, for the
first time, we are not looking at a scenario where we are trying
to determine if differences exist between groups of data.
Rather, we are trying to decide if a relationship exists between two sets of data. Here,
we are trying to determine if it is true that absences from class are directly related to
scores on the final exam.
As we said earlier, when we look at relationships of this type, we are looking at the
correlation between the two sets of values. In this case, we are trying to determine if
a negative correlation exists; is it true that a higher number of absences is related to a
lower grade on the final exam, or vice versa? In other cases, we are trying to decide if
a positive correlation exists. We saw a good example in the earlier chapters where we
looked at the relationship between the amount of tuition money spent and the num-
ber of resignations each year. In both cases we are trying to see if we can predict the
values in one dataset by knowing the values in another.
Since, in these cases, we are not looking at cause and effect, we must get away from
the idea of having an independent variable and a dependent variable. Since we are try-
ing to determine if a relationship exists between two variables, instead of independent
and dependent variables, we use the terms predictor (i.e., what we are using to help us
make a prediction) variable and criterion (i.e., what we are trying to predict) variable.
In this case, our predictor variable is the number of absences for each student.
348 | STATISTICS TRANSLATED

S TEP 4 Identify and Describe the Dependent Variable

Again, since we technically do not have a dependent variable,
our criterion variable will be each student’s grade on the final
examination. Usually at this point we would show the descrip-
tive statistics for that variable, but there is something else that is
going to be a bit different about this procedure. Since both our
predictor variable and our criterion variable are quantitative, we can show the descrip-
tive statistics for both. Let’s use Table 11.11 showing data from 20 students.

TABLE 11.11.  Absences and Score Data

Absences Score Absences Score Absences Score Absences Score
1 100 4 85 7 70 12 70
2 95 6 82 9 90 6 80
4 90 7 80 11 55 7 80
5 90 2 97 2 94 2 93
6 80 7 77 1 90 3 88

Just by looking closely at this table, it appears that our hypothesis may be right
on target; it does seem as if the kids with the lower number of absences have higher
test scores. Since we have already seen, step by step, how to calculate the descriptive
statistics, let’s just look at the results from SPSS shown in Figure 11.7.

Descriptive Statistics

Mean Std. Deviation N

Number of Absences 5.2000 3.18880 20
Final Exam Score 84.3000 10.84872 20
FIGURE 11.7.  Descriptive statistics for absence and final exam data.

Overall, these do not look too bad; we have an average of 5.2 absences and an aver-
age grade of 84.30. Remember, though, these descriptive statistics are not as meaning-
ful to us as when we were looking for significant differences. Let’s move forward and
try to resolve our hypothesis.

S TEP 5 Choose the Right Statistical Test

Since we have two sets of quantitative data and we want to deter-
mine if a relationship exists between the two, we will use Pear-
son’s r to test our hypothesis.
 The Correlational Procedures  |  349

S TEP 6 Use Data Analysis Software to Test the Hypothesis

In Figure 11.8, we can see we have a Pearson r value of –.836,
which indicates a strong negative correlation; it appears that
the number of absences is inversely related to a student’s final
exam score.

Correlations

Number of Final Exam

Absences Score

Number of Absences Pearson Correlation 1 -.836

Sig. (2-tailed) .000

N 20 20
Final Exam Score Pearson Correlation -.836 1

Sig. (2-tailed) .000

N 20 20

FIGURE 11.8.  Negative Pearson correlation coefficients for absence and exam scores.

In Figure 11.9, the data values have been plotted on the scatterplot with grades
plotted across the x (bottom) axis and the corresponding number of absences plot-
ted along the y (left) axis. We instruct-
ed SPSS to plot a line of best fit and it
shows a nearly 45-degree angle upward
to the left. When we see a line such as
this, we immediately know the relation-
ship we have plotted is negative; when
one value goes up, the other goes down.
In this case, since the line is nearly at a
45-degree angle, we know the relation-
ship is fairly strong. This verifies what
we saw with Pearson’s r.
Always keep in mind, however, that
while both the computed value of r and
the scatterplot indicate a strong relation-
ship, do not be fooled. Correlations are
just descriptive measures of the relation-
FIGURE 11.9.  Scatterplot showing the negative
correlation between absences and exam scores.
ship; that doesn’t necessarily mean that
one caused the other to happen.
350 | STATISTICS TRANSLATED

Another Example: The Case against Sleep

We are not, in every instance, looking for a strong negative or positive relationship;
in some cases, we want to make sure a relationship doesn’t exist. Let’s suppose, for
example, a group of overzealous parents has determined the number of hours chil-
dren sleep is directly related to their performance in class. After all, they surmise,
how can a sleepy child do well in class? Knowing this, the parents have set up a rigid
standard that requires their high school students to get at least 9 hours of sleep per
night. Naturally, the students do three things. First, they get aggravated about this
“stupid” requirement; second, many of them disobey the edicts of their parents; and
third, they will do anything they can, even resort to the use of statistics, to help show
there is no relationship between hours slept and class performance!

S TEP 1 Identify the Problem

In this case, the students are trying to do something I’ve
already warned you about—they want to try to “prove” there is
no relationship between the number of hours they sleep and
their performance in class. Remembering that correlations are
both descriptive and in many instances do not represent a caus-
al relationship, they should be very careful with what they discover! At the same time,
the students do have the ability to investigate this problem, and to them it is impor-
tant, ethical, and so on. Here is their problem statement:

„ A group of parents believes there is a strong relationship between the number

of hours per night a student sleeps and the student’s academic performance;
because of that, they have mandated a minimum number of hours their
students must sleep per night. Students do not think this is realistic or fair
and want to investigate the validity of their parents’ claim.

S TEP 2 State a Hypothesis

To begin their argument, the students start with the following
hypothesis:

„ There will be no correlation between the number of

hours slept per night and grade point average.

S TEP 3 Identify the Independent Variable

Again, we do not have an independent variable. Instead we are
looking at a predictor variable called “number of hours slept.”
 The Correlational Procedures  |  351

S TEP 4 Identify and Describe the Dependent Variable

Our criterion variable is “grade point average.” In Table 11.12,
we have data for 32 students. In Figure 11.10, we can use our
software to describe both the predictor (i.e., hours) and the cri-
terion (i.e., GPA) variables.

TABLE 11.12.  Hours Slept and GPA Data

Hours GPA Hours GPA Hours GPA Hours GPA
4 4.00 5.5 3.00 6.5 3.88  7 3.60
4 2.00 6 4.00 7 1.99  7 3.21
4.5 3.98 6 3.01 7 3.98  7 2.88
4.5 1.15 6 3.12 7 4.00  8 4.00
5 3.33 6 2.88 7 2.98  8 1.84
5 3.46 6 2.75 7 1.15 10 2.83
5 3.80 6.5 2.00 7 3.10 12 3.87
5.5 2.15 6.5 3.45 7 3.84 13 3.09

Descriptive Statistics

Mean Std. Deviation N

GPA 3.0725 .83556 32
Hours Slept 6.6719 1.97814 32

FIGURE 11.10.  Descriptive statistics for GPA and hours slept data.

Here we have a B grade point average (i.e., 3.0725). The average hours slept is
somewhat low, but remember, those two things aren’t important. We are interested
in the relationship between the two variables. By looking at the dataset, we can see
that there is no apparent relationship between the two. For example, some students
who slept 4 hours had a grade point average of 4.00, while others who slept the same
amount of time had a grade point average of 2.00. The lowest grade point average
(1.15) was shared by two students; one of them slept 4.5 hours, while the other slept
over 7 hours. Given this, we will be expecting a small correlation coefficient and a flat
line on the scatterplot. Let’s see what happens.

S TEP 5 Choose the Right Statistical Test

We have two datasets, each of which contains quantitative data,
and we are looking to see if a relationship exists between the
two datasets. Given this, we will use Pearson’s r.
352 | STATISTICS TRANSLATED

S TEP 6 Use Data Analysis Software to Test the Hypothesis

When we enter the data, we will get the output shown in Figures
11.11 and 11.12.
This doesn’t look very promising for the parents, does it?
We can see the line of best fit is almost parallel with the bottom
of the graph and is nowhere near the 45-degree angle we have
seen in the other examples. We can also see, just as we predicted, a very low value for
Pearson’s r of .084. This means there is almost no correlation between the number of
hours students sleep and their grade point average. By using this output and looking
back at the data table, this conclusion becomes obvious. Perhaps some of those who
slept 4 hours used their time to study, while some others used the time to watch televi-
sion; we will never know by looking at this data. The only thing we do know is that,

Correlations

GPA Hours Slept

GPA Pearson Correlation 1 .084

Sig. (2-tailed) .646

N 32 32
Hours Slept Pearson Correlation .084 1

Sig. (2-tailed) .646

N 32 32
FIGURE 11.11.  Pearson correlation coefficient showing a weak correlation between GPA and
hours slept.

FIGURE 11.12.  Scatterplot showing the weak correlation between GPA and hours slept.
 The Correlational Procedures  |  353

even though correlations never prove cause and effect, the students will be using these
results as “scientific proof” that they should be able to stay up later.

The Case of Height versus Weight

Let’s look at one last example that is dear to many of us, the relationship between
height and weight. We all know that, generally speaking, taller people weigh more
than shorter people, but that doesn’t always hold true.

S TEP 1 Identify the Problem

In this case, we are interested in looking at something that is
usually pretty obvious: tall people generally weigh more than
short people. While this is not really a problem per se, it is the
type of thing that correlations are used for a lot of times.

„ The purpose of this study is to examine the relationship between height and
weight.

S TEP 2 State a Hypothesis

This one is easy.

„ There will be a positive correlation between a

person’s height and weight.

S TEP 3 Identify the Independent Variable

In this case our predictor variable (i.e., the independent vari-
able) is height. It is quantitative, and, for our purposes, it ranges
from 50 inches to 80 inches (i.e., 4′ 2″ to 6′ 8″).

S TEP 4 Identify and Describe the Dependent Variable

In this case, our criterion variable is weight. Again, let’s arbi-
trarily use 80 to 320 pounds. We can use the data in Table 11.13
to show the heights and weights of 20 people. The resultant
descriptive statistics are shown in Figure 11.13.
354 | STATISTICS TRANSLATED

TABLE 11.13.  Height and Weight Data

Weight Height Weight Height Weight Height Weight Height
100 60 190 70 120 58 145 68
120 62 150 65 110 63 200 72
130 58 135 69 170 60 180 74
150 70 130 62 160 68 210 76
140 78 170 66 150 69 170 68

Descriptive Statistics

Mean Std. Deviation N

Weight 151.5000 29.82846 20
Height in Inches 66.8000 5.78200 20

FIGURE 11.13.  Descriptive statistics for weight and height.

We can see that the 20 participants have an average height of 66.8 inches (i.e., 5′ 6.8″)
and an average weight of 151.5 pounds.

S TEP 5 Choose the Right Statistical Test

We have two variables, both representing quantitative data, and
we are looking for the relationship between them. Obviously,
we are going to use Pearson’s r.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

SPSS would produce the results shown in Figure 11.14.
It appears we have a fairly strong positive relationship (i.e.,
r = .629); this is verified by the scatterplot where we have a line
sloping up to the right at an approximately 60-degree angle
(Figure 11.15).

Correlations

Weight Height in Inches

Weight Pearson Correlation 1 .629
Sig. (2-tailed) .003

N 20 20
Height in Inches Pearson Correlation .629 1

Sig. (2-tailed) .003

N 20 20

FIGURE 11.14.  Pearson correlation coefficient showing a moderately high correlation between
weight and height.
 The Correlational Procedures  |  355

Based on what we just saw, it seems

we can somewhat support our hypoth-
esis. We do have a moderately strong r
value, and our line of best fit is sloping
up to the right. Again, it is not a perfect
45-degree angle, but it is a lot closer to
that than it is to being a flat line. The
bottom line is that height seems to be a
good predictor of weight but there are
obviously other factors that contribute
as well.

FIGURE 11.15.  Scatterplot showing a moderately

high correlation between weight and height.

The Case of Different Tastes

Recently I visited the Coca-Cola museum in Atlanta. During the tour, I was amazed
to learn that the formula for Coke differs throughout the world; the museum had
dispensers that allowed us to taste different varieties of their product from different
countries. I tried some and I was astonished at how bad some of them tasted. “How,” I
asked, “could anyone enjoy drinking this? Wouldn’t it be better if everyone drank the
same brand of Coke that we enjoy in the United States?”
Of course, my statistical mind went immediately to work. I thought, “If I could
only get two groups of people, some from the United States and some from foreign
countries, I could get them to rank samples from different parts of the world. I will
bet, if everyone were exposed to OUR formula, they would agree that it is best!”

S TEP 1 Identify the Problem

Here, again, there is not a problem from my perspective. I’m
just interested in knowing if people from throughout the world
rank the different varieties of Coke in the same way. It might,
however, be in the best interests of Coca-Cola to investigate my
idea.

„ Developing, manufacturing, and distributing soft drinks throughout the

world is a timely and expensive proposition. If there is no difference in
the rankings of different products, it is possible that manufacturing and
356 | STATISTICS TRANSLATED

advertising costs could be lessened by focusing on a smaller number of

products.

S TEP 2 State a Hypothesis

If I really wanted to conduct such a study, I would use the fol-
lowing hypothesis:

„ There will be no difference in the preference rankings

for persons from the United States and persons from
outside the United States.

S TEP 3 Identify the Independent Variable

In this case, our predictor variable is going to be a person’s
country of residence. Either he is from the United States or he
is from a foreign country.

S TEP 4 Identify and Describe the Dependent Variable

Let’s suppose I asked each group to rank their preferences for
10 different formulas of Coke. Our dependent variable is the
rankings for each of the groups. Let’s use the data in Table
11.14 to get started.

TABLE 11.14.  Ranking of Coke Products from

Different Countries
Formula United States Foreign
A  5 10
B  3  5
C  1  6
D  1  3
E  6  7
F  7  9
G  9  1
H 10  2
I  4  8
J  8  4

The chart has three columns. The first column represents the 10 different for-
mulas we are interested in ranking. Just for the sake of our discussion, we will assume
that formula A is that from the United States. The second column shows the average
ranking for each brand by persons from the United States; the third column shows the
same thing for persons from outside the United States.
 The Correlational Procedures  |  357

S TEP 5 Choose the Right Statistical Test

Because we are using ordinal data (i.e., the rankings from the
two groups) and we are interested in computing a correlation
between the two groups, we will be using the Spearman rho
formula.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

SPSS would generate Figures 11.16 and 11.17.
Our value of rho (i.e., –.316) indicates that a small negative
correlation exists; people from foreign countries have a slight
tendency to rank the sodas in exactly the opposite order from
those in the United States.

Correlations

Foreign
US Ranking Ranking

Spearman's rho US Ranking Correlation Coefficient 1.000 -.316

Sig. (2-tailed) . .374
N 10 10
Foreign Ranking Correlation Coefficient -.316 1.000

Sig. (2-tailed) .374 .

N 10 10

FIGURE 11.16.  Spearman’s rho showing a moderate negative correlation in rankings between
the United States and foreign countries.

The scatterplot verifies this; the

line is going slightly down from left to
right. If the same line was flat, it would
indicate a great deal of disagreement
in the rankings (e.g., one group ranked
an item tenth and the other first, one
ranked an item second, weight and the
other ninth, etc.). A line with a perfect
45-degree angle would indicate a nearly
perfect agreement on the rankings. As
we have said all along, that will probably
never happen.

FIGURE 11.17.  Scatterplot showing a moderate

negative correlation in rankings between the
United States and foreign countries.
358 | STATISTICS TRANSLATED

Once We Have a Linear Relationship, What Can We Do with It?

We have learned how to determine whether a linear relationship exists, now it is time
to put it to work. To do so, think back to our use of Pearson’s r in the last example. It
was easy to use it to determine if a relationship existed between two datasets. Again,
the relationship wasn’t perfect, but there was a relationship between height and weight.
Given that, what would happen if we wanted to use the relationship based on our
sample data to try to predict values for other people in the population? For example,
suppose we wanted to estimate a person’s weight, as closely as possible, based on their
height? Fortunately for us, we have another correlational tool, linear regression; that
allows us to do exactly that.

Linear Regression
If we had a dataset of height and weight data and wound up with an r value of 1.00, we
would be able to state, with 100% accuracy, either a person’s height based on knowing
his weight or his weight based on knowing his height. Unfortunately, we rarely, if ever,
have that degree of accuracy.
At the same time, unless the r value is zero, we know there is some degree of
relationship between height and weight. Knowing that, and depending on the value
of r, we should be able to predict, with some accuracy, one of the variables based on
knowing the other variable. The key to doing so is a thorough understanding of the
line of best fit. Up to this point we have referred to it as a line that shows the trend
of a data distribution, but now we need to know exactly what it is, how to compute it,
and how to use it.

The Regression Equation

Those of you who paid more attention than I did in algebra might remember the for-
mula for plotting a line on a graph; this is known as a regression equation.

y = a + bx

First, y is the value we are trying to predict. For example, if we wanted to predict
weight based on height, we would change our equation to read:

Weight = a + bx

Lower-case x represents the value of the predictor variable. In this case, we know
a person’s height, so let’s enter that into our equation.

Weight = a + b(height)

The next symbol, b, represents the slope; this tells us how steep or flat the best-fit
line is. We can enter it into our equation.
 The Correlational Procedures  |  359

Weight = a + slope(height)

Finally, a is the point on the y-axis where the line of best fit crosses for a given
value of x; we call this the intercept.

Weight = intercept + slope (height)

Since we do not know the value for weight but we do know the value for height,
all we need to get this equation to work is to determine how to compute the intercept
and the slope.

Computing the Slope

Before we compute the slope, it is probably best to understand exactly what we are try-
ing to compute. In order to do this, let’s define the slope as the change in the y value
on a graph based on every incremental change in the x value on the same graph. In
order to get a better conceptual understanding of this, look at Figure 11.18.
You can see the line labeled A starts
at the zero and runs up, to the right, at
a 45-degree angle (i.e., a positive slope).
In this case, as x goes up 1 point, so
does y.
For example, you can see the line
passes through the points (x = 1, y = 1),
(x = 2, y = 2), and so on.
In the following picture, we have
the opposite. The slope is negative; for
each value that x goes up, the value for
y goes down. This means, as shown in
Figure 11.19, that there is a negative cor-
relation.
FIGURE 11.18.  Positive slope. Finally, in Figure 11.20, you can see
the slope is equal to zero; as the value of
x goes up, the value of y stays the same.
It stands to reason then that if we
want to use regression to estimate one
value based on another, we need to
know the slope of the line. Since we have
moved beyond the point where we can
refer to it in general terms, we need a
formula to compute the slope of a line
based on data values we have collected.
Luckily for us, such a formula exists:
n ( Sxy) − ( Sx ) ( Sy)
     r =
n( S x 2 ) − ( S x )
2

  FIGURE 11.19.  Negative slope.

360 | STATISTICS TRANSLATED

As you can see, this formula looks

quite a bit like the formula for Pearson’s r.
Since we have seen how to do all of
those computations, this will prove to be
no trouble at all. First, let’s set up Table
11.15 for us to use.

FIGURE 11.20.  Zero slope.

TABLE 11.15.  Values Needed to Compute the Slope

Height x2 Weight y2 xy
60 3600 100 10000  6000
62 3844 120 14400  7440
58 3364 130 16900  7540
70 4900 150 22500 10500
78 6084 140 19600 10920
70 4900 190 36100 13300
65 4225 150 22500  9750
69 4761 135 18225  9315
62 3844 130 16900  8060
66 4356 170 28900 11220
58 3364 120 14400  6960
63 3969 110 12100  6930
60 3600 170 28900 10200
68 4624 160 25600 10880
69 4761 150 22500 10350
68 4624 145 21025  9860
72 5184 200 40000 14400
74 5476 180 32400 13320
76 5776 210 44100 15960
68 4624 170 29000 11560
Sx = 1336 Sx2 = 89880 Sy = 3030 Sy2 = 475950 Sxy = 204465

You can see that I have already done some of the basic math (e.g., squared each
of the height and weight values) that we will need to use in our equation. We can put
these into our regression equation and work through the following four steps:
 The Correlational Procedures  |  361

20(204465) − (1336)(3030)
1. slope =
20(89880) − (1336)2
4089300 − 4048080
2. slope =
1797600 − 1784896
41220
3. slope =
12704

4. slope = 3.24

Computing the Intercept

As was the case with the slope, it is important to understand what the intercept is prior
to using it in our formula. By definition, it is the value of y when the value of x is zero.
Let’s use our pictures from above to help us better understand this. In Figure 11.21,
when x is zero, so is y; this means our intercept is zero. In Figure 11.22, when x is zero,
y (i.e., our intercept) is 6. Finally, in Figure 11.23, the slope in the diagram is 3; when
x is zero, the line of best fit crosses the y-axis at 3.
As was the case with the slope, it is not always easy to look at a scatterplot and
determine the intercept. We can calculate it, however, so that does not present a prob-
lem.
We have the slope value and using the data from our height and weight table, we
can easily compute the mean of x and the mean of y, so let’s put these values into our
equation.

