Learn R For Applied Statistics

Applied Statistics with R
David Dalpiaz
2
Contents
1 Introduction 11
1.1 About This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 License . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 Introduction to R 15
2.1 Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Basic Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Getting Help . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Installing Packages . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Data and Programming 21

3.1 Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.1 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.2 Vectorization . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.3 Logical Operators . . . . . . . . . . . . . . . . . . . . . . 27
3.2.4 More Vectorization . . . . . . . . . . . . . . . . . . . . . . 29
3.2.5 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.6 Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.7 Data Frames . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3 Programming Basics . . . . . . . . . . . . . . . . . . . . . . . . . 51
3
4 CONTENTS
3.3.1 Control Flow . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.3.2 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4 Summarizing Data 57
4.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2 Plotting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2.1 Histograms . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2.2 Barplots . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2.3 Boxplots . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.4 Scatterplots . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5 Probability and Statistics in R 67

5.1 Probability in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1.1 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Hypothesis Tests in R . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.1 One Sample t-Test: Review . . . . . . . . . . . . . . . . . 69
5.2.2 One Sample t-Test: Example . . . . . . . . . . . . . . . . 70
5.2.3 Two Sample t-Test: Review . . . . . . . . . . . . . . . . . 73
5.2.4 Two Sample t-Test: Example . . . . . . . . . . . . . . . . 73
5.3 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.3.1 Paired Differences . . . . . . . . . . . . . . . . . . . . . . 77
5.3.2 Distribution of a Sample Mean . . . . . . . . . . . . . . . 80
6 R Resources 85
6.1 Beginner Tutorials and References . . . . . . . . . . . . . . . . . 85
6.2 Intermediate References . . . . . . . . . . . . . . . . . . . . . . . 85
6.3 Advanced References . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.4 Quick Comparisons to Other Languages . . . . . . . . . . . . . . 86
6.5 RStudio and RMarkdown Videos . . . . . . . . . . . . . . . . . . 86
6.6 RMarkdown Template . . . . . . . . . . . . . . . . . . . . . . . . 87
CONTENTS 5
7 Simple Linear Regression 89

7.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.1.1 Simple Linear Regression Model . . . . . . . . . . . . . . 94
7.2 Least Squares Approach . . . . . . . . . . . . . . . . . . . . . . . 97
7.2.1 Making Predictions . . . . . . . . . . . . . . . . . . . . . . 99
7.2.2 Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.2.3 Variance Estimation . . . . . . . . . . . . . . . . . . . . . 103
7.3 Decomposition of Variation . . . . . . . . . . . . . . . . . . . . . 104
7.3.1 Coefficient of Determination . . . . . . . . . . . . . . . . . 106
7.4 The lm Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.5 Maximum Likelihood Estimation (MLE) Approach . . . . . . . . 115
7.6 Simulating SLR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.7 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.8 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
8 Inference for Simple Linear Regression 123

8.1 Gauss–Markov Theorem . . . . . . . . . . . . . . . . . . . . . . . 126
8.2 Sampling Distributions . . . . . . . . . . . . . . . . . . . . . . . . 127
8.2.1 Simulating Sampling Distributions . . . . . . . . . . . . . 128
8.3 Standard Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.4 Confidence Intervals for Slope and Intercept . . . . . . . . . . . . 137
8.5 Hypothesis Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.6 cars Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.6.1 Tests in R . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.6.2 Significance of Regression, t-Test . . . . . . . . . . . . . . 142
8.6.3 Confidence Intervals in R . . . . . . . . . . . . . . . . . . . 143
8.7 Confidence Interval for Mean Response . . . . . . . . . . . . . . . 145
8.8 Prediction Interval for New Observations . . . . . . . . . . . . . . 146
8.9 Confidence and Prediction Bands . . . . . . . . . . . . . . . . . . 147
8.10 Significance of Regression, F-Test . . . . . . . . . . . . . . . . . . 149
8.11 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6 CONTENTS
9 Multiple Linear Regression 153

9.1 Matrix Approach to Regression . . . . . . . . . . . . . . . . . . . 157
9.2 Sampling Distribution . . . . . . . . . . . . . . . . . . . . . . . . 161
9.2.1 Single Parameter Tests . . . . . . . . . . . . . . . . . . . . 163
9.2.2 Confidence Intervals . . . . . . . . . . . . . . . . . . . . . 165
9.2.3 Confidence Intervals for Mean Response . . . . . . . . . . 165
9.2.4 Prediction Intervals . . . . . . . . . . . . . . . . . . . . . 169
9.3 Significance of Regression . . . . . . . . . . . . . . . . . . . . . . 170
9.4 Nested Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
9.5 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
9.6 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10 Model Building 185

10.1 Family, Form, and Fit . . . . . . . . . . . . . . . . . . . . . . . . 186
10.1.1 Fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
10.1.2 Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
10.1.3 Family . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
10.1.4 Assumed Model, Fitted Model . . . . . . . . . . . . . . . 188
10.2 Explanation versus Prediction . . . . . . . . . . . . . . . . . . . . 189
10.2.1 Explanation . . . . . . . . . . . . . . . . . . . . . . . . . . 189
10.2.2 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
10.4 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
11 Categorical Predictors and Interactions 195

11.1 Dummy Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
11.2 Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
11.3 Factor Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
11.3.1 Factors with More Than Two Levels . . . . . . . . . . . . 215
11.4 Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
11.5 Building Larger Models . . . . . . . . . . . . . . . . . . . . . . . 225
11.6 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
CONTENTS 7
12 Analysis of Variance 231
12.1 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
12.2 Two-Sample t-Test . . . . . . . . . . . . . . . . . . . . . . . . . . 232
12.3 One-Way ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . 235
12.3.1 Factor Variables . . . . . . . . . . . . . . . . . . . . . . . 242
12.3.2 Some Simulation . . . . . . . . . . . . . . . . . . . . . . . 243
12.3.3 Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
12.4 Post Hoc Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
12.5 Two-Way ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . 249
12.6 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
13 Model Diagnostics 261
13.1 Model Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . 261
13.2 Checking Assumptions . . . . . . . . . . . . . . . . . . . . . . . . 263
13.2.1 Fitted versus Residuals Plot . . . . . . . . . . . . . . . . . 264
13.2.2 Breusch-Pagan Test . . . . . . . . . . . . . . . . . . . . . 270
13.2.3 Histograms . . . . . . . . . . . . . . . . . . . . . . . . . . 272
13.2.4 Q-Q Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
13.2.5 Shapiro-Wilk Test . . . . . . . . . . . . . . . . . . . . . . 280
13.3 Unusual Observations . . . . . . . . . . . . . . . . . . . . . . . . 282
13.3.1 Leverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
13.3.2 Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
13.3.3 Influence . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
13.4 Data Analysis Examples . . . . . . . . . . . . . . . . . . . . . . . 294
13.4.1 Good Diagnostics . . . . . . . . . . . . . . . . . . . . . . . 294
13.4.2 Suspect Diagnostics . . . . . . . . . . . . . . . . . . . . . 298
13.5 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301

8 CONTENTS
14 Transformations 303
14.1 Response Transformation . . . . . . . . . . . . . . . . . . . . . . 303
14.1.1 Variance Stabilizing Transformations . . . . . . . . . . . . 306
14.1.2 Box-Cox Transformations . . . . . . . . . . . . . . . . . . 311
14.2 Predictor Transformation . . . . . . . . . . . . . . . . . . . . . . 319
14.2.1 Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . 322
14.2.2 A Quadratic Model . . . . . . . . . . . . . . . . . . . . . . 345
14.2.3 Overfitting and Extrapolation . . . . . . . . . . . . . . . . 350
14.2.4 Comparing Polynomial Models . . . . . . . . . . . . . . . 351
14.2.5 poly() Function and Orthogonal Polynomials . . . . . . . 354
14.2.6 Inhibit Function . . . . . . . . . . . . . . . . . . . . . . . 356
14.2.7 Data Example . . . . . . . . . . . . . . . . . . . . . . . . 357
14.3 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
15 Collinearity 365
15.1 Exact Collinearity . . . . . . . . . . . . . . . . . . . . . . . . . . 365
15.2 Collinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
15.2.1 Variance Inflation Factor. . . . . . . . . . . . . . . . . . . 371
15.3 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
15.4 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
16 Variable Selection and Model Building 383

16.1 Quality Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
16.1.1 Akaike Information Criterion . . . . . . . . . . . . . . . . 384
16.1.2 Bayesian Information Criterion . . . . . . . . . . . . . . . 385
16.1.3 Adjusted R-Squared . . . . . . . . . . . . . . . . . . . . . 386
16.1.4 Cross-Validated RMSE . . . . . . . . . . . . . . . . . . . 386
16.2 Selection Procedures . . . . . . . . . . . . . . . . . . . . . . . . . 390
16.2.1 Backward Search . . . . . . . . . . . . . . . . . . . . . . . 391
16.2.2 Forward Search . . . . . . . . . . . . . . . . . . . . . . . . 397
16.2.3 Stepwise Search . . . . . . . . . . . . . . . . . . . . . . . . 400
16.2.4 Exhaustive Search . . . . . . . . . . . . . . . . . . . . . . 403
CONTENTS 9
16.3 Higher Order Terms . . . . . . . . . . . . . . . . . . . . . . . . . 408

16.4 Explanation versus Prediction . . . . . . . . . . . . . . . . . . . . 413
16.4.1 Explanation . . . . . . . . . . . . . . . . . . . . . . . . . . 413
16.4.2 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
16.5 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
17 Logistic Regression 417

17.1 Generalized Linear Models . . . . . . . . . . . . . . . . . . . . . . 417
17.2 Binary Response . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
17.2.1 Fitting Logistic Regression . . . . . . . . . . . . . . . . . 421
17.2.2 Fitting Issues . . . . . . . . . . . . . . . . . . . . . . . . . 422
17.2.3 Simulation Examples . . . . . . . . . . . . . . . . . . . . . 422
17.3 Working with Logistic Regression . . . . . . . . . . . . . . . . . . 430
17.3.1 Testing with GLMs . . . . . . . . . . . . . . . . . . . . . . 431
17.3.2 Wald Test . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
17.3.3 Likelihood-Ratio Test . . . . . . . . . . . . . . . . . . . . 431
17.3.4 SAheart Example . . . . . . . . . . . . . . . . . . . . . . 432
17.3.5 Confidence Intervals . . . . . . . . . . . . . . . . . . . . . 436
17.3.6 Confidence Intervals for Mean Response . . . . . . . . . . 437
17.3.7 Formula Syntax . . . . . . . . . . . . . . . . . . . . . . . . 439
17.3.8 Deviance . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
17.4 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
17.4.1 spam Example . . . . . . . . . . . . . . . . . . . . . . . . 443
17.4.2 Evaluating Classifiers . . . . . . . . . . . . . . . . . . . . 445
17.5 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
18 Beyond 453
18.1 What’s Next . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
18.2 RStudio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
18.3 Tidy Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
18.4 Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
18.5 Web Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
10 CONTENTS
18.6 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . 454

18.7 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
18.7.1 Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . 455
18.8 Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
18.9 Bayesianism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
18.10High Performance Computing . . . . . . . . . . . . . . . . . . . . 456
18.11Further R Resources . . . . . . . . . . . . . . . . . . . . . . . . . 456
19 Appendix 457
Chapter 1
Introduction
Welcome to Applied Statistics with R!
1.1 About This Book
This book was originally (and currently) designed for use with STAT 420,
Methods of Applied Statistics, at the University of Illinois Urbana-Champaign.
It may certainly be used elsewhere, but any references to “this course” in this
book specifically refer to STAT 420.
This book is under active development. When possible, it would be best to
always access the text online to be sure you are using the most up-to-date
version. Also, the html version provides additional features such as changing
text size, font, and colors. If you are in need of a local copy, a pdf version
is continuously maintained, however, because a pdf uses pages, the formatting
may not be as functional. (In other words, the author needs to go back and
spend some time working on the pdf formatting.)
Since this book is under active development you may encounter errors ranging
from typos, to broken code, to poorly explained topics. If you do, please let us
know! Simply send an email and we will make the changes as soon as possible.
(dalpiaz2 AT illinois DOT edu) Or, if you know RMarkdown and are famil-
iar with GitHub, make a pull request and fix an issue yourself! This process is
partially automated by the edit button in the top-left corner of the html version.
If your suggestion or fix becomes part of the book, you will be added to the list
at the end of this chapter. We’ll also link to your GitHub account, or personal
website upon request.
This text uses MathJax to render mathematical notation for the web. Occa-
sionally, but rarely, a JavaScript error will prevent MathJax from rendering
11
12 CHAPTER 1. INTRODUCTION
correctly. In this case, you will see the “code” instead of the expected math-
ematical equations. From experience, this is almost always fixed by simply
refreshing the page. You’ll also notice that if you right-click any equation you
can obtain the MathML Code (for copying into Microsoft Word) or the TeX
command used to generate the equation.
𝑎2 + 𝑏2 = 𝑐2
1.2 Conventions
R code will be typeset using a monospace font which is syntax highlighted.
a = 3
b = 4
sqrt(a ^ 2 + b ^ 2)
R output lines, which would appear in the console will begin with ##. They will
generally not be syntax highlighted.
## [1] 5
We use the quantity 𝑝 to refer to the number of 𝛽 parameters in a linear model,

not the number of predictors. Don’t worry if you don’t know what this means
yet!
1.3 Acknowledgements
Material in this book was heavily influenced by:
• Alex Stepanov
– Longtime instructor of STAT 420 at the University of Illinois at
Urbana-Champaign. The author of this book actually took Alex’s
STAT 420 class many years ago! Alex provided or inspired many of
the examples in the text.
• David Unger
– Another STAT 420 instructor at the University of Illinois at Urbana-
Champaign. Co-taught with the author during the summer of 2016
while this book was first being developed. Provided endless hours of
copy editing and countless suggestions.
1.3. ACKNOWLEDGEMENTS 13
• James Balamuta
– Current graduate student at the University of Illinois at Urbana-
Champaign. Provided the initial push to write this book by intro-
ducing the author to the bookdown package in R. Also a frequent
contributor via GitHub.
Your name could be here! Suggest an edit! Correct a typo! If you submit a
correction and would like to be listed below, please provide your name as you
would like it to appear, as well as a link to a GitHub, LinkedIn, or personal
website.
• Daniel McQuillan
• Mason Rubenstein
• Yuhang Wang
• Zhao Liu
• Jinfeng Xiao
• Somu Palaniappan
• Michael Hung-Yiu Chan
• Eloise Rosen
• Kiomars Nassiri
• Jeff Gerlach
• Brandon Ching
• Ray Fix
• Tyler Kim
• Yeongho Kim
• Elmar Langholz
• Thai Duy Cuong Nguyen
• Junyoung Kim
• Sezgin Kucukcoban
• Tony Ma
• Radu Manolescu
• Dileep Pasumarthi
• Sihun Wang
• Joseph Wilson
• Yingkui Lin
• Andy Siddall
• Nishant Balepur
• Durga Krovi
• Raj Krishnan
• Ed Pureza
• Siddharth Singh
• Schillaci Mcinnis
• Ivan Valdes Castillo
• Tony Mu
• Salman Yousaf
14 CHAPTER 1. INTRODUCTION
• Yutaro Nishiyama
• Regina Sahani Goonetilleke
• Paul Zuradzki
• Will Tsai
• Ellen Veomett
• David Newman
• Yixing Zheng
1.4 License
Figure 1.1: This work is licensed under a Creative Commons Attribution-

NonCommercial-ShareAlike 4.0 International License.
Chapter 2
Introduction to R
2.1 Getting Started

R is both a programming language and software environment for statistical com-
puting, which is free and open-source. To get started, you will need to install
two pieces of software:
• R, the actual programming language.

– Chose your operating system, and select the most recent version,
4.2.1.
• RStudio, an excellent IDE for working with R.
– Note, you must have R installed to use RStudio. RStudio is simply
an interface used to interact with R.
The popularity of R is on the rise, and every day it becomes a better tool for
statistical analysis. It even generated this book! (A skill you will learn in this
course.) There are many good resources for learning R.
The following few chapters will serve as a whirlwind introduction to R. They are
by no means meant to be a complete reference for the R language, but simply an
introduction to the basics that we will need along the way. Several of the more
important topics will be re-stressed as they are actually needed for analyses.
These introductory R chapters may feel like an overwhelming amount of infor-
mation. You are not expected to pick up everything the first time through. You
should try all of the code from these chapters, then return to them a number of
times as you return to the concepts when performing analyses.
R is used both for software development and data analysis. We will operate in a
grey area, somewhere between these two tasks. Our main goal will be to analyze
15
16 CHAPTER 2. INTRODUCTION TO R
data, but we will also perform programming exercises that help illustrate certain
concepts.
RStudio has a large number of useful keyboard shortcuts. A list of these can be
found using a keyboard shortcut – the keyboard shortcut to rule them all:
• On Windows: Alt + Shift + K

• On Mac: Option + Shift + K
The RStudio team has developed a number of “cheatsheets” for working with
both R and RStudio. This particular cheatsheet for “Base” R will summarize
many of the concepts in this document. (“Base” R is a name used to differentiate
the practice of using built-in R functions, as opposed to using functions from
outside packages, in particular, those from the tidyverse. More on this later.)
When programming, it is often a good practice to follow a style guide. (Where do
spaces go? Tabs or spaces? Underscores or CamelCase when naming variables?)
No style guide is “correct,” but it helps to be aware of what others do. The
more important thing is to be consistent within your own code.
• Hadley Wickham Style Guide from Advanced R

• Google Style Guide
For this course, our main deviation from these two guides is the use of = in place
of <-. (More on that later.)
2.2 Basic Calculations

To get started, we’ll use R like a simple calculator.
Addition, Subtraction, Multiplication, and Division
Math R Result
3+2 3 + 2 5
3−2 3 - 2 1
3⋅2 3 * 2 6
3/2 3 / 2 1.5
Exponents
2.3. GETTING HELP 17
Math R Result
32 3 ^ 2 9
2(−3) 2 ^ (-3) 0.125
1/2
100
√ 100 ^ (1 / 2) 10
100 sqrt(100) 10
Mathematical Constants
Math R Result
𝜋 pi 3.1415927
𝑒 exp(1) 2.7182818
Logarithms
Note that we will use ln and log interchangeably to mean the natural logarithm.
There is no ln() in R, instead it uses log() to mean the natural logarithm.
Math R Result
log(𝑒) log(exp(1)) 1
log10 (1000) log10(1000) 3
log2 (8) log2(8) 3
log4 (16) log(16, base = 4) 2
Trigonometry
Math R Result
sin(𝜋/2) sin(pi / 2) 1
cos(0) cos(0) 1
2.3 Getting Help

In using R as a calculator, we have seen a number of functions: sqrt(), exp(),
log() and sin(). To get documentation about a function in R, simply put
a question mark in front of the function name and RStudio will display the
documentation, for example:
?log
?sin
?paste
?lm
Frequently one of the most difficult things to do when learning R is asking for
help. First, you need to decide to ask for help, then you need to know how
to ask for help. Your very first line of defense should be to Google your error
message or a short description of your issue. (The ability to solve problems
using this method is quickly becoming an extremely valuable skill.) If that fails,
and it eventually will, you should ask for help. There are a number of things
you should include when emailing an instructor, or posting to a help website
such as Stack Exchange.
• Describe what you expect the code to do.

• State the end goal you are trying to achieve. (Sometimes what you expect
the code to do, is not what you want to actually do.)
• Provide the full text of any errors you have received.
• Provide enough code to recreate the error. Often for the purpose of this
course, you could simply email your entire .R or .Rmd file.
• Sometimes it is also helpful to include a screenshot of your entire RStudio
window when the error occurs.
If you follow these steps, you will get your issue resolved much quicker, and
possibly learn more in the process. Do not be discouraged by running into
errors and difficulties when learning R. (Or any technical skill.) It is simply part
of the learning process.
2.4 Installing Packages

R comes with a number of built-in functions and datasets, but one of the main
strengths of R as an open-source project is its package system. Packages add
additional functions and data. Frequently if you want to do something in R,
and it is not available by default, there is a good chance that there is a package
that will fulfill your needs.
To install a package, use the install.packages() function. Think of this as
buying a recipe book from the store, bringing it home, and putting it on your
shelf.
install.packages("ggplot2")
Once a package is installed, it must be loaded into your current R session before
being used. Think of this as taking the book off of the shelf and opening it up
to read.
2.4. INSTALLING PACKAGES 19
library(ggplot2)
Once you close R, all the packages are closed and put back on the imaginary
shelf. The next time you open R, you do not have to install the package again,
but you do have to load any packages you intend to use by invoking library().
Chapter 3
Data and Programming
3.1 Data Types

R has a number of basic data types.
• Numeric
– Also known as Double. The default type when dealing with numbers.
– Examples: 1, 1.0, 42.5
• Integer
– Examples: 1L, 2L, 42L
• Complex
– Example: 4 + 2i
• Logical
– Two possible values: TRUE and FALSE
– You can also use T and F, but this is not recommended.
– NA is also considered logical.
• Character
– Examples: "a", "Statistics", "1 plus 2."
3.2 Data Structures

R also has a number of basic data structures. A data structure is either homoge-
neous (all elements are of the same data type) or heterogeneous (elements can
be of more than one data type).
21
22 CHAPTER 3. DATA AND PROGRAMMING
Dimension Homogeneous Heterogeneous

1 Vector List
2 Matrix Data Frame
3+ Array
3.2.1 Vectors
Many operations in R make heavy use of vectors. Vectors in R are indexed

starting at 1. That is what the [1] in the output is indicating, that the first
element of the row being displayed is the first element of the vector. Larger
vectors will start additional rows with [*] where * is the index of the first
element of the row.
Possibly the most common way to create a vector in R is using the c() function,
which is short for “combine.” As the name suggests, it combines a list of elements
separated by commas.
c(1, 3, 5, 7, 8, 9)
## [1] 1 3 5 7 8 9
Here R simply outputs this vector. If we would like to store this vector in a
variable, we can do so with the assignment operator =. In this case the
variable x now holds the vector we just created, and we can access the vector
by typing x.
x = c(1, 3, 5, 7, 8, 9)
x
## [1] 1 3 5 7 8 9
As an aside, there is a long history of the assignment operator in R, partially

due to the keys available on the keyboards of the creators of the S language.
(Which preceded R.) For simplicity we will use =, but know that often you will
see <- as the assignment operator.
The pros and cons of these two are well beyond the scope of this book, but
know that for our purposes you will have no issue if you simply use =. If you
are interested in the weird cases where the difference matters, check out The R
Inferno.
If you wish to use <-, you will still need to use =, however only for argument
passing. Some users like to keep assignment (<-) and argument passing (=)
separate. No matter what you choose, the more important thing is that you
3.2. DATA STRUCTURES 23
stay consistent. Also, if working on a larger collaborative project, you should

use whatever style is already in place.
Because vectors must contain elements that are all the same type, R will au-
tomatically coerce to a single type when attempting to create a vector that
combines multiple types.
c(42, "Statistics", TRUE)
## [1] "42" "Statistics" "TRUE"
c(42, TRUE)
## [1] 42 1
Frequently you may wish to create a vector based on a sequence of numbers.

The quickest and easiest way to do this is with the : operator, which creates a
sequence of integers between two specified integers.
(y = 1:100)
## [1] 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
## [19] 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
## [37] 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54
## [55] 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
## [73] 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
## [91] 91 92 93 94 95 96 97 98 99 100
Here we see R labeling the rows after the first since this is a large vector. Also,
we see that by putting parentheses around the assignment, R both stores the
vector in a variable called y and automatically outputs y to the console.
Note that scalars do not exist in R. They are simply vectors of length 1.
## [1] 2
To create a sequence that is not limited to consecutive integers, use the seq()
function to define a sequence by its start, end, and increment.
seq(from = 1.5, to = 4.2, by = 0.1)
## [1] 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0 3.1 3.2 3.3
## [20] 3.4 3.5 3.6 3.7 3.8 3.9 4.0 4.1 4.2
We will discuss functions in detail later, but note here that the input labels
from, to, and by are optional.
seq(1.5, 4.2, 0.1)
## [1] 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0 3.1 3.2 3.3
## [20] 3.4 3.5 3.6 3.7 3.8 3.9 4.0 4.1 4.2
Another common operation to create a vector is rep(), which can repeat a

single value a number of times.
rep("A", times = 10)
## [1] "A" "A" "A" "A" "A" "A" "A" "A" "A" "A"
The rep() function can be used to repeat a vector some number of times.
rep(x, times = 3)
## [1] 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9
We have now seen four different ways to create vectors:
• c()
• :
• seq()
• rep()
So far we have mostly used them in isolation, but they are often used together.
c(x, rep(seq(1, 9, 2), 3), c(1, 2, 3), 42, 2:4)
## [1] 1 3 5 7 8 9 1 3 5 7 9 1 3 5 7 9 1 3 5 7 9 1 2 3 42
## [26] 2 3 4
The length of a vector can be obtained with the length() function.

length(x)
## [1] 6
length(y)
## [1] 100
3.2.1.1 Subsetting
To subset a vector, we use square brackets, [].

x
## [1] 1 3 5 7 8 9
x[1]
## [1] 1
x[3]
## [1] 5
We see that x[1] returns the first element, and x[3] returns the third element.
x[-2]
## [1] 1 5 7 8 9
We can also exclude certain indexes, in this case the second element.
x[1:3]
## [1] 1 3 5
x[c(1,3,4)]
## [1] 1 5 7
Lastly we see that we can subset based on a vector of indices.

All of the above are subsetting a vector using a vector of indexes. (Remember a
single number is still a vector.) We could instead use a vector of logical values.
z = c(TRUE, TRUE, FALSE, TRUE, TRUE, FALSE)

z
## [1] TRUE TRUE FALSE TRUE TRUE FALSE
x[z]
## [1] 1 3 7 8
3.2.2 Vectorization
One of the biggest strengths of R is its use of vectorized operations. (Frequently

the lack of understanding of this concept leads to a belief that R is slow. R is
not the fastest language, but it has a reputation for being slower than it really
is.)
x = 1:10
x + 1
## [1] 2 3 4 5 6 7 8 9 10 11
2 * x
## [1] 2 4 6 8 10 12 14 16 18 20
2 ^ x
## [1] 2 4 8 16 32 64 128 256 512 1024
sqrt(x)
## [1] 1.000000 1.414214 1.732051 2.000000 2.236068 2.449490 2.645751 2.828427

## [9] 3.000000 3.162278
log(x)
## [1] 0.0000000 0.6931472 1.0986123 1.3862944 1.6094379 1.7917595 1.9459101

## [8] 2.0794415 2.1972246 2.3025851
We see that when a function like log() is called on a vector x, a vector is

returned which has applied the function to each element of the vector x.
3.2.3 Logical Operators
Operator Summary Example Result

x < y x less than y 3 < 42 TRUE
x > y x greater than y 3 > 42 FALSE
x <= y x less than or equal to y 3 <= 42 TRUE
x >= y x greater than or equal to y 3 >= 42 FALSE
x == y xequal to y 3 == 42 FALSE
x != y x not equal to y 3 != 42 TRUE
!x not x !(3 > 42) TRUE
x | y x or y (3 > 42) | TRUE TRUE
x & y x and y (3 < 4) & ( 42 > 13) TRUE
In R, logical operators are vectorized.
x = c(1, 3, 5, 7, 8, 9)
x > 3
## [1] FALSE FALSE TRUE TRUE TRUE TRUE
x < 3
## [1] TRUE FALSE FALSE FALSE FALSE FALSE
x == 3
## [1] FALSE TRUE FALSE FALSE FALSE FALSE
x != 3
## [1] TRUE FALSE TRUE TRUE TRUE TRUE
x == 3 & x != 3
## [1] FALSE FALSE FALSE FALSE FALSE FALSE
x == 3 | x != 3
## [1] TRUE TRUE TRUE TRUE TRUE TRUE
This is extremely useful for subsetting.

x[x > 3]
## [1] 5 7 8 9
x[x != 3]
## [1] 1 5 7 8 9
sum(x > 3)
## [1] 4
as.numeric(x > 3)
## [1] 0 0 1 1 1 1
Here we see that using the sum() function on a vector of logical TRUE and FALSE
values that is the result of x > 3 results in a numeric result. R is first auto-
matically coercing the logical to numeric where TRUE is 1 and FALSE is 0. This
coercion from logical to numeric happens for most mathematical operations. If
you are interested in more detail, check out Advanced R.
which(x > 3)
## [1] 3 4 5 6
x[which(x > 3)]
## [1] 5 7 8 9
max(x)
## [1] 9
which(x == max(x))
## [1] 6
which.max(x)
## [1] 6
3.2.4 More Vectorization
x = c(1, 3, 5, 7, 8, 9)
y = 1:100
x + 2
## [1] 3 5 7 9 10 11
x + rep(2, 6)
## [1] 3 5 7 9 10 11
x > 3
x > rep(3, 6)
x + y
## Warning in x + y: longer object length is not a multiple of shorter object

## length
## [1] 2 5 8 11 13 15 8 11 14 17 19 21 14 17 20 23 25 27
## [19] 20 23 26 29 31 33 26 29 32 35 37 39 32 35 38 41 43 45
## [37] 38 41 44 47 49 51 44 47 50 53 55 57 50 53 56 59 61 63
## [55] 56 59 62 65 67 69 62 65 68 71 73 75 68 71 74 77 79 81
## [73] 74 77 80 83 85 87 80 83 86 89 91 93 86 89 92 95 97 99
## [91] 92 95 98 101 103 105 98 101 104 107
length(x)
## [1] 6
length(y)
## [1] 100
length(y) / length(x)
## [1] 16.66667
(x + y) - y
## Warning in x + y: longer object length is not a multiple of shorter object

## length
## [1] 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1
## [38] 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3
## [75] 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7 8 9 1 3 5 7
y = 1:60
x + y
## [1] 2 5 8 11 13 15 8 11 14 17 19 21 14 17 20 23 25 27 20 23 26 29 31 33 26
## [26] 29 32 35 37 39 32 35 38 41 43 45 38 41 44 47 49 51 44 47 50 53 55 57 50 53
## [51] 56 59 61 63 56 59 62 65 67 69
length(y) / length(x)
## [1] 10
rep(x, 10) + y
## [1] 2 5 8 11 13 15 8 11 14 17 19 21 14 17 20 23 25 27 20 23 26 29 31 33 26
## [26] 29 32 35 37 39 32 35 38 41 43 45 38 41 44 47 49 51 44 47 50 53 55 57 50 53
## [51] 56 59 61 63 56 59 62 65 67 69
all(x + y == rep(x, 10) + y)
## [1] TRUE
identical(x + y, rep(x, 10) + y)
## [1] TRUE
# ?any
# ?all.equal
x = c(1, 3, 5)
y = c(1, 2, 4)
x == y
## [1] TRUE FALSE FALSE
all(x == y)
## [1] FALSE
any(x == y)
## [1] TRUE
While all returns TRUE only when all of its arguments are TRUE, any returns
TRUE when at least one of its arguments is TRUE.
x = c(10 ^ (-8))
y = c(10 ^ (-9))
all(x == y)
## [1] FALSE
all.equal(x, y)
## [1] TRUE
The all.equal function tests “near equality” with a default tolerance value
around 1.5e-8 and returns TRUE if all of its arguments have differences smaller
than the tolerance.
3.2.5 Matrices
R can also be used for matrix calculations. Matrices have rows and columns
containing a single data type. In a matrix, the order of rows and columns is
important. (This is not true of data frames, which we will see later.)
Matrices can be created using the matrix function.
x = 1:9
x
## [1] 1 2 3 4 5 6 7 8 9
X = matrix(x, nrow = 3, ncol = 3)

X
## [,1] [,2] [,3]

## [1,] 1 4 7
## [2,] 2 5 8
## [3,] 3 6 9
Note here that we are using two different variables: lower case x, which stores a
vector and capital X, which stores a matrix. (Following the usual mathematical
convention.) We can do this because R is case sensitive.
By default the matrix function reorders a vector into columns, but we can also
tell R to use rows instead.
Y = matrix(x, nrow = 3, ncol = 3, byrow = TRUE)

Y
## [,1] [,2] [,3]

## [1,] 1 2 3
## [2,] 4 5 6
## [3,] 7 8 9
We can also create a matrix of a specified dimension where every element is the
same, in this case 0.
Z = matrix(0, 2, 4)
Z
## [,1] [,2] [,3] [,4]

## [1,] 0 0 0 0
## [2,] 0 0 0 0
Like vectors, matrices can be subsetted using square brackets, []. However,
since matrices are two-dimensional, we need to specify both a row and a column
when subsetting.
## [,1] [,2] [,3]

## [1,] 1 4 7
## [2,] 2 5 8
## [3,] 3 6 9
X[1, 2]
## [1] 4
Here we accessed the element in the first row and the second column. We could
also subset an entire row or column.
X[1, ]
## [1] 1 4 7
X[, 2]
## [1] 4 5 6
We can also use vectors to subset more than one row or column at a time. Here
we subset to the first and third column of the second row.
X[2, c(1, 3)]
## [1] 2 8
Matrices can also be created by combining vectors as columns, using cbind, or

combining vectors as rows, using rbind.
x = 1:9
rev(x)
## [1] 9 8 7 6 5 4 3 2 1
rep(1, 9)
## [1] 1 1 1 1 1 1 1 1 1
rbind(x, rev(x), rep(1, 9))
## [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9]

## x 1 2 3 4 5 6 7 8 9
## 9 8 7 6 5 4 3 2 1
## 1 1 1 1 1 1 1 1 1
cbind(col_1 = x, col_2 = rev(x), col_3 = rep(1, 9))
## col_1 col_2 col_3

## [1,] 1 9 1
## [2,] 2 8 1
## [3,] 3 7 1
## [4,] 4 6 1
## [5,] 5 5 1
## [6,] 6 4 1
## [7,] 7 3 1
## [8,] 8 2 1
## [9,] 9 1 1
When using rbind and cbind you can specify “argument” names that will be
used as column names.
R can then be used to perform matrix calculations.
x = 1:9
y = 9:1
X = matrix(x, 3, 3)
Y = matrix(y, 3, 3)
X
## [,1] [,2] [,3]

## [1,] 1 4 7
## [2,] 2 5 8
## [3,] 3 6 9
Y
## [,1] [,2] [,3]

## [1,] 9 6 3
## [2,] 8 5 2
## [3,] 7 4 1
X + Y
## [,1] [,2] [,3]

## [1,] 10 10 10
## [2,] 10 10 10
## [3,] 10 10 10
X - Y
## [,1] [,2] [,3]

## [1,] -8 -2 4
## [2,] -6 0 6
## [3,] -4 2 8
X * Y
## [,1] [,2] [,3]

## [1,] 9 24 21
## [2,] 16 25 16
## [3,] 21 24 9
X / Y
## [,1] [,2] [,3]

## [1,] 0.1111111 0.6666667 2.333333
## [2,] 0.2500000 1.0000000 4.000000
## [3,] 0.4285714 1.5000000 9.000000
Note that X * Y is not matrix multiplication. It is element by element mul-

tiplication. (Same for X / Y). Instead, matrix multiplication uses %*%. Other
matrix functions include t() which gives the transpose of a matrix and solve()
which returns the inverse of a square matrix if it is invertible.
X %*% Y
## [,1] [,2] [,3]

## [1,] 90 54 18
## [2,] 114 69 24
## [3,] 138 84 30
t(X)
## [,1] [,2] [,3]

## [1,] 1 2 3
## [2,] 4 5 6
## [3,] 7 8 9
Z = matrix(c(9, 2, -3, 2, 4, -2, -3, -2, 16), 3, byrow = TRUE)

Z
## [,1] [,2] [,3]

## [1,] 9 2 -3
## [2,] 2 4 -2
## [3,] -3 -2 16
solve(Z)
## [,1] [,2] [,3]

## [1,] 0.12931034 -0.05603448 0.01724138
## [2,] -0.05603448 0.29094828 0.02586207
## [3,] 0.01724138 0.02586207 0.06896552
To verify that solve(Z) returns the inverse, we multiply it by Z. We would

expect this to return the identity matrix, however we see that this is not the
case due to some computational issues. However, R also has the all.equal()
function which checks for equality, with some small tolerance which accounts
for some computational issues. The identical() function is used to check for
exact equality.
solve(Z) %*% Z
## [,1] [,2] [,3]

## [1,] 1.000000e+00 -6.245005e-17 0.000000e+00
## [2,] 8.326673e-17 1.000000e+00 5.551115e-17
## [3,] 2.775558e-17 0.000000e+00 1.000000e+00
diag(3)
## [,1] [,2] [,3]

## [1,] 1 0 0
## [2,] 0 1 0
## [3,] 0 0 1
all.equal(solve(Z) %*% Z, diag(3))
## [1] TRUE
R has a number of matrix specific functions for obtaining dimension and sum-
mary information.
X = matrix(1:6, 2, 3)
X
## [,1] [,2] [,3]

## [1,] 1 3 5
## [2,] 2 4 6
dim(X)
## [1] 2 3
rowSums(X)
## [1] 9 12
colSums(X)
## [1] 3 7 11
rowMeans(X)
## [1] 3 4
colMeans(X)
## [1] 1.5 3.5 5.5
The diag() function can be used in a number of ways. We can extract the
diagonal of a matrix.
diag(Z)
## [1] 9 4 16
Or create a matrix with specified elements on the diagonal. (And 0 on the

off-diagonals.)
diag(1:5)
## [,1] [,2] [,3] [,4] [,5]

## [1,] 1 0 0 0 0
## [2,] 0 2 0 0 0
## [3,] 0 0 3 0 0
## [4,] 0 0 0 4 0
## [5,] 0 0 0 0 5
Or, lastly, create a square matrix of a certain dimension with 1 for every element
of the diagonal and 0 for the off-diagonals.
diag(5)
## [,1] [,2] [,3] [,4] [,5]

## [1,] 1 0 0 0 0
## [2,] 0 1 0 0 0
## [3,] 0 0 1 0 0
## [4,] 0 0 0 1 0
## [5,] 0 0 0 0 1
Calculations with Vectors and Matrices
Certain operations in R, for example %*% have different behavior on vectors and
matrices. To illustrate this, we will first create two vectors.
a_vec = c(1, 2, 3)
b_vec = c(2, 2, 2)
Note that these are indeed vectors. They are not matrices.
c(is.vector(a_vec), is.vector(b_vec))
## [1] TRUE TRUE
c(is.matrix(a_vec), is.matrix(b_vec))
## [1] FALSE FALSE