1. Intercept = y − slope(x )
2. Intercept = 151.5 – 3.24(66.8)

This gives us an intercept value of –65.242. Notice that this number is negative;
that is perfectly normal and happens from time to time. Using it, we have everything
we need for our regression formula so let’s pick a value of x (i.e., the height), 65 for
example, and predict what our y value (i.e., the weight) should be.

FIGURE 11.21.  Intercept of zero.   FIGURE 11.22.  Intercept of 6.

362 | STATISTICS TRANSLATED

Weight = –65.242 + 3.24(65)

When we do the math, we find that,

for a person who is 65 inches tall, we are
predicting a weight of 145.68. In order
to verify this figure, look at the point on
the scatterplot in Figure 11.24 where 65
inches and 145 inches intersect. When
you find that point, you will see that it is
very near the line of best fit. This means
our regression formula is a pretty good
estimator of weight based on height (see
FIGURE 11.23.  Intercept of 3 with zero slope. Figure 11.24).
Let’s use this same data as input
into the SPSS regression function,
shown in Figure 11.25, to make sure we
get the same results. First, we will create
two variables, Height and Weight. Fol-
lowing that, we will enter the same data
we used to compute the correlation; this
is shown in Figure 11.26. As seen in Fig-
ure 11.27, in order to run the regression,
we will select Analyze, Regression, and
Linear. We follow this by identifying our
independent and dependent variables as
well as the statistics we want to have gen-
erated (Figure 11.28).
Figure 11.29 shows that SPSS would
provide the same descriptive statistics as
FIGURE 11.24.  Regression plot for height and for the correlation procedure and would
weight. create the following output from the
regression calculation. On the left, you
can see a box containing the words “Constant” and “Height in inches”; their actual
values fall into the column labeled “Unstandardized Coefficients,” meaning they are
in their original scale. Within that column, there are two smaller columns. The “B”
column contains the actual coefficients that go into the regression equation; the inter-
cept is –65.24 and the slope is 3.245. In both instances, the computed values are the
same as those we computed by hand. To the right of these is the standard error for
both values.
The next column is labeled “Standardized Coefficients”; these are more common-
ly known as “Beta coefficients.” In this column, we will see the standardized values for
all independent (i.e., predictor) values in our regression equation. Since, in this case,
we are dealing with only one independent variable, there is only the one value. To
the right of that, the t values and p values shown are used to test the hypothesis that
the slope and intercept values are zero in the population from which the data points
are drawn. They help the user estimate how likely it is that the computed coefficients
occurred by chance.
 The Correlational Procedures  |  363

FIGURE 11.25.  Creating the height and weight variables in the Variable View spreadsheet.

FIGURE 11.26.  Height and weight data in the Data View spreadsheet.
364 | STATISTICS TRANSLATED

FIGURE 11.27.  Selecting the Linear Regression command.

FIGURE 11.28.  Identifying the independent variable, the dependent variable, and the statis-
tics in the Linear Regression command.
 The Correlational Procedures  |  365

a
Coefficients

Model Standardized
Unstandardized Coefficients Coefficients
B Std. Error Beta T Sig.

1 (Constant) -65.242 63.373 -1.030 .317

Height in Inches 3.245 .945 .629 3.432 .003

a. Dependent Variable: Weight

FIGURE 11.29.  Descriptive statistics from the linear regression.

Why Wasn’t It Exactly Right?

A lot of beginning statisticians are surprised when we get to this point. After all the
computations, they want to know why our regression formula did not exactly predict
the person’s weight based on their height. The answer is simple: the correlation coeffi-
cient between height and weight is less than 1.00. Because of that, error is introduced
into the process, just as it was in the other statistical tests we have used. This is verified
by three things.
First, we can look at the scatterplot and see that all the values do not fall directly
onto the line of best fit. Second, we can see that the standard errors for both the inde-
pendent variable and the slope are greater than zero. Third, we can compute a coef-
ficient of determination by squaring the Pearson’s r value from the correlation (i.e.,
.629 * .629 = .395). This coefficient tells us that 39.5% of the change in the criterion
variable (i.e., weight) is caused by the predictor variable (i.e., height). The remaining
61.5% is either due to error or other variables not considered in the equation.
To better understand this concept, let’s imagine we had a dataset where there
was a perfect correlation between the independent and dependent variables. In Table
11.16, you can see that, for every 1-inch difference in height, there is a 5-pound increase
in weight.

TABLE 11.16.  Height and Weight Data

Height Weight Height Weight Height Weight Height Weight
60 100 65 125 70 150 75 175
61 105 66 130 71 155 76 180
62 110 67 135 72 160 77 185
63 115 68 140 73 165 78 190
64 120 69 145 74 170 79 195

If we computed Pearson’s r for the two values, as shown in Figure 11.30, this would
mean it would be exactly 1.00, indicating a perfect positive linear relationship. If we
look at Figure 11.31, we first see that our standard error is zero; given that, we do
not expect error to interfere with our ability to use the regression procedure to accu-
rately use the predictor variable to predict the criterion variable. We can also see our
366 | STATISTICS TRANSLATED

Correlations

Height in inches Weight

Height in inches Pearson Correlation 1 1.000

Sig. (2-tailed) .000

N 20 20
Weight Pearson Correlation 1.000 1

Sig. (2-tailed) .000

N 20 20

FIGURE 11.30.  Pearson correlation coefficients between height and weight.

a
Coefficients

Model Standardized
Unstandardized Coefficients Coefficients

B Std. Error Beta t Sig.

1 (Constant) -200.000 .000 . .

Height in inches 5.000 .000 1.000 . .
a. Dependent Variable: Weight

FIGURE 11.31.  Inferential statistics from the linear regression.

intercept (i.e., constant) and our slope (i.e., height). In this case if we know a person’s
height, we can accurately state their weight.
If everything we have said up to this point is true, then we should be able to use
the slope and height, along with a given value of the predictor variable, to accurately
predict a criterion variable. Let’s use a height of 64 and enter all these data into our
equation:

1. Weight = –200 + 5(64)

2. Weight = –200 + 320
3. Weight = 120

As we can see, for a height of 64

inches, we are expecting the weight to be
120; we can verify that by looking back
at the table. We could further verify it by
computing the coefficient of determina-
tion (i.e., 1 * 1 = 1). This tells us 100%
of the change in the criterion variable is
caused by the predictor variable. We can
also see this quite clearly by looking at
the scatterplot in Figure 11.32.
As I said, however, this perfect pre-
FIGURE 11.32.  Scatterplot of the height and diction can only come true if there is a
weight data.
 The Correlational Procedures  |  367

perfect linear relationship between the predictor and criterion variables. This perfect
relationship, of course, happens very rarely. In most cases, we can use our regression
equation to help predict values, but we must be aware of the error inherent in the
process.

Using the Six-Step Model: The Case of Age and Driving

Like all the statistical tools we have talked about, we can use regression as part of our
six-step process. In order to do that, let’s imagine an argument between an elderly
rights group and a group interested in highway safety. The safety group wants driv-
ers to be tested every year after they reached the age of 65; they contend this will cut
down on the number of highway accidents. The elderly rights group claims there is no
relationship between drivers’ age and their safety record. They feel that drivers who
are 85 years old are just as safe as those who are 65. You decide to collect data and to
help settle the argument.

S TEP 1 Identify the Problem

Here again we have an implied problem. The safety group
believes one thing, the elderly rights group another. Since both
sides are basing their contentions on feelings rather than fact,
we need a generic problem statement.

„ Safety groups have stated they believe that drivers reaching the age of 65
should be tested each year; their reasoning is that, once a driver reaches that
age, their ability to safely drive an automobile diminishes. Elderly rights
groups feel there is no manner by which the given number of accidents in a
year’s time can be predicted by age. This study will collect and analyze data
to see if such a relationship exists.

S TEP 2 State a Hypothesis

Remember, we are going to leave the word “significant” out of
this.

„ There is no relationship between the age of elderly drivers and the number of
automobile accidents they are involved in during a given year.
368 | STATISTICS TRANSLATED

S TEP 3 Identify the Independent Variable

Our predictor variable is age. Since we are dealing with “elder-
ly” people, let’s use a range from 65 to 85.

S TEP 4 Identify and Describe the Dependent Variable

Our criterion variable is the number of accidents an elderly
driver is involved in for a given year. Let’s use Table 11.17 for our
example. The descriptive statistics and the correlation matrix
for our dataset are shown in Figures 11.33 and 11.34.

TABLE 11.17.  Age and Accident Data

Age Accidents Age Accidents Age Accidents Age Accidents
67 4 66 4 82 3 85 2
72 3 82 0 83 2 76 4
65 6 79 2 70 5 65 3
78 2 77 5 73 4 74 3
80 1 73 3 67 5 79 3

Descriptive Statistics

Mean Std. Deviation N

Number of Accidents 3.2000 1.47256 20

Driver's Age 74.6500 6.41770 20

FIGURE 11.33.  Descriptive statistics for accident and age data.

Correlations

Number of
Accidents Driver's Age

Pearson Correlation Number of Accidents 1.000 -.694

Driver's Age -.694 1.000

Sig. (1-tailed) Number of Accidents . .000
Driver's Age .000 .
N Number of Accidents 20 20

Driver's Age 20 20

FIGURE 11.34.  Pearson correlation coefficients between accidents and age.

 The Correlational Procedures  |  369

An average of 3.2 accidents per year is somewhat disturbing. What is even more
worrisome is that our correlation coefficient is negative; it seems there is an inverse
relationship between age and the number of accidents. Remember, although we are
getting a good picture of what seems to be happening, these are descriptive statistics.
Let’s move forward and see what happens from here.

S TEP 5 Choose the Right Statistical Test

Given the type of data we have collected and our desire to pre-
dict one value given the other, we will use the linear regression
function to analyze our data.

S TEP 6 Use Data Analysis Software to Test the Hypothesis

Using the linear regression function, SPSS will produce Figures
11.35 and 11.36.
The bad feeling we had after Step 3 is getting worse. We
have a negative correlation coefficient; this means there is an
inverse relationship. This idea is supported by the best-fit line
sloping downward from right to left. Let’s move to the next step
and see what happens.

Coefficients

Model Standardized
Unstandardized Coefficients Coefficients
B Std. Error Beta t Sig.

1 (Constant) 15.086 2.917 5.171 .000

Driver's Age -.159 .039 -.694 -4.089 .001

FIGURE 11.35.  Inferential statistics from the linear regression.

Based on the correlation coeffi-

cient, we have a linear, albeit negative,
relationship between age and the num-
ber of accidents in which a driver is
involved. Let’s investigate that further
by picking a couple of ages from our
dataset, 82 and 65, insert them into our
regression formula, and see what hap-
pens. First, for the 82-year-old driver:

 1.
Accidents = 15.086 + –.159(82)
 2.
Accidents = 15.086 – 13.038
 3.
Accidents = 2.084
FIGURE 11.36.  Scatterplot of the age and
­accident data.
370 | STATISTICS TRANSLATED

Then for the 65-year-old driver:

1. Accidents = 15.086 + –.159(65)

2. Accidents = 15.086 – 10.335
3. Accidents = 4.751

Unfortunately, for the highway safety personnel, the results were exactly the
opposite of what they believed: the 85-yearold driver is estimated to be in about two
accidents per year, while the 65-year-old drivers will be in over four accidents during
that same time period!
Remember, although it is implied, we are not proving cause and effect here. Our
coefficient of determination is only .481. This tells us that only 48.1% of the change
in the criterion variable is attributable to age. This means there is error involved;
obviously, other things predict the number of accidents. In this case, we might have to
consider other factors such as number of miles driven per year, number of other driv-
ers on the road, and prevalent weather conditions in each driver’s location.

Summary
Most of my students agree that the correlational procedures are extremely easy to
understand. We used basic math skills to help us decide if a linear relationship exists
between two sets of data. If we do find a linear relationship, we then determine if we
can use one set of data to predict values in the other set. Although we did not talk
about them, we could consider many other tools of this type: correlations where there
are more than two variables; regression formulas with more than one predictor variable
(i.e., multiple regression); and regression equations where you can use your data to
predict a nominal value (i.e., logistic regression). Use of these tools is beyond the scope
of this book, but the good news is that once you understand the purpose and use of
the tools in this chapter, understanding the others is made a lot easier.

Do You Understand These Key Words and Phrases?

coefficient of determination correlation coefficient
criterion variable intercept
line of best fit Pearson’s r
predictor variable slope
Spearman’s rho x-axis
y-axis
 The Correlational Procedures  |  371

Do You Understand These Formulas?

Intercept:
Intercept = y − slope( x )

Regression equation (equation of a line):

y = a + bx

Slope of a line:

slope =
n (∑ xy) − (∑ x )(∑ y)
n (∑ x 2 ) − (∑ x )
2

Quiz Time!

We covered everything we need to cover in this chapter but, before we go, let’s make sure you
completely understand everything we have done. Answer each of these questions and then
turn to the end of the book to check your work.

The Case of “Like Father, Like Son”

In Figures 11.37, 11.38, and 11.39 there is output from a regression analysis trying to predict a
son’s age of death based on the father’s age of death:

Descriptive Statistics

Mean Std. Deviation N

Son's Age 79.1000 3.75500 10

Father's Age 74.3000 4.52278 10

FIGURE 11.37.  Descriptive statistics for father and son age data.

Correlations

Son's Age Father's Age

Pearson Correlation Son's Age 1.000 .790

Father's Age .790 1.000

Sig. (1-tailed) Son's Age . .003
Father's Age .003 .
N Son's Age 10 10

Father's Age 10 10

FIGURE 11.38.  Positive Pearson correlation coefficients between father and son age data.
372 | STATISTICS TRANSLATED

a
Coefficients

Model Standardized
Unstandardized Coefficients Coefficients

B Std. Error Beta t Sig.

1 (Constant) 30.387 13.403 2.267 .053

Father's Age .656 .180 .790 3.641 .007

a. Dependent Variable: Son's Age

FIGURE 11.39.  Inferential statistics from the linear regression.

1. What does Pearson’s r tell us?

2. What is the coefficient of determination? What does it tell us?
3. Given a father’s age at death of 65, how long would we expect the son to live?
4. Generally speaking, in which direction would the line of best fit flow?

The Case of “Can’t We All Just Get Along?”

Figure 11.40 shows the correlation between how Democrats and Republicans ranked the items
in a list of the most pressing issues in America:

Correlations

Democrat Republican

Spearman's rho Democrat Correlation Coefficient 1.000 -.515

Sig. (2-tailed) . .128

N 10 10
Republican Correlation Coefficient -.515 1.000
Sig. (2-tailed) .128 .

N 10 10

FIGURE 11.40.  Spearman rho correlation coefficients for rankings between political parties.

1. Why would you use Spearman’s rho to analyze this data?

2. What does Spearman’s rho tell us?
3. What would you say to the president if he asked you the results?
 The Correlational Procedures  |  373

The Case of More Is Better

Figures 11.41, 11.42, and 11.43 show the output of a regression procedure using highest level of
education attained and average salary in the workplace.

Descriptive Statistics

Mean Std. Deviation N

Income 14,692.31 6.5774 1430

Highest Year 13.07 3.090 1430

FIGURE 11.41.  Descriptive statistics for income and highest year of education data.

Correlations

Income Highest Year

Pearson Correlation Income 1.000 .437

Highest year .437 1.000

Sig. (1-tailed) Income . .000
Highest year .000 .
N Income 1430 1430

Highest year 1430 1430

FIGURE 11.42.  Moderate Pearson correlation coefficient between income and highest year of
education.
a
Coefficients

Model Standardized
Unstandardized Coefficients Coefficients

B Std. Error Beta T Sig.

1 (Constant) 4721.569 563.170 8.206 .000

Highest Year 770.527 41.934 .437 18.375 .000
a. Dependent Variable: Income

FIGURE 11.43.  Inferential statistics from the linear regression.

1. If you were asked to speak to a group of potential high school dropouts, how would
you interpret Pearson’s r?

2. In comparison to the overall average, what would you say to someone who was
thinking about dropping out in the ninth grade?

3. What is the coefficient of determination? What does it tell us?

4. Generally speaking, in which direction would the line of best fit flow?
374 | STATISTICS TRANSLATED

The Case of More Is Better Still

Suppose you use Figures 11.44, 11.45, and 11.46 to follow up the prior case. At this point, you
tell potential dropouts that students who drop out of school earlier tend to get married earlier
and have more children.

Descriptive Statistics

Mean Std. Deviation N

Children 1.85 1.68 1430

Highest Year 13.07 3.090 1430

FIGURE 11.44.  Descriptive statistics for the number of children and highest year of education.

Correlations

Children Highest Year

Pearson Correlation Children 1.000 -.259

Highest year -.259 1.000

Sig. (1-tailed) Children . .000
Highest year .000 .
N Children 1430 1430

Highest year 1430 1430

FIGURE 11.45.  Small negative Pearson correlation coefficient between number of children and
highest year of education.
a
Coefficients

Model Standardized
Unstandardized Coefficients Coefficients

B Std. Error Beta T Sig.

1 (Constant) 3.357 .184 19.195 .000

Highest Year -.129 .014 -.237 -9.405 .000

a. Dependent Variable: Children

FIGURE 11.46.  Inferential statistics from the linear regression.

1. What would you tell the students about Pearson’s r?

2. What is the coefficient of determination? What does it tell us?
3. In comparison to the overall average, what would you say to someone who was
thinking about dropping out in the eighth grade?

4. Generally speaking, in which direction would the line of best fit flow?
5. Combining what you know from Scenario 3, what would you say to the students, in
terms of living in poverty?
 Conclusion: Have We Accomplished
What We Set Out to Do?

When I wrote the first edition of this book, I had several goals in mind. First, I wanted
the reader to put any past statistical prejudices behind them and look at this topic
from a new perspective. Second, I wanted the reader to understand that, in most
cases, you are going to use a relatively easy-to-understand set of statistical tools. Third,
even though we worked through the manual calculations, we also saw we could use
SPSS, one of the many easily used software packages, for our calculations. Lastly, I
wanted the reader to understand there is a straightforward way they can approach
these cases. People that read that edition were very complimentary, I was glad that I
accomplished what I set out to do. I asked then did the book accomplish the goals I set
out for it. Did I accomplish these goals? Let’s take a moment and see.

Statistics in a New Light

Many of you may have approached this book with some trepidation. Simply put, you
really do not like statistics, and, if it were up to you, you would have avoided even
thinking about taking a statistics course. Do you still feel that way? I hope not because,
hopefully, you’ve seen that statistics are no more than two things. First, they are a set
of tools that help us “describe” data we collect. Second, they are tools that help us
make decisions about that same data. As you saw, every statistical tool you’ve learned
about is based on basic math: addition, subtraction, multiplication, and division. You
can do those basic things, and now you can do this.

A Limited Set of Statistical Techniques

There are literally hundreds of statistical tests and techniques we could have talked
about in this book. You may have heard of tools such as multivariate regression,

375
376 | STATISTICS TRANSLATED

canonical correlations, structural equation modeling, and the Komolgorov–Smirnoff

test tossed around, but we did not talk about them here. Why not? Easy, as I said ear-
lier, as a beginning statistician and a consumer of statistics, you do not need many of
these tools except in limited situations. What you do need, in most cases, is the group
of statistical tests we talked about in this book. I firmly believe these will suit your
needs at least 90% of the time. Again, if you need something else, there are plenty
of people willing and able to help you. With their expert guidance, you will find that
most statistical techniques are straightforward—just like those we have talked about.

The Use of Statistical Software Packages

A lot of us statistical “old-timers” remember the days when we had no choice; we had
to calculate these statistics manually. Although the underlying math is basic, it is easy
to make a mistake, especially when you are working with large groups of numbers.
Because of that, there are many statistical software packages, usually based on a com-
mon spreadsheet, that can be used to run these different tests. Here we used SPSS
but, as I said, the differences between the packages are not so great that you cannot
quickly learn them.