When this is the case, the %*% operator is used to calculate the dot product,
also known as the inner product of the two vectors.
The dot product of vectors 𝑎 = [𝑎1 , 𝑎2 , ⋯ 𝑎𝑛 ] and 𝑏 = [𝑏1 , 𝑏2 , ⋯ 𝑏𝑛 ] is defined to
be
𝑛
𝑎 ⋅ 𝑏 = ∑ 𝑎𝑖 𝑏𝑖 = 𝑎1 𝑏1 + 𝑎2 𝑏2 + ⋯ 𝑎𝑛 𝑏𝑛 .
𝑖=1
a_vec %*% b_vec # inner product
## [,1]
## [1,] 12
a_vec %o% b_vec # outer product
## [,1] [,2] [,3]

## [1,] 2 2 2
## [2,] 4 4 4
## [3,] 6 6 6
The %o% operator is used to calculate the outer product of the two vectors.
When vectors are coerced to become matrices, they are column vectors. So a
vector of length 𝑛 becomes an 𝑛 × 1 matrix after coercion.
as.matrix(a_vec)
## [,1]
## [1,] 1
## [2,] 2
## [3,] 3
If we use the %*% operator on matrices, %*% again performs the expected matrix
multiplication. So you might expect the following to produce an error, because
the dimensions are incorrect.
as.matrix(a_vec) %*% b_vec
## [,1] [,2] [,3]

## [1,] 2 2 2
## [2,] 4 4 4
## [3,] 6 6 6
At face value this is a 3 × 1 matrix, multiplied by a 3 × 1 matrix. However,

when b_vec is automatically coerced to be a matrix, R decided to make it a “row
vector,” a 1 × 3 matrix, so that the multiplication has conformable dimensions.
If we had coerced both, then R would produce an error.
as.matrix(a_vec) %*% as.matrix(b_vec)
Another way to calculate a dot product is with the crossprod() function. Given
two vectors, the crossprod() function calculates their dot product. The func-
tion has a rather misleading name.
crossprod(a_vec, b_vec) # inner product
## [,1]
## [1,] 12
tcrossprod(a_vec, b_vec) # outer product
## [,1] [,2] [,3]

## [1,] 2 2 2
## [2,] 4 4 4
## [3,] 6 6 6
These functions could be very useful later. When used with matrices 𝑋 and 𝑌
as arguments, it calculates
𝑋⊤𝑌 .
When dealing with linear models, the calculation
𝑋⊤𝑋
is used repeatedly.
C_mat = matrix(c(1, 2, 3, 4, 5, 6), 2, 3)

D_mat = matrix(c(2, 2, 2, 2, 2, 2), 2, 3)
This is useful both as a shortcut for a frequent calculation and as a more efficient
implementation than using t() and %*%.
crossprod(C_mat, D_mat)
## [,1] [,2] [,3]

## [1,] 6 6 6
## [2,] 14 14 14
## [3,] 22 22 22
t(C_mat) %*% D_mat
## [,1] [,2] [,3]

## [1,] 6 6 6
## [2,] 14 14 14
## [3,] 22 22 22
all.equal(crossprod(C_mat, D_mat), t(C_mat) %*% D_mat)
## [1] TRUE
crossprod(C_mat, C_mat)
## [,1] [,2] [,3]

## [1,] 5 11 17
## [2,] 11 25 39
## [3,] 17 39 61
t(C_mat) %*% C_mat
## [,1] [,2] [,3]

## [1,] 5 11 17
## [2,] 11 25 39
## [3,] 17 39 61
all.equal(crossprod(C_mat, C_mat), t(C_mat) %*% C_mat)
## [1] TRUE
3.2.6 Lists
A list is a one-dimensional heterogeneous data structure. So it is indexed like a
vector with a single integer value, but each element can contain an element of
any type.
# creation
list(42, "Hello", TRUE)
## [[1]]
## [1] 42
##
## [[2]]
## [1] "Hello"
##
## [[3]]
## [1] TRUE
ex_list = list(
a = c(1, 2, 3, 4),
b = TRUE,
c = "Hello!",
d = function(arg = 42) {print("Hello World!")},
e = diag(5)
)
Lists can be subset using two syntaxes, the $ operator, and square brackets [].
The $ operator returns a named element of a list. The [] syntax returns a
list, while the [[]] returns an element of a list.
• ex_list[1] returns a list containing the first element.

• ex_list[[1]] returns the first element of the list, in this case, a vector.
# subsetting
ex_list$e
## [,1] [,2] [,3] [,4] [,5]

## [1,] 1 0 0 0 0
## [2,] 0 1 0 0 0
## [3,] 0 0 1 0 0
## [4,] 0 0 0 1 0
## [5,] 0 0 0 0 1
ex_list[1:2]
## $a
## [1] 1 2 3 4
##
## $b
## [1] TRUE
ex_list[1]
## $a
## [1] 1 2 3 4
ex_list[[1]]
## [1] 1 2 3 4
ex_list[c("e", "a")]
## $e
## [,1] [,2] [,3] [,4] [,5]
## [1,] 1 0 0 0 0
## [2,] 0 1 0 0 0
## [3,] 0 0 1 0 0
## [4,] 0 0 0 1 0
## [5,] 0 0 0 0 1
##
## $a
## [1] 1 2 3 4
ex_list["e"]
## $e
## [,1] [,2] [,3] [,4] [,5]
## [1,] 1 0 0 0 0
## [2,] 0 1 0 0 0
## [3,] 0 0 1 0 0
## [4,] 0 0 0 1 0
## [5,] 0 0 0 0 1
ex_list[["e"]]
## [,1] [,2] [,3] [,4] [,5]

## [1,] 1 0 0 0 0
## [2,] 0 1 0 0 0
## [3,] 0 0 1 0 0
## [4,] 0 0 0 1 0
## [5,] 0 0 0 0 1
ex_list$d
## function(arg = 42) {print("Hello World!")}
ex_list$d(arg = 1)
## [1] "Hello World!"
3.2.7 Data Frames
We have previously seen vectors and matrices for storing data as we introduced
R. We will now introduce a data frame which will be the most common way
that we store and interact with data in this course.
example_data = data.frame(x = c(1, 3, 5, 7, 9, 1, 3, 5, 7, 9),

y = c(rep("Hello", 9), "Goodbye"),
z = rep(c(TRUE, FALSE), 5))
Unlike a matrix, which can be thought of as a vector rearranged into rows and
columns, a data frame is not required to have the same data type for each
element. A data frame is a list of vectors. So, each vector must contain the
same data type, but the different vectors can store different data types.
example_data
## x y z
## 1 1 Hello TRUE
## 2 3 Hello FALSE
## 3 5 Hello TRUE
## 4 7 Hello FALSE
## 5 9 Hello TRUE
## 6 1 Hello FALSE
## 7 3 Hello TRUE
## 8 5 Hello FALSE
## 9 7 Hello TRUE
## 10 9 Goodbye FALSE
Unlike a list which has more flexibility, the elements of a data frame must all
be vectors, and have the same length.
example_data$x
## [1] 1 3 5 7 9 1 3 5 7 9
all.equal(length(example_data$x),
length(example_data$y),
length(example_data$z))
## [1] TRUE
str(example_data)
## 'data.frame': 10 obs. of 3 variables:

## $ x: num 1 3 5 7 9 1 3 5 7 9
## $ y: chr "Hello" "Hello" "Hello" "Hello" ...
## $ z: logi TRUE FALSE TRUE FALSE TRUE FALSE ...
nrow(example_data)
## [1] 10
ncol(example_data)
## [1] 3
dim(example_data)
## [1] 10 3
The data.frame() function above is one way to create a data frame. We can
also import data from various file types into R, as well as use data stored in
packages.
The example data above can also be found here as a .csv file. To read this
data into R, we would use the read_csv() function from the readr package.
Note that R has a built-in function read.csv() that operates very similarly.
The readr function read_csv() has a number of advantages. For example, it
is much faster reading larger data. It also uses the tibble package to read the
data as a tibble.
library(readr)
example_data_from_csv = read_csv("data/example-data.csv")
This particular line of code assumes that the file example_data.csv exists in a
folder called data in your current working directory.
example_data_from_csv
## # A tibble: 10 x 3
## x y z
## <dbl> <chr> <lgl>
## 1 1 Hello TRUE
## 2 3 Hello FALSE
## 3 5 Hello TRUE
## 4 7 Hello FALSE
## 5 9 Hello TRUE
## 6 1 Hello FALSE
## 7 3 Hello TRUE
## 8 5 Hello FALSE
## 9 7 Hello TRUE
A tibble is simply a data frame that prints with sanity. Notice in the output
above that we are given additional information such as dimension and variable
type.
The as_tibble() function can be used to coerce a regular data frame to a
tibble.
library(tibble)
example_data = as_tibble(example_data)
example_data
## # A tibble: 10 x 3
## x y z
## <dbl> <chr> <lgl>
## 1 1 Hello TRUE
## 2 3 Hello FALSE
## 3 5 Hello TRUE
## 4 7 Hello FALSE
## 5 9 Hello TRUE
## 6 1 Hello FALSE
## 7 3 Hello TRUE
## 8 5 Hello FALSE
## 9 7 Hello TRUE
Alternatively, we could use the “Import Dataset” feature in RStudio which

can be found in the environment window. (By default, the top-right pane of
RStudio.) Once completed, this process will automatically generate the code to
import a file. The resulting code will be shown in the console window. In recent
versions of RStudio, read_csv() is used by default, thus reading in a tibble.
Earlier we looked at installing packages, in particular the ggplot2 package.
(A package for visualization. While not necessary for this course, it is quickly
growing in popularity.)
library(ggplot2)
Inside the ggplot2 package is a dataset called mpg. By loading the package
using the library() function, we can now access mpg.
When using data from inside a package, there are three things we would gener-
ally like to do:
• Look at the raw data.

• Understand the data. (Where did it come from? What are the variables?
Etc.)
• Visualize the data.
To look at the data, we have two useful commands: head() and str().
head(mpg, n = 10)
## # A tibble: 10 x 11
## manufacturer model displ year cyl trans drv cty hwy fl class
## <chr> <chr> <dbl> <int> <int> <chr> <chr> <int> <int> <chr> <chr>
## 1 audi a4 1.8 1999 4 auto~ f 18 29 p comp~
## 2 audi a4 1.8 1999 4 manu~ f 21 29 p comp~
## 3 audi a4 2 2008 4 manu~ f 20 31 p comp~
## 4 audi a4 2 2008 4 auto~ f 21 30 p comp~
## 8 audi a4 quattro 1.8 1999 4 manu~ 4 18 26 p comp~
## 9 audi a4 quattro 1.8 1999 4 auto~ 4 16 25 p comp~
## 10 audi a4 quattro 2 2008 4 manu~ 4 20 28 p comp~
The function head() will display the first n observations of the data frame. The
head() function was more useful before tibbles. Notice that mpg is a tibble
already, so the output from head() indicates there are only 10 observations.
Note that this applies to head(mpg, n = 10) and not mpg itself. Also note that
tibbles print a limited number of rows and columns by default. The last line of
the printed output indicates which rows and columns were omitted.
mpg
## # A tibble: 234 x 11
## manufacturer model displ year cyl trans drv cty hwy fl class
## <chr> <chr> <dbl> <int> <int> <chr> <chr> <int> <int> <chr> <chr>
## 3 audi a4 2 2008 4 manu~ f 20 31 p comp~
## 4 audi a4 2 2008 4 auto~ f 21 30 p comp~
## 8 audi a4 quattro 1.8 1999 4 manu~ 4 18 26 p comp~
## 9 audi a4 quattro 1.8 1999 4 auto~ 4 16 25 p comp~
## 10 audi a4 quattro 2 2008 4 manu~ 4 20 28 p comp~
## # ... with 224 more rows
The function str() will display the “structure” of the data frame. It will display
the number of observations and variables, list the variables, give the type of
each variable, and show some elements of each variable. This information can
also be found in the “Environment” window in RStudio.
str(mpg)
## tibble [234 x 11] (S3: tbl_df/tbl/data.frame)

## $ manufacturer: chr [1:234] "audi" "audi" "audi" "audi" ...
## $ model : chr [1:234] "a4" "a4" "a4" "a4" ...
## $ displ : num [1:234] 1.8 1.8 2 2 2.8 2.8 3.1 1.8 1.8 2 ...
## $ year : int [1:234] 1999 1999 2008 2008 1999 1999 2008 1999 1999 2008 ...
## $ cyl : int [1:234] 4 4 4 4 6 6 6 4 4 4 ...
## $ trans : chr [1:234] "auto(l5)" "manual(m5)" "manual(m6)" "auto(av)" ...
## $ drv : chr [1:234] "f" "f" "f" "f" ...
## $ cty : int [1:234] 18 21 20 21 16 18 18 18 16 20 ...
## $ hwy : int [1:234] 29 29 31 30 26 26 27 26 25 28 ...
## $ fl : chr [1:234] "p" "p" "p" "p" ...
## $ class : chr [1:234] "compact" "compact" "compact" "compact" ...
It is important to note that while matrices have rows and columns, data frames
(tibbles) instead have observations and variables. When displayed in the console
or viewer, each row is an observation and each column is a variable. However
generally speaking, their order does not matter, it is simply a side-effect of how
the data was entered or stored.
In this dataset an observation is for a particular model-year of a car, and the
variables describe attributes of the car, for example its highway fuel efficiency.
To understand more about the data set, we use the ? operator to pull up the
documentation for the data.
?mpg
R has a number of functions for quickly working with and extracting basic in-
formation from data frames. To quickly obtain a vector of the variable names,
we use the names() function.
names(mpg)
## [1] "manufacturer" "model" "displ" "year" "cyl"

## [6] "trans" "drv" "cty" "hwy" "fl"
## [11] "class"
To access one of the variables as a vector, we use the $ operator.
mpg$year
## [1] 1999 1999 2008 2008 1999 1999 2008 1999 1999 2008 2008 1999 1999 2008 2008
## [16] 1999 2008 2008 2008 2008 2008 1999 2008 1999 1999 2008 2008 2008 2008 2008
## [31] 1999 1999 1999 2008 1999 2008 2008 1999 1999 1999 1999 2008 2008 2008 1999
## [46] 1999 2008 2008 2008 2008 1999 1999 2008 2008 2008 1999 1999 1999 2008 2008
## [61] 2008 1999 2008 1999 2008 2008 2008 2008 2008 2008 1999 1999 2008 1999 1999
## [76] 1999 2008 1999 1999 1999 2008 2008 1999 1999 1999 1999 1999 2008 1999 2008
## [91] 1999 1999 2008 2008 1999 1999 2008 2008 2008 1999 1999 1999 1999 1999 2008
## [106] 2008 2008 2008 1999 1999 2008 2008 1999 1999 2008 1999 1999 2008 2008 2008
## [121] 2008 2008 2008 2008 1999 1999 2008 2008 2008 2008 1999 2008 2008 1999 1999
## [136] 1999 2008 1999 2008 2008 1999 1999 1999 2008 2008 2008 2008 1999 1999 2008
## [151] 1999 1999 2008 2008 1999 1999 1999 2008 2008 1999 1999 2008 2008 2008 2008
## [166] 1999 1999 1999 1999 2008 2008 2008 2008 1999 1999 1999 1999 2008 2008 1999
## [181] 1999 2008 2008 1999 1999 2008 1999 1999 2008 2008 1999 1999 2008 1999 1999
## [196] 1999 2008 2008 1999 2008 1999 1999 2008 1999 1999 2008 2008 1999 1999 2008
## [211] 2008 1999 1999 1999 1999 2008 2008 2008 2008 1999 1999 1999 1999 1999 1999
## [226] 2008 2008 1999 1999 2008 2008 1999 1999 2008
mpg$hwy
## [1] 29 29 31 30 26 26 27 26 25 28 27 25 25 25 25 24 25 23 20 15 20 17 17 26 23
## [26] 26 25 24 19 14 15 17 27 30 26 29 26 24 24 22 22 24 24 17 22 21 23 23 19 18
## [51] 17 17 19 19 12 17 15 17 17 12 17 16 18 15 16 12 17 17 16 12 15 16 17 15 17
## [76] 17 18 17 19 17 19 19 17 17 17 16 16 17 15 17 26 25 26 24 21 22 23 22 20 33
## [101] 32 32 29 32 34 36 36 29 26 27 30 31 26 26 28 26 29 28 27 24 24 24 22 19 20
## [126] 17 12 19 18 14 15 18 18 15 17 16 18 17 19 19 17 29 27 31 32 27 26 26 25 25
## [151] 17 17 20 18 26 26 27 28 25 25 24 27 25 26 23 26 26 26 26 25 27 25 27 20 20
## [176] 19 17 20 17 29 27 31 31 26 26 28 27 29 31 31 26 26 27 30 33 35 37 35 15 18
## [201] 20 20 22 17 19 18 20 29 26 29 29 24 44 29 26 29 29 29 29 23 24 44 41 29 26
## [226] 28 29 29 29 28 29 26 26 26
We can use the dim(), nrow() and ncol() functions to obtain information
about the dimension of the data frame.
dim(mpg)
## [1] 234 11
nrow(mpg)
## [1] 234
ncol(mpg)
## [1] 11
Here nrow() is also the number of observations, which in most cases is the
sample size.
Subsetting data frames can work much like subsetting matrices using square
brackets, [,]. Here, we find fuel efficient vehicles earning over 35 miles per
gallon and only display manufacturer, model and year.
mpg[mpg$hwy > 35, c("manufacturer", "model", "year")]
## # A tibble: 6 x 3
## manufacturer model year
## <chr> <chr> <int>
## 1 honda civic 2008
## 2 honda civic 2008
3.3. PROGRAMMING BASICS 51
## 3 toyota corolla 2008

## 4 volkswagen jetta 1999
## 5 volkswagen new beetle 1999
## 6 volkswagen new beetle 1999
An alternative would be to use the subset() function, which has a much more
readable syntax.
subset(mpg, subset = hwy > 35, select = c("manufacturer", "model", "year"))
Lastly, we could use the filter and select functions from the dplyr package
which introduces the %>% operator from the magrittr package. This is not
necessary for this course, however the dplyr package is something you should
be aware of as it is becoming a popular tool in the R world.
library(dplyr)
mpg %>% filter(hwy > 35) %>% select(manufacturer, model, year)
All three approaches produce the same results. Which you use will be largely
based on a given situation as well as user preference.
When subsetting a data frame, be aware of what is being returned, as sometimes
it may be a vector instead of a data frame. Also note that there are differences
between subsetting a data frame and a tibble. A data frame operates more like
a matrix where it is possible to reduce the subset to a vector. A tibble operates
more like a list where it always subsets to another tibble.
3.3 Programming Basics
3.3.1 Control Flow
In R, the if/else syntax is:
if (...) {
some R code
} else {
more R code
}
For example,
x = 1
y = 3
if (x > y) {
z = x * y
print("x is larger than y")
} else {
z = x + 5 * y
print("x is less than or equal to y")
}
## [1] "x is less than or equal to y"
## [1] 16
R also has a special function ifelse() which is very useful. It returns one of
two specified values based on a conditional statement.
ifelse(4 > 3, 1, 0)
## [1] 1
The real power of ifelse() comes from its ability to be applied to vectors.
fib = c(1, 1, 2, 3, 5, 8, 13, 21)

ifelse(fib > 6, "Foo", "Bar")
## [1] "Bar" "Bar" "Bar" "Bar" "Bar" "Foo" "Foo" "Foo"
Now a for loop example,
x = 11:15
for (i in 1:5) {
x[i] = x[i] * 2
}
## [1] 22 24 26 28 30
Note that this for loop is very normal in many programming languages, but
not in R. In R we would not use a loop, instead we would simply use a vectorized
operation.
x = 11:15
x = x * 2
x
## [1] 22 24 26 28 30
3.3.2 Functions
So far we have been using functions, but haven’t actually discussed some of their
details.
function_name(arg1 = 10, arg2 = 20)
To use a function, you simply type its name, followed by an open parenthesis,
then specify values of its arguments, then finish with a closing parenthesis.
An argument is a variable which is used in the body of the function. Specifying
the values of the arguments is essentially providing the inputs to the function.
We can also write our own functions in R. For example, we often like to “stan-
dardize” variables, that is, subtracting the sample mean, and dividing by the
sample standard deviation.
𝑥 − 𝑥̄
𝑠
In R we would write a function to do this. When writing a function, there are

three things you must do.
• Give the function a name. Preferably something that is short, but de-
scriptive.
• Specify the arguments using function().
• Write the body of the function within curly braces, {}.
standardize = function(x) {
m = mean(x)
std = sd(x)
result = (x - m) / std
result
}
Here the name of the function is standardize, and the function has a single
argument x which is used in the body of function. Note that the output of
the final line of the body is what is returned by the function. In this case the
function returns the vector stored in the variable result.
To test our function, we will take a random sample of size n = 10 from a normal
distribution with a mean of 2 and a standard deviation of 5.
(test_sample = rnorm(n = 10, mean = 2, sd = 5))
## [1] 1.0011119 -0.8924812 1.3952092 -5.7941583 12.9114152 -1.5795204