A Straightforward Approach
As I said at the beginning of the book, I can remember being faced with situations
when I first started learning about statistics where I was not sure which statistical tech-
nique I should be using. This anxiety, combined with the other issues I have already
discussed, led to a great deal of reluctance on my part, as well as that of my classmates,
to get excited about what we were supposed to be doing. As you saw in this text, how-
ever, we were able to identify a set of six steps that helped us navigate through the sta-
tistical fog! As I have said before, these six steps do not cover every statistical situation
you might find yourself in. They should, however, cover most cases.

At Long Last
So, there we have it, it seems we have met our goals. I hope you are now able to not
only identify and interpret the statistics you will need and, just as importantly, you
are far more comfortable as a consumer of statistics. We have taken what many of you
considered to be a huge, anxiety-provoking problem, and we have shown how easy it
is solving it if we only approach it using small, logic-driven steps. Why can’t the rest of
life be so simple?
APPENDIX A

Area under the Normal Curve Table

(Critical Values of z)

z score 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.0 .0000 .004 .0080 .0120 .0160 .0199 .0239 .0279 .0319 .0359
0.1 .0398 .0438 .0478 .0517 .0557 .0596 .0636 .0675 .0714 .0753
0.2 .0793 .0832 .0871 .0910 .0948 .0987 .1026 .1075 .1103 .1141
0.3 .1179 .1217 .1255 .1293 .1331 .1368 .1406 .1443 .1480 .1517
0.4 .1554 .1591 .1628 .1664 .1700 .1736 .1772 .1808 .1844 .1879
0.5 .1915 .1950 .1985 .2019 .2054 .2088 .2123 .2157 .2190 .2224
0.6 .2257 .2291 .2324 .2357 .2389 .2422 .2454 .2486 .2517 .2549
0.7 .2580 .2611 .2642 .2673 .2704 .2734 .2764 .2794 .2823 .2852
0.8 .2881 .2910 .2939 .2967 .2995 .3023 .3051 .3078 .3106 .3133
0.9 .3159 .3186 .3212 .3238 .3264 .3289 .3315 .3340 .3365 .3389
1.0 .3413 .3438 .3461 .3485 .3508 .3531 .3554 .3577 .3599 .3621
1.1 .3643 .3665 .3686 .3708 .3729 .3745 .3770 .3790 .3810 .3830
1.2 .3849 .3869 .3888 .3907 .3925 .3944 .3962 .3980 .3997 .4015
1.3 .4032 .4049 .4066 .4082 .4099 .4115 .4131 .4147 .4162 .4177
1.4 .4192 .4207 .4222 .4236 .4261 .4265 .4279 .4291 .4306 .4319
1.5 .4332 .4345 .4357 .4370 .4382 .4394 .4406 .4418 .4429 .4441
1.6 .4452 .4463 .4474 .4484 .4495 .4505 .4515 .4525 .4535 .4545
1.7 .4554 .4564 .4573 .4582 .4591 .4590 .4608 .4616 .4625 .4633
1.8 .4641 .4649 .4645 .4664 .4671 .4678 .4686 .4693 .4699 .4706
1.9 .4713 .4719 .4726 .4732 .4738 .4744 .4750 .4756 .4761 .4767
2.0 .4772 .4778 .4783 .4788 .4793 .4798 .4803 .4808 .4812 .4817
2.1 .4821 .2826 .4830 .4834 .4838 .4842 .4846 .4850 .4854 .4857
2.2 .4861 .4874 .4868 .4871 .4875 .4887 .4881 .4884 .4887 .4890
2.3 .4893 .4896 .4998 .4901 .4904 .4906 .4909 .4911 .4913 .4916
2.4 .4918 .4920 .4922 .4925 .4927 .4929 .4931 .4932 .4934 .4936
2.5 .4938 .4940 .4941 .4943 .4945 .4946 .4948 .4949 .4951 .4952
2.6 .4953 .4955 .4956 .4957 .4959 .4960 .4961 .4962 .4963 .4964
2.7 .4965 .4966 .4967 .4468 .4969 .4970 .4971 .4972 .4973 .4974
2.8 .4974 .4975 .4976 .4977 .4977 .4978 .4979 .4979 .4980 .4981
2.9 .4981 .4982 .4982 .4983 .4985 .4984 .4985 .4985 .4986 .4986
3.0 .4987 .4987 .4987 .4988 .4988 .4989 .4989 .4989 .4990 .4990

377
APPENDIX B

Critical Values of t

df Alpha = .10 Alpha = .05 Alpha = .025 df Alpha = .10 Alpha = .05 Alpha = .025
 1 3.078 6.314 12.706 18 1.330 1.734 2.101
 2 1.886 2.920 4.303 19 1.328 1.729 2.093
 3 1.638 2.353 3.182 20 1.325 1.725 2.086
 4 1.533 2.132 2.776 21 1.323 1.721 2.080
 5 1.476 2.015 2.571 22 1.321 1.717 2.074
 6 1.440 1.943 2.447 23 1.319 1.714 2.069
 7 1.415 1.895 2.365 24 1.318 1.711 2.064
 8 1.397 1.860 2.306 25 1.316 1.708 2.060
 9 1.383 1.833 2.262 26 1.315 1.706 2.056
10 1.372 1.812 2.228 27 1.314 1.703 2.052
11 1.363 1.796 2.201 28 1.313 1.701 2.048
12 1.356 1.782 2.179 29 1.311 1.699 2.045
13 1.350 1.771 2.160 30 1.310 1.697 2.042
14 1.345 1.761 2.145 40 1.303 1.684 2.021
15 1.341 1.753 2.131 60 1.296 1.671 2.000
16 1.337 1.746 2.120 120 1.289 1.658 1.980
17 1.333 1.740 2.110 ∞ 1.282 1.645 1.960

378
APPENDIX C

Critical Values of F When Alpha = .01

Within-group
degrees of Between-groups degrees of freedom
freedom 1 2 3 4 5 6 7 8 9 10
  1 4052 5000 5403 5625 5764 5859 5928 5982 6022 6056
  2 98.50 99 99.17 99.25 99.30 99.33 99.36 99.37 99.39 99.40
  3 34.12 30.82 29.46 28.71 28.24 27.91 27.67 27.49 27.35 27.23
  4 21.20 18.00 16.69 15.98 15.52 15.21 14.98 14.80 14.66 14.55
  5 16.26 13.27 12.06 11.39 10.97 10.67 10.46 10.29 10.16 10.05
  6 13.75 10.92 9.78 9.15 8.75 8.47 8.26 8.10 7.98 7.87
  7 12.25 9.55 8.45 7.85 7.46 7.19 6.99 6.84 6.72 6.62
  8 11.26 8.65 7.59 7.01 6.63 6.37 6.18 6.03 5.91 5.81
  9 10.56 8.02 6.99 6.42 6.06 5.80 5.61 5.47 5.35 5.26
 10 10.04 7.56 6.55 5.99 5.64 5.39 5.20 5.06 4.94 4.85
 11 9.65 7.21 6.22 5.67 5.32 5.07 4.89 4.74 4.63 4.54
 12 9.33 6.93 5.95 5.41 5.06 4.82 4.64 4.50 4.39 4.30
 13 9.07 9.7 5.74 5.21 4.86 4.62 4.44 4.30 4.19 4.10
 14 8.86 6.51 5.56 5.04 4.69 4.46 4.28 4.14 4.03 3.94
 15 8.68 6.36 5.42 4.89 4.56 4.32 4.14 4.00 3.89 3.80
 16 8.53 6.23 5.29 4.77 4.44 4.20 4.03 3.89 3.78 3.69
 17 8.40 6.11 5.18 4.67 4.34 4.10 3.93 3.79 3.68 3.59
 18 8.29 6.01 5.09 4.58 4.25 4.01 3.84 3.71 3.60 3.51
 19 8.18 5.93 5.01 4.50 4.17 3.94 3.77 3.63 3.52 3.43
 20 8.10 5.85 4.94 4.43 4.10 3.87 3.70 3.56 3.46 3.37
 21 8.02 5.78 4.87 4.37 4.04 3.81 3.64 3.51 3.40 3.31
 22 7.95 5.72 4.82 4.31 3.99 3.76 3.59 3.45 3.35 3.26
 23 7.88 5.66 4.76 4.26 3.94 3.71 3.54 3.41 3.30 3.21
 24 7.82 5.61 4.72 4.22 3.90 3.67 3.50 3.36 3.26 3.17
 25 7.77 5.57 4.68 4.18 3.85 3.63 3.46 3.32 3.22 3.13
 26 7.72 5.53 4.64 4.14 3.82 3.59 3.42 3.29 3.18 3.09
 27 7.68 5.49 4.60 4.11 3.78 3.56 3.39 3.26 3.15 3.06
 28 7.64 5.45 4.57 4.07 3.75 3.53 3.36 3.23 3.12 3.03
 29 7.60 5.42 4.54 4.04 3.73 3.50 3.33 3.20 3.09 3.00
 30 7.56 5.39 4.51 4.02 3.70 3.47 3.30 3.17 3.07 2.98
 40 7.31 5.18 4.31 3.83 3.51 3.29 3.12 2.99 2.89 2.80
 60 7.08 4.98 4.13 3.65 3.34 3.12 2.95 2.82 2.72 2.63
120 6.85 4.79 3.95 3.48 3.17 2.96 2.79 2.66 2.56 2.47
∞ 6.63 4.61 3.78 3.32 3.02 2.80 2.64 2.51 2.41 2.32

379
380 | Appendix C

Within-group
degrees of Between-groups degrees of freedom
freedom 12 15 20 24 30 40 60 120 ∞
  1 6106 6157 6209 6235 6261 6287 6313 6339 6366
  2 99.42 99.43 99.45 99.46 99.47 99.47 99.48 99.49 99.50
  3 27.05 26.87 26.69 26.60 26.50 26.41 26.32 26.62 26.13
  4 14.37 14.20 14.02 13.93 13.84 13.75 13.65 13.56 13.46
  5 9.89 9.72 9.55 9.47 9.38 9.29 9.20 9.11 9.02
  6 7.72 7.56 7.40 7.31 7.23 7.14 7.06 6.97 6.88
  7 6.47 6.31 6.16 6.07 5.99 5.91 5.82 5.74 5.65
  8 5.67 5.52 5.36 5.28 5.20 5.12 5.03 4.95 4.86
  9 5.11 4.96 4.81 4.73 4.65 4.57 4.48 4.40 4.31
 10 4.71 4.56 4.41 4.33 4.25 4.17 4.08 4.00 3.91
 11 4.40 4.25 4.10 4.02 3.94 3.86 3.78 3.69 3.60
 12 4.16 4.01 3.86 3.78 3.70 3.62 3.54 3.45 3.36
 13 3.96 3.82 3.66 3.59 3.51 3.43 3.34 3.25 3.17
 14 3.80 3.66 3.51 3.43 3.35 3.27 3.18 3.09 3.00
 15 3.67 3.52 3.37 3.29 3.21 3.13 3.05 2.96 2.87
 16 3.55 3.41 3.26 3.18 3.10 3.02 2.93 3.84 2.75
 17 3.46 3.31 3.16 3.08 3.00 2.92 2.83 2.75 2.65
 18 3.37 3.23 3.08 3.00 2.92 2.84 2.75 2.66 2.57
 19 3.30 3.15 3.00 2.92 2.84 2.76 2.67 2.58 2.49
 20 3.23 3.09 2.94 2.86 2.78 2.69 2.61 2.52 2.42
 21 3.17 3.03 2.88 2.80 2.72 2.64 2.55 2.46 2.36
 22 3.123 2.98 2.83 2.75 2.67 2.58 2.50 2.40 2.31
 23 3.07 2.93 2.78 2.70 2.62 2.54 2.45 2.35 2.26
 24 3.03 2.89 2.74 2.66 2.58 2.49 2.40 2.31 2.21
 25 2.99 2.85 2.70 2.62 2.54 2.45 2.36 2.27 2.17
 26 2.96 2.81 2.66 2.58 2.50 2.42 2.33 2.23 2.13
 27 2.93 2.78 2.63 2.55 2.47 2.38 2.29 2.20 2.10
 28 2.90 2.75 2.60 2.52 2.44 2.35 2.26 2.17 2.06
 29 2.87 2.73 2.57 2.49 2.41 2.33 2.23 2.14 2.03
 30 2.84 2.70 2.55 2.47 2.39 3.30 2.21 2.11 2.01
 40 2.66 2.52 2.37 2.29 2.20 2.11 2.02 1.92 1.80
 60 2.50 2.35 2.20 2.12 2.03 1.94 1.84 1.73 1.60
120 2.34 2.19 2.03 1.95 1.86 1.76 1.66 1.53 1.38
∞ 2.18 2.04 1.88 1.79 1.70 1.59 1.47 1.32 1.00
APPENDIX D

Critical Values of F When Alpha = .05

Within-group
degrees of Between-groups degrees of freedom
freedom 1 2 3 4 5 6 7 8 9 10
  1 161.4 199.5 215.7 224.6 230.2 234.0 236.8 238.9 240.5 241.9
  2 18.51 19.0 19.16 19.25 19.30 19.33 19.35 19.37 19.38 19.40
  3 10.13 9.55 9.28 9.12 9.01 8.94 8.89 8.85 8.81 8.79
  4 7.71 6.94 6.59 6.39 6.26 6.16 6.09 6.04 6.00 5.96
  5 6.61 5.79 5.41 5.19 5.05 4.95 4.88 4.82 4.77 4.74
  6 5.99 5.14 4.76 4.53 4.39 4.28 4.21 4.15 4.10 4.06
  7 5.59 4.74 4.35 4.12 3.97 3.87 3.79 3.73 3.68 3.64
  8 5.32 4.46 4.07 3.84 3.69 3.58 3.5 3.44 3.39 3.35
  9 5.12 4.26 3.86 3.63 3.48 3.37 3.29 3.23 3.18 3.14
 10 4.96 4.10 3.71 3.48 3.33 3.22 3.14 3.07 3.02 2.98
 11 4.84 3.98 3.59 3.36 3.20 3.09 3.01 2.95 2.90 2.85
 12 4.75 3.89 3.49 3.26 3.11 3.00 2.91 2.85 2.80 2.75
 13 4.67 3.81 3.41 3.18 3.03 2.92 2.83 2.77 2.71 2.67
 14 4.60 3.74 3.34 3.11 2.96 2.85 2.76 2.70 2.65 2.60
 15 4.54 3.68 3.29 3.06 2.90 2.79 2.71 2.64 2.59 2.54
 16 4.49 3.63 3.24 3.01 2.85 2.74 2.66 2.59 2.54 2.49
 17 4.45 3.59 3.20 2.96 2.81 2.70 2.61 2.55 2.49 2.45
 18 4.41 3.55 3.16 2.93 2.77 2.66 2.58 2.51 2.46 2.41
 19 4.38 3.52 3.13 2.90 2.74 2.63 2.54 2.48 2.42 2.38
 20 4.35 3.49 3.10 2.87 2.71 2.60 2.51 2.45 2.39 2.35
 21 4.32 3.47 3.07 2.84 2.68 2.57 2.49 2.42 2.37 2.32
 22 4.30 3.44 3.05 2.82 2.66 2.55 2.46 2.40 2.34 2.30
 23 4.28 3.42 3.03 2.80 2.64 2.53 2.44 2.37 2.32 2.27
 24 4.26 3.40 3.01 2.78 2.62 2.51 2.42 2.36 2.30 2.25
 25 4.24 3.39 2.99 2.76 2.60 2.49 2.40 2.34 2.28 2.24
 26 4.23 3.37 2.98 2.74 2.59 2.47 2.39 2.32 2.27 2.22
 27 4.21 3.35 2.96 2.73 2.57 2.46 2.37 2.31 2.25 2.20
 28 4.20 3.34 2.95 2.71 2.56 2.45 2.36 2.29 2.24 2.19
 29 4.18 3.33 2.93 2.70 2.55 2.43 2.35 2.28 2.22 2.18
 30 4.17 3.32 2.92 2.69 2.53 2.42 2.33 2.27 2.21 2.16
 40 4.08 3.23 2.84 2.61 2.45 2.34 2.25 2.18 2.12 2.08
 60 4.00 3.15 2.76 2.53 2.37 2.25 2.17 2.10 2.04 1.99
120 3.92 3.07 2.68 2.45 2.29 2.17 2.09 2.02 1.96 1.91
∞ 3.84 3.00 2.60 2.37 2.21 2.10 2.01 1.94 1.88 1.83

381
382 | Appendix D

Within-group
degrees of Between-groups degrees of freedom
freedom 12 15 20 24 30 40 60 120 ∞
  1 243.9 245.9 248.0 249.1 250.1 251.1 252.2 253.3 254.3
  2 19.41 19.43 19.45 19.45 19.46 19.47 19.48 19.49 19.5
  3 8.74 8.70 8.66 8.64 8.62 8.59 8.57 8.55 8.53
  4 5.91 5.86 5.80 5.77 5.75 5.72 5.69 5.66 5.63
  5 4.68 4.62 4.56 4.53 4.50 4.46 4.43 4.40 4.36
  6 4.00 3.94 3.87 3.84 3.81 3.77 3.74 3.70 3.67
  7 3.57 3.51 3.44 3.41 3.38 3.34 3.30 3.27 3.23
  8 3.28 3.22 3.15 3.12 3.08 3.04 3.01 2.97 2.93
  9 3.07 3.01 2.94 2.90 2.86 2.83 2.79 2.75 2.71
 10 2.91 2.85 2.77 2.74 2.70 2.66 2.62 2.58 2.54
 11 2.79 2.72 2.65 2.61 2.57 2.53 2.49 2.45 2.40
 12 2.69 2.62 2.54 2.51 2.47 2.43 2.38 2.34 2.30
 13 2.60 2.53 2.46 2.42 2.38 2.34 2.30 2.25 2.21
 14 2.53 2.46 2.39 2.35 2.31 2.27 2.22 2.18 2.13
 15 2.48 2.40 2.33 2.29 2.25 2.20 2.16 2.11 2.07
 16 2.42 2.35 2.28 2.24 2.19 2.15 2.11 2.06 2.01
 17 2.38 2.31 2.23 2.19 2.15 2.10 2.06 2.01 1.96
 18 2.34 2.27 2.19 2.15 2.11 2.06 2.02 1.97 1.92
 19 2.31 2.23 2.16 2.11 2.07 2.03 1.98 1.93 1.88
 20 2.28 2.20 2.12 2.08 2.04 1.99 1.95 1.90 1.84
 21 2.25 2.18 2.10 2.05 2.01 1.96 1.92 1.87 1.81
 22 2.23 2.15 2.07 2.03 1.98 1.94 1.89 1.84 1.78
 23 2.20 2.13 2.05 2.01 1.96 1.91 1.86 1.81 1.76
 24 2.18 2.11 2.03 1.98 1.94 1.89 1.84 1.79 1.73
 25 2.16 2.09 2.01 1.96 1.92 1.87 1.82 1.77 1.71
 26 2.15 2.07 1.99 1.95 1.90 1.85 1.80 1.75 1.69
 27 2.13 2.06 1.97 1.93 1.88 1.84 1.79 1.73 1.67
 28 2.12 2.04 1.96 1.91 1.87 1.82 1.77 1.71 1.65
 29 2.10 2.03 1.94 1.90 1.85 1.81 1.75 1.70 1.64
 30 2.09 2.01 1.93 1.89 1.84 1.79 1.74 1.68 1.62
 40 2.00 1.92 1.84 1.79 1..74 1.69 1.64 1.58 1.51
 60 1.92 1.84 1.75 1.70 1.65 1.59 1.53 1.47 1.39
120 1.83 1.75 1.66 1.61 1.55 1.50 1.43 1.35 1.25
∞ 1.75 1.67 1.57 1.52 1.46 1.39 1.32 1.22 1.00
APPENDIX E

Critical Values of F When Alpha = .10

Within-group
degrees of Between-groups degrees of freedom
freedom 1 2 3 4 5 6 7 8 9 10
  1 39.86 49.50 53.59 55.83 57.24 58.20 58.91 59.44 59.86 60.19
  2 8.53 9.00 9.16 9.24 9.29 9.33 9.35 9.37 9.38 9.39
  3 5.54 5.46 5.39 5.34 5.31 5.28 5.27 5.25 5.24 5.23
  4 4.54 4.32 4.19 4.11 4.05 4.01 3.98 3.95 3.94 3.92
  5 4.06 3.78 3.62 3.52 3.45 3.40 3.37 3.34 3.32 3.30
  6 3.78 3.46 3.29 3.18 3.11 3.05 3.01 2.98 2.96 2.94
  7 3.59 3.26 3.07 2.96 2.88 2.83 2.78 2.75 2.72 2.70
  8 3.46 3.11 2.92 2.81 2.73 2.67 2.62 2.59 2.56 2.54
  9 3.36 3.01 2.81 2.69 2.61 2.55 2.51 2.47 2.44 2.42
 10 3.29 2.92 2.73 2.61 2.52 2.46 2.41 2.38 2.35 2.32
 11 3.23 2.86 2.66 2.54 2.45 2.39 2.34 2.30 2.27 2.25
 12 3.18 2.81 2.61 2.48 2.39 2.33 2.28 2.24 2.21 2.19
 13 3.14 2.76 2.56 2.43 2.35 2.28 2.23 2.20 2.16 2.14
 14 3.10 2.73 2.52 2.39 2.31 2.24 2.19 2.15 2.12 2.10
 15 3.07 2.70 2.49 2.36 2.27 2.21 2.16 2.12 2.09 2.06
 16 3.05 2.67 2.46 2.33 2.24 2.18 2.13 2.09 2.06 2.03
 17 3.03 2.64 2.44 2.31 2.22 2.15 2.10 2.06 2.03 2.00
 18 3.01 2.62 2.42 2.29 2.20 2.13 2.08 2.04 2.00 1.98
 19 2.99 2.61 2.40 2.27 2.18 2.11 2.06 2.02 1.98 1.96
 20 2.97 2.59 2.38 2.25 2.16 2.09 2.04 2.00 1.96 1.94
 21 2.96 2.57 2.36 2.23 2.14 2.08 2.02 1.98 1.95 1.92
 22 2.95 2.56 2.35 2.22 2.13 2.06 2.01 1.97 1.93 1.90
 23 2.94 2.55 2.34 2.21 2.11 2.05 1.99 1.95 1.92 1.89
 24 2.93 2.54 2.33 2.19 2.10 2.04 1.98 1.94 1.91 1.88
 25 2.92 2.53 2.32 2.18 2.09 2.02 1.97 1.93 1.89 1.87
 26 2.91 2.52 2.31 2.17 2.08 2.01 1.96 1.92 1.88 1.86
 27 2.90 2.51 2.30 2.17 2.07 2.00 1.95 1.91 1.87 1.85
 28 2.89 2.50 2.29 2.16 2.06 2.00 1.94 1.90 1.87 1.84
 29 2.89 2.50 2.28 2.15 2.06 1.99 1.93 1.89 1.86 1.83
 30 2.88 2.49 2.28 2.14 2.05 1.98 1.93 1.88 1.85 1.82
 40 2.84 2.44 2.23 2.09 2.00 1.93 1.87 1.83 1.79 1.76
 60 2.79 2.39 2.18 2.04 1.95 1.87 1.82 1.77 1.74 1.71
120 2.75 2.35 2.13 1.99 1.90 1.82 1.77 1.72 1.68 1.65
∞ 2.71 2.30 2.08 1.94 1.85 1.77 1.72 1.67 1.63 1.60

383
384 | Appendix E

Within-group
degrees of Between-groups degrees of freedom
freedom 12 15 20 24 30 40 60 120 ∞
  1 60.71 61.22 61.74 62.00 62.26 62.53 62.79 63.06 63.33
  2 9.41 9.42 9.44 9.45 9.46 9.47 9.47 9.48 9.49
  3 5.22 5.20 5.18 5.18 5.17 5.16 5.15 5.14 5.13
  4 3.90 3.87 3.84 3.83 3.82 3.80 3.79 3.78 3.76
  5 3.27 3.24 3.21 3.19 3.17 3.16 3.14 3.12 3.10
  6 2.90 2.87 2.84 2.82 2.80 2.78 2.76 2.74 2.72
  7 2.67 2.63 2.59 2.58 2.56 2.54 2.51 2.49 2.47
  8 2.50 2.46 2.42 2.40 2.38 2.36 2.34 2.32 2.29
  9 2.38 2.34 2.30 2.28 2.25 2.23 2.21 2.18 2.16
 10 2.28 2.24 2.20 2.18 2.16 2.13 2.11 2.08 2.06
 11 2.21 2.17 2.12 2.10 2.08 2.05 2.03 2.00 1.97
 12 2.15 2.10 2.06 2.04 2.01 1.99 1.96 1.93 1.90
 13 2.10 2.05 2.01 1.98 1.96 1.93 1.90 1.88 1.85
 14 2.05 2.01 1.96 1.94 1.91 1.89 1.86 1.83 1.80
 15 2.02 1.97 1.92 1.90 1.87 1.85 1.82 1.79 1.76
 16 1.99 1.94 1.89 1.87 1.84 1.81 1.78 1.75 1.72
 17 1.96 1.91 1.86 1.84 1.81 1.78 1.75 1.72 1.69
 18 1.93 1.89 1.84 1.81 1.78 1.75 1.72 1.69 1.66
 19 1.91 1.86 1.81 1.79 1.76 1.73 1.70 1.67 1.63
 20 1.89 1.84 1.79 1.77 1.74 1.71 1.68 1.64 1.61
 21 1.87 1.83 1.78 1.75 1.72 1.69 1.66 1.62 1.59
 22 1.86 1.81 1.76 1.73 1.70 1.67 1.64 1.60 1.57
 23 1.84 1.80 1.74 1.72 1.69 1.66 1.62 1.59 1.55
 24 1.83 1.78 1.73 1.70 1.67 1.64 1.61 1.57 1.53
 25 1.82 1.77 1.72 1.69 1.66 1.63 1.59 2.56 1.52
 26 1.81 1.76 1.71 1.68 1.65 1.61 1.58 1.54 1.50
 27 1.80 1.75 1.70 1.67 1.64 1.60 1.57 1.53 1.49
 28 1.79 1.74 1.69 1.66 1.63 1.59 1.56 1.52 1.48
 29 1.78 1.73 1.68 1.65 1.62 1.58 1.55 1.51 1.47
 30 1.77 1.72 1.67 1.64 1.61 1.57 1.54 1.50 1.46
 40 1.71 1.66 1.61 1.57 1.54 1.51 1.47 1.42 1.38
 60 1.66 1.60 1.54 1.51 1.48 1.44 1.40 1.35 1.29
120 1.60 1.55 1.48 1.45 1.41 1.37 1.32 1.26 1.19
∞ 1.55 1.49 1.42 1.38 1.34 1.30 1.24 1.17 1.00
APPENDIX F

Critical Values of Chi-Square

Degrees of
freedom Alpha = .01 Alpha = .05 Alpha = .10
 1 6.63 3.84 2.71
 2 9.21 5.99 4.61
 3 11.34 7.82 6.25
 4 13.28 9.49 7.78
 5 15.09 11.07 9.24
 6 16.81 12.59 10.65
 7 18.48 14.07 12.02
 8 20.09 15.51 13.36
 9 21.66 16.92 14.68
10 23.21 18.31 15.99
11 24.73 19.68 17.28
12 26.22 21.03 18.55
13 27.69 22.36 19.81
14 29.25 23.68 21.06
15 30.58 25.00 22.31

385
APPENDIX G

Selecting the Right Statistical Test

Number of Type of data

Number of levels in the Number of the dependent Alternate
independent independent dependent variable Statistical test statistical
variables variable variables represents to use test**
N/A* N/A* 1 Quantitative One-sample N/A
z-test
or
one-sample
t-test

1 2 1 Quantitative Dependent- Wilcoxon t-test

sample t-test or
or Mann–
independent- Whitney U test
sample t-test

1 3 or more 1 Quantitative Analysis of Kruskal–Wallis

variance H test
(ANOVA)

2 2 or more 1 Quantitative Factorial N/A

ANOVA

1 or more 2 or more 2 or more Quantitative Multivariate N/A

ANOVA
(MANOVA)

0*** 0 1*** Quantitative Pearson N/A

correlation

0*** 0 1*** Ordinal Spearman N/A

correlation

1 or more 2 or more 1 or more Nominal Chi-square N/A

*Not all parametric tests have a nonparametric equivalent, or the test mentioned is already a nonparametric
test.
**When you’re working with one-sample z tests or one-sample t-tests, it is really not logical to point out the
independent variable or its levels. As we have just seen with the one-sample z-test, we are just comparing a
sample statistic to a population parameter. We will see the same holds true, albeit with smaller sample sizes,
when we visit the one-sample t-test in the next chapter.
***Independent variables are called predictor variables in correlations. Dependent variables are known as
criterion variables in correlations.

386
 Glossary

a priori:  Before an event occurs. For example, we decide our alpha value a priori.
This means we decide our alpha value prior to the start of a study.
alpha value:  The degree of risk we are willing to take when computing inferential
statistics. Sometimes referred to as the Type I error rate.
alternate hypothesis:  See “research hypothesis.”
analysis of covariance (ANCOVA):  A version of the analysis of variance where ini-
tial differences in the dependent variable are taken into account prior to final calcula-
tions.
analysis of variance (ANOVA):  A statistical tool based on an f distribution. Varieties
of the ANOVA include the one-way ANOVA, the factorial ANOVA, and the multivari-
ate ANOVA (MANOVA).
area under the curve:  A value ranging from 0 to 100% representing the percentage
of values in a given range under one of the data distributions (e.g., the z distribution
and the t distribution).
area under the normal curve table:  See “z table.”
assumptions:  Characteristics of a dataset we assume to be true prior to using a given
statistical procedure.
bar chart:  A graph showing bars rising from the bottom that are proportionate to
the number of occurrences of a given value in a dataset.
beta:  Annotation used to designate the probability of making a Type II error.
between-groups variance:  The total amount of variance between datasets represent-
ing the levels of an independent variable.

387
388 | Glossary

bimodal:  A data distribution with two modes.

bivariate:  Involving two datasets. For example, a correlation procedure using two
datasets is a bivariate correlation.
Bonferroni test:  A post-hoc test used with the analysis of variance. When significant
overall differences are found, this test runs a series of modified t-tests to ascertain
which levels of the independent variable are significantly different from one another.
categorical data:  See “nominal data.”
cell:  The space formed at the intersection of a row and a column in a table. The cell
is used to record frequencies of values meeting the criteria of the cell.
central limit theorem:  Statement telling us that “If a sample size is sufficiently large,
the sampling distribution of the mean will be approximately normal.” A theorem is
nothing more than a statement of an idea that has proven to be true.
chi-square goodness-of-fit test:  A nonparametric test that compares an observed
distribution of nominal data to an expected distribution of nominal data.
chi-square test of independence:  A nonparametric test of association that deter-
mines if similar numbers of occurrences appear in groups of nominal-level data;
sometimes called a “factorial chi-square test.”
coefficient of determination:  The result of squaring Pearson’s r in a correlational
procedure. It represents the strength of a correlation directly attributable to the inde-
pendent variable.
computed range:  The range of a quantitative dataset computed by subtracting the
lowest observed value from the highest observed value.
computed value of F:  The result of an analysis of variance. This value can be com-
pared to a critical value of F to determine if a null hypothesis should be rejected.
computed value of p:  The probability that datasets being compared are significantly
different. This is done by comparing p to a predetermined alpha value.
computed value of t:  An output value from both independent- and dependent-sam-
ple t-tests. In the days prior to the use of computers, this value was compared to a table
(critical) value of t to determine if a null hypothesis should be rejected.
computed value of z:  The value of a data point on a normal curve representative of
the number of standard deviations above or below the mean.
confidence interval:  A range of sample statistics in which there is a given probability
that a population parameter exists.
consumer of statistics:  A person who uses statistics to interpret data as well as un-
derstands data that others are presenting.
contingency table:  A multidimensional table used in the chi-square tests when the
rows and columns of the table are defined by the levels of the independent variables.
 Glossary | 389

continuous data:  A generic name for interval- or ratio-level data; also called “quan-
titative data.”
correlation coefficient:  An output value from any of the correlational procedures. It
represents the strength of the relationship between two or more sets of data.
criterion variable:  The value that is being predicted in a correlation procedure.
critical value:  A table value to which one of the computed statistics (e.g., t, z, or F) is
compared in order to test a null hypothesis. There are specific tables for each of the
respective distributions.
degrees of freedom:  The number of scores that are free to vary when describing a
sample. The method by which they are calculated changes depending on the statistical
test you are using, but for most calculations, the value is defined as one less than the
number of data values.
dependent-sample t-test:  Statistical tool used when a study involves one indepen-
dent variable with two levels and one dependent variable that is measured with quan-
titative data. In this case, the levels of the independent variable must be related to or
correlated with one another.
dependent variable:  The “effect” that is being measured in a study.
descriptive statistics:  Numeric and graphical statistical tools that help us “describe”
the data so it can be better used for our decision making. Examples include the mean
of a dataset or a pie chart.
deviation from the mean:  The amount any given value in a dataset is from the mean
of the dataset.
directional hypothesis:  A hypothesis that implies a “greater than” or a “less than”
relationship between the variables being studied; also called a “one-tailed hypoth-
esis.”
disordinal interaction:  An interaction of two independent variables wherein the
values of one independent variable have a dramatically opposite effect on a dependent
variable than do values of the second independent variable.
effect size:  In parametric statistics, a measure of the degree to which an indepen-
dent variable affects a dependent variable.
empirical rule:  The rule stating that, in a normal distribution, approximately 68%
of values are within ±1 standard deviation from the mean, 95% of values are ­w ithin ±2
standard deviations of the mean, and nearly all (99.7%) of all values are within ±3
standard deviations of the mean.
equal variances assumed/not assumed:  A test used in parametric statistics to en-
sure that the variability within sets of data being compared is equitable. Significant
differences in variance call for modification of how computed values in parametric
statistics are calculated.
390 | Glossary

ethical research:  A research study in which the researcher ensures that participa-
tion by subjects is voluntary; they should not be harmed in any way—socially, physi-
cally, or mentally.
expected value:  Value used in chi-square tests as a measure of the number of occur-
rences of each cell value that the researcher believes should appear. These expected
values are compared to the actual number of occurrences in each cell.
experimental independent variable:  See “manipulated (experimental) independent
variable.”
F distribution:  The plot of F values computed from repeated samples of data.
F value:  The value computed in an analysis of variance. It is compared to a critical
value of F to interpret hypotheses.
factorial analysis of variance:  Statistical tool used when a study involves more than
one independent variable as well as one dependent variable that represents quantita-
tive data.
fail to reject the null hypothesis:  Your inability to reject the null hypothesis based
on the results of your statistical test. This means you are unable to support your re-
search hypothesis.
frequency distribution table:  A table showing the number of occurrences of the
various values in a dataset.
goodness of fit:  The degree to which an observed distribution of data values fits the
distribution that was expected.
graphical descriptive statistics:  The use of tools such as pie charts, bar charts, and
relative frequency diagrams to illustrate the characteristics of a dataset.
histogram:  A graph showing the number of occurrences of the various values in a
dataset.
hypothesis:  A statement that reflects the researcher’s beliefs about an event that has
occurred or will occur.
independent-sample t-test:  A statistical tool used when a study involves one inde-
pendent variable with two levels and one dependent variable that is measured with
quantitative data. In this case, the levels of the independent variable cannot be related
to or correlated with one another.
independent variable:  The “cause” being investigated in a study.
inferential statistics:  Statistical tools used to make decisions or draw inferences
about the data we have collected.
interaction effect:  The simultaneous effect on a dependent variable of two or more
independent variables.
intercept:  The point at which the line of best fit crosses the x- or y-axis on a scat-
terplot.
 Glossary | 391

interquartile range:  The range of values between the first and third quartiles in a
data distribution.
interval data:  One of two types of data that are called quantitative or continuous.
Interval data can theoretically fall anywhere within the range of a given dataset. The
range can be divided into equal intervals, but this does not imply that the intervals
can be directly compared. For example, a student scoring in the 80s on an examina-
tion is not twice as smart as a student scoring in the 40s. There is no absolute zero
point in an interval-level dataset.
kurtosis:  The degree to which a data distribution deviates from normal by being too
“peaked” or too “flat.”
latent independent variable:  An independent variable that is examined “as is” and
is not manipulated by the researcher. Examples include gender and ethnic group.
leptokurtosis:  A bell-shaped distribution that is too peaked (i.e., too many values
around the mean of a distribution) to be perfectly normally distributed.
levels of the independent variable:  Different values of the independent variable that
are investigated to determine if there is a differing effect on the dependent variable.
Levene’s test:  A statistical tool for testing the equality of variance in datasets being
compared.
line of best fit:  A graphical technique used in correlational procedures to show the
trend of correlations being plotted. A line of best fit that emulates a 45-degree angle
demonstrates a strong positive correlation. A line of best fit that appears flatter indi-
cates a lesser degree of correlation.
logistic regression:  A linear regression procedure used to predict the values of a
nominal dependent variable.
lower confidence limit:  The smallest value in a confidence interval.
main effect:  The effect of a single independent variable on a dependent variable.
manipulated (experimental) independent variable:  An independent variable where
the researcher defines membership into the factors or levels. An example would show
a researcher placing students in different groups to measure the effect of technology
on learning. Sometimes called “experimental independent variable.”
Mann–Whitney U test:  Nonparametric alternative to the independent-sample t-test.
mean:  A measure of central tendency reflecting the average value in a dataset. Only
used with quantitative data (interval and ratio).
mean of means:  The average score of a dataset created by computing and plotting
the mean of repeated samples of a population.
mean square:  A constant used in the calculation of an F value computed by dividing
the sum of squares value for a specific group by the degrees of freedom for the same
group.
392 | Glossary

measures of central tendency:  Descriptive statistics that help us determine the mid-
dle of a dataset. Examples include the mean, the median, and the mode.
measures of dispersion:  Descriptive statistics that help determine how spread out
a data distribution is. Examples include the range, the standard deviation, and the
variance.
measures of relative standing:  Measures used to compare data points in terms of
their relationship within a given dataset. Examples include z scores and percentiles.
median:  A measure of central tendency that describes the midpoint of data that are
ordinal, interval, or ratio level and have been sorted into ascending or descending
sequence.
mode:  A measure of central tendency reflecting the value that occurs most often in
a dataset. Can be used with all types of data.
multimodal:  A dataset having more than one mode.
multiple-comparison test:  Tests used after the analysis of variance to determine
exactly which means are significantly different from one another. An example is the
Bonferroni test.
multiple regression:  A regression procedure that uses more than one predictor vari-
able.
multivariate:  Referring to a statistical test that has more than one independent or
dependent variable.
multivariate analysis of variance (MANOVA):  Statistical tool used when a study
involves any number of independent variables and more than one dependent variable
that measures quantitative data.
negatively skewed:  Skewed to the left (i.e., there are more values below the mean of
a quantitative dataset than there are above the mean).
nominal data:  Data that are categorical in nature. Examples include gender, ethnic
group, and grade in school.
nondirectional hypothesis:  A hypothesis that implies a difference will exist between
the variables being studied but no direction is implied. Also called a “two-tailed hy-
pothesis.”
nonmanipulated (quasi-) independent variable:  An independent variable where the
levels are preexisting. An example would include a situation where the independent
variable was gender; the levels are male and female.
nonparametric statistics:  Inferential statistical tools used with nominal or ordinal
data or with quantitative data where the distribution is very abnormally distributed
(e.g., skewed or kurtotic).
normal distribution:  A quantitative distribution wherein the distribution is bell-
shaped.
 Glossary | 393