## [7] 5.0909231 -7.6582372 9.1962799 3.7950510
standardize(x = test_sample)
## [1] -0.11793891 -0.41752851 -0.05558789 -1.19303383 1.76641685 -0.52622653

## [7] 0.52911929 -1.48795393 1.17863699 0.32409646
This function could be written much more succinctly, simply performing all the
operations on one line and immediately returning the result, without storing
any of the intermediate results.
standardize = function(x) {
(x - mean(x)) / sd(x)
}
When specifying arguments, you can provide default arguments.
power_of_num = function(num, power = 2) {

num ^ power
}
Let’s look at a number of ways that we could run this function to perform the
operation 10^2 resulting in 100.
power_of_num(10)
## [1] 100
power_of_num(10, 2)
## [1] 100
power_of_num(num = 10, power = 2)
## [1] 100
power_of_num(power = 2, num = 10)
## [1] 100
Note that without using the argument names, the order matters. The following
code will not evaluate to the same output as the previous example.
power_of_num(2, 10)
## [1] 1024
Also, the following line of code would produce an error since arguments without
a default value must be specified.
power_of_num(power = 5)
To further illustrate a function with a default argument, we will write a function

that calculates sample variance two ways.
By default, it will calculate the unbiased estimate of 𝜎2 , which we will call 𝑠2 .
𝑛
1
𝑠2 = ∑(𝑥 − 𝑥)̄ 2
𝑛 − 1 𝑖=1
It will also have the ability to return the biased estimate (based on maximum
likelihood) which we will call 𝜎̂ 2 .
1 𝑛
𝜎̂ 2 = ∑(𝑥 − 𝑥)̄ 2
𝑛 𝑖=1
get_var = function(x, biased = FALSE) {

n = length(x) - 1 * !biased
(1 / n) * sum((x - mean(x)) ^ 2)
}
get_var(test_sample)
## [1] 39.95028
get_var(test_sample, biased = FALSE)
## [1] 39.95028
var(test_sample)
## [1] 39.95028
We see the function is working as expected, and when returning the unbiased
estimate it matches R’s built-in function var(). Finally, let’s examine the biased
estimate of 𝜎2 .
get_var(test_sample, biased = TRUE)
## [1] 35.95525
Chapter 4
Summarizing Data
4.1 Summary Statistics

R has built-in functions for a large number of summary statistics. For numeric
variables, we can summarize data with the center and spread. We’ll again look
at the mpg dataset from the ggplot2 package.
library(ggplot2)
Central Tendency
Measure R Result
Mean mean(mpg$cty) 16.8589744
Median median(mpg$cty) 17
Spread
Measure R Result
Variance var(mpg$cty) 18.1130736
Standard Deviation sd(mpg$cty) 4.2559457
IQR IQR(mpg$cty) 5
Minimum min(mpg$cty) 9
Maximum max(mpg$cty) 35
Range range(mpg$cty) 9, 35
57
58 CHAPTER 4. SUMMARIZING DATA
Categorical
For categorical variables, counts and percentages can be used for summary.
table(mpg$drv)
##
## 4 f r
## 103 106 25
table(mpg$drv) / nrow(mpg)
##
## 4 f r
## 0.4401709 0.4529915 0.1068376
4.2 Plotting
Now that we have some data to work with, and we have learned about the
data at the most basic level, our next task is to visualize the data. Often, a
proper visualization can illuminate features of the data that can inform further
analysis.
We will look at four methods of visualizing data that we will use throughout
the course:
• Histograms
• Barplots
• Boxplots
• Scatterplots
4.2.1 Histograms
When visualizing a single numerical variable, a histogram will be our go-to

tool, which can be created in R using the hist() function.
hist(mpg$cty)
4.2. PLOTTING 59
Histogram of mpg$cty
80
60
Frequency
40
20
0
5 10 15 20 25 30 35
mpg$cty
The histogram function has a number of parameters which can be changed to

make our plot look much nicer. Use the ? operator to read the documentation
for the hist() to see a full list of these parameters.
hist(mpg$cty,
xlab = "Miles Per Gallon (City)",
main = "Histogram of MPG (City)",
breaks = 12,
col = "dodgerblue",
border = "darkorange")
Histogram of MPG (City)
40
30
Frequency
20
10
0
10 15 20 25 30 35
Miles Per Gallon (City)
Importantly, you should always be sure to label your axes and give the plot a
title. The argument breaks is specific to hist(). Entering an integer will give
a suggestion to R for how many bars to use for the histogram. By default R will
attempt to intelligently guess a good number of breaks, but as we can see here,
it is sometimes useful to modify this yourself.
4.2.2 Barplots
Somewhat similar to a histogram, a barplot can provide a visual summary of a

categorical variable, or a numeric variable with a finite number of values, like a
ranking from 1 to 10.
barplot(table(mpg$drv))
4.2. PLOTTING 61
100
80
60
40
20
0
4 f r
barplot(table(mpg$drv),
xlab = "Drivetrain (f = FWD, r = RWD, 4 = 4WD)",
ylab = "Frequency",
main = "Drivetrains",
col = "dodgerblue",
Drivetrains
100
80
Frequency
60
40
20
0
4 f r
Drivetrain (f = FWD, r = RWD, 4 = 4WD)
4.2.3 Boxplots
To visualize the relationship between a numerical and categorical variable, we

will use a boxplot. In the mpg dataset, the drv variable takes a small, finite
number of values. A car can only be front wheel drive, 4 wheel drive, or rear
wheel drive.
unique(mpg$drv)
## [1] "f" "4" "r"
First note that we can use a single boxplot as an alternative to a histogram for
visualizing a single numerical variable. To do so in R, we use the boxplot()
function.
boxplot(mpg$hwy)
4.2. PLOTTING 63
45
40
35
30
25
20
15
However, more often we will use boxplots to compare a numerical variable for
different values of a categorical variable.
boxplot(hwy ~ drv, data = mpg)

45
40
35
30
hwy
25
20
15
4 f r
drv
Here we used the boxplot() command to create side-by-side boxplots. However,

since we are now dealing with two variables, the syntax has changed. The R
syntax hwy ~ drv, data = mpg reads “Plot the hwy variable against the drv
variable using the dataset mpg.” We see the use of a ~ (which specifies a formula)
and also a data = argument. This will be a syntax that is common to many
functions we will use in this course.
boxplot(hwy ~ drv, data = mpg,

xlab = "Drivetrain (f = FWD, r = RWD, 4 = 4WD)",
ylab = "Miles Per Gallon (Highway)",
main = "MPG (Highway) vs Drivetrain",
pch = 20,
cex = 2,
col = "darkorange",
border = "dodgerblue")
MPG (Highway) vs Drivetrain

45
40
Miles Per Gallon (Highway)
35
30
25
20
15
4 f r
Drivetrain (f = FWD, r = RWD, 4 = 4WD)
Again, boxplot() has a number of additional arguments which have the ability
to make our plot more visually appealing.
4.2.4 Scatterplots
Lastly, to visualize the relationship between two numeric variables we will use a
scatterplot. This can be done with the plot() function and the ~ syntax we
4.2. PLOTTING 65
just used with a boxplot. (The function plot() can also be used more generally;
see the documentation for details.)
plot(hwy ~ displ, data = mpg)

45
40
35
30
hwy
25
20
15
2 3 4 5 6 7
displ
plot(hwy ~ displ, data = mpg,

xlab = "Engine Displacement (in Liters)",
ylab = "Miles Per Gallon (Highway)",
main = "MPG (Highway) vs Engine Displacement",
pch = 20,
cex = 2,
col = "dodgerblue")
MPG (Highway) vs Engine Displacement
45
40
Miles Per Gallon (Highway)
35
30
25
20
15
2 3 4 5 6 7
Engine Displacement (in Liters)

Chapter 5
Probability and Statistics in

R
5.1 Probability in R
5.1.1 Distributions
When working with different statistical distributions, we often want to make

probabilistic statements based on the distribution.
We typically want to know one of four things:
• The density (pdf) at a particular value.

• The distribution (cdf) at a particular value.
• The quantile value corresponding to a particular probability.
• A random draw of values from a particular distribution.
This used to be done with statistical tables printed in the back of textbooks.
Now, R has functions for obtaining density, distribution, quantile and random
values.
The general naming structure of the relevant R functions is:
• dname calculates density (pdf) at input x.

• pname calculates distribution (cdf) at input x.
• qname calculates the quantile at an input probability.
• rname generates a random draw from a particular distribution.
Note that name represents the name of the given distribution.
67
68 CHAPTER 5. PROBABILITY AND STATISTICS IN R
For example, consider a random variable 𝑋 which is 𝑁 (𝜇 = 2, 𝜎2 = 25). (Note,

we are parameterizing using the variance 𝜎2 . R however uses the standard
deviation.)
To calculate the value of the pdf at x = 3, that is, the height of the curve at x
= 3, use:
dnorm(x = 3, mean = 2, sd = 5)
## [1] 0.07820854
To calculate the value of the cdf at x = 3, that is, 𝑃 (𝑋 ≤ 3), the probability
that 𝑋 is less than or equal to 3, use:
pnorm(q = 3, mean = 2, sd = 5)
## [1] 0.5792597
Or, to calculate the quantile for probability 0.975, use:
qnorm(p = 0.975, mean = 2, sd = 5)
## [1] 11.79982
Lastly, to generate a random sample of size n = 10, use:
rnorm(n = 10, mean = 2, sd = 5)
## [1] 2.6951209 5.8123383 11.4235778 4.4769069 0.2720456 0.8973757

## [7] 5.4764981 4.5156775 8.5026252 -3.4875814
These functions exist for many other distributions, including but not limited to:
Command Distribution
*binom Binomial
*t t
*pois Poisson
*f F
*chisq Chi-Squared
Where * can be d, p, q, and r. Each distribution will have its own set of
parameters which need to be passed to the functions as arguments. For ex-
ample, dbinom() would not have arguments for mean and sd, since those are
5.2. HYPOTHESIS TESTS IN R 69
not parameters of the distribution. Instead a binomial distribution is usually

parameterized by 𝑛 and 𝑝, however R chooses to call them something else. To
find the names that R uses we would use ?dbinom and see that R instead calls
the arguments size and prob. For example:
dbinom(x = 6, size = 10, prob = 0.75)
## [1] 0.145998
Also note that, when using the dname functions with discrete distributions, they
are the pmf of the distribution. For example, the above command is 𝑃 (𝑌 = 6)
if 𝑌 ∼ 𝑏(𝑛 = 10, 𝑝 = 0.75). (The probability of flipping an unfair coin 10 times
and seeing 6 heads, if the probability of heads is 0.75.)
5.2 Hypothesis Tests in R

A prerequisite for STAT 420 is an understanding of the basics of hypothesis
testing. Recall the basic structure of hypothesis tests:
• An overall model and related assumptions are made. (The most common
being observations following a normal distribution.)
• The null (𝐻0 ) and alternative (𝐻1 or 𝐻𝐴 ) hypotheses are specified.
Usually the null specifies a particular value of a parameter.
• With given data, the value of the test statistic is calculated.
• Under the general assumptions, as well as assuming the null hypothesis is
true, the distribution of the test statistic is known.
• Given the distribution and value of the test statistic, as well as the form
of the alternative hypothesis, we can calculate a p-value of the test.
• Based on the p-value and pre-specified level of significance, we make a
decision. One of:
– Fail to reject the null hypothesis.
– Reject the null hypothesis.
We’ll do some quick review of two of the most common tests to show how they
are performed using R.
5.2.1 One Sample t-Test: Review
Suppose 𝑥𝑖 ∼ N(𝜇, 𝜎2 ) and we want to test 𝐻0 ∶ 𝜇 = 𝜇0 versus 𝐻1 ∶ 𝜇 ≠ 𝜇0 .

Assuming 𝜎 is unknown, we use the one-sample Student’s 𝑡 test statistic:
𝑥 ̄ − 𝜇0
𝑡= √ ∼ 𝑡𝑛−1 ,
𝑠/ 𝑛
𝑛 𝑛
∑ 𝑥𝑖 1
where 𝑥̄ = 𝑖=1 and 𝑠 = √ ∑(𝑥𝑖 − 𝑥)̄ 2 .
𝑛 𝑛 − 1 𝑖=1
A 100(1 − 𝛼)% confidence interval for 𝜇 is given by,
𝑠
𝑥̄ ± 𝑡𝛼/2,𝑛−1 √
𝑛
where 𝑡𝛼/2,𝑛−1 is the critical value such that 𝑃 (𝑡 > 𝑡𝛼/2,𝑛−1 ) = 𝛼/2 for 𝑛 − 1
degrees of freedom.
5.2.2 One Sample t-Test: Example
Suppose a grocery store sells “16 ounce” boxes of Captain Crisp cereal. A
random sample of 9 boxes was taken and weighed. The weight in ounces is
stored in the data frame capt_crisp.
capt_crisp = data.frame(weight = c(15.5, 16.2, 16.1, 15.8, 15.6, 16.0, 15.8, 15.9, 16.2
The company that makes Captain Crisp cereal claims that the average weight
of a box is at least 16 ounces. We will assume the weight of cereal in a box is
normally distributed and use a 0.05 level of significance to test the company’s
claim.
To test 𝐻0 ∶ 𝜇 ≥ 16 versus 𝐻1 ∶ 𝜇 < 16, the test statistic is
𝑥 ̄ − 𝜇0
𝑡= √
𝑠/ 𝑛
The sample mean 𝑥̄ and the sample standard deviation 𝑠 can be easily computed
using R. We also create variables which store the hypothesized mean and the
sample size.
x_bar = mean(capt_crisp$weight)
s = sd(capt_crisp$weight)
mu_0 = 16
n = 9
We can then easily compute the test statistic.

t = (x_bar - mu_0) / (s / sqrt(n))

t
## [1] -1.2
Under the null hypothesis, the test statistic has a 𝑡 distribution with 𝑛 − 1
degrees of freedom, in this case 8.
To complete the test, we need to obtain the p-value of the test. Since this is a
one-sided test with a less-than alternative, we need the area to the left of -1.2
for a 𝑡 distribution with 8 degrees of freedom. That is,
𝑃 (𝑡8 < −1.2)
pt(t, df = n - 1)
## [1] 0.1322336
We now have the p-value of our test, which is greater than our significance level
(0.05), so we fail to reject the null hypothesis.
Alternatively, this entire process could have been completed using one line of R
code.
t.test(x = capt_crisp$weight, mu = 16, alternative = c("less"), conf.level = 0.95)
##
## One Sample t-test
##
## data: capt_crisp$weight
## t = -1.2, df = 8, p-value = 0.1322
## alternative hypothesis: true mean is less than 16
## 95 percent confidence interval:
## -Inf 16.05496
## sample estimates:
## mean of x
## 15.9
We supply R with the data, the hypothesized value of 𝜇, the alternative, and
the confidence level. R then returns a wealth of information including:
• The value of the test statistic.

• The degrees of freedom of the distribution under the null hypothesis.
• The p-value of the test.

• The confidence interval which corresponds to the test.
• An estimate of 𝜇.
Since the test was one-sided, R returned a one-sided confidence interval. If

instead we wanted a two-sided interval for the mean weight of boxes of Captain
Crisp cereal we could modify our code.
capt_test_results = t.test(capt_crisp$weight, mu = 16,

alternative = c("two.sided"), conf.level = 0.95)
This time we have stored the results. By doing so, we can directly access portions
of the output from t.test(). To see what information is available we use the
names() function.
names(capt_test_results)
## [1] "statistic" "parameter" "p.value" "conf.int" "estimate"

## [6] "null.value" "stderr" "alternative" "method" "data.name"
We are interested in the confidence interval which is stored in conf.int.
capt_test_results$conf.int
## [1] 15.70783 16.09217

## attr(,"conf.level")
## [1] 0.95
Let’s check this interval “by hand.” The one piece of information we are missing
is the critical value, 𝑡𝛼/2,𝑛−1 = 𝑡8 (0.025), which can be calculated in R using the
qt() function.
qt(0.975, df = 8)
## [1] 2.306004
So, the 95% CI for the mean weight of a cereal box is calculated by plugging
into the formula,
𝑠
𝑥̄ ± 𝑡𝛼/2,𝑛−1 √
𝑛
c(mean(capt_crisp$weight) - qt(0.975, df = 8) * sd(capt_crisp$weight) / sqrt(9),

mean(capt_crisp$weight) + qt(0.975, df = 8) * sd(capt_crisp$weight) / sqrt(9))
## [1] 15.70783 16.09217
5.2.3 Two Sample t-Test: Review
Suppose 𝑥𝑖 ∼ N(𝜇𝑥 , 𝜎2 ) and 𝑦𝑖 ∼ N(𝜇𝑦 , 𝜎2 ).

Want to test 𝐻0 ∶ 𝜇𝑥 − 𝜇𝑦 = 𝜇0 versus 𝐻1 ∶ 𝜇𝑥 − 𝜇𝑦 ≠ 𝜇0 .
Assuming 𝜎 is unknown, use the two-sample Student’s 𝑡 test statistic:
(𝑥̄ − 𝑦)̄ − 𝜇0
𝑡= ∼ 𝑡𝑛+𝑚−2 ,
𝑠𝑝 √ 𝑛1 + 1
𝑚
𝑛 𝑚
∑𝑖=1 𝑥𝑖 ∑ 𝑦𝑖 (𝑛 − 1)𝑠2𝑥 + (𝑚 − 1)𝑠2𝑦
where 𝑥̄ = , 𝑦 ̄ = 𝑖=1 , and 𝑠2𝑝 = .
𝑛 𝑚 𝑛+𝑚−2
A 100(1 − 𝛼)% CI for 𝜇𝑥 − 𝜇𝑦 is given by
(𝑥̄ − 𝑦)̄ ± 𝑡𝛼/2,𝑛+𝑚−2 (𝑠𝑝 √ 𝑛1 + 1

𝑚) ,
where 𝑡𝛼/2,𝑛+𝑚−2 is the critical value such that 𝑃 (𝑡 > 𝑡𝛼/2,𝑛+𝑚−2 ) = 𝛼/2.
5.2.4 Two Sample t-Test: Example
Assume that the distributions of 𝑋 and 𝑌 are N(𝜇1 , 𝜎2 ) and N(𝜇2 , 𝜎2 ), respec-
tively. Given the 𝑛 = 6 observations of 𝑋,
x = c(70, 82, 78, 74, 94, 82)

n = length(x)
and the 𝑚 = 8 observations of 𝑌 ,
y = c(64, 72, 60, 76, 72, 80, 84, 68)

m = length(y)
we will test 𝐻0 ∶ 𝜇1 = 𝜇2 versus 𝐻1 ∶ 𝜇1 > 𝜇2 .