null hypothesis:  A hypothesis that states there will be no relationship between the
variables being studied. The null hypothesis is the antithesis of the research hypoth-
esis.
numeric descriptive statistics:  The use of tools such as measures of central ten-
dency, measures of dispersion, and measures of relative standing to illustrate the char-
acteristics of a dataset.
observed value:  The actual count of occurrences in a cell of nominal value. It is used
in chi-square tests to compare to expected values.
one-sample chi-square test:  The comparison of cell frequencies in a nominal distri-
bution to those that would be expected according to a previously defined criterion. See
also “chi-square goodness-of-fit test.”
one-tailed hypothesis:  See “directional hypothesis.”
one-way analysis of variance (ANOVA):  Statistical tool used when a study involves
one independent variable with three or more levels and one dependent variable that
is measured with quantitative data.
ordinal data:  Data that are rank-ordered. Examples include order of finish in a race
and class standing.
ordinal interaction:  An interaction effect where the influence of one independent
variable remains in the same direction but varies in magnitude across levels of an-
other independent variable
p value:  The probability that groups being compared came from the same popula-
tion.
paired-samples t-test:  See “dependent-sample t-test.”
parameter:  Any value known about a dataset representing an entire population.
parametric:  Pertaining to the use of quantitative data that form a mound-shaped
distribution.
parametric statistics:  Inferential statistical tools used with quantitative data.
Pearson’s r:  Output from a correlation procedure using quantitative datasets. Pear-
son’s r can range from –1.00 to +1.00 with extremely high values indicating a strong
positive correlation; extremely low values indicate a strong negative correlation. Val-
ues around zero indicate a weak correlation.
percentile:  A measure of relative standing that describes the percentage of other
values in the dataset falling below it. For example, a test score of 50 that falls into the
80th percentile means that 80% of the other scores are less than 50.
pie chart:  A circular chart divided into segments, with each representing the per-
centage of a given value in a dataset.
platykurtosis:  A bell-shaped distribution that is too flat (i.e., fewer values around the
mean of the distribution than is expected) to be perfectly normally distributed.
394 | Glossary

pooled standard deviation:  An estimate of the population standard deviation based

on a weighted average from a set of sample standard deviations.
population:  All members of a group being investigated.
population parameter:  Any value known about a population.
positively skewed:  Skewed to the right (i.e., there are more values above the mean of
a quantitative dataset than there are below the mean).
possible range:  The set of values a dataset theoretically covers. For example, on
most examinations, the possible range is 0 to 100.
post-hoc test:  See “multiple-comparison test.”
power:  The ability of a statistical test to identify when there is a true significant
difference in the values being compared. It is computed by subtracting beta from
1.00; this means that as power increases, the probability of making a Type II error
decreases.
practical significance:  The determination that the results of a research study are
practically important or that they can be important or useful in real life.
predictor variable:  The independent variable in a correlation procedure. It is used
to predict values of the criterion variable.
problem statement:  The specific issue, concern, or controversy that we want to in-
vestigate using inferential statistics.
qualitative data:  Non-numeric data such as interviews and recordings.
quantitative data:  A generic label for any data that are on an interval or ratio scale.
quartile:  The range of data representing 25% of the distribution.
range:  The set of values that a dataset actually covers. For example, if the lowest
value in our dataset is 30 and the highest value is 70, our range is 40 (highest minus
lowest).
rank data:  See “ordinal data.”
ratio data:  One of two types of data (see “interval data”) that can be classified as
quantitative. Ratio-level data can be divided into equal intervals and allow for direct
comparison. For example, a person who weighs 150 pounds is twice as heavy as a per-
son who weighs 75 pounds. There is an absolute zero point in ratio-level data.
rejecting the null hypothesis:  A decision made, based on the computed and critical
value of a statistic, that indicates that differences between values being compared are
true and not due to chance.
relative frequency:  The percentage of occurrences of various values in a dataset.
research hypothesis:  A synonym for the directional or nondirectional hypothesis
being investigated. The research hypothesis is the antithesis of the null hypothesis.
 Glossary | 395

sample:  A subset of a population being studied.

sample statistic:  Any value known about a sample of data.
sampling distribution of the means:  The resulting distribution when repeated sam-
ples are drawn from a population and their mean is calculated and plotted. Given
enough samples, this will result in a perfect normal distribution.
sampling error:  The error inherent in measuring only a selected group (sample)
from a population.
scatterplot:  A graph that shows the relationship between sets of quantitative data.
significantly different:  A difference between two variables being measured that
cannot be attributed to sampling error. The values are different due to reasons other
than chance and would cause a null hypothesis to be rejected.
simple frequency:  The actual count of the number of occurrences of a given vari-
able in a dataset.
skewed left:  See “negatively skewed.”
skewed right:  See “positively skewed.”
skewness:  Degree to which a data distribution deviates from being a normal distri-
bution (i.e., mound shaped) by having too many values in one end of the distribution
or the other.
slope:  The degree of inclination of a line of best fit.
Spearman’s rho:  A correlational procedure used with ordinal-level data. Rho can
range from zero to +1.00. The closer rho is to +1.00, the stronger the correlation.
standard deviation:  Measurement used to describe how spread out a dataset is. It
can be thought of as the average distance of any data point in a distribution from the
mean of the dataset. Obviously, the larger the standard deviation, the more spread out
the dataset is. The standard deviation is the square root of the variance.
standard error of the mean:  The standard deviation of a sampling distribution of
the means.
standard error of mean difference:  The standard deviation of the sampling distri-
bution of the mean differences used in a t-test.
stanine:  A measure of relative standing involving the division of a distribution of
quantitative scores into nine parts.
statistic:  Any value we know about a sample of data.
sum of squares:  The sum of the within-group and between-groups sum of squares;
used in the analysis of variance.
t distribution:  The resulting distribution when repeated samples are drawn from a
population and their t value is calculated and plotted. Given enough samples, this will
result in a perfect normal distribution.
396 | Glossary

T-score:  A measure of relative standing based on a scale of 20 to 80. Caution: This

is not the t value computed as part of a t-test.
t table:  A table showing the critical values of t for various alpha values and degrees
of freedom.
theoretical significance:  Research results that may be useful in testing a current
theory or the development of a new theory.
two-tailed hypothesis:  See “nondirectional hypothesis.”
Type I error:  Rejecting the null hypothesis when we should not reject it.
Type II error:  Failing to reject the null hypothesis when we should reject it.
Type I error rate:  The probability of rejecting the null hypothesis when we should
not reject it (i.e., rejecting the null hypothesis when it is true). The Type I error rate is
synonymous with our alpha value.
Type II error rate:  The probability of failing to reject the null hypothesis when we
should reject it; sometimes referred to as our “beta value.”
univariate:  Dealing with one independent and one dependent variable.
upper confidence limit:  The highest value in a confidence interval.
variance:  A measure of dispersion that is determined by squaring the standard devi-
ation. Used in computations such as the analysis of variance where negative numbers
are not allowed.
Wilcoxon t-test:  Nonparametric alternative to the dependent-sample t-test.
within-group degrees of freedom:  An integral part of the calculation of F values;
computed by subtracting one from the number of groups (i.e., levels of the indepen-
dent variable) in the analysis of variance.
within-group variance:  The amount of variance that occurs within one of the groups
of data represented by a level of an independent variable.
x-axis:  The horizontal axis of a plot.
y-axis:  The vertical axis of a plot.
z score:  A measure of relative standing computed by subtracting the mean of a bell-
shaped dataset from the observed score and dividing by the standard deviation.
z table:  A table showing the critical value of z for all values in a normal distribu-
tion.
 Answers to Quiz Time!

Chapter 1 Quiz Time!

1. This statement is not clear and concise. By reading it, it appears that the researcher
wants to investigate some relationship between absenteeism and the particular
school, but one cannot say for certain.

2. In this case, there are two issues. First, this study is not within the scope of the
teachers’ expertise; mold inspection and the like is best left to experts. Second, all
of the variables are not considered. The problem statement indicates that potential
health hazards are to be investigated but it does not mention mold and asbestos.
While they may be part of what could be considered health issues, there are quite a
few other problems that might be identified as well.

3. This appears to be a valid problem. It meets all of the criteria and is stated in an
appropriate manner.

4. This appears to be a good problem statement given that the attitudes of ranchers
could be measured by an instrument representing numeric data. All other charac-
teristics of the problem statement are met or assumed.

5. This has the potential to be a good problem statement, but there appears to be
one major flaw; what is the scope of the problem? It would be better if the author
included characteristics such as age, gender, and ethnic group that might be appli-
cable to the study.

6. The major issue with this problem statement is one of scope. While the problem
is clear, the goals of the lesson plan are clearly beyond their reach. Very few teach-
ers at that level would have access to the types of tools necessary to conduct the
research that is being suggested.

7. The problem statement seems to be very well written, but I worry about the scope.
It would be better if the scope was limited to different makes of foreign cars. For
example, the study might not be valid if the researcher were comparing large Euro-
pean luxury sedans to economical imports from Japan or Korea.

397
398  |  Answers to Quiz Time!

8. Again, there is the potential for a very good problem statement here. The author
might better state the relationship between the variables, however. Given, weight
gain is a part of prenatal lifestyle, but it is not the only component; other issues
such as exercise, alcohol consumption, and tobacco use could also contribute to the
health of newborn infants.

9. This problem statement begs to ask, “What is the problem?” The scope is not men-
tioned, it is not clear and concise, and not all of the variables seem to be included.
In addition, it’s not clear whether the researcher is comfortable with the subject
area. In short, this is not a good problem statement.

The Case of Distance Therapy

Notice, in this case, that no direction is implied. The case study says “no one really knows if
either one works better or worse than the other.” Because of that, we need to state a nondirec-
tional research hypothesis:

„ There will be a significant difference in the number of sessions for treating

depression between clients in a traditional setting and clients in a distance setting.
The null hypothesis would be:

„ There will not be a significant difference in the number of sessions for treating
depression between clients in a traditional setting and clients in a distance setting.

The Case of the New Teacher

In this case, a direction is implied. The teacher wants to determine if there are fewer parent–
teacher conferences in schools in a less affluent area than in schools from a more affluent sec-
tion of town. Her research hypothesis would be:

„ There will be significantly fewer parent–teacher conferences in nonaffluent schools

than in affluent schools.
The null hypothesis would read:

„ There will be no significant difference in the number of parent–teacher conferences

between nonaffluent schools and affluent schools.

The Case of Being Exactly Right

In this case, the engineers want to develop software to monitor the process to ensure the tem-
perature is maintained at 155°. Because of that, the null and research hypotheses will reflect a
comparison to an exact value. The engineer’s research hypothesis would be:

„ There will be a significant difference between the temperature of our process and
155°.
The null hypothesis would read:
 Answers to Quiz Time!  |  399

„ There will be no significant difference between the temperature of our process and
155°.

The Case of “Does It Really Work?”

In this case, no direction is implied; the dean only wants to know if the students taking anat-
omy and physiology online do as well as their on-campus counterparts. Knowing that, the
faculty would test the following research hypothesis:

„ There will be a significant difference in achievement between students taking

anatomy and physiology in an online environment and students taking the same
course on-campus.
The null hypothesis would be:

„ There will be no significant difference in achievement between students taking

anatomy and physiology in an online environment and students taking the same
course on-campus.

The Case of Advertising

In this case a direction is implied; the superintendent is interested in finding out if online ad-
vertising leads to a larger number of job applicants. Her research hypothesis would be:

„ There will be a significantly larger number of job applicants from Internet-based

advertisements than from newspaper advertisements.
The null hypothesis would be:

„ There will be no significant difference in the number of job applicants from

Internet-based advertisements than from newspaper advertisements.

The Case of Learning to Speak

In this case, teachers want to investigate the difference between two types of instruction: lan-
guage immersion and formal instruction in the classroom. There is no direction implied, so
our research hypothesis would read:

„ There will be a significant difference in language skills after one year of instruction
between students taught in an immersion environment and students taught in the
classroom.
The null hypothesis would read:

„ There will be no significant difference in language skills after one year of

instruction between students taught in an immersion environment and students
taught in the classroom.
400  |  Answers to Quiz Time!

The Case of Kids on Cruises

Here a direction is implied; one group feels that the number of potential customers that would
like to book a cruise with no children on board is larger than the number of customers who
think that children should be allowed on all cruises. Because of that, we need to state a direc-
tional research hypothesis:

„ The number of customers who would like to take a cruise where no children are
allowed on board is significantly larger than the number of customers who feel that
children should be allowed on all cruises.
The null hypothesis would read:

„ There is no significant difference in the number of customers who would like to take
a cruise where no children are allowed on board, and the number of customers who
feel that children should be allowed on all cruises.

Chapter 2 Quiz Time!

Independent Dependent
Hypothesis variable and levels variable Data type
1. There will be a significant difference Students: Those Motivation. Quantitative
in motivation scores between students in traditional (interval). There
in online programs and students in programs and has to be some level
traditional programs. those in online of motivation, but
programs. ratios cannot be
made.
2. There will be a significant difference Programmers: Count of Nominal.
in the number of males and females Males and each gender.
working as computer programmers in females.
corporate America.
3. There will be a significant difference Country: United Count of Nominal.
in the number of females in computer States, France, females
science classes between students in the and Russia. in each
United States, France, and Russia. country.

4. There will be a significant difference Residence: At Weight gain. Quantitative

in weight gained during their fresh- home or away (ratio). There is a
man year between students who live at from home. real zero point, and
home and students who live away from ratios can be made.
home.
5. There will be a significant difference University: Ivy First-year Quantitative
in first-year salaries between graduates League and state. salary. (ratio). There
of Ivy League schools and graduates of is a real zero
state universities. (unemployment),
and ratios can be
made.
6. Administrative assistants who work in Work location: Productivity. Quantitative (ratio)
cubicles are significantly less produc- Cubicles or if measured on a
tive than administrative assistants who enclosed office. numeric scale.
work in enclosed offices.
 Answers to Quiz Time!  |  401

7. Primary-care patients who are treated Patients: Those Health Quantitative

by an osteopathic physician will have with osteopathic problems. (ratio). There is a
significantly fewer health problems doctors and those real zero, and ratios
than primary-care patients who are with allopathic can be made.
treated by an allopathic physician. doctors.

8. Truck drivers who check their tire Truck drivers: Miles per Quantitative
pressure frequently will have signifi- Those who check gallon. (ratio). There is a
cantly higher miles per gallon than their tire pressure real zero, and ratios
truck drivers who do not check their frequently and can be made.
tire pressure frequently. those who do not.

9. Insurance companies that use Method of Number of Quantitative

computer-dialing services to call pro- solicitation: sales. (ratio). There is a
spective clients will have a significantly Computer-dialing real zero, and ratios
lower number of sales than insurance and live agent. can be made.
companies that use live agents to call
prospective clients.
10. The rankings of favorite sporting Country: Mexico, Ranking Ordinal.
activities will be significantly different the United States, of sporting
between Mexico, the United States, and Canada. activity.
and Canada.

1. The average height is 68.07 inches.

2. The median class rank is 8; remember, this tells us nothing really. We need to see
what grade or other measurement that relates to.

3. The mode of shoe size is 8.

4. There is no average class; it is a nominal value.
5. The median height is 69 inches.
6. The mode of height is 69.
7. The median shoe size is 9.
8. The mode class rank is 6; it occurs twice. Remember, we need to see what grade or
other measurement that relates to.

9. Class is multimodal; there are four each of freshman, sophomore, and senior.
10. Class is a nominal value; the median should not be computed.
11. Class rank is ordinal; the average should not be computed.
12. The average shoe size is 9.6.

Chapter 3 Quiz Time!

1. The possible range for company A is 47; the possible range for company B is 47; the
possible range for company C is 47.

2. The possible range for all companies taken together is 47.

402  |  Answers to Quiz Time!

3. The computed range of company A is 43; the computed range of company B is 25;
and the computed range of company C is 40.

4. The range of all companies combined is 45.

5. For company A, the data are spread out since the computed range of 43 is 91.4% of
the possible range. For company B, the data are less spread out since the computed
range of 25 is only 53.2% of the possible range. For company C, the data are again
very spread out since the computed range of 40 is 85.1% of the possible range. The
computed overall range of 45 represents 95.7% of the possible range.

6. The standard deviation of company A is 14.89, the standard deviation for company
B is 7.56, and the standard deviation for company C is 11.34.

7. For company A, on average, each value in the dataset is approximately 14.89

standard deviations away from the mean. For company B, on average, each value
in the dataset is approximately 7.56 standard deviations away from the mean. For
company C, on average, each value in the dataset is approximately 11.34 standard
deviations away from the mean.

8. The standard deviation for age across all companies is 12.48.

9. For company A, the variance is 221.82, the variance for company B is 57.12, and the
variance for company C is 128.67.

10. The variance of all companies combined is 155.82.

11. Since the variance is not in the scale of the original data, the variance is not very
meaningful. A large value would indicate a broader range of data; smaller variance
values would indicate a smaller range.

1.
TABLE 3.27.  Data for Computation of z Scores, T-Scores, Stanines, and
Ranked Mean Scores
Observed Standard Ranked
Mean value deviation z score T-score Stanine mean score
 30 33 2.00 1.50 65 8 100
 48 52 5.00 .80 58 7  81
 55 54 3.00 –.33 47 5  71
 71 77 7.00 –.86 59 7  61
 14 8 2.70 –2.22 28 1  55
 23 35 5.00 2.40 74 9  48
 61 48.6 2.90 –4.28  7 1  47
100 114 6.33 2.21 72 9  30
 81 78.5 1.55 –1.61 34 2  23
 47 60.00 12.0 1.08 61 8  14

25th percentile = 30
50th percentile = 51.5
75th percentile = 71
 Answers to Quiz Time!  |  403

2. The z score for a score of 130 with a mean of 100 and a standard deviation of 15 is
2.00.

3. In this case, we would first have to determine the standard deviation. Since we
know that it is the square root of the variance, it is easy to determine. The square
root of $9,000 is $94.87. Since we are interested in a point 2 standard deviations
below the mean, we then multiply that by 2: $94.87 * 2 = $189.74. In order to deter-
mine our cutoff point, we subtract that from $22,000 and get $21,810.26. Since the
family we are interested in has an income of $19,500, they would qualify for govern-
mental assistance.

4. We compute the z score by first subtracting the sample mean from the observed
value we are investigating; in this case 14 – 10 = 4; we then divide that by the sample
standard deviation (i.e., 3), giving us a z score of 1.33. For an observed value of 5,
the z score is –1.67 and for the observed value of 20, the z score is 3.33.

Chapter 4 Quiz Time!

Since we said that we always use a computer to help us when we’re graphically describing our
dependent variable, we are in somewhat of a jam. Knowing we can’t do that, let’s use the fol-
lowing questions to understand conceptually where we are.

1. The easiest way to show the frequency of the values would be with a bar chart. Each
bar would represent the actual number of each value in the dataset.

2. In this case there would seem to be no consistent relationship between the data val-
ues (i.e., the line of best fit would tend to be flat). This is based on the observation
that, as some of the reading scores go up, their corresponding math scores do the
same. In other cases, the opposite is true; as some of the reading scores go up, their
corresponding math score goes down.

3. In this case, it appears to be a positive relationship. As each value of education goes

up, so does the salary. The line of best fit would be sloped up from the left to the
right.

4. The histogram is used with quantitative data in order to allow for fractional values;
given this, nominal data are not plotted using a histogram. A histogram is mound-
shaped if the mean, median, and mode are approximately equal and has a sym-
metrical distribution of values on either side of the center. If there are more values
in either end of the distribution than would generally be found in a bell-shaped dis-
tribution, we say the distribution is skewed. If the mean is greater than the median,
the distribution is positively skewed; if the mean is less than the median, negative
skewness occurs. If there are more data values than expected in the middle of the
dataset, we say the distribution is leptokurtotic and the plotted distribution would
be more peaked than normal. Fewer values in the center cause the distribution to
“flatten out”; we call this distribution platykurtotic.

5. A perfect normal distribution has an equal mean, median, and mode with data
values distributed symmetrically around the center of the dataset.
404  |  Answers to Quiz Time!

6. Since the mean is greater than the median, the distribution would be positively
skewed.

7. The distribution will be leptokurtotic; it will appear more peaked than a normal
distribution.

8. No, there are many reasons a dataset with an equal mean and median does not
have to be normally distributed. One primary reason would be the possibility of
too few data values to give the symmetrical, mound-shaped appearance of a normal
distribution.

9. No, most parametric inferential statistical tests are powerful enough to work with
minor problems with skewness or kurtosis.

10. If the mean is less than the median, negative skewness occurs. Statistical software
can be used to determine if this is problematic by returning a skewness coefficient.
This value is only problematic if it is less than –2.00. If the mean is greater than the
median, positive skewness occurs. Again, this is detrimental to decision making
only if the software returns a value greater than +2.00.

11.
TABLE 4.10.  Values for Quiz Time Question 11
Area under the Area under the
z value 1 curve for z 1 z value 2 curve for z 2 Difference
 3.01 49.87%  2.00 47.72% 2.15%
–2.50 49.38%  3.00 49.87% 99.25%
–1.99 47.67% –0.09 3.59% 44.08%
 1.50 43.32% –1.50 43.32% 86.64%

Chapter 5 – In-Chapter Exercises

How Are We Doing So Far?