First, note that we can calculate the sample means and standard deviations.
x_bar = mean(x)
s_x = sd(x)
y_bar = mean(y)
s_y = sd(y)
We can then calculate the pooled standard deviation.
(𝑛 − 1)𝑠2𝑥 + (𝑚 − 1)𝑠2𝑦
𝑠𝑝 = √
𝑛+𝑚−2
s_p = sqrt(((n - 1) * s_x ^ 2 + (m - 1) * s_y ^ 2) / (n + m - 2))
Thus, the relevant 𝑡 test statistic is given by
(𝑥̄ − 𝑦)̄ − 𝜇0
𝑡= .
𝑠𝑝 √ 𝑛1 + 1
𝑚
t = ((x_bar - y_bar) - 0) / (s_p * sqrt(1 / n + 1 / m))

t
## [1] 1.823369
Note that 𝑡 ∼ 𝑡𝑛+𝑚−2 = 𝑡12 , so we can calculate the p-value, which is
𝑃 (𝑡12 > 1.8233692).
1 - pt(t, df = n + m - 2)
## [1] 0.04661961
But, then again, we could have simply performed this test in one line of R.
t.test(x, y, alternative = c("greater"), var.equal = TRUE)
##
## Two Sample t-test
##
## data: x and y
## t = 1.8234, df = 12, p-value = 0.04662
## alternative hypothesis: true difference in means is greater than 0

## 0.1802451 Inf
## mean of x mean of y
## 80 72
Recall that a two-sample 𝑡-test can be done with or without an equal variance
assumption. Here var.equal = TRUE tells R we would like to perform the test
under the equal variance assumption.
Above we carried out the analysis using two vectors x and y. In general, we will
have a preference for using data frames.
t_test_data = data.frame(values = c(x, y),

group = c(rep("A", length(x)), rep("B", length(y))))
We now have the data stored in a single variables (values) and have created a
second variable (group) which indicates which “sample” the value belongs to.
t_test_data
## values group
## 1 70 A
## 2 82 A
## 3 78 A
## 4 74 A
## 5 94 A
## 6 82 A
## 7 64 B
## 8 72 B
## 9 60 B
## 10 76 B
## 11 72 B
## 12 80 B
## 13 84 B
## 14 68 B
Now to perform the test, we still use the t.test() function but with the ~
syntax and a data argument.
t.test(values ~ group, data = t_test_data,

alternative = c("greater"), var.equal = TRUE)
##

##
## data: values by group
## t = 1.8234, df = 12, p-value = 0.04662
## alternative hypothesis: true difference in means between group A and group B is grea
## 0.1802451 Inf
## mean in group A mean in group B
## 80 72
5.3 Simulation
Simulation and model fitting are related but opposite processes.
• In simulation, the data generating process is known. We will know the

form of the model as well as the value of each of the parameters. In
particular, we will often control the distribution and parameters which
define the randomness, or noise in the data.
• In model fitting, the data is known. We will then assume a certain form
of model and find the best possible values of the parameters given the
observed data. Essentially we are seeking to uncover the truth. Often we
will attempt to fit many models, and we will learn metrics to assess which
model fits best.
Figure 5.1: Simulation vs Modeling
Often we will simulate data according to a process we decide, then use a model-
ing method seen in class. We can then verify how well the method works, since
we know the data generating process.
One of the biggest strengths of R is its ability to carry out simulations using
built-in functions for generating random samples from certain distributions.
5.3. SIMULATION 77
We’ll look at two very simple examples here, however simulation will be a topic
we revisit several times throughout the course.
5.3.1 Paired Differences
Consider the model:
𝑋11 , 𝑋12 , … , 𝑋1𝑛 ∼ 𝑁 (𝜇1 , 𝜎2 )

𝑋21 , 𝑋22 , … , 𝑋2𝑛 ∼ 𝑁 (𝜇2 , 𝜎2 )
Assume that 𝜇1 = 6, 𝜇2 = 5, 𝜎2 = 4 and 𝑛 = 25.

Let
1 𝑛
𝑋̄ 1 = ∑ 𝑋1𝑖
𝑛 𝑖=1
1 𝑛
𝑋̄ 2 = ∑ 𝑋2𝑖
𝑛 𝑖=1
𝐷 = 𝑋̄ 1 − 𝑋̄ 2 .
Suppose we would like to calculate 𝑃 (0 < 𝐷 < 2). First we will need to obtain
the distribution of 𝐷.
Recall,
2
𝜎
𝑋̄ 1 ∼ 𝑁 (𝜇1 , )
𝑛
and
2
𝜎
𝑋̄ 2 ∼ 𝑁 (𝜇2 , ) .
𝑛
Then,
𝜎2 𝜎2 4 4
𝐷 = 𝑋̄ 1 − 𝑋̄ 2 ∼ 𝑁 (𝜇1 − 𝜇2 , + ) = 𝑁 (6 − 5, + ).
𝑛 𝑛 25 25
So,
𝐷 ∼ 𝑁 (𝜇 = 1, 𝜎2 = 0.32).
Thus,
𝑃 (0 < 𝐷 < 2) = 𝑃 (𝐷 < 2) − 𝑃 (𝐷 < 0).
This can then be calculated using R without a need to first standardize, or use
a table.
pnorm(2, mean = 1, sd = sqrt(0.32)) - pnorm(0, mean = 1, sd = sqrt(0.32))
## [1] 0.9229001
An alternative approach, would be to simulate a large number of observations

of 𝐷 then use the empirical distribution to calculate the probability.
Our strategy will be to repeatedly:
• Generate a sample of 25 random observations from 𝑁 (𝜇1 = 6, 𝜎2 = 4).

Call the mean of this sample 𝑥1𝑠̄ .
• Generate a sample of 25 random observations from 𝑁 (𝜇1 = 5, 𝜎2 = 4).
Call the mean of this sample 𝑥2𝑠 ̄ .
• Calculate the differences in the means, 𝑑𝑠 = 𝑥1𝑠
̄ − 𝑥2𝑠
̄ .
We will repeat the process a large number of times. Then we will use the
distribution of the simulated observations of 𝑑𝑠 as an estimate for the true
distribution of 𝐷.
set.seed(42)
num_samples = 10000
differences = rep(0, num_samples)
Before starting our for loop to perform the operation, we set a seed for repro-
ducibility, create and set a variable num_samples which will define the number
of repetitions, and lastly create a variables differences which will store the
simulate values, 𝑑𝑠 .
By using set.seed() we can reproduce the random results of rnorm() each
time starting from that line.
for (s in 1:num_samples) {
x1 = rnorm(n = 25, mean = 6, sd = 2)
x2 = rnorm(n = 25, mean = 5, sd = 2)
differences[s] = mean(x1) - mean(x2)
}
5.3. SIMULATION 79
To estimate 𝑃 (0 < 𝐷 < 2) we will find the proportion of values of 𝑑𝑠 (among

the 104 values of 𝑑𝑠 generated) that are between 0 and 2.
mean(0 < differences & differences < 2)
## [1] 0.9222
Recall that above we derived the distribution of 𝐷 to be 𝑁 (𝜇 = 1, 𝜎2 = 0.32)

If we look at a histogram of the differences, we find that it looks very much like
a normal distribution.
hist(differences, breaks = 20,

main = "Empirical Distribution of D",
xlab = "Simulated Values of D",
col = "dodgerblue",
Empirical Distribution of D
1400
1000
Frequency
600
200
0
-1 0 1 2 3
Simulated Values of D
Also the sample mean and variance are very close to what we would expect.
mean(differences)
## [1] 1.001423
var(differences)
## [1] 0.3230183
We could have also accomplished this task with a single line of more “idiomatic”
R.
set.seed(42)
diffs = replicate(10000, mean(rnorm(25, 6, 2)) - mean(rnorm(25, 5, 2)))
Use ?replicate to take a look at the documentation for the replicate function
and see if you can understand how this line performs the same operations that
our for loop above executed.
mean(differences == diffs)
## [1] 1
We see that by setting the same seed for the randomization, we actually obtain
identical results!
5.3.2 Distribution of a Sample Mean
For another example of simulation, we will simulate observations from a Poisson

distribution, and examine the empirical distribution of the sample mean of these
observations.
Recall, if
𝑋 ∼ 𝑃 𝑜𝑖𝑠(𝜇)
then
𝐸[𝑋] = 𝜇
and
𝑉 𝑎𝑟[𝑋] = 𝜇.
Also, recall that for a random variable 𝑋 with finite mean 𝜇 and finite variance
𝜎2 , the central limit theorem tells us that the mean, 𝑋̄ of a random sample of
size 𝑛 is approximately normal for large values of 𝑛. Specifically, as 𝑛 → ∞,
5.3. SIMULATION 81
2
𝑑 𝜎
𝑋̄ → 𝑁 (𝜇, ) .
𝑛
The following verifies this result for a Poisson distribution with 𝜇 = 10 and a
sample size of 𝑛 = 50.
set.seed(1337)
mu = 10
sample_size = 50
samples = 100000
x_bars = rep(0, samples)
for(i in 1:samples){
x_bars[i] = mean(rpois(sample_size, lambda = mu))
}
x_bar_hist = hist(x_bars, breaks = 50,

main = "Histogram of Sample Means",
xlab = "Sample Means")
Histogram of Sample Means

8000
6000
Frequency
4000
2000
0
8 9 10 11 12
Sample Means
Now we will compare sample statistics from the empirical distribution with their
known values based on the parent distribution.
c(mean(x_bars), mu)
## [1] 10.00008 10.00000
c(var(x_bars), mu / sample_size)
## [1] 0.1989732 0.2000000
c(sd(x_bars), sqrt(mu) / sqrt(sample_size))
## [1] 0.4460641 0.4472136
And here, we will calculate the proportion of sample means that are within 2
standard deviations of the population mean.
mean(x_bars > mu - 2 * sqrt(mu) / sqrt(sample_size) &

x_bars < mu + 2 * sqrt(mu) / sqrt(sample_size))
## [1] 0.95429
This last histogram uses a bit of a trick to approximately shade the bars that
are within two standard deviations of the mean.)
shading = ifelse(x_bar_hist$breaks > mu - 2 * sqrt(mu) / sqrt(sample_size) &

x_bar_hist$breaks < mu + 2 * sqrt(mu) / sqrt(sample_size),
"darkorange", "dodgerblue")
x_bar_hist = hist(x_bars, breaks = 50, col = shading,

main = "Histogram of Sample Means, Two Standard Deviations",
xlab = "Sample Means")
5.3. SIMULATION 83
Histogram of Sample Means, Two Standard Deviations

8000
6000
Frequency
4000
2000
0
8 9 10 11 12
Sample Means
Chapter 6
R Resources
So far, we have seen a lot of R, and a lot of R quickly. Again, the preceding
chapters were in no way meant to be a complete reference for the R language,
but rather an introduction to many of the concepts we will need in this text.
The following resources are not necessary for the remainder of this text, but you
may find them useful if you would like a deeper understanding of R:
6.1 Beginner Tutorials and References

• Try R from Code School.
– An interactive introduction to the basics of R. Useful for getting up
to speed on R’s syntax.
• Quick-R by Robert Kabacoff.
– A good reference for R basics.
• R Tutorial by Chi Yau.
– A combination reference and tutorial for R basics.
• R Programming for Data Science by Roger Peng
– A great text for R programming beginners. Discusses R from the
ground up, highlighting programming details we might not discuss.
6.2 Intermediate References

• R for Data Science by Hadley Wickham and Garrett Grolemund.
85
86 CHAPTER 6. R RESOURCES
– Similar to Advanced R, but focuses more on data analysis, while still

introducing programming concepts. Especially useful for working in
the tidyverse.
• The Art of R Programming by Norman Matloff.
– Gentle introduction to the programming side of R. (Whereas we will
focus more on the data analysis side.) A free electronic version is
available through the Illinois library.
6.3 Advanced References

• Advanced R by Hadley Wickham.
– From the author of several extremely popular R packages. Good
follow-up to The Art of R Programming. (And more up-to-date ma-
terial.)
• The R Inferno by Patrick Burns.
– Likens learning the tricks of R to descending through the levels of
hell. Very advanced material, but may be important if R becomes a
part of your everyday toolkit.
• Efficient R Programming by Colin Gillespie and Robin Lovelace
– Discusses both efficient R programs, as well as programming in R
efficiently.
6.4 Quick Comparisons to Other Languages

Those who are familiar with other languages may find the following “cheat-
sheets” helpful for transitioning to R.
• MATLAB, NumPy, Julia

• Python pandas comparison with R/R Libraries
• Stata
• SAS - Look for a resource still! Suggestions welcome.
6.5 RStudio and RMarkdown Videos

The following video playlists were made as an introduction to R, RStudio, and
RMarkdown for STAT 420 at UIUC. If you are currently using this text for a
Coursera course, you can also find updated videos there.
6.6. RMARKDOWN TEMPLATE 87
• R and RStudio
• Data in R
• RMarkdown
Note that RStudio and RMarkdown are constantly receiving excellent support
and updates, so these videos may already contain some outdated information.
RStudio provides their own tutorial for RMarkdown. They also have an excellent
RStudio “cheatsheet” which visually identifies many of the features available in
the IDE. You may also explore some additional “cheatsheets” on their website.
6.6 RMarkdown Template

This .zip file contains the files necessary to produce this rendered document.
This document is a more complete version of a template than what is seen in
the above videos.
88 CHAPTER 6. R RESOURCES
Chapter 7
Simple Linear Regression
“All models are wrong, but some are useful.”

— George E. P. Box
After reading this chapter you will be able to:
• Understand the concept of a model.

• Describe two ways in which regression coefficients are derived.
• Estimate and visualize a regression model using R.
• Interpret regression coefficients and statistics in the context of real-world
problems.
• Use a regression model to make predictions.
7.1 Modeling
Let’s consider a simple example of how the speed of a car affects its stopping
distance, that is, how far it travels before it comes to a stop. To examine this
relationship, we will use the cars dataset, which is a default R dataset. Thus,
we don’t need to load a package first; it is immediately available.
To get a first look at the data you can use the View() function inside RStudio.
View(cars)
We could also take a look at the variable names, the dimension of the data
frame, and some sample observations with str().
89
90 CHAPTER 7. SIMPLE LINEAR REGRESSION
str(cars)

## $ speed: num 4 4 7 7 8 9 10 10 10 11 ...
## $ dist : num 2 10 4 22 16 10 18 26 34 17 ...
As we have seen before with data frames, there are a number of additional
functions to access some of this information directly.
dim(cars)
## [1] 50 2
nrow(cars)
## [1] 50
ncol(cars)
## [1] 2
Other than the two variable names and the number of observations, this data
is still just a bunch of numbers, so we should probably obtain some context.
?cars
Reading the documentation we learn that this is data gathered during the 1920s
about the speed of cars and the resulting distance it takes for the car to come
to a stop. The interesting task here is to determine how far a car travels before
stopping, when traveling at a certain speed. So, we will first plot the stopping
distance against the speed.
plot(dist ~ speed, data = cars,

xlab = "Speed (in Miles Per Hour)",
ylab = "Stopping Distance (in Feet)",
main = "Stopping Distance vs Speed",
pch = 20,
cex = 2,
col = "grey")
7.1. MODELING 91
Stopping Distance vs Speed

100 120
Stopping Distance (in Feet)
80
60
40
20
0
5 10 15 20 25
Speed (in Miles Per Hour)
Let’s now define some terminology. We have pairs of data, (𝑥𝑖 , 𝑦𝑖 ), for 𝑖 =
1, 2, … 𝑛, where 𝑛 is the sample size of the dataset.
We use 𝑖 as an index, simply for notation. We use 𝑥𝑖 as the predictor (ex-
planatory) variable. The predictor variable is used to help predict or explain
the response (target, outcome) variable, 𝑦𝑖 .
Other texts may use the term independent variable instead of predictor and
dependent variable in place of response. However, those monikers imply math-
ematical characteristics that might not be true. While these other terms are
not incorrect, independence is already a strictly defined concept in probability.
For example, when trying to predict a person’s weight given their height, would
it be accurate to say that height is independent of weight? Certainly not, but
that is an unintended implication of saying “independent variable.” We prefer
to stay away from this nomenclature.
In the cars example, we are interested in using the predictor variable speed to
predict and explain the response variable dist.
Broadly speaking, we would like to model the relationship between 𝑋 and 𝑌
using the form
𝑌 = 𝑓(𝑋) + 𝜖.
The function 𝑓 describes the functional relationship between the two variables,
and the 𝜖 term is used to account for error. This indicates that if we plug in a
given value of 𝑋 as input, our output is a value of 𝑌 , within a certain range of

error. You could think of this a number of ways:
• Response = Prediction + Error

• Response = Signal + Noise
• Response = Model + Unexplained
• Response = Deterministic + Random
• Response = Explainable + Unexplainable
What sort of function should we use for 𝑓(𝑋) for the cars data?
We could try to model the data with a horizontal line. That is, the model
for 𝑦 does not depend on the value of 𝑥. (Some function 𝑓(𝑋) = 𝑐.) In the
plot below, we see this doesn’t seem to do a very good job. Many of the data
points are very far from the orange line representing 𝑐. This is an example of
underfitting. The obvious fix is to make the function 𝑓(𝑋) actually depend
on 𝑥.