1. z score:
a. 1.347219
b. –2.57196
c. 1.837116
d. 0.367423
e. –0.36742
f. –1.22474
g. –0.24495
h. –0.9798
 Answers to Quiz Time!  |  405

2.
TABLE 5.6.  Computing the Distance between z Scores
Area under the Area under the Area between
Value 1 z1 curve for z 1 Value 2 z2 curve for z 2 z 1 and z 2
90 –1 .3413 95 1.5 .4332 .7745
89 –1.5 .4332 90 –1 .3413 0.0919
91.5 –.25 .0987 93.5 .75 .2734 .3721
92 0 0 96 2 .4772 .4772

Do We Understand Everything We Need to Know about Confidence Intervals?

TABLE 5.13.  Learning to Compute the Width and Limits of a Confidence Interval
Lower limit Upper limit Width
x Alpha n s of CI of CI of CI
100 .10 25  5 98.36 101.65 3.29
500 .05 50 25 493.07 506.93 13.86
 20 .01 15  3 18.01 21.99 3.99
 55 .05 20  7 51.93 58.07 6.14
 70 .01 22  6 66.71 73.29 6.59
220 .10 40 10 217.40 222.60 5.20

Great News—We Will Always Use Software to Compute Our p Value

TABLE 5.19.  Intelligence—Is There a Significant Difference?

Null hypothesis: There will be no significant difference in the intelligence
level between our class and the overall university level.
Research hypothesis: The average intelligence level of our class will be
significantly higher than the average intelligence level of
all other classes in the university.
Population average: 100
Class Average: 105
SEM: 3
Alpha: .05
Computed z value: 1.67
Critical z value: 1.645
Reject null hypothesis? Yes
How do you know? The computed value of z is greater than the critical value
of z.
Is p less than .05? Yes
How do you know? Because the null is rejected, p must be less than alpha.
406  |  Answers to Quiz Time!

TABLE 5.20.  Rainfall—Is There a Significant Difference?

Null hypothesis: There will be no significant difference in the yearly
amount of rainfall between Texas and Florida.
Research hypothesis: There will be a significant difference in the yearly
amount of rainfall between Texas and Florida.
Texas average: 70 inches
Florida average: 62 inches
SEM: 10 inches
Alpha: .05
Computed z value: –.8
Critical z value: 1.96
Reject null hypothesis? No
How do you know? The computed value of z is not less than the critical
value of z.
Is p less than .05? Yes
How do you know? Because the null is not rejected, p must be equal to or
greater than alpha.

TABLE 5.21.  Truancy—Is There a Significant Difference?

Null hypothesis: There will be no significant difference in the number
of truancy cases at our school and the district average
number of truancy cases.
Research hypothesis: There will be a significant difference in the number
of truancy cases in our school and the district average
number of truancy cases.
District average: 12 cases
Our school’s average: 17 cases
SEM: 2 cases
Alpha: .05
Computed z value: 2.5
Critical z value: 1.96
Reject null hypothesis? Yes
How do you know? The computed value of z is greater than the critical z
value.
Is p less than .05? Yes
How do you know? Because the null is rejected, p must be less than alpha.

Chapter 5 Quiz Time! Answers

1. There is one independent variable, fish, with three levels—lakes, streams, and rivers.
The number of fish caught is the dependent variable. Because you are collecting
quantitative data, you would use a one-way ANOVA.

2. There is one independent variable with two levels—young females before the science
 Answers to Quiz Time!  |  407

program and the same young females after the science program. Their interest level
is the dependent variable. If, for example, we measured interest on a 1 to 5 scale the
data would be quantitative. Because of this, we would use a dependent sample t-test
to test the hypothesis.

3. There is one independent variable, office, with two levels—well-lighted and dimly
lighted. The level of productivity is the dependent variable. Because you are collect-
ing quantitative data, you would use an independent-sample t-test.

4. Here there is one independent variable—country, with three levels—Canada,

Mexico, and the United States. The number of health problems is the dependent
variable; this would call for the use of a one-way ANOVA.

5. In this case, you have one independent variable with two levels—persons raised in
an urban setting and persons raised in a rural setting. There are two dependent
variables, height and weight. Because of this, a MANOVA would be used to investi-
gate this hypothesis.

6. In this case, you have an independent variable with one level—foreign-manufactured

cars. Since you want to compare miles per gallon, the dependent variable, to a
known value of 19, you would use a one-sample t-test.

7. The number of children is the dependent variable; family is the independent vari-
able and has three levels—lower, middle, and high socioeconomic status. In this
case, you would use a one-way ANOVA.

8. The location of the show is the independent variable; the two levels are “on Broad-
way” and “touring.” The degree of audience appreciation is a quantitative depen-
dent variable; in this case, you would use an independent-sample t-test.

9. In this case, you have a predictor variable, number of siblings, and a criterion vari-
able, annual income. Because of this, you would use the Pearson correlation.

10. Here there is one independent variable, political party, with three levels—Repub-
licans, Democrats, and Liberals. Since we have one dependent variable that is
numeric, we would use a one-way ANOVA.

11. Here we have one independent variable, gender, with two levels—males and females.
We are comparing the number of each in our office (i.e., the observed value) to the
percentage of each nationwide (i.e., the expected value). Because of this, we would
use a one-way chi-square test.

12. Herein, we’re comparing the degree of procrastination of one group of authors, be-
fore and after a call from their editor. Because it’s the same group being compared
at two different times, we would use a dependent-sample t-test.

Chapter 6 Quiz Time!

1. In this case, there is a one-tailed “greater than” hypothesis stating that children
without fluoride will have a significantly higher number of cavities than the na-
tional average of 2.0. The mean number of cavities for our sample group is 2.33
resulting in a mean difference of .333. While this is greater, given the p value of
408  |  Answers to Quiz Time!

.215 (remember, you must divide the two-tailed p value of .430 by 2 when you have
a directional hypothesis), the difference is not significant. We will fail to reject the
null hypothesis and will not support the research hypothesis.

2. Here we have a one-tailed “less than” hypothesis stating that Ivy League-trained
physicians will have significantly fewer malpractice lawsuits than the national mean
of 11. In this case, the Ivy League physicians have an average of 5.20, giving a mean
difference of –5.8. Our p value of .000 indicates this difference is significantly
lower. We will reject the null hypothesis and support the research hypothesis.

3. In this case, we have a one-tailed “greater than” hypothesis in that we believe the
mean graduation rate of students who went to community college for their first two
years will be greater than the national average of 82%. The graduation rate of our
sample is 79.6% and results in a p value of .0215; because of this, we will reject the
null hypothesis. However, since the mean difference is opposite of what we hypoth-
esized, we will not support the research hypothesis.

4. In this case, we’ve hypothesized that adult turkeys in the wild will weigh significant-
ly less than 12.0 pounds. Given a mean difference of –1.467 and a p value of .008,
we reject the null hypothesis and support the research hypothesis.

5. We can see that the time to restore power to lower socioeconomic sections of town
is 5.15 hours, the average time in more affluent neighborhoods is 4.0 hours. Our
p value (.009) shows that we reject our null hypothesis and support our research
hypothesis.

6. Here we tested the number of holes-in-one in the current season (i.e., 25) against
those from past years. We can see there is hardly any difference, 25 vs. 24.85. This
results in a very large p value of .901 meaning we would fail to support our research
hypothesis and fail to support the research hypothesis.

Chapter 7 Quiz Time!

The Case of the Homesick Blues

1. The researchers do not state that they expect a difference between the groups, so
the research hypothesis must be nondirectional:

„ There will be a significant difference in the number of phone calls made

during basic training between males and females.
The null hypothesis states:

„ There will be no significant difference in the number of phone calls made

between males and females in basic training.
2. The independent variable is gender; the two levels are male and female.
3. The dependent variable is “number of calls.”
4. The mean value of male soldiers is much larger than that of the female soldiers.
 Answers to Quiz Time!  |  409

The p value for the Levene statistic is less than .05, so we have to use the “Equal
variances not assumed” column. Since we have a nondirectional hypothesis, we
have to divide the p value of .002 by two. In this case, the very low p value of .001
indicates there is a significant difference; male trainees call home a significantly
greater number of times than female trainees.

The Case of the Cold Call

1. The management wants to see if working in cubicles leads to a higher number of

phone calls, so they must state a “greater than” directional research hypothesis:

„ Employees in private offices will make significantly more phone calls than
employees working in cubicles.
The null hypothesis would read:

„ There will be no significant difference in the number of phone calls made

between employees working in cubicles and employees working in private
offices.
2. The independent variable is “location,” and the two levels are “cubicle” and “office.”
3. The dependent variable is “number of calls.”
4. In this case, the mean number of calls for employees in cubicles appears to be
much greater than for those working in offices. Because of the significant Levene
value, the “Equal variances not assumed” column must be used. The p value of .000
indicates that the null hypothesis must be rejected. Remember, check the means. In
this case, contrary to the hypothesis, employees in cubicles, despite the noise level,
make significantly more phone calls than their peers in private offices.

The Case of the Prima Donnas

1. The owners are investigating the female actresses’ statement, so the research hy-
pothesis must be a “greater than” directional research hypothesis:

„ Females, when placed on the marquee first, will lead to significantly higher
ticket sales than when males are placed first.
The null hypothesis would be:

„ There will be no significant difference in ticket sales when females are first on
the marquee or when males are first on the marquee.
2. The independent variable is “Name,” and the two levels are “Female” and “Male.”
3. The dependent variable is “ticket sales.”
4. Here the females do have a higher number of ticket sales, but we need to determine
if the difference is significant. Using the “Equal variances assumed” column, the
low p value (i.e., 439) shows that the differences between male and female actors are
different but not significantly so.
410  |  Answers to Quiz Time!

The Case of the Wrong Side of the Road

1. In this case, I am hypothesizing that the gas stations on the way out of town have
prices that are significantly higher than those coming into town. Because of that, I
state a one-tailed directional research hypothesis:

„ Gas stations going out of town have significantly higher prices than gas
stations going into town.
The null hypothesis would read:

„ There will be no significant difference in gas prices between stations going into
town and those coming out of town.
2. The independent variable is “Gas Station,” and the levels are “Into town” and “Out
of town.”

3. The dependent variable is “Price.”

4. Here we have mean values that are very close; despite that, we must test them for
significance. Here we can use the “Equal variances assumed” column and see that
we have a very large p value, .721. Given that, we cannot reject the null hypothesis;
even though the prices are different, they are not significantly different.

The Case of Workplace Satisfaction

1. Here the manager is testing the research hypothesis:

„ Female employees have significantly lower workplace satisfaction than their
male coworkers.
The null hypothesis is:

„ There will be no significant difference in workplace satisfaction between

female employees and male employees.
2. The independent variable is gender; the two levels are females and males.
3. The dependent variable is workplace satisfaction.
4. The mean score for workplace satisfaction is 5.40 for females, and 4.60 for males.
Right away, since the females’ mean score is greater than the male score, we know
that we will not support the research hypothesis. In either case, the p value of .3605
(i.e., .721/2) indicates that we would not reject the null hypothesis. Bonus question:
What did she learn from the overall descriptive statistics? Easy, given a range from
zero to 20 on the survey and based on the results, neither males nor females are
very satisfied!

The Case of the Flower Show

1. The gardener is testing the research hypothesis:

 Answers to Quiz Time!  |  411

„ There will be a significant difference in beauty between roses grown using

commercial fertilizer and roses grown with fertilizer coming from the earth.
The null hypothesis is:

„ There will be no significant difference in beauty between roses grown using

commercial fertilizer and roses grown with fertilizer coming from the earth.
2. The independent variable is roses, the two levels are those grown with commercial
fertilizer and those grown with organic fertilizer.

3. The dependent variable is perceived beauty as judged by a group of florists.

4. The mean score for flowers grown with natural fertilizer (i.e., 3.2), is slightly larger
than the mean score for roses grown with commercial fertilizer. The p value of .530
indicates that this difference is not significant.

Chapter 8 Quiz Time!

The Case of Technology and Achievement

1. In this case, the proponents of technology believe that computers are beneficial to
students. Because of that, they want to test a directional “greater than” research
hypothesis:

„ Students will have significantly higher grades after the use of technology in the
classroom than before they used technology in the classroom.
The null hypothesis will be:

„ There will be no significant difference in grades before the technology was

introduced and after the technology was introduced.
2. The independent variable is “Instructional Method,” and the two levels are “Before
technology” and “After technology.”

3. The dependent variable is Student Grades.

4. Here the mean scores are very close; the grades actually went down slightly over the
term. The p value of .434 shows that there is no significant difference.

The Case of Worrying about Our Neighbors

1. Here the citizens are concerned that their property values will go down. Because of
that, we must state a directional “less than” research hypothesis:

„ Values after annexation will be significantly lower than values before

annexation.
The null hypothesis would read:
412  |  Answers to Quiz Time!

„ There will be no significant difference in property values prior to the

annexation and after the annexation.
2. The independent variable is “Annexation”; the levels are “Before” and “After.”
3. The dependent variable is Property Value.
4. We can see that the citizens may have reason for concern; the property values
dropped nearly $10,000. The p value of .024 shows the difference to be significant.

The Case of SPAM

1. The entrepreneur is attempting to sell his service by advertising fewer junk e-mails;
he is stating a one-tailed “less than” research hypothesis:

„ There will be significantly fewer junk e-mails after customers start using his
service than prior to using his service.
The null hypothesis would be:

„ There is no significant difference in the amount of SPAM between the

entrepreneur’s service and that of other services.
2. The independent variable is “Internet Provider,” and there are two levels, “New
Service” and “Old Service.”

3. The dependent variable is the amount of SPAM received.

4. It looks as if the customers might have a valid complaint; they are receiving nearly
20 more junk e-mails per month with the new service than with the older service.
The p value of .000 indicates this difference is very significant.

The Case of “We Can’t Get No Satisfaction”

1. While the members of the management team were happy when the new President
was hired, apparently their satisfaction went down over time. That means they need
to state a directional “less than” research hypothesis to determine if the drop was
statistically significant:

„ Management satisfaction after the President has been employed for 3 months
is significantly lower than when the President was hired.
The null hypothesis is:

„ There is no significant difference in management satisfaction when the new

President was hired and after he had been there 3 months.
2. We could call the independent variable “President” with two levels, “At Hiring” and
“After Three Months.”

3. The dependent variable is Management Satisfaction.

4. It’s apparent that there is a large drop in satisfaction; the average of 60.13 when
 Answers to Quiz Time!  |  413

the new President was hired is dramatically higher than the mean of 42.27 three
months later. We can see from the p value of .000 that this is a very significant dif-
ference.

The Case of “Winning at the Lottery”

1. I’m interested in looking at the amount of money spent by people who play a weekly
lottery. I feel that using a video to show these players that the odds of winning the
lottery are probably far less than they imagined. To investigate this, I will state a
directional “less than” research hypothesis:

„ Money spent playing the lottery will be significantly less after watching the
video than was spent prior to watching the video.
The null hypothesis is:

„ There is no significant difference in the amount of money spent prior to

watching the video and after watching the video.
2. We could call the independent variable “Lottery Player” with two levels, “Before the
Video” and “After the Video.”

3. The dependent variable is amount of money spent playing the lottery.

4. The average money spent dropped from $15.80 to $12.00. Our p value of .010 shows
a significant drop in money spent.

Chapter 9 Quiz Time!

The Case of Degree Completion

1. The university is interested in determining whether age is a predictor of the time it

takes to complete a doctoral degree; our research hypothesis is:

„ There will be a significant difference in the time it takes to complete a doctoral

degree between various age groups.
The null hypothesis would be:

„ There is no significant difference in the time it takes to complete a doctoral

degree between various age groups.
2. The independent variable is “Age Group.”
3. The four levels are “20–29,” “30–39,” “40–49,” and “50 and older.”
4. The dependent variable is the number of years it takes to complete the degree.
5. The descriptive statistics show that the average number of years for degree comple-
tion are relatively close, ranging from 3.6 for students 40–49 to 4.8 for those 50
or older. The Levene test shows no problems with the homogeneity of variance,
and the p value of .162 indicates this difference in mean scores is not significant.
414  |  Answers to Quiz Time!

Because of that, the null hypothesis cannot be rejected, and the research hypothesis
cannot be supported.

The Case of Seasonal Depression

1. The research hypothesis is very straightforward:

„ There will be a significant difference in the number of anxiety attacks during
the four seasons of the year.
This means our null hypothesis is:

„ There will be no significant difference in the number of anxiety attacks during

the four seasons of the year.
2. The independent variable is “Season of the Year.”
3. There are four levels: Winter, Spring, Summer, and Fall.
4. The dependent variable is the number of anxiety attacks.
5. The descriptive statistics show a much larger number of anxiety attacks in the win-
ter than in the other seasons and there is no problem with the Levene test. The p
value indicates that a significant difference does exist, and the Bonferroni post-hoc
test supports our observation that the number of anxiety attacks in the winter is
significantly higher than in the other three seasons.

The Case of Driving Away

1. Since the teachers are interested in determining if there is a difference in the num-
ber of absences between the three groups, the null hypothesis would read:

„ There will be no significant difference in the number of absences between teens

with cars, teens with access to cars, and teens without a car.
The research hypothesis would state:

„ There will be a significant difference in the number of absences between teens

with cars, teens with access to cars, and teens without a car.
2. The independent variable could be something like “Transportation Status.”
3. The levels are “owns car,” “has access to car,” and “doesn’t have access to a car.”
4. The dependent variable is “Number of Absences.”
5. According to the Levene statistics, there is no problem with the variance between
the groups. It does appear, however, that the average number of absences varies
from group to group. The low p value indicates there is a significant difference, and
the post-hoc tests show that each group is significantly different from the other. In-
terestingly, the students who own their cars have the fewest absences; those with no
access have the greatest number of absences. This makes sense—maybe they can’t
get to school!
 Answers to Quiz Time!  |  415

The Case of Climbing

1. There are three null hypotheses:

a. The interaction null hypothesis would read:
„ There will be no significant difference in altitude climbed and the
interaction between the route climbers take and whether or not they used a
professional guide.
b. The first main effect null hypothesis would read:
„ There will be no significant difference in the altitude climbed between
climbers who used a professional guide and those who did not.
c. The second main effect hypothesis would read:
„ There will be no significant difference in altitude climbed between the
different routes climbers used.
2. There are three research hypotheses:
a. The interaction research hypothesis would read:
„ There will be a significant difference in altitude climbed and the
interaction between the route climbers take and whether or not they used a
professional guide.
b. The first main effect research hypothesis would read:
„ There will be a significant difference in the altitude climbed between
climbers who used a professional guide and those who did not.
c. The second main research hypothesis would read:
„ There will be a significant difference in altitude climbed between the
different routes climbers used.
3. For the independent variables we could label them “Guide Service Used” and
“Route.”

4. “Guide Service Used” has two levels: “yes” and “no.” Route has four levels: “Ingra-
ham Direct,” “Gibraltar Rock,” “Disappointment Cleaver,” and “Nisqually Ice Cliff.”

5. The dependent variable is the altitude the climbers reached.

6. There seem to be huge differences in some of the variables; we can see that climb-
ers who went up the Nisqually Ice Cliff with a guide averaged over 11,000 feet, while
those without a guide only managed about 8,000 feet. Before we make decisions,
though, let’s look further at our output. We can see the p value for the interac-
tion between whether or not a guide was used and the route taken is insignificant
(p = .356). Given that, we have to look at the main-effect hypotheses. We can see
that the altitude differences between climbers who used a guide and those who
didn’t is insignificant (p = .299). There is also no significant difference in altitude
416  |  Answers to Quiz Time!

gained between the four routes (p = .754). In short, climbers can sit around the fire
and debate all night, but it really doesn’t make a difference.

The Case of Employee Productivity

Throughout the United States, the number of companies hiring employees to work from home
is growing. Many administrators who support these arrangements believe this will lead to high-
er productivity than that of employees working in a traditional office environment. In order to
investigate this, administrators from one company collected data from 60 employees working
as telephone technical support personnel—30 working from home and 30 from a traditional
workplace. The data collected measured employee satisfaction and the average number of cus-
tomers an employee was able to work with over time.

1. The interaction hypothesis is:

„ There will be no significant difference in employee satisfaction and the average
number of customers worked with, when considered concurrently, between
employees who work at home, and those working in the traditional office
environment.
2. What are the two main-effect research hypotheses the managers are investigating?
„ There will be no significant difference in employee satisfaction between
employees who work at home, and those working in the traditional office
environment.