100 120
80
60
40
20
0
5 10 15 20 25
We could also try to model the data with a very “wiggly” function that tries to
go through as many of the data points as possible. This also doesn’t seem to
work very well. The stopping distance for a speed of 5 mph shouldn’t be off the
chart! (Even in 1920.) This is an example of overfitting. (Note that in this
example no function will go through every point, since there are some 𝑥 values
that have several possible 𝑦 values in the data.)
7.1. MODELING 93

100 120
80
60
40
20
0
5 10 15 20 25
Lastly, we could try to model the data with a well-chosen line rather than one
of the two extremes previously attempted. The line on the plot below seems to
summarize the relationship between stopping distance and speed quite well. As
speed increases, the distance required to come to a stop increases. There is still
some variation about this line, but it seems to capture the overall trend.
100 120
80
60
40
20
0
5 10 15 20 25
With this in mind, we would like to restrict our choice of 𝑓(𝑋) to linear functions
of 𝑋. We will write our model using 𝛽1 for the slope, and 𝛽0 for the intercept,
𝑌 = 𝛽0 + 𝛽1 𝑋 + 𝜖.
7.1.1 Simple Linear Regression Model
We now define what we will call the simple linear regression model,
𝑌 𝑖 = 𝛽 0 + 𝛽 1 𝑥𝑖 + 𝜖 𝑖
where
𝜖𝑖 ∼ 𝑁 (0, 𝜎2 ).
That is, the 𝜖𝑖 are independent and identically distributed (iid) normal random
variables with mean 0 and variance 𝜎2 . This model has three parameters to be
estimated: 𝛽0 , 𝛽1 , and 𝜎2 , which are fixed, but unknown constants.
We have slightly modified our notation here. We are now using 𝑌𝑖 and 𝑥𝑖 , since
we will be fitting this model to a set of 𝑛 data points, for 𝑖 = 1, 2, … 𝑛.
7.1. MODELING 95
Recall that we use capital 𝑌 to indicate a random variable, and lower case
𝑦 to denote a potential value of the random variable. Since we will have 𝑛
observations, we have 𝑛 random variables 𝑌𝑖 and their possible values 𝑦𝑖 .
In the simple linear regression model, the 𝑥𝑖 are assumed to be fixed, known
constants, and are thus notated with a lower case variable. The response 𝑌𝑖
remains a random variable because of the random behavior of the error vari-
able, 𝜖𝑖 . That is, each response 𝑌𝑖 is tied to an observable 𝑥𝑖 and a random,
unobservable, 𝜖𝑖 .
Essentially, we could explicitly think of the 𝑌𝑖 as having a different distribution
for each 𝑋𝑖 . In other words, 𝑌𝑖 has a conditional distribution dependent on the
value of 𝑋𝑖 , written 𝑥𝑖 . Doing so, we still make no distributional assumptions of
the 𝑋𝑖 , since we are only interested in the distribution of the 𝑌𝑖 for a particular
value 𝑥𝑖 .
𝑌𝑖 ∣ 𝑋𝑖 ∼ 𝑁 (𝛽0 + 𝛽1 𝑥𝑖 , 𝜎2 )
The random 𝑌𝑖 are a function of 𝑥𝑖 , thus we can write its mean as a function of
𝑥𝑖 ,
E[𝑌𝑖 ∣ 𝑋𝑖 = 𝑥𝑖 ] = 𝛽0 + 𝛽1 𝑥𝑖 .
However, its variance remains constant for each 𝑥𝑖 ,
Var[𝑌𝑖 ∣ 𝑋𝑖 = 𝑥𝑖 ] = 𝜎2 .
This is visually displayed in the image below. We see that for any value 𝑥, the
expected value of 𝑌 is 𝛽0 + 𝛽1 𝑥. At each value of 𝑥, 𝑌 has the same variance
𝜎2 .
Often, we directly talk about the assumptions that this model makes. They can
be cleverly shortened to LINE.
• Linear. The relationship between 𝑌 and 𝑥 is linear, of the form 𝛽0 + 𝛽1 𝑥.

• Independent. The errors 𝜖 are independent.
• Normal. The errors, 𝜖 are normally distributed. That is the “error”
around the line follows a normal distribution.
• Equal Variance. At each value of 𝑥, the variance of 𝑌 is the same, 𝜎2 .
We are also assuming that the values of 𝑥 are fixed, that is, not random. We
do not make a distributional assumption about the predictor variable.
As a side note, we will often refer to simple linear regression as SLR. Some
explanation of the name SLR:
Figure 7.1: Simple Linear Regression Model Introductory Statistics (Shafer and
Zhang), UC Davis Stat Wiki
• Simple refers to the fact that we are using a single predictor variable.
Later we will use multiple predictor variables.
• Linear tells us that our model for 𝑌 is a linear combination of the pre-
dictors 𝑋. (In this case just the one.) Right now, this always results in a
model that is a line, but later we will see how this is not always the case.
• Regression simply means that we are attempting to measure the rela-
tionship between a response variable and (one or more) predictor variables.
In the case of SLR, both the response and the predictor are numeric vari-
ables.
So SLR models 𝑌 as a linear function of 𝑋, but how do we actually define a

good line? There are an infinite number of lines we could use, so we will attempt
to find one with “small errors.” That is a line with as many points as close to
it as possible. The question now becomes, how do we find such a line? There
are many approaches we could take.
We could find the line that has the smallest maximum distance from any of the
points to the line. That is,
argmin max |𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖 )|.

𝛽0 ,𝛽1
We could find the line that minimizes the sum of all the distances from the
points to the line. That is,
𝑛
argmin ∑ |𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖 )|.
𝛽0 ,𝛽1 𝑖=1
7.2. LEAST SQUARES APPROACH 97
We could find the line that minimizes the sum of all the squared distances from
the points to the line. That is,
𝑛
argmin ∑(𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖 ))2 .
𝛽0 ,𝛽1 𝑖=1
This last option is called the method of least squares. It is essentially the
de-facto method for fitting a line to data. (You may have even seen it before
in a linear algebra course.) Its popularity is largely due to the fact that it is
mathematically “easy.” (Which was important historically, as computers are a
modern contraption.) It is also very popular because many relationships are
well approximated by a linear function.
7.2 Least Squares Approach

Given observations (𝑥𝑖 , 𝑦𝑖 ), for 𝑖 = 1, 2, … 𝑛, we want to find values of 𝛽0 and
𝛽1 which minimize
𝑛 𝑛
𝑓(𝛽0 , 𝛽1 ) = ∑(𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖 ))2 = ∑(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 )2 .
𝑖=1 𝑖=1
We will call these values 𝛽0̂ and 𝛽1̂ .

First, we take a partial derivative with respect to both 𝛽0 and 𝛽1 .
𝑛
𝜕𝑓
= −2 ∑(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 )
𝜕𝛽0 𝑖=1
𝑛
𝜕𝑓
= −2 ∑(𝑥𝑖 )(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 )
𝜕𝛽1 𝑖=1
We then set each of the partial derivatives equal to zero and solve the resulting
system of equations.
𝑛
∑(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 ) = 0
𝑖=1
𝑛
∑(𝑥𝑖 )(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 ) = 0
𝑖=1
While solving the system of equations, one common algebraic rearrangement

results in the normal equations.
𝑛 𝑛
𝑛𝛽0 + 𝛽1 ∑ 𝑥𝑖 = ∑ 𝑦𝑖
𝑖=1 𝑖=1
𝑛 𝑛 𝑛
𝛽0 ∑ 𝑥𝑖 + 𝛽1 ∑ 𝑥2𝑖 = ∑ 𝑥𝑖 𝑦𝑖
𝑖=1 𝑖=1 𝑖=1
Finally, we finish solving the system of equations.
𝑛 𝑛
𝑛 (∑𝑖=1 𝑥𝑖 )(∑𝑖=1 𝑦𝑖 )
∑𝑖=1 𝑥𝑖 𝑦𝑖 − 𝑆𝑥𝑦
𝛽1̂ = 𝑛
𝑛
(∑𝑖=1 𝑥𝑖 )2
=
𝑛 2
∑𝑖=1 𝑥𝑖 − 𝑆𝑥𝑥
𝑛
𝛽0̂ = 𝑦 ̄ − 𝛽1̂ 𝑥̄
Here, we have defined some notation for the expression we’ve obtained. Note
that they have alternative forms which are much easier to work with. (We won’t
do it here, but you can try to prove the equalities below on your own, for “fun”).
We use the capital letter 𝑆 to denote “summation” which replaces the capital
letter Σ when we calculate these values based on observed data, (𝑥𝑖 , 𝑦𝑖 ). The
subscripts such as 𝑥𝑦 denote over which variables the function (𝑧 − 𝑧)̄ is applied.
𝑛 𝑛 𝑛
𝑛
(∑𝑖=1 𝑥𝑖 )(∑𝑖=1 𝑦𝑖 )
𝑆𝑥𝑦 = ∑ 𝑥𝑖 𝑦𝑖 − = ∑(𝑥𝑖 − 𝑥)(𝑦
̄ 𝑖 − 𝑦)̄
𝑖=1
𝑛 𝑖=1
𝑛 𝑛 𝑛
(∑𝑖=1 𝑥𝑖 )2
𝑆𝑥𝑥 = ∑ 𝑥2𝑖 − = ∑(𝑥𝑖 − 𝑥)̄ 2
𝑖=1
𝑛 𝑖=1
𝑛 𝑛 𝑛
(∑𝑖=1 𝑦𝑖 )2
𝑆𝑦𝑦 = ∑ 𝑦𝑖2 − = ∑(𝑦𝑖 − 𝑦)̄ 2
𝑖=1
𝑛 𝑖=1
Note that these summations 𝑆 are not to be confused with sample standard
deviation 𝑠.
By using the above alternative expressions for 𝑆𝑥𝑦 and 𝑆𝑥𝑥 , we arrive at a
cleaner, more useful expression for 𝛽1̂ .
𝑛
𝑆𝑥𝑦 ∑ (𝑥𝑖 − 𝑥)(𝑦
̄ 𝑖 − 𝑦)̄
𝛽1̂ = = 𝑖=1 𝑛
𝑆𝑥𝑥 ∑𝑖=1 (𝑥𝑖 − 𝑥)̄ 2
Traditionally we would now calculate 𝛽0̂ and 𝛽1̂ by hand for the cars dataset.
However because we are living in the 21st century and are intelligent (or lazy
or efficient, depending on your perspective), we will utilize R to do the number
crunching for us.
To keep some notation consistent with the above mathematics, we will store the
response variable as y and the predictor variable as x.
x = cars$speed
y = cars$dist
We then calculate the three sums of squares defined above.
Sxy = sum((x - mean(x)) * (y - mean(y)))

Sxx = sum((x - mean(x)) ^ 2)
Syy = sum((y - mean(y)) ^ 2)
c(Sxy, Sxx, Syy)
## [1] 5387.40 1370.00 32538.98
Then finally calculate 𝛽0̂ and 𝛽1̂ .
beta_1_hat = Sxy / Sxx

beta_0_hat = mean(y) - beta_1_hat * mean(x)
c(beta_0_hat, beta_1_hat)
## [1] -17.579095 3.932409
What do these values tell us about our dataset?

The slope parameter 𝛽1 tells us that for an increase in speed of one mile per
hour, the mean stopping distance increases by 𝛽1 . It is important to specify
that we are talking about the mean. Recall that 𝛽0 + 𝛽1 𝑥 is the mean of 𝑌 , in
this case stopping distance, for a particular value of 𝑥. (In this case speed.) So
𝛽1 tells us how the mean of 𝑌 is affected by a change in 𝑥.
Similarly, the estimate 𝛽1̂ = 3.93 tells us that for an increase in speed of one mile
per hour, the estimated mean stopping distance increases by 3.93 feet. Here
we should be sure to specify we are discussing an estimated quantity. Recall
that 𝑦 ̂ is the estimated mean of 𝑌 , so 𝛽1̂ tells us how the estimated mean of 𝑌
is affected by changing 𝑥.
The intercept parameter 𝛽0 tells us the mean stopping distance for a car trav-
eling zero miles per hour. (Not moving.) The estimate 𝛽0̂ = −17.58 tells us
that the estimated mean stopping distance for a car traveling zero miles per
hour is −17.58 feet. So when you apply the brakes to a car that is not moving,
it moves backwards? This doesn’t seem right. (Extrapolation, which we will see
later, is the issue here.)
7.2.1 Making Predictions
We can now write the fitted or estimated line,

𝑦 ̂ = 𝛽0̂ + 𝛽1̂ 𝑥.
In this case,
𝑦 ̂ = −17.58 + 3.93𝑥.
We can now use this line to make predictions. First, let’s see the possible 𝑥
values in the cars dataset. Since some 𝑥 values may appear more than once,
we use the unique() to return each unique value only once.
unique(cars$speed)
## [1] 4 7 8 9 10 11 12 13 14 15 16 17 18 19 20 22 23 24 25
Let’s make a prediction for the stopping distance of a car traveling at 8 miles
per hour.
𝑦 ̂ = −17.58 + 3.93 × 8
beta_0_hat + beta_1_hat * 8
## [1] 13.88018
This tells us that the estimated mean stopping distance of a car traveling at 8
miles per hour is 13.88.
Now let’s make a prediction for the stopping distance of a car traveling at 21
miles per hour. This is considered interpolation as 21 is not an observed value
of 𝑥. (But is in the data range.) We can use the special %in% operator to quickly
verify this in R.
8 %in% unique(cars$speed)
## [1] TRUE
21 %in% unique(cars$speed)
## [1] FALSE
min(cars$speed) < 21 & 21 < max(cars$speed)
## [1] TRUE
𝑦 ̂ = −17.58 + 3.93 × 21
## [1] 65.00149
Lastly, we can make a prediction for the stopping distance of a car traveling at
50 miles per hour. This is considered extrapolation as 50 is not an observed
value of 𝑥 and is outside data range. We should be less confident in predictions
of this type.
range(cars$speed)
## [1] 4 25
range(cars$speed)[1] < 50 & 50 < range(cars$speed)[2]
## [1] FALSE
𝑦 ̂ = −17.58 + 3.93 × 50
## [1] 179.0413
Cars travel 50 miles per hour rather easily today, but not in the 1920s!
This is also an issue we saw when interpreting 𝛽0̂ = −17.58, which is equivalent
to making a prediction at 𝑥 = 0. We should not be confident in the estimated
linear relationship outside of the range of data we have observed.
7.2.2 Residuals
If we think of our model as “Response = Prediction + Error,” we can then write

it as
𝑦 = 𝑦 ̂ + 𝑒.
We then define a residual to be the observed value minus the predicted value.
𝑒𝑖 = 𝑦𝑖 − 𝑦𝑖̂
Let’s calculate the residual for the prediction we made for a car traveling 8 miles
per hour. First, we need to obtain the observed value of 𝑦 for this 𝑥 value.
which(cars$speed == 8)
## [1] 5
cars[5, ]
## speed dist
## 5 8 16
cars[which(cars$speed == 8), ]
## speed dist
## 5 8 16
We can then calculate the residual.
𝑒 = 16 − 13.88 = 2.12
16 - (beta_0_hat + beta_1_hat * 8)
## [1] 2.119825
The positive residual value indicates that the observed stopping distance is
actually 2.12 feet more than what was predicted.
7.2.3 Variance Estimation
We’ll now use the residuals for each of the points to create an estimate for the
variance, 𝜎2 .
Recall that,
E[𝑌𝑖 ∣ 𝑋𝑖 = 𝑥𝑖 ] = 𝛽0 + 𝛽1 𝑥𝑖 .
So,
𝑦𝑖̂ = 𝛽0̂ + 𝛽1̂ 𝑥𝑖
is a natural estimate for the mean of 𝑌𝑖 for a given value of 𝑥𝑖 .

Also, recall that when we specified the model, we had three unknown parame-
ters; 𝛽0 , 𝛽1 , and 𝜎2 . The method of least squares gave us estimates for 𝛽0 and
𝛽1 , however, we have yet to see an estimate for 𝜎2 . We will now define 𝑠2𝑒 which
will be an estimate for 𝜎2 .
𝑛
1
𝑠2𝑒 = ∑(𝑦𝑖 − (𝛽0̂ + 𝛽1̂ 𝑥𝑖 ))2
𝑛 − 2 𝑖=1
𝑛
1
= ∑(𝑦𝑖 − 𝑦𝑖̂ )2
𝑛 − 2 𝑖=1
𝑛
1
= ∑ 𝑒2𝑖
𝑛 − 2 𝑖=1
This probably seems like a natural estimate, aside from the use of 𝑛 − 2, which
we will put off explaining until the next chapter. It should actually look rather
similar to something we have seen before.
𝑛
1
𝑠2 = ∑(𝑥𝑖 − 𝑥)̄ 2
𝑛 − 1 𝑖=1
Here, 𝑠2 is the estimate of 𝜎2 when we have a single random variable 𝑋. In this

case 𝑥̄ is an estimate of 𝜇 which is assumed to be the same for each 𝑥.
Now, in the regression case, with 𝑠2𝑒 each 𝑦 has a different mean because of the
relationship with 𝑥. Thus, for each 𝑦𝑖 , we use a different estimate of the mean,
that is 𝑦𝑖̂ .
y_hat = beta_0_hat + beta_1_hat * x

e = y - y_hat
n = length(e)
s2_e = sum(e^2) / (n - 2)
s2_e
## [1] 236.5317
Just as with the univariate measure of variance, this value of 236.53 doesn’t
have a practical interpretation in terms of stopping distance. Taking the square
root, however, computes the standard deviation of the residuals, also known as
residual standard error.
s_e = sqrt(s2_e)
s_e
## [1] 15.37959
This tells us that our estimates of mean stopping distance are “typically” off by
15.38 feet.
7.3 Decomposition of Variation

We can re-express 𝑦𝑖 − 𝑦,̄ which measures the deviation of an observation from
the sample mean, in the following way,
𝑦𝑖 − 𝑦 ̄ = (𝑦𝑖 − 𝑦𝑖̂ ) + (𝑦𝑖̂ − 𝑦).

̄
This is the common mathematical trick of “adding zero.” In this case we both
added and subtracted 𝑦𝑖̂ .
Here, 𝑦𝑖 − 𝑦𝑖̂ measures the deviation of an observation from the fitted regression
line and 𝑦𝑖̂ − 𝑦 ̄ measures the deviation of the fitted regression line from the
sample mean.
If we square then sum both sides of the equation above, we can obtain the
following,
𝑛 𝑛 𝑛
∑(𝑦𝑖 − 𝑦)̄ 2 = ∑(𝑦𝑖 − 𝑦𝑖̂ )2 + ∑(𝑦𝑖̂ − 𝑦)̄ 2 .
𝑖=1 𝑖=1 𝑖=1
This should be somewhat alarming or amazing. How is this true? For now we
will leave this question unanswered. (Think about this, and maybe try to prove
it.) We will now define three of the quantities seen in this equation.
7.3. DECOMPOSITION OF VARIATION 105
Sum of Squares Total

𝑛
SST = ∑(𝑦𝑖 − 𝑦)̄ 2
𝑖=1
The quantity “Sum of Squares Total,” or SST, represents the total variation
of the observed 𝑦 values. This should be a familiar-looking expression. Note
that,
𝑛
1 1
𝑠2 = ∑(𝑦𝑖 − 𝑦)̄ 2 = SST.
𝑛 − 1 𝑖=1 𝑛−1
Sum of Squares Regression

𝑛
SSReg = ∑(𝑦𝑖̂ − 𝑦)̄ 2
𝑖=1
The quantity “Sum of Squares Regression,” SSReg, represents the explained

variation of the observed 𝑦 values.
Sum of Squares Error

𝑛
SSE = RSS = ∑(𝑦𝑖 − 𝑦𝑖̂ )2
𝑖=1
The quantity “Sum of Squares Error,” SSE, represents the unexplained vari-
ation of the observed 𝑦 values. You will often see SSE written as RSS, or
“Residual Sum of Squares.”
SST = sum((y - mean(y)) ^ 2)

SSReg = sum((y_hat - mean(y)) ^ 2)
SSE = sum((y - y_hat) ^ 2)
c(SST = SST, SSReg = SSReg, SSE = SSE)
## SST SSReg SSE

## 32538.98 21185.46 11353.52
Note that,
SSE
𝑠2𝑒 = .
𝑛−2
SSE / (n - 2)
## [1] 236.5317
We can use R to verify that this matches our previous calculation of 𝑠2𝑒 .
s2_e == SSE / (n - 2)
## [1] TRUE
These three measures also do not have an important practical interpretation in-
dividually. But together, they’re about to reveal a new statistic to help measure
the strength of an SLR model.
7.3.1 Coefficient of Determination
The coefficient of determination, 𝑅2 , is defined as
𝑛
SSReg ∑ (𝑦𝑖̂ − 𝑦)̄ 2
𝑅2 = = 𝑖=1
𝑛
SST ∑𝑖=1 (𝑦𝑖 − 𝑦)̄ 2
SST − SSE SSE