„ There will be no significant difference in the average number of customers

worked with between employees who work at home, and those working in the
traditional office environment.
3. The independent variable is work location.
4. The levels of the independent variable are employees who work at home, and those
who work in a traditional office environment.

5. The dependent variables are employee satisfaction and the average number
of customers worked with.
6. The difference in mean scores in satisfaction and number of customers between
the two groups is small. Box’s M is not significant, so we can examine the results
of the Multivariate tests. Therein you can see that Pillai’s Trace is not significant
(i.e., p = .784.). This means there is no significant interaction effect, therefore we do
not need to evaluate the Between-Subjects Effects tests.
 Answers to Quiz Time!  |  417

Chapter 10 Quiz Time!

The Case of Prerequisites and Performance

1. Since the professor is interested in determining the relationship between major and
grade, he would be investigating this null hypothesis:

„ College major has no significant effect on grade earned in a graduate

psychology class.
The research hypothesis would read:

„ College major has a significant effect on the grade earned in a graduate

psychology class.
2. The independent variables are grades, “A” through “F,” and whether or not the
student’s undergraduate degree was in psychology, “Yes” or “No.”

3. The dependent variable is the frequency of each grade.

4. Since there are two independent variables and the professor wants to see if an inter-
action exists between the undergraduate degree and letter grade, he would use the
chi-square test of independence.

5. Because of the p value of .027, he would reject the null hypothesis. While there
isn’t a post-hoc test, it appears that those students with undergraduate degrees in
psychology are far more likely to make an A grade than their classmates with other
majors.

The Case of Getting What You Asked For

1. The null hypothesis is:

„ There will be no significant difference in the actual number of each grade
assigned and the expected number of each grade.
The research hypothesis is:

„ There will be a significant difference in the number of each grade assigned

and the expected number of each grade.
2. The independent variable is grade, and there are three levels, “A,” “B,” and “C.”
3. The dependent variable is the frequency of each grade.
4. Because the professor is comparing the observed number of each grade to the
expected number of each grade, he will use the chi-square goodness of fit.

5. While there are fewer “A” grades and more “C” grades than expected, the p value of
.07 shows there is not a significant difference in the number of each grade that was
expected and the number of each grade that was actually received.
418  |  Answers to Quiz Time!

The Case of Money Meaning Nothing

1. In this case, we are interested in looking at the relationship between drug use and
socioeconomic status. Our null hypothesis would be:

„ Socioeconomic status has no significant effect on whether or not a person uses

drugs.
Our research hypothesis would be:

„ Socioeconomic status has a significant effect on whether or not a person uses

drugs.
2. The first independent variable is “Socioeconomic” status, and there are five ­levels:
“less than $20,000,” “$20,000 to $39,999,” “$40,000 to $59,999,” “$60,000 to
$79,999,” and “$80,000 and above.” The second independent variable is “Drug
Use,” and the two levels are “Yes” and “No.”

3. The dependent variable is the number of each occurrence within the cells (e.g., the
number of drug users who make below $20,000).

4. Since there are two independent variables and we want to see if an interaction ex-
ists between socioeconomic status and drug use, we would use the chi-square test of
independence.

5. In this case, we would fail to reject the null hypothesis given the p value of .726.
Simply put, it doesn’t make any difference how much money someone makes when
it comes to whether or not they use drugs.

The Case of Equal Opportunity

1. The null hypothesis is:

„ There will be no significant difference in the actual number of each gender
and the expected number of each gender.
The research hypothesis would state that a significant difference exists.

2. The independent variable is “Gender,” and the two levels are “Male” and “Female.”
3. The dependent variable is the actual count of males and females.
4. Because the manager is comparing the number of each gender to the expected
number of each gender, he will use the chi-square goodness of fit.

5. The p value of .421 indicates that the number of males and females in the company
is not significantly different from the national average.
 Answers to Quiz Time!  |  419

Chapter 11 Quiz Time!

The Case of “Like Father, Like Son”

1. Pearson’s r says there is a strong relationship (i.e., r = .790) between a father’s age at
death and the son’s age at death.

2. The coefficient of determination is .6241. This tells us that about 62.41% of change
in the criterion variable (i.e., the son’s age) is due to the predictor variable (i.e., the
father’s age).

3. Given a father’s age at death of 65, the son would be expected to live about 73 years.
4. The line-of-best-fit flow would move from the bottom left to the upper right at
slightly less than a 45-degree angle.

The Case of “Can’t We All Just Get Along?”

1. We would use Spearman’s rho because we are dealing with rank (i.e., ordinal) data.
2. The rho value of –.515 means there is a moderate negative relationship between the
Democrats’ rankings and the Republicans’ rankings.

3. I would tell the president that I was surprised the correlation wasn’t closer to –1.00!
After all, aren’t these two parties supposed to vote exactly the opposite of one an-
other?

The Case of More Is Better

Following is the output of a regression procedure using data showing highest level of education
attained and average salary in the workplace:

1. I would tell a group of potential high school dropouts that there is a fairly strong
relationship (i.e., r = .437) between the number of years they go to school and how
much money they can expect to make.

2. I would tell a student that the average income is nearly $15,000 annually, while a
ninth-grade dropout can only expect to make about $11,600.

3. The coefficient of determination is .1909. This tells us that about 19.09% of change
in the criterion variable (i.e., salary) is due to the predictor variable (i.e., the highest
grade completed). While the number of years of education is a predictor, over 80%
of the time, other factors are involved.

4. The line of best fit would rise, from left to right, at about a 25- to 30-degree angle.

The Case of More Is Better Still

Suppose you followed up the prior case by telling the potential dropouts that students who
drop out of school earlier tend to get married earlier and have more children.

1. In this case, Pearson’s r of –.259 shows a small negative correlation; people with
more education tend to have a slightly smaller number of children.
420  |  Answers to Quiz Time!

2. The coefficient of determination is .0670. This tells us that about 6.7% of change in
the criterion variable (i.e., number of children) is due to the predictor variable (i.e.,
the highest grade completed). While the number of years of education is a slight
predictor, other factors are much more likely to determine the number of children.

3. I would tell the potential dropouts that students who drop out after the eighth
grade have, on average, more children than the overall average.

4. The line of best fit would drop slightly from the left upper corner down to the right.
5. I would point out to the students considering leaving school that the combination
of lower wages and more children makes for a difficult life. Stay in school!
 Index

Note: bold entries indicate definitions in the Glossary.

A priori, 126, 387 normal distribution and, 101

Alpha value. See also Level of significance; Type overview, 245–248, 247t, 295
I error quiz regarding, 296–303, 413–416
analysis of variance (ANOVA) and, 262 six-step model and, 259–275
changing, 135–136, 146–147 types of data and, 42
definition, 387 Area under the curve, 111, 123, 124, 143, 159,
independent-sample t-test and, 190–191, 196, 201, 257–258, 387
201–202 Area under the normal curve table. See also
overview, 125–127 Critical values tables
predicting a population parameter based analysis of variance (ANOVA) and, 257–258
on a sample statistic using confidence complete, 375
intervals and, 129–131, 132–136 definition, 387
probability (p) values and, 149–150 empirical rule and, 109–112
sampling distribution of mean differences predicting a population parameter based
and, 185 on a sample statistic using confidence
three or more comparisons and, 245 intervals and, 131
Alternative hypothesis. See Hypothesis; sampling distribution of the means and, 123
Research hypothesis testing hypotheses about a population
Analysis of covariance (ANCOVA), 154, 387 parameter based on a sample statistic
Analysis of data, 2–3, 11, 20. See also Statistical and, 143
software packages Assumptions, 248–249, 288, 387
Analysis of variance (ANOVA). See also Factori- Average. See Mean
al ANOVA; Multivariate ANOVA (MANO-
VA); Parametric statistics; Variance
assumptions of, 248–249 B
calculating, 249–259
choosing the right statistical test and, 153 Bar charts, 90–91, 92, 387. See also Graphical
definition, 387 descriptive statistics
independent-sample t-test and, 201 Beta value, 128, 387. See also Type II error

421
422 | Index

Between sum of squares, 251–252, 253 Prerequisites and Performance, 331, 417
Between-group degrees of freedom, 249, Prima Donnas, 212–213, 409
257–258 Quality Time, 268–271
Between-groups variance, 249, 387 Regional Discrepancies, 271–275
Bimodal dataset, 51, 388. See also Mode Report Cards, 202–206
Bivariate, 339, 388 Seasonal Depression, 297–298, 414
Bonferroni procedure Seniors Skipping School, 265–268
analysis of variance (ANOVA) and, 263–264, Slow Response Time, 168–171
267, 268, 271, 275 SPAM, 240–241, 412
definition, 388 Stopping Sneezing, 171–173
overview, 3 Technology and Achievement, 238–239,
411
Type of Instruction and Learning Style,
C 327–330
Unexcused Students, 232–234
Cases “We Can’t Get No Satisfaction,” 241–242,
Absent Students, 346–349 412–413
Advertising, 31–32, 399 “Winning at the Lottery,” 242–243, 413
Affecting Ability, 275–284 Workplace Satisfaction, 214–215, 410
Age and Driving, 367–360 Worrying about Our Neighbors, 239–240,
Anxious Athletes, 206–209 411–412
Balancing Time, 292–295 Wrong Side of the Road, 213–214, 410
Being Exactly Right, 31, 398–399 Categorical data. See Nominal data
Belligerent Bus Drivers, 313–315 Causal relationships, compared to correlations,
“Can’t We All Just Get Along?” 372, 419 341–342
Case Against Sleep, 350–353 Cell, 388
Cavernous Lab, 197–202 Central limit theorem
Climbing, 300–301, 415–416 definition, 388
Coach, 284–287 estimating population parameters using con-
Cold Call, 211–212, 409 fidence intervals and, 125–128
Corporal Punishment, 324–326 overview, 116–117
Degree Completion, 296–297, 413–414 sampling distribution of mean differences
Different Tastes, 355–357 and, 184
Distance Therapy, 30, 398 sampling distribution of the means and,
“Does It Really Work?” 31, 399 117–124
Driving Away, 299–300, 414 Central tendency measures. See Measures of
Employee Productivity, 301–303, 416 central tendency
Equal Opportunity, 333–334, 418 Chi-square test. See also Chi-square test of
Flower Shop, 215–216, 410–411 independence; Factorial chi-square test;
Getting What You Asked For, 332, 417 Nonparametric statistics; One-way chi-
Growing Tomatoes, 174–176 square test
Height versus Weight, 353–355 chi-square distribution, 307–309, 310–311
Homesick Blues, 211, 408–409 choosing the right statistical test and, 153
Irate Parents, 316–317 computing, 306–307, 319–321
Kids on Cruises, 32, 400 definition, 388
Learning to Speak, 32, 399 hypothesis testing and, 321–326
“Like Father, Like Son,” 371–372, 419 normal distribution and, 101
Money Meaning Nothing, 332–333, 418 overview, 304–305, 330
More Is Better, 372–373, 419 post-hoc tests and, 327
More Is Better Still, 373–374, 419–420 quiz regarding, 331–334, 419–420
Multiple Means of Math Mastery, 259–265 six-step model and, 312–317, 322–326,
Never Saying Never, 234–237 327–330
New Teacher, 30–31, 398 types of data and, 42
 Index | 423

Chi-square test of independence, 317–326, 388. Critical value

See also Chi-square test; Factorial chi- of chi-square, 308–309, 310, 385
square test definition, 390
Choosing the right statistical test step. See also of F, 262–263
Six-step model of t, 161–162, 164–165, 166–167, 185, 191,
analysis of variance (ANOVA) and, 261, 267, 196–197, 201–202
270, 273 Critical values tables
chart to assist with, 386 f, 257–258
chi-square test and, 315, 317, 325, 329 F, 379–384
correlations and, 348, 351, 354, 357, 369 t, 159, 161–162, 190–191, 377
dependent-sample t-test and, 231, 233, 236 z, 109, 111, 377
factorial ANOVA and, 280, 286
independent-sample t-test and, 199, 204, 208
multivariate ANOVA (MANOVA) and, 294 D
one-sample t-test and, 170, 173, 175
overview, 115, 152–154 Data analysis software. See Statistical software
quiz regarding, 155, 404–407 packages
Coefficient of determination, 388 Data collection, 11, 20, 39–42, 43. See also Col-
Cohen’s delta, 163, 182, 192 lection of data
Collection of data. See Data collection Data distribution. See also Normal distribution
Comparisons analysis of variance (ANOVA) and, 248
standard deviation and, 64–65 chi-square distribution, 307–309, 310
three or more comparisons and, 244–245 empirical rule and, 107–112
T-scores and, 77–78 kurtosis and, 105–107
z scores and, 76–77 overview, 112–113
Computed range, 61, 388. See also Range sampling distribution of mean differences
Computed value of F, 262–263, 388 and, 182–185
Computed value of p, 201, 209, 312, 388 sampling distribution of the means and,
Computed value of z, 139, 145, 157–158, 388 117–124
Computed values of t, 196, 197, 201–202, 388 skewness and, 102–105
Confidence intervals t distribution and, 157–163
changing alpha values and, 135–136, 146– Data types, 39–43, 57–58, 400–401. See also
147 Interval data; Nominal data; Ordinal data;
definition, 388 Ratio data
estimating population parameters using, Decisions based on data
125–128 alpha value and, 127
overview, 127 choosing the right statistical test and, 115–116
predicting a population parameter based on probability (p) values and, 148–152
a sample statistic using, 129–136 testing hypotheses about a population
Consumers of statistics, 2–3, 388 parameter based on a sample statistic
Contingency table, 319–320, 388 and, 138–141
Continuous data, 40, 41, 390. See also Interval use of statistics in, 2–3
data; Ratio data Degrees of freedom
Correlation coefficient, 93, 335, 390 analysis of variance (ANOVA) and, 249, 255,
Correlations. See also Linear regression; Pear- 257–258
son product–moment correlation coeffi- chi-square test and, 308–309, 310
cient (Pearson’s r); Spearman rank-differ- chi-square test of independence and, 321
ence correlation (Spearman’s rho) definition, 390
overview, 335–339, 370 independent-sample t-test and, 190–191,
probability (p) values and, 345–346 201–202
quiz regarding, 371–374, 419–420 overview, 159–162
six-step model and, 346–358, 366–370 sampling distribution of mean differences
Criterion variable, 347, 390 and, 185
424 | Index

Dependent variable. See also Dependent vari- Disordinal interactions, 283–284, 390
able identification and description step Dispersion, measures of. See Measures of dis-
choosing the right statistical test and, persion
152–154 Distance from the mean, 68–70
definition, 390 Distribution of data. See Data distribution
overview, 5, 33
quiz regarding, 57–58, 400–401
relationship between independent variables E
and, 37–38
Dependent variable identification and descrip- Effect size
tion step. See also Dependent variable; analysis of variance (ANOVA) and, 258–259,
Measures of dispersion; Measures of rela- 274
tive standing; Six-step model definition, 390
analysis of variance (ANOVA) and, 260–261, dependent-sample t-test and, 222–223
266, 269, 272–273 independent-sample t-test and, 192–196, 205
chi-square test and, 314–315, 317, 325, t distribution and, 163
328–329 Empirical rule
correlations and, 348, 351, 353–354, 356, 368 definition, 390
dependent-sample t-test and, 230, 231, 233, overview, 107–112, 113
235, 236 sampling distribution of mean differences
factorial ANOVA and, 277–278, 285, 286 and, 184
independent-sample t-test and, 199, 203–204, sampling distribution of the means and,
207 121–123, 124
multivariate ANOVA (MANOVA) and, 293, t distribution and, 157
294 Equal variances assumed, 190, 191–192, 210,
one-sample t-test and, 170, 172, 173, 175 201, 390
overview, 5, 37–56 Equal variances not assumed, 192, 201, 390
quiz regarding, 57–58, 400–401 Error, sampling. See Sampling error
Dependent-sample t-test. See also t-test Error probability, 135–136, 146–147
choosing the right statistical test and, 153 Estimations, 64–65, 125–128, 129–136
computing the t value for, 218–221, 222, 223 Ethical research, 12–13, 390
definition, 390 Examples. See Cases
effect size and, 222–223 Expected values
independent-sample t-test and, 218 chi-square test and, 311–312
one-tailed hypothesis testing and, 223–226 chi-square test of independence and,
overview, 181–182, 217–218, 237 319–320
quiz regarding, 238–243, 411–413 definition, 390
six-step model and, 229–237 factorial chi-square test and, 306
two-tailed hypothesis testing and, 226–228 overview, 305
Descriptive statistics. See also Graphical descrip- Experimental independent variables, 35–36,
tive statistics 390. See also Independent variables
analysis of variance (ANOVA) and, 249–251
definition, 390
one-sample t-test and, 166, 167 F
overview, 2–3, 5
Deviation from the mean, 67, 390. See also Stan- F distribution, 256–258, 390
dard deviation F value, 249, 255–256, 262–263, 390
Directional hypothesis. See also Hypothesis Factorial ANOVA. See also Analysis of variance
definition, 390 (ANOVA)
independent-sample t-test and, 195–196, 202 choosing the right statistical test and, 153
null hypotheses for, 20–22 definition, 390
one-sample t-test and, 163–166 overview, 246–247, 274–275, 295
overview, 16, 17–19 six-step model and, 275–287
 Index | 425

Factorial chi-square test, 305–312. See also Chi- analysis of variance (ANOVA) and, 260, 265,
square test 269, 272
Failure to reject the null hypothesis, 26–27, case studies to consider in, 30–32
127, 390. See also Null hypothesis chi-square test and, 314, 316, 324–325, 328
Frequency distribution table, 88, 390 correlations and, 347, 350, 353, 356, 367
Friedman’s ANOVA, 275. See also Analysis of dependent-sample t-test and, 229–230, 232,
variance (ANOVA); Factorial ANOVA 235
factorial ANOVA and, 276–277, 285
independent-sample t-test and, 198, 203,
G 206–207
multivariate ANOVA (MANOVA) and, 293
Goodness of fit, 304–305, 306, 311–312, one-sample t-test and, 169, 172, 174
317–318, 390. See also Chi-square test overview, 4, 7, 16–28
Graphical descriptive statistics. See also Descrip- quiz regarding, 29–30, 397–400
tive statistics; Scatterplots Hypothesis testing. See also Hypothesis;
bar charts and, 90–91 Hypothesis testing step
definition, 390 changing alpha values and, 135–136, 146–147
empirical rule and, 107–112 chi-square test and, 305, 321–326
histograms, 97, 98, 99 dependent-sample t-test and, 217–218
kurtosis and, 105–107 independent-sample t-test and, 190–191
nominal data and, 88, 89, 90 population parameter based on a sample
normal distribution and, 99–113 statistic, 137–146
overview, 87, 97–99, 100, 112–113 probability (p) values and, 148–152
pie charts, 88, 90, 91 quiz regarding, 155, 404–407
quantitative data and, 92 t distribution and, 157–163
quiz regarding, 113–114, 403–404 three or more comparisons and, 244–245
skewness and, 102–105 Hypothesis testing step. See also Hypothesis;
“Greater than” relationship testing, 18. See also Hypothesis testing; Six-step model
One-tailed hypothesis testing analysis of variance (ANOVA) and, 261–265,
267, 268, 270, 271, 273–274, 275
chi-square test and, 315, 317, 325–326,
H 329–330
correlations and, 349, 352–353, 354–355,
Histograms. See also Graphical descriptive 357, 369–370
statistics dependent-sample t-test and, 231, 233–234,
definition, 390 236–237
normal distribution and, 100–101 factorial ANOVA and, 280–284, 286–287
overview, 91, 97, 98, 99 independent-sample t-test and, 200–202,
Homogeneity of covariance matrices, 288, 291, 204–206, 208–209
292 multivariate ANOVA (MANOVA) and,
Homogeneity of variance, 248–249, 261–262, 294–295
267, 270, 273–274, 275 one-sample t-test and, 170–171, 173, 175–176
Hypothesis. See also Directional hypothesis; testing the null hypothesis and, 23–25
Hypothesis development/stating step;
Hypothesis testing; Hypothesis testing
step; Nondirectional hypothesis; Null I
hypothesis
definition, 390 Independence of scores, 248
overview, 4, 16–17 Independent variable identification step. See
rejecting and failing to reject, 26–27 also Independent variables; Six-step model
significance and, 27 analysis of variance (ANOVA) and, 260, 266,
Hypothesis development/stating step. See also 269, 272
Data analysis software; Hypothesis; Six- chi-square test and, 314, 316, 325, 328
step model
426 | Index