= =1−
SST SST
𝑛 𝑛
∑𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2 ∑𝑖=1 𝑒2𝑖
=1− 𝑛 =1− 𝑛
∑𝑖=1 (𝑦𝑖 − 𝑦)̄ 2 ∑𝑖=1 (𝑦𝑖 − 𝑦)̄ 2
The coefficient of determination is interpreted as the proportion of observed

variation in 𝑦 that can be explained by the simple linear regression model.
R2 = SSReg / SST
R2
## [1] 0.6510794
For the cars example, we calculate 𝑅2 = 0.65. We then say that 65% of the
observed variability in stopping distance is explained by the linear relationship
with speed.
The following plots visually demonstrate the three “sums of squares” for a sim-
ulated dataset which has 𝑅2 = 0.92 which is a somewhat high value. Notice
in the final plot, that the orange arrows account for a larger proportion of the
total arrow.
7.3. DECOMPOSITION OF VARIATION 107
SSReg (Sum of Squares Regression) SSE (Sum of Squares Error)

25
25
20
20
15
15
y
y
10
10
5
5
0
0
0 2 4 6 8 10 0 2 4 6 8 10
x x
SST (Sum of Squares Total) SST (Sum of Squares Total)

25
25
20
20
15
15
y
y
10
10
5
5
0
0 2 4 6 8 10 0 2 4 6 8 10
x x
The next plots again visually demonstrate the three “sums of squares,” this time
for a simulated dataset which has 𝑅2 = 0.19. Notice in the final plot, that now
the blue arrows account for a larger proportion of the total arrow.
SSReg (Sum of Squares Regression) SSE (Sum of Squares Error)
10
10
0
0
-10
-10
y
y
-20
-20
-30
-30
0 2 4 6 8 10 0 2 4 6 8 10
x x
SST (Sum of Squares Total) SST (Sum of Squares Total)

10
10
0
0
-10
-10
y
y
-20
-20
-30
-30
0 2 4 6 8 10 0 2 4 6 8 10
x x
7.4 The lm Function
So far we have done regression by deriving the least squares estimates, then
writing simple R commands to perform the necessary calculations. Since this is
such a common task, this is functionality that is built directly into R via the
lm() command.
The lm() command is used to fit linear models which actually account for
a broader class of models than simple linear regression, but we will use SLR
as our first demonstration of lm(). The lm() function will be one of our most
commonly used tools, so you may want to take a look at the documentation by
using ?lm. You’ll notice there is a lot of information there, but we will start
with just the very basics. This is documentation you will want to return to
often.
7.4. THE LM FUNCTION 109
We’ll continue using the cars data, and essentially use the lm() function to
check the work we had previously done.
stop_dist_model = lm(dist ~ speed, data = cars)
This line of code fits our very first linear model. The syntax should look some-
what familiar. We use the dist ~ speed syntax to tell R we would like to model
the response variable dist as a linear function of the predictor variable speed.
In general, you should think of the syntax as response ~ predictor. The
data = cars argument then tells R that the dist and speed variables are from
the dataset cars. We then store this result in a variable stop_dist_model.
The variable stop_dist_model now contains a wealth of information, and we
will now see how to extract and use that information. The first thing we will do is
simply output whatever is stored immediately in the variable stop_dist_model.
stop_dist_model
##
## Call:
## lm(formula = dist ~ speed, data = cars)
##
## Coefficients:
## (Intercept) speed
## -17.579 3.932
We see that it first tells us the formula we input into R, that is lm(formula
= dist ~ speed, data = cars). We also see the coefficients of the model.
We can check that these are what we had calculated previously. (Minus some
rounding that R is doing when displaying the results. They are stored with full
precision.)
c(beta_0_hat, beta_1_hat)
## [1] -17.579095 3.932409
Next, it would be nice to add the fitted line to the scatterplot. To do so we will
use the abline() function.

pch = 20,
cex = 2,
col = "grey")
abline(stop_dist_model, lwd = 3, col = "darkorange")

100 120
80
60
40
20
0
5 10 15 20 25
The abline() function is used to add lines of the form 𝑎 + 𝑏𝑥 to a plot. (Hence
abline.) When we give it stop_dist_model as an argument, it automatically
extracts the regression coefficient estimates (𝛽0̂ and 𝛽1̂ ) and uses them as the
intercept and slope of the line. Here we also use lwd to modify the width of the
line, as well as col to modify the color of the line.
The “thing” that is returned by the lm() function is actually an object of class
lm which is a list. The exact details of this are unimportant unless you are
seriously interested in the inner-workings of R, but know that we can determine
the names of the elements of the list using the names() command.
names(stop_dist_model)
## [1] "coefficients" "residuals" "effects" "rank"

## [5] "fitted.values" "assign" "qr" "df.residual"
## [9] "xlevels" "call" "terms" "model"
We can then use this information to, for example, access the residuals using the
$ operator.
stop_dist_model$residuals
## 1 2 3 4 5 6 7
## 3.849460 11.849460 -5.947766 12.052234 2.119825 -7.812584 -3.744993
## 8 9 10 11 12 13 14
## 4.255007 12.255007 -8.677401 2.322599 -15.609810 -9.609810 -5.609810
## 15 16 17 18 19 20 21
## -1.609810 -7.542219 0.457781 0.457781 12.457781 -11.474628 -1.474628
## 22 23 24 25 26 27 28
## 22.525372 42.525372 -21.407036 -15.407036 12.592964 -13.339445 -5.339445
## 29 30 31 32 33 34 35
## -17.271854 -9.271854 0.728146 -11.204263 2.795737 22.795737 30.795737
## 36 37 38 39 40 41 42
## -21.136672 -11.136672 10.863328 -29.069080 -13.069080 -9.069080 -5.069080
## 43 44 45 46 47 48 49
## 2.930920 -2.933898 -18.866307 -6.798715 15.201285 16.201285 43.201285
## 50
## 4.268876
Another way to access stored information in stop_dist_model are the coef(),

resid(), and fitted() functions. These return the coefficients, residuals, and
fitted values, respectively.
coef(stop_dist_model)
## (Intercept) speed
## -17.579095 3.932409
resid(stop_dist_model)
## 1 2 3 4 5 6 7
## 3.849460 11.849460 -5.947766 12.052234 2.119825 -7.812584 -3.744993
## 8 9 10 11 12 13 14
## 4.255007 12.255007 -8.677401 2.322599 -15.609810 -9.609810 -5.609810
## 15 16 17 18 19 20 21
## -1.609810 -7.542219 0.457781 0.457781 12.457781 -11.474628 -1.474628
## 22 23 24 25 26 27 28
## 22.525372 42.525372 -21.407036 -15.407036 12.592964 -13.339445 -5.339445
## 29 30 31 32 33 34 35
## -17.271854 -9.271854 0.728146 -11.204263 2.795737 22.795737 30.795737
## 36 37 38 39 40 41 42
## -21.136672 -11.136672 10.863328 -29.069080 -13.069080 -9.069080 -5.069080
## 43 44 45 46 47 48 49
## 2.930920 -2.933898 -18.866307 -6.798715 15.201285 16.201285 43.201285

## 50
## 4.268876
fitted(stop_dist_model)
## 1 2 3 4 5 6 7 8
## -1.849460 -1.849460 9.947766 9.947766 13.880175 17.812584 21.744993 21.744993
## 9 10 11 12 13 14 15 16
## 21.744993 25.677401 25.677401 29.609810 29.609810 29.609810 29.609810 33.542219
## 17 18 19 20 21 22 23 24
## 33.542219 33.542219 33.542219 37.474628 37.474628 37.474628 37.474628 41.407036
## 25 26 27 28 29 30 31 32
## 41.407036 41.407036 45.339445 45.339445 49.271854 49.271854 49.271854 53.204263
## 33 34 35 36 37 38 39 40
## 53.204263 53.204263 53.204263 57.136672 57.136672 57.136672 61.069080 61.069080
## 41 42 43 44 45 46 47 48
## 61.069080 61.069080 61.069080 68.933898 72.866307 76.798715 76.798715 76.798715
## 49 50
## 76.798715 80.731124
An R function that is useful in many situations is summary(). We see that when

it is called on our model, it returns a good deal of information. By the end
of the course, you will know what every value here is used for. For now, you
should immediately notice the coefficient estimates, and you may recognize the
𝑅2 value we saw earlier.
summary(stop_dist_model)
##
## Call:
##
## Residuals:
## Min 1Q Median 3Q Max
## -29.069 -9.525 -2.272 9.215 43.201
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -17.5791 6.7584 -2.601 0.0123 *
## speed 3.9324 0.4155 9.464 1.49e-12 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 15.38 on 48 degrees of freedom

## Multiple R-squared: 0.6511, Adjusted R-squared: 0.6438
## F-statistic: 89.57 on 1 and 48 DF, p-value: 1.49e-12
The summary() command also returns a list, and we can again use names() to
learn what about the elements of this list.
names(summary(stop_dist_model))
## [1] "call" "terms" "residuals" "coefficients"

## [5] "aliased" "sigma" "df" "r.squared"
## [9] "adj.r.squared" "fstatistic" "cov.unscaled"
So, for example, if we wanted to directly access the value of 𝑅2 , instead of copy
and pasting it out of the printed statement from summary(), we could do so.
summary(stop_dist_model)$r.squared
## [1] 0.6510794
Another value we may want to access is 𝑠𝑒 , which R calls sigma.
summary(stop_dist_model)$sigma
## [1] 15.37959
Note that this is the same result seen earlier as s_e. You may also notice that
this value was displayed above as a result of the summary() command, which R
labeled the “Residual Standard Error.”
𝑛
1
𝑠𝑒 = RSE = √ ∑ 𝑒2𝑖
𝑛 − 2 𝑖=1
Often it is useful to talk about 𝑠𝑒 (or RSE) instead of 𝑠2𝑒 because of their units.
The unit of 𝑠𝑒 in the cars example is feet, while the unit of 𝑠2𝑒 is feet-squared.
Another useful function, which we will use almost as often as lm() is the
predict() function.
predict(stop_dist_model, newdata = data.frame(speed = 8))
## 1
## 13.88018
The above code reads “predict the stopping distance of a car traveling 8 miles
per hour using the stop_dist_model.” Importantly, the second argument to
predict() is a data frame that we make in place. We do this so that we can
specify that 8 is a value of speed, so that predict knows how to use it with
the model stored in stop_dist_model. We see that this result is what we had
calculated “by hand” previously.
We could also predict multiple values at once.
predict(stop_dist_model, newdata = data.frame(speed = c(8, 21, 50)))
## 1 2 3
## 13.88018 65.00149 179.04134
𝑦 ̂ = −17.58 + 3.93 × 8 = 13.88

𝑦 ̂ = −17.58 + 3.93 × 21 = 65
𝑦 ̂ = −17.58 + 3.93 × 50 = 179.04
Or we could calculate the fitted value for each of the original data points. We
can simply supply the original data frame, cars, since it contains a variables
called speed which has the values we would like to predict at.
predict(stop_dist_model, newdata = cars)
## 1 2 3 4 5 6 7 8
## -1.849460 -1.849460 9.947766 9.947766 13.880175 17.812584 21.744993 21.744993
## 9 10 11 12 13 14 15 16
## 21.744993 25.677401 25.677401 29.609810 29.609810 29.609810 29.609810 33.542219
## 17 18 19 20 21 22 23 24
## 33.542219 33.542219 33.542219 37.474628 37.474628 37.474628 37.474628 41.407036
## 25 26 27 28 29 30 31 32
## 41.407036 41.407036 45.339445 45.339445 49.271854 49.271854 49.271854 53.204263
## 33 34 35 36 37 38 39 40
## 53.204263 53.204263 53.204263 57.136672 57.136672 57.136672 61.069080 61.069080
## 41 42 43 44 45 46 47 48
## 61.069080 61.069080 61.069080 68.933898 72.866307 76.798715 76.798715 76.798715
## 49 50
## 76.798715 80.731124
# predict(stop_dist_model, newdata = data.frame(speed = cars$speed))
This is actually equivalent to simply calling predict() on stop_dist_model

without a second argument.
7.5. MAXIMUM LIKELIHOOD ESTIMATION (MLE) APPROACH 115
predict(stop_dist_model)
## 1 2 3 4 5 6 7 8
## -1.849460 -1.849460 9.947766 9.947766 13.880175 17.812584 21.744993 21.744993
## 9 10 11 12 13 14 15 16
## 21.744993 25.677401 25.677401 29.609810 29.609810 29.609810 29.609810 33.542219
## 17 18 19 20 21 22 23 24
## 33.542219 33.542219 33.542219 37.474628 37.474628 37.474628 37.474628 41.407036
## 25 26 27 28 29 30 31 32
## 41.407036 41.407036 45.339445 45.339445 49.271854 49.271854 49.271854 53.204263
## 33 34 35 36 37 38 39 40
## 53.204263 53.204263 53.204263 57.136672 57.136672 57.136672 61.069080 61.069080
## 41 42 43 44 45 46 47 48
## 61.069080 61.069080 61.069080 68.933898 72.866307 76.798715 76.798715 76.798715
## 49 50
## 76.798715 80.731124
Note that then in this case, this is the same as using fitted().
fitted(stop_dist_model)
## 1 2 3 4 5 6 7 8
## -1.849460 -1.849460 9.947766 9.947766 13.880175 17.812584 21.744993 21.744993
## 9 10 11 12 13 14 15 16
## 21.744993 25.677401 25.677401 29.609810 29.609810 29.609810 29.609810 33.542219
## 17 18 19 20 21 22 23 24
## 33.542219 33.542219 33.542219 37.474628 37.474628 37.474628 37.474628 41.407036
## 25 26 27 28 29 30 31 32
## 41.407036 41.407036 45.339445 45.339445 49.271854 49.271854 49.271854 53.204263
## 33 34 35 36 37 38 39 40
## 53.204263 53.204263 53.204263 57.136672 57.136672 57.136672 61.069080 61.069080
## 41 42 43 44 45 46 47 48
## 61.069080 61.069080 61.069080 68.933898 72.866307 76.798715 76.798715 76.798715
## 49 50
## 76.798715 80.731124
7.5 Maximum Likelihood Estimation (MLE)

Approach
Recall the model,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝜖𝑖
where 𝜖𝑖 ∼ 𝑁 (0, 𝜎2 ).
Then we can find the mean and variance of each 𝑌𝑖 .
E[𝑌𝑖 ∣ 𝑋𝑖 = 𝑥𝑖 ] = 𝛽0 + 𝛽1 𝑥𝑖
and
Var[𝑌𝑖 ∣ 𝑋𝑖 = 𝑥𝑖 ] = 𝜎2 .
Additionally, the 𝑌𝑖 follow a normal distribution conditioned on the 𝑥𝑖 .
𝑌𝑖 ∣ 𝑋𝑖 ∼ 𝑁 (𝛽0 + 𝛽1 𝑥𝑖 , 𝜎2 )
Recall that the pdf of a random variable 𝑋 ∼ 𝑁 (𝜇, 𝜎2 ) is given by
1 1 𝑥−𝜇 2
𝑓𝑋 (𝑥; 𝜇, 𝜎2 ) = √ exp [− ( ) ].
2𝜋𝜎2 2 𝜎
Then we can write the pdf of each of the 𝑌𝑖 as
2
1 1 𝑦 − (𝛽0 + 𝛽1 𝑥𝑖 )
𝑓𝑌𝑖 (𝑦𝑖 ; 𝑥𝑖 , 𝛽0 , 𝛽1 , 𝜎2 ) = √ exp [− ( 𝑖 ) ].
2𝜋𝜎2 2 𝜎
Given 𝑛 data points (𝑥𝑖 , 𝑦𝑖 ) we can write the likelihood, which is a function of
the three parameters 𝛽0 , 𝛽1 , and 𝜎2 . Since the data have been observed, we use
lower case 𝑦𝑖 to denote that these values are no longer random.
𝑛 2
1 1 𝑦 − 𝛽0 − 𝛽1 𝑥𝑖
𝐿(𝛽0 , 𝛽1 , 𝜎2 ) = ∏ √ exp [− ( 𝑖 ) ]
𝑖=1 2𝜋𝜎 2 2 𝜎
Our goal is to find values of 𝛽0 , 𝛽1 , and 𝜎2 which maximize this function, which
is a straightforward multivariate calculus problem.
We’ll start by doing a bit of rearranging to make our task easier.
𝑛
1 1 𝑛
𝐿(𝛽0 , 𝛽1 , 𝜎2 ) = ( √ ) exp [− ∑(𝑦 − 𝛽0 − 𝛽1 𝑥𝑖 )2 ]
2𝜋𝜎2 2𝜎2 𝑖=1 𝑖
Then, as is often the case when finding MLEs, for mathematical convenience we
will take the natural logarithm of the likelihood function since log is a monoton-
ically increasing function. Then we will proceed to maximize the log-likelihood,
and the resulting estimates will be the same as if we had not taken the log.
7.5. MAXIMUM LIKELIHOOD ESTIMATION (MLE) APPROACH 117
𝑛 𝑛 1 𝑛
log 𝐿(𝛽0 , 𝛽1 , 𝜎2 ) = − log(2𝜋) − log(𝜎2 ) − 2 ∑(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 )2
2 2 2𝜎 𝑖=1
Note that we use log to mean the natural logarithm. We now take a partial
derivative with respect to each of the parameters.
𝜕 log 𝐿(𝛽0 , 𝛽1 , 𝜎2 ) 1 𝑛
= 2 ∑(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 )
𝜕𝛽0 𝜎 𝑖=1
𝜕 log 𝐿(𝛽0 , 𝛽1 , 𝜎2 ) 1 𝑛
= 2 ∑(𝑥𝑖 )(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 )
𝜕𝛽1 𝜎 𝑖=1
𝑛
𝜕 log 𝐿(𝛽0 , 𝛽1 , 𝜎2 ) 𝑛 1
= − + ∑ (𝑦 − 𝛽0 − 𝛽1 𝑥𝑖 )2
𝜕𝜎2 2𝜎2 2(𝜎2 )2 𝑖=1 𝑖
We then set each of the partial derivatives equal to zero and solve the resulting
system of equations.
𝑛
∑(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 ) = 0
𝑖=1
𝑛
∑(𝑥𝑖 )(𝑦𝑖 − 𝛽0 − 𝛽1 𝑥𝑖 ) = 0
𝑖=1
𝑛
𝑛 1
− + ∑ (𝑦 − 𝛽0 − 𝛽1 𝑥𝑖 )2 = 0
2𝜎2 2(𝜎2 )2 𝑖=1 𝑖
You may notice that the first two equations also appear in the least squares
approach. Then, skipping the issue of actually checking if we have found a
maximum, we then arrive at our estimates. We call these estimates the maxi-
mum likelihood estimates.
𝑛 𝑛
𝑛 (∑𝑖=1 𝑥𝑖 )(∑𝑖=1 𝑦𝑖 )
∑𝑖=1 𝑥𝑖 𝑦𝑖 − 𝑆𝑥𝑦
𝛽1̂ = 𝑛
𝑛
𝑥 )2
=
𝑛 (∑
∑𝑖=1 𝑥2𝑖 − 𝑖=1𝑛 𝑖 𝑆𝑥𝑥
𝛽0̂ = 𝑦 ̄ − 𝛽1̂ 𝑥̄
1 𝑛
𝜎̂ 2 = ∑(𝑦 − 𝑦𝑖̂ )2
𝑛 𝑖=1 𝑖
Note that 𝛽0̂ and 𝛽1̂ are the same as the least squares estimates. However
we now have a new estimate of 𝜎2 , that is 𝜎̂ 2 . So we now have two different
estimates of 𝜎2 .
𝑛 𝑛
1 1
𝑠2𝑒 = ∑(𝑦𝑖 − 𝑦𝑖̂ )2 = ∑ 𝑒2𝑖 Least Squares
𝑛 − 2 𝑖=1 𝑛 − 2 𝑖=1
1 𝑛 1 𝑛
𝜎̂ 2 = ∑(𝑦𝑖 − 𝑦𝑖̂ )2 = ∑ 𝑒2𝑖 MLE
𝑛 𝑖=1 𝑛 𝑖=1
In the next chapter, we will discuss in detail the difference between these two
estimates, which involves biasedness.
7.6 Simulating SLR