Independent variable identification step (cont.) definition, 391

correlations and, 347, 350, 353, 356, 368 measures of central tendency and, 43–45
dependent-sample t-test and, 230, 233, 235 overview, 40–41, 42, 43
factorial ANOVA and, 277, 285
independent-sample t-test and, 198–199, 203,
207 K
multivariate ANOVA (MANOVA) and, 293
one-sample t-test and, 169, 172, 175 Kruskal–Wallis H test, 262, 274–275
overview, 4, 33–37 Kurtosis, 105–107, 112–113, 117, 391
quiz regarding, 57–58, 400–401
Independent variables. See also Independent
variable identification step L
analysis of variance (ANOVA) and, 245
chi-square test of independence and, 318 Latent independent variable, 391
choosing the right statistical test and, Leptokurtosis, 106–107, 112–113, 391. See also
152–154 Kurtosis
definition, 390 “Less than” relationship testing, 18–19. See also
levels of, 36 One-tailed hypothesis testing
manipulated independent variables, 35–36 Level of significance, 126. See also Alpha value
nonmanipulated independent variables, Levels of the independent variable, 4, 391
34–35, 36 Levene’s test
overview, 4, 33 analysis of variance (ANOVA) and, 249, 270,
quiz regarding, 57–58, 400–401 273–275
relationship between dependent variables definition, 391
and, 37–38 factorial ANOVA and, 280, 281, 286–287
Independent-sample t-test. See also t-test independent-sample t-test and, 201, 204, 208
calculating the t value for, 185–190 Line of best fit, 93, 391
choosing the right statistical test and, 153 Linear regression. See also Correlations
definition, 390 computing the intercept and, 361–362, 363,
effect size and, 192–196 364, 365–366
overview, 181–182, 209 computing the slope and, 359–361, 365–
quiz regarding, 211–216, 408–411 366
sampling distribution of mean differences overview, 358–366, 370
and, 182–185 quiz regarding, 371–374, 419–420
six-step model and, 197–209 regression equation and, 358–359
Inferential statistics six-step model and, 366–370
definition, 390 Logistic regression, 370, 391
description of, 2–3 Lower confidence limit, 391
measures of dispersion and, 60
normal distribution and, 112–113
one-sample t-test and, 166, 167 M
overview, 115–116, 147–148
testability of hypotheses and, 20 Main effects, 277, 282, 289–290, 391
Interaction effect, 277, 289, 390 Manipulated (experimental) independent vari-
Interaction p value, 282–284. See also Probabil- ables, 35–36, 57–58, 391, 400–401
ity (p) values Mann–Whitney U test, 209, 237, 391
Intercept, 361–362, 363, 364, 365–366, 390. See Math skills required in statistics, 2
also Linear regression Mean. See also Measures of central tendency;
Interpreting data, 3, 171 Measures of dispersion
Interquartile range, 391 definition, 391
Interval data. See also Continuous data; Data empirical rule and, 107–112
types; Quantitative data normal distribution and, 100–101
choosing the right statistical test and, 152–153 overview, 45–47
 Index | 427

quiz regarding, 57–58, 400–401 Multiple-comparison tests, 263–264, 267, 268,

sampling distribution of the means and, 124 270, 271, 273–274, 275, 392
SPSS and, 54–55, 56 Multivariate ANOVA (MANOVA). See also
standard deviation and, 63, 79, 80 Analysis of variance (ANOVA)
stanines and, 79, 80 assumptions of, 288
T-scores and, 77–78, 79, 80 definition, 392
z scores and, 79, 80 overview, 247–248, 287–292, 295
Mean of means, 119, 391. See also Sampling six-step model and, 292–295
distribution of the means Multivariate statistics, 36, 288, 392
Mean scores, 165–166
Mean square values, 249, 255, 391
Measures of central tendency. See also Mean; N
Median; Mode
definition, 392 Negative correlations, 340–341. See also Cor-
overview, 43–45 relations
quiz regarding, 57–58, 400–401 Negative skewness, 103–105, 392. See also Skew-
sampling distribution of the means and, 124 ness
SPSS and, 54–55 Nominal data. See also Data types; Nonpara-
Measures of dispersion. See also Dependent metric statistics
variable identification and description chi-square test of independence and, 318
step; Quantitative data; Range; Standard choosing the right statistical test and,
deviation; Variance 152–153
definition, 392 definition, 392
overview, 60–70, 71, 83 graphical descriptions of, 88, 89, 90, 101
quiz regarding, 85–86, 401–403 mean and, 46–47
sampling distribution of mean differences measures of central tendency and, 43–45
and, 184 median and, 50
sampling distribution of the means and, 124 mode and, 50–51
SPSS and, 70, 71 overview, 39, 42, 43
Measures of relative standing. See also Depen- Nondirectional hypothesis. See also Hypothesis
dent variable identification and descrip- analysis of variance (ANOVA) and, 245
tion step; Percentiles; Quartiles; Stanines; definition, 392
T-scores; z scores independent-sample t-test and, 196
definition, 392 null hypotheses for, 22–23
overview, 60, 70–80, 83 overview, 16, 18, 19–20
quiz regarding, 85–86, 401–403 Nonmanipulated (quasi-) independent vari-
sampling distribution of the means and, 124 ables, 34–35, 36, 57–58, 392, 400–401. See
SPSS and, 80–84 also Independent variables
Median. See also Measures of central tendency Nonparametric statistics. See also Chi-square
definition, 392 test
normal distribution and, 100–101 choosing the right statistical test and,
overview, 47–50 152–153
as a percentile, 73–75 definition, 392
quiz regarding, 57–58, 400–401 dependent-sample t-test and, 237
SPSS and, 54–55, 56 independent-sample t-test and, 209
Mode. See also Measures of central tendency normal distribution and, 101, 113
definition, 392 overview, 42
normal distribution and, 100–101 Spearman rank difference correlation and,
overview, 50–51 344–346
quiz regarding, 57–58, 400–401 types of data and, 43
SPSS and, 54–55, 56 Normal distribution. See also Data distribution
Multimodal dataset, 51, 392. See also Mode analysis of variance (ANOVA) and, 248
Multiple regressions, 370, 392 definition, 392
428 | Index

Normal distribution (cont.) One-way ANOVA, 246, 393. See also Analysis of
empirical rule and, 107–112 variance (ANOVA)
graphical descriptions of, 99–113 One-way chi-square test, 304–305. See also Chi-
histograms and, 100–101 square test
hypothesis testing and, 138–139 Ordinal data. See also Data types; Nonparamet-
overview, 100–101, 112–113 ric statistics
quiz regarding, 113–114, 403–404 choosing the right statistical test and,
sampling distribution of the means and, 152–153
118–119 definition, 393
Null hypothesis. See also Hypothesis; Hypoth- graphical descriptions of, 101
esis testing measures of central tendency and, 43–45,
case studies to consider in, 30–32 46–47, 50–51
definition, 393 overview, 39–40, 42, 43
for directional research hypotheses and, 20–22 Spearman rank difference correlation and,
for nondirectional research hypotheses and, 344
22–23 Ordinal interactions, 283–284, 287, 393
one-tailed hypothesis testing and, 142–146
overview, 16, 20–21
probability (p) values and, 148–152 P
rejecting and failing to reject, 26–28, 127
research hypotheses and, 26–27 p values. See Probability (p) values
significance and, 27, 28 Paired-samples t-test. See Dependent-sample
testing, 23–25 t-test
Type I and II errors and, 128 Parametric statistics. See also Analysis of vari-
Numeric descriptive statistics, 393 ance (ANOVA); t-test
choosing the right statistical test and,
152–153
O definition, 393
normal distribution and, 101, 112–113
Observed values overview, 24, 42
chi-square test and, 312 types of data and, 43
chi-square test of independence and, 319 Pearson product–moment correlation coeffi-
definition, 393 cient (Pearson’s r). See also Correlations
factorial chi-square test and, 306 choosing the right statistical test and, 153
overview, 304–305 definition, 393
One-sample chi-square test, 393 graphical descriptive statistics and, 92
One-sample t-test. See also t-test interpreting, 339–343
choosing the right statistical test and, 153 linear regression and, 365–366
directional hypothesis and, 163–166 overview, 336–339
one-tailed hypothesis testing and, 166–168 quiz regarding, 371–374, 417–418
overview, 156–157, 176, 182 Percentiles. See also Measures of relative stand-
quiz regarding, 177–180, 407–408 ing; Quartiles
six-step model and, 168–176 definition, 393
t distribution and, 157–163 median as, 73–75
One-sample z-test, 153. See also z-test overview, 71–75
One-tailed hypothesis testing. See also Direc- SPSS and, 80–83, 84
tional hypothesis; Hypothesis testing Pie charts, 90, 91, 393. See also Graphical
dependent-sample t-test and, 219–221, 222, descriptive statistics
223–226 Pillai’s Trace Sig, 294, 295
one-sample t-test and, 166–167, 168 Platykurtosis. See also Kurtosis
testing hypotheses about a population definition, 393
parameter based on a sample statistic overview, 105, 106, 112–113
and, 141–146 sampling distribution of the means and, 117
 Index | 429

Plot line, 93, 96 chi-square test and, 313–314, 316, 324,

Pooled standard deviation, 192, 394. See also 327–328
Standard deviation correlations and, 346, 350, 353, 355–356, 367
Population dependent-sample t-test and, 229, 232, 235
computing the standard deviation for, 67 factorial ANOVA and, 276, 284–285
definition, 394 finding a research problem and, 8–13
mean and, 46 independent-sample t-test and, 198, 202–203,
overview, 24, 43 206
sampling distribution of the means and, 124 multivariate ANOVA (MANOVA) and, 293
variance of, 69–70 one-sample t-test and, 168–169, 171–172, 174
Population distribution, 120–121 overview, 4, 7–16
Population parameter quiz regarding, 29–30, 397–400
definition, 394 Problem statement. See also Problem identifica-
estimating using confidence intervals, tion step
125–128 characteristics of, 8, 13–14
overview, 69–70 definition, 394
predicting based on a sample statistic using finding a research problem and, 8–9
confidence intervals and, 129–136 narrowing down and focusing, 10
sampling distribution of the means and, overview, 4
116–117 quiz regarding, 29–30, 397–400
testing hypotheses about based on a sample writing, 13–15
statistic and, 137–146
Positive correlations, 340–341. See also Correla-
tions Q
Positive skewness, 102–103, 394. See also Skewness
Possible range, 40–41, 394. See also Interval Qualitative data, 11, 394
data; Range Quantitative data. See also Interval data; Mea-
Post-hoc tests sures of dispersion; Parametric statistics;
analysis of variance (ANOVA) and, 270, 271 Ratio data
chi-square test and, 309, 310, 327 choosing the right statistical test and,
definition, 394 152–153
Power, statistical, 128, 394 definition, 394
Practical significance, 11–12, 163, 394 graphical descriptions of, 92, 101–112
Prediction, 129–136 mean and, 46–47
Predictor variable, 347, 394 median and, 50
Probability (p) values mode and, 50–51
analysis of variance (ANOVA) and, 258, 262, normal distribution and, 112–113
270, 274 overview, 40, 41
of a correlation, 345–346 sampling distribution of the means and, 124
definition, 393 Quartiles, 73–75, 80–83, 84, 394. See also Mea-
directional hypothesis and, 165–166 sures of relative standing; Percentiles
factorial ANOVA and, 282–284, 286–287
independent-sample t-test and, 182, 191, 196,
201–202, 204 R
multivariate ANOVA (MANOVA) and,
294–295 Random samples, 248
one-tailed hypothesis testing and, 167, 168 Range. See also Measures of dispersion
overview, 148–152 definition, 394
t distribution and, 162–163 overview, 61, 63, 64
Problem identification step. See also Problem quiz regarding, 85–86, 401–403
statement; Six-step model Rank data. See Ordinal data
analysis of variance (ANOVA) and, 259–260, Ratio data. See also Continuous data; Data
265, 268, 272 types; Quantitative data
430 | Index

Ratio data (cont.) overview, 26

choosing the right statistical test and, testing the null hypothesis and, 23–25, 24t,
152–153 25t
definition, 394 Type I and II errors and, 128
measures of central tendency and, 43–45 Scatterplots. See also Graphical descriptive
overview, 41, 42, 43 statistics
Regression, linear. See Linear regression correlations and, 357
Regression equation, 358–359 definition, 395
Rejecting the null hypothesis, 26–27, 26t, 28t, linear regression and, 365, 366
See also Null hypothesis overview, 92–96, 97, 335
Relative frequency, 394 Scheffé test, 3
Relative standing, measures of. See Measures of Significance, 27, 28, 126
relative standing Significantly different, 28, 395
Research hypothesis. See also Hypothesis Simple frequency, 395
case studies to consider in, 30–32 Six-step model. See also Cases; Choosing the
definition, 394 right statistical test step; Dependent vari-
null hypotheses and, 26–27 able identification and description step;
overview, 20 Hypothesis development/stating step;
rejecting and failing to reject, 26–27, 28 Hypothesis testing step; Independent vari-
significance and, 27, 28 able identification step; Problem identifica-
Research planning, 2–3 tion step; Steps in using statistics
Research problems, 8–13, 10f. See also Problem analysis of variance (ANOVA) and, 259–275
identification step chi-square test and, 312–317, 322–326,
Researcher characteristics, 10–11 327–330
correlations and, 346–358
dependent-sample t-test and, 229–237
S factorial ANOVA and, 275–287
independent-sample t-test and, 197–209
Sample. See also Sample statistic linear regression and, 367–370
computing the standard deviation for, 65–67 multivariate ANOVA (MANOVA) and, 292–295
definition, 395 one-sample t-test and, 168–166, 170, 171, 172,
overview, 24, 43 173, 175, 176
predicting a population parameter based Skewed left. See Negative skewness
on a sample statistic using confidence Skewed right. See Positive skewness
intervals and, 129–136 Skewness, 102–105, 112–113, 395
sampling distribution of the means and, Slope, 359–361, 365–366, 395. See also Linear
120–121 regression
Sample statistic, 24, 137–146, 395. See also Sample Software packages. See Statistical software
Sampling distribution of mean differences, packages
182–185 Spearman rank-difference correlation (Spear-
Sampling distribution of the means man’s rho), 92, 153, 344–346, 395. See also
definition, 395 Correlations
estimating population parameters using con- SPSS (Statistical Package for the Social Scienc-
fidence intervals and, 125–128 es). See also Statistical software packages
independent-sample t-test and, 182 analysis of variance (ANOVA) and, 249–251,
overview, 117–124 261–265
quiz regarding, 124–125 bar charts and, 90–91, 92
Sampling error. See also Type I error; Type II chi-square test and, 308, 309, 310, 311–312,
error 322–326
changing alpha values and, 146–147 correlations and, 349, 357, 369–370
definition, 395 dependent-sample t-test and, 221, 222, 223,
estimating population parameters using con- 225–226
fidence intervals and, 126–128 directional hypothesis and, 163–166
 Index | 431

entering data into, 52–55 factorial ANOVA and, 286–287

factorial ANOVA and, 277–278, 279 independent-sample t-test and, 200–202,
histograms and, 97, 98, 99 208–209
independent-sample t-test and, 188–190, overview, 51–55, 56, 376
192–193, 194, 195, 200–202, 204–206, probability (p) values and, 150–152
208–209 Statistical tests, 5
kurtosis and, 105, 106 Statistics, 2–3, 375–376, 395. See also Descrip-
linear regression and, 362, 363, 364, 365, 366 tive statistics; Inferential statistics; Steps in
measures of dispersion and, 70, 71 using statistics
measures of relative standing and, 80–83, 84 Steps in using statistics, 3–5, 6. See also Statis-
multivariate ANOVA (MANOVA) and, tics
290–292 Sum of squares, 186–188, 395
nominal data and, 88, 89, 90
one-sample t-test and, 167, 168
overview, 51–55, 56 T
Pearson correlation coefficient and, 339, 340,
343 t distribution, 157–163, 201–202, 395
pie charts and, 88, 90, 91 t table, 396
sampling distribution of mean differences t value
and, 184–185 compared to a T-score, 77
scatterplots and, 92–96, 97 dependent-sample t-test and, 218–221, 222,
skewness and, 103 223, 224–226
t distribution and, 158–160, 162 independent-sample t-test and, 185–190
Standard deviation. See also Measures of disper- sampling distribution of mean differences
sion; Pooled standard deviation and, 185
definition, 395 Testability of hypotheses, 20, 23–25, 24t, 25t
empirical rule and, 107–112 Theoretical significance, 11–12, 396
overview, 63–67, 79, 80 Total degrees of freedom, 249. See also Degrees
for a population, 67 of freedom
quiz regarding, 85–86, 124–125, 401–403 Total sum of squares, 251–252. See also Sum of
for a sample, 65–67 squares
sampling distribution of the means and, 119, Total variance, 249, 251–255. See also Variance
124 T-scores. See also Measures of relative standing;
variance and, 69 z scores
Standard error of mean differences, 184, 395 definition, 396
Standard error of the mean, 124, 395 overview, 77–78, 79, 80
Stanines. See also Measures of relative standing quiz regarding, 85–86, 401–403
definition, 395 SPSS and, 80–83, 84
overview, 78–79, 80 t-test. See also Dependent-sample t-test; Inde-
quiz regarding, 85–86, 401–403 pendent-sample t-test; One-sample t-test;
SPSS and, 80–83, 84 Parametric statistics
Statistical power. See Power, statistical choosing the right statistical test and,
Statistical software packages. See also Analysis 153–154
of data; Hypothesis testing step; Six-step compared to a T-score, 77
model; SPSS (Statistical Package for the normal distribution and, 101
Social Sciences) three or more comparisons and, 244–245
analysis of variance (ANOVA) and, 261–265, types of data and, 42
267, 268, 273–274, 275 Tukey’s HSD test, 3
chi-square test and, 315, 317, 325–326 “2 x 2” analysis of variance, 280
correlations and, 349, 352–353, 354–355, Two-tailed hypothesis testing, 226–228. See also
357, 369–370 Nondirectional hypothesis
dependent-sample t-test and, 233–234, Two-way chi-square test. See Chi-square test;
236–237 Chi-square test of independence
432 | Index

Type I error. See also Alpha value; Sampling Within-group degrees of freedom, 249,
error 257–258, 396
analysis of variance (ANOVA) and, 263–264 Within-group variance, 249, 396
changing alpha values and, 146–147
definition, 396
multivariate ANOVA (MANOVA) and, 289
overview, 127–128, 245
X
three or more comparisons and, 244–245
Type I error rate, 396 x-axis, 93, 282, 396
Type II error. See also Beta value; Sampling error
analysis of variance (ANOVA) and, 263–264
changing alpha values and, 146–147 Y
definition, 396
overview, 128, 245 y-axis, 93, 282, 396
three or more comparisons and, 244–245
Type II error rate, 396

Z
U
z scores. See also Measures of relative standing;
Univariate statistics, 36, 396 T-scores
Upper confidence limit, 396 changing alpha values and, 135–136, 146–147
definition, 396
empirical rule and, 111–112, 113
V overview, 75–78, 79, 80, 147–148
predicting a population parameter based
Variance. See also Analysis of variance on a sample statistic using confidence
(ANOVA); Measures of dispersion intervals and, 131–136
definition, 396 probability (p) values and, 148–152
independent-sample t-test and, 192 quiz regarding, 85–86, 124–125, 401–403
overview, 67–70 SPSS and, 80–83, 84
of a population, 69–70 stanines and, 78–79
quiz regarding, 85–86, 401–403 testing hypotheses about a population
parameter based on, 145, 146
testing hypotheses about a population
W parameter based on a sample statistic
and, 139–141, 142–146
Wilcoxon t-test, 237, 396 z table, 396
Within sum of squares, 251–252, 253–255 z-test, 153–154. See also One-sample z-test
 About the Author

Dr. Steven R. Terrell is Professor at Nova Southeastern University. He has taught

quantitative and qualitative research methods for over 30 years, and is the author of
over 130 journal articles, conference presentations, and two books. Dr. Terrell is active
in the American Psychological Association and the American Counseling Associa-
tion and served as Chair of the American Educational Research Association’s Online
Teaching and Learning Special Interest Group.

433