We return again to more examples of simulation. This will be a common theme!
In practice you will almost never have a true model, and you will use data to at-
tempt to recover information about the unknown true model. With simulation,
we decide the true model and simulate data from it. Then, we apply a method
to the data, in this case least squares. Now, since we know the true model, we
can assess how well it did.
For this example, we will simulate 𝑛 = 21 observations from the model
𝑌 = 5 − 2𝑥 + 𝜖.
That is 𝛽0 = 5, 𝛽1 = −2, and let 𝜖 ∼ 𝑁 (𝜇 = 0, 𝜎2 = 9). Or, even more

succinctly we could write
𝑌 ∣ 𝑋 ∼ 𝑁 (𝜇 = 5 − 2𝑥, 𝜎2 = 9).
We first set the true parameters of the model to be simulated.
num_obs = 21
beta_0 = 5
beta_1 = -2
sigma = 3
Next, we obtain simulated values of 𝜖𝑖 after setting a seed for reproducibility.
set.seed(1)
epsilon = rnorm(n = num_obs, mean = 0, sd = sigma)
Now, since the 𝑥𝑖 values in SLR are considered fixed and known, we simply
specify 21 values. Another common practice is to generate them from a uniform
distribution, and then use them for the remainder of the analysis.
7.6. SIMULATING SLR 119
x_vals = seq(from = 0, to = 10, length.out = num_obs)

# set.seed(1)
# x_vals = runif(num_obs, 0, 10)
We then generate the 𝑦 values according to the specified functional relationship.
y_vals = beta_0 + beta_1 * x_vals + epsilon
The data, (𝑥𝑖 , 𝑦𝑖 ), represent a possible sample from the true distribution. Now
to check how well the method of least squares works, we use lm() to fit the
model to our simulated data, then take a look at the estimated coefficients.
sim_fit = lm(y_vals ~ x_vals)

coef(sim_fit)
## (Intercept) x_vals
## 4.832639 -1.831401
And look at that, they aren’t too far from the true parameters we specified!
plot(y_vals ~ x_vals)
abline(sim_fit)
5
0
y_vals
-5
-10
-15
0 2 4 6 8 10
x_vals
We should say here, that we’re being sort of lazy, and not the good kind of
lazy that could be considered efficient. Any time you simulate data, you should
consider doing two things: writing a function, and storing the data in a data
frame.
The function below, sim_slr(), can be used for the same task as above, but
is much more flexible. Notice that we provide x to the function, instead of
generating x inside the function. In the SLR model, the 𝑥𝑖 are considered known
values. That is, they are not random, so we do not assume a distribution for
the 𝑥𝑖 . Because of this, we will repeatedly use the same x values across all
simulations.
sim_slr = function(x, beta_0 = 10, beta_1 = 5, sigma = 1) {

n = length(x)
epsilon = rnorm(n, mean = 0, sd = sigma)
y = beta_0 + beta_1 * x + epsilon
data.frame(predictor = x, response = y)
}
Here, we use the function to repeat the analysis above.
set.seed(1)
sim_data = sim_slr(x = x_vals, beta_0 = 5, beta_1 = -2, sigma = 3)
This time, the simulated observations are stored in a data frame.
head(sim_data)
## predictor response
## 1 0.0 3.1206386
## 2 0.5 4.5509300
## 3 1.0 0.4931142
## 4 1.5 6.7858424
## 5 2.0 1.9885233
## 6 2.5 -2.4614052
Now when we fit the model with lm() we can use a data argument, a very good
practice.
sim_fit = lm(response ~ predictor, data = sim_data)

coef(sim_fit)
## (Intercept) predictor
## 4.832639 -1.831401
7.7. HISTORY 121
And this time, we’ll make the plot look a lot nicer.
plot(response ~ predictor, data = sim_data,

xlab = "Simulated Predictor Variable",
ylab = "Simulated Response Variable",
main = "Simulated Regression Data",
pch = 20,
cex = 2,
col = "grey")
abline(sim_fit, lwd = 3, lty = 1, col = "darkorange")
abline(beta_0, beta_1, lwd = 3, lty = 2, col = "dodgerblue")
legend("topright", c("Estimate", "Truth"), lty = c(1, 2), lwd = 2,
col = c("darkorange", "dodgerblue"))
Simulated Regression Data
Estimate
Simulated Response Variable
Truth
0
-5
-10
-15
0 2 4 6 8 10
Simulated Predictor Variable
7.7 History
For some brief background on the history of linear regression, see “Galton,
Pearson, and the Peas: A Brief History of Linear Regression for Statistics In-
structors” from the Journal of Statistics Education as well as the Wikipedia
page on the history of regression analysis and lastly the article for regression
toward the mean which details the origins of the term “regression.”
7.8 R Markdown
The R Markdown file for this chapter can be found here:
• slr.Rmd
The file was created using R version 4.2.1.

Chapter 8
Inference for Simple Linear

Regression
“There are three types of lies: lies, damn lies, and statistics.”
— Benjamin Disraeli
• Understand the distributions of regression estimates.

• Create interval estimates for regression parameters, mean response, and
predictions.
• Test for significance of regression.
Last chapter we defined the simple linear regression model,
where 𝜖𝑖 ∼ 𝑁 (0, 𝜎2 ). We then used observations (𝑥𝑖 , 𝑦𝑖 ), for 𝑖 = 1, 2, … 𝑛, to

find values of 𝛽0 and 𝛽1 which minimized
𝑛
𝑓(𝛽0 , 𝛽1 ) = ∑(𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖 ))2 .
𝑖=1
We called these values 𝛽0̂ and 𝛽1̂ , which we found to be
𝑛
𝑆𝑥𝑦 ∑ (𝑥𝑖 − 𝑥)(𝑦
̄ 𝑖 − 𝑦)̄
𝛽1̂ = = 𝑖=1 𝑛
𝑆𝑥𝑥 ∑𝑖=1 (𝑥𝑖 − 𝑥)̄ 2
𝛽0̂ = 𝑦 ̄ − 𝛽1̂ 𝑥.̄
123
124 CHAPTER 8. INFERENCE FOR SIMPLE LINEAR REGRESSION
We also estimated 𝜎2 using 𝑠2𝑒 . In other words, we found that 𝑠𝑒 is an estimate

of 𝜎, where;
𝑛
1
𝑠𝑒 = RSE = √ ∑ 𝑒2𝑖
𝑛 − 2 𝑖=1
which we also called RSE, for “Residual Standard Error.”

When applied to the cars data, we obtained the following results:

##
## Call:
##
## Residuals:
## -29.069 -9.525 -2.272 9.215 43.201
##
## Coefficients:
## (Intercept) -17.5791 6.7584 -2.601 0.0123 *
## speed 3.9324 0.4155 9.464 1.49e-12 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
Last chapter, we only discussed the Estimate, Residual standard error, and
Multiple R-squared values. In this chapter, we will discuss all of the informa-
tion under Coefficients as well as F-statistic.

pch = 20,
cex = 2,
col = "grey")
125

100 120
80
60
40
20
0
5 10 15 20 25
To get started, we’ll note that there is another equivalent expression for 𝑆𝑥𝑦
which we did not see last chapter,
𝑛 𝑛
𝑆𝑥𝑦 = ∑(𝑥𝑖 − 𝑥)(𝑦
̄ 𝑖 − 𝑦)̄ = ∑(𝑥𝑖 − 𝑥)𝑦
̄ 𝑖.
𝑖=1 𝑖=1
This may be a surprising equivalence. (Maybe try to prove it.) However, it will
be useful for illustrating concepts in this chapter.
Note that 𝛽1̂ is a sample statistic when calculated with observed data as
written above, as is 𝛽 ̂ . 0
However, in this chapter it will often be convenient to use both 𝛽1̂ and 𝛽0̂ as
random variables, that is, we have not yet observed the values for each 𝑌𝑖 .
When this is the case, we will use a slightly different notation, substituting in
capital 𝑌𝑖 for lower case 𝑦𝑖 .
𝑛
∑𝑖=1 (𝑥𝑖 − 𝑥)𝑌
̄ 𝑖
𝛽1̂ = 𝑛
∑𝑖=1 (𝑥𝑖 − 𝑥)̄ 2
𝛽0̂ = 𝑌 ̄ − 𝛽1̂ 𝑥̄
Last chapter we argued that these estimates of unknown model parameters 𝛽0

and 𝛽1 were good because we obtained them by minimizing errors. We will now
discuss the Gauss–Markov theorem, which takes this idea further, showing that
these estimates are actually the “best” estimates, from a certain point of view.
8.1 Gauss–Markov Theorem
The Gauss–Markov theorem tells us that when estimating the parameters

of the simple linear regression model 𝛽0 and 𝛽1 , the 𝛽0̂ and 𝛽1̂ which we derived
are the best linear unbiased estimates, or BLUE for short. (The actual con-
ditions for the Gauss–Markov theorem are more relaxed than the SLR model.)
We will now discuss linear, unbiased, and best as they relate to these estimates.
Linear
Recall, in the SLR setup that the 𝑥𝑖 values are considered fixed and known
quantities. Then a linear estimate is one which can be written as a linear
combination of the 𝑌𝑖 . In the case of 𝛽1̂ we see
𝑛 𝑛
∑𝑖=1 (𝑥𝑖 − 𝑥)𝑌
̄ 𝑖
𝛽1̂ = 𝑛 = ∑ 𝑘𝑖 𝑌𝑖 = 𝑘1 𝑌1 + 𝑘2 𝑌2 + ⋯ 𝑘𝑛 𝑌𝑛
∑𝑖=1 (𝑥𝑖 − 𝑥)̄ 2 𝑖=1
(𝑥𝑖 − 𝑥)̄
where 𝑘𝑖 = 𝑛 .
∑𝑖=1 (𝑥𝑖 − 𝑥)̄ 2
In a similar fashion, we could show that 𝛽0̂ can be written as a linear combination
of the 𝑌𝑖 . Thus both 𝛽0̂ and 𝛽1̂ are linear estimators.
Unbiased
Now that we know our estimates are linear, how good are these estimates? One
measure of the “goodness” of an estimate is its bias. Specifically, we prefer
estimates that are unbiased, meaning their expected value is the parameter
being estimated.
In the case of the regression estimates, we have,
E[𝛽0̂ ] = 𝛽0
E[𝛽 ̂ ] = 𝛽 .
1 1
This tells us that, when the conditions of the SLR model are met, on average
our estimates will be correct. However, as we saw last chapter when simulating
from the SLR model, that does not mean that each individual estimate will be
correct. Only that, if we repeated the process an infinite number of times, on
average the estimate would be correct.
8.2. SAMPLING DISTRIBUTIONS 127
Best
Now, if we restrict ourselves to both linear and unbiased estimates, how do we

define the best estimate? The estimate with the minimum variance.
First note that it is very easy to create an estimate for 𝛽1 that has very low
variance, but is not unbiased. For example, define:
̂
𝜃𝐵𝐴𝐷 = 5.
̂
Then, since 𝜃𝐵𝐴𝐷 is a constant value,
̂
Var[𝜃𝐵𝐴𝐷 ] = 0.
However since,
̂
E[𝜃𝐵𝐴𝐷 ]=5
we say that 𝜃𝐵𝐴𝐷̂ is a biased estimator unless 𝛽1 = 5, which we would not

know ahead of time. For this reason, it is a terrible estimate (unless by chance
𝛽1 = 5) even though it has the smallest possible variance. This is part of the
reason we restrict ourselves to unbiased estimates. What good is an estimate,
if it estimates the wrong quantity?
So now, the natural question is, what are the variances of 𝛽0̂ and 𝛽1̂ ? They are,
1 𝑥2̄
Var[𝛽0̂ ] = 𝜎2 ( + )
𝑛 𝑆𝑥𝑥
𝜎2
Var[𝛽1̂ ] = .
𝑆𝑥𝑥
These quantify the variability of the estimates due to random chance during
sampling. Are these “the best”? Are these variances as small as we can possibly
get? You’ll just have to take our word for it that they are because showing that
this is true is beyond the scope of this course.
8.2 Sampling Distributions
Now that we have “redefined” the estimates for 𝛽0̂ and 𝛽1̂ as random variables,
we can discuss their sampling distribution, which is the distribution when a
statistic is considered a random variable.
Since both 𝛽0̂ and 𝛽1̂ are a linear combination of the 𝑌𝑖 and each 𝑌𝑖 is normally
distributed, then both 𝛽0̂ and 𝛽1̂ also follow a normal distribution.
Then, putting all of the above together, we arrive at the distributions of 𝛽0̂ and
𝛽1̂ .
For 𝛽1̂ we say,
𝑛
𝑆𝑥𝑦 ∑ (𝑥𝑖 − 𝑥)𝑌 ̄ 𝑖 𝜎2
𝛽1̂ = = 𝑖=1
𝑛 ∼ 𝑁 (𝛽1 , 𝑛 ).
𝑆𝑥𝑥 ∑𝑖=1 (𝑥𝑖 − 𝑥)̄2 ∑𝑖=1 (𝑥𝑖 − 𝑥)̄ 2
Or more succinctly,
𝜎2
𝛽1̂ ∼ 𝑁 (𝛽1 , ).
𝑆𝑥𝑥
And for 𝛽0̂ ,
𝑛
𝜎2 ∑𝑖=1 𝑥2𝑖
𝛽0̂ = 𝑌 ̄ − 𝛽1̂ 𝑥̄ ∼ 𝑁 (𝛽0 , 𝑛 ).
𝑛 ∑𝑖=1 (𝑥𝑖 − 𝑥)̄ 2
Or more succinctly,
1 𝑥2̄
𝛽0̂ ∼ 𝑁 (𝛽0 , 𝜎2 ( + ))
𝑛 𝑆𝑥𝑥
At this point we have neglected to prove a number of these results. Instead of

working through the tedious derivations of these sampling distributions, we will
instead justify these results to ourselves using simulation.
A note to current readers: These derivations and proofs may be added to an
appendix at a later time. You can also find these results in nearly any standard
linear regression textbook. At UIUC, these results will likely be presented in
both STAT 424 and STAT 425. However, since you will not be asked to perform
derivations of this type in this course, they are for now omitted.
8.2.1 Simulating Sampling Distributions
To verify the above results, we will simulate samples of size 𝑛 = 100 from the
model
𝑌 𝑖 = 𝛽 0 + 𝛽 1 𝑥𝑖 + 𝜖 𝑖
where 𝜖𝑖 ∼ 𝑁 (0, 𝜎2 ). In this case, the parameters are known to be:
• 𝛽0 = 3
• 𝛽1 = 6
• 𝜎2 = 4
Then, based on the above, we should find that
𝜎2
𝛽1̂ ∼ 𝑁 (𝛽1 , )
𝑆𝑥𝑥
and
1 𝑥2̄
𝛽0̂ ∼ 𝑁 (𝛽0 , 𝜎2 ( + )) .
𝑛 𝑆𝑥𝑥
First we need to decide ahead of time what our 𝑥 values will be for this sim-
ulation, since the 𝑥 values in SLR are also considered known quantities. The
choice of 𝑥 values is arbitrary. Here we also set a seed for randomization, and
calculate 𝑆𝑥𝑥 which we will need going forward.
set.seed(42)
sample_size = 100 # this is n
x = seq(-1, 1, length = sample_size)
Sxx = sum((x - mean(x)) ^ 2)
We also fix our parameter values.
beta_0 = 3
beta_1 = 6
sigma = 2
With this information, we know the sampling distributions should be:
(var_beta_1_hat = sigma ^ 2 / Sxx)
## [1] 0.1176238
(var_beta_0_hat = sigma ^ 2 * (1 / sample_size + mean(x) ^ 2 / Sxx))
## [1] 0.04
𝛽1̂ ∼ 𝑁 (6, 0.1176238)
and
𝛽0̂ ∼ 𝑁 (3, 0.04).
That is,
E[𝛽1̂ ] = 6
Var[𝛽 ̂ ] = 0.1176238
1
and
E[𝛽0̂ ] = 3
Var[𝛽 ̂ ] = 0.04.
0
We now simulate data from this model 10,000 times. Note this may not be the
most R way of doing the simulation. We perform the simulation in this manner
in an attempt at clarity. For example, we could have used the sim_slr()
function from the previous chapter. We also simply store variables in the global
environment instead of creating a data frame for each new simulated dataset.
num_samples = 10000
beta_0_hats = rep(0, num_samples)
beta_1_hats = rep(0, num_samples)
for (i in 1:num_samples) {
eps = rnorm(sample_size, mean = 0, sd = sigma)
y = beta_0 + beta_1 * x + eps
sim_model = lm(y ~ x)
beta_0_hats[i] = coef(sim_model)[1]
beta_1_hats[i] = coef(sim_model)[2]
}
Each time we simulated the data, we obtained values of the estimated coeffi-
cients. The variables beta_0_hats and beta_1_hats now store 10,000 simu-
lated values of 𝛽0̂ and 𝛽1̂ respectively.
We first verify the distribution of 𝛽1̂ .
mean(beta_1_hats) # empirical mean
## [1] 6.001998
beta_1 # true mean
## [1] 6
var(beta_1_hats) # empirical variance
## [1] 0.11899
var_beta_1_hat # true variance
## [1] 0.1176238
We see that the empirical and true means and variances are very similar. We also
verify that the empirical distribution is normal. To do so, we plot a histogram
of the beta_1_hats, and add the curve for the true distribution of 𝛽1̂ . We use
prob = TRUE to put the histogram on the same scale as the normal curve.
# note need to use prob = TRUE

hist(beta_1_hats, prob = TRUE, breaks = 20,
xlab = expression(hat(beta)[1]), main = "", border = "dodgerblue")
curve(dnorm(x, mean = beta_1, sd = sqrt(var_beta_1_hat)),
col = "darkorange", add = TRUE, lwd = 3)
1.0
0.8
Density
0.6
0.4
0.2
0.0
5.0 5.5 6.0 6.5 7.0 7.5

^
β 1
We then repeat the process for 𝛽0̂ .
mean(beta_0_hats) # empirical mean
## [1] 3.001147
beta_0 # true mean
## [1] 3
var(beta_0_hats) # empirical variance
## [1] 0.04017924
var_beta_0_hat # true variance
## [1] 0.04
hist(beta_0_hats, prob = TRUE, breaks = 25,

curve(dnorm(x, mean = beta_0, sd = sqrt(var_beta_0_hat)),
2.0
1.5
Density
1.0
0.5
0.0
2.5 3.0 3.5

^
β 0
In this simulation study, we have only simulated a finite number of samples.

To truly verify the distributional results, we would need to observe an infinite
number of samples. However, the following plot should make it clear that if we
continued simulating, the empirical results would get closer and closer to what
we should expect.
par(mar = c(5, 5, 1, 1)) # adjusted plot margins, otherwise the "hat" does not display
plot(cumsum(beta_1_hats) / (1:length(beta_1_hats)), type = "l", ylim = c(5.95, 6.05),
xlab = "Number of Simulations",
ylab = expression("Empirical Mean of " ~ hat(beta)[1]),
col = "dodgerblue")
abline(h = 6, col = "darkorange", lwd = 2)
6.04
1
6.02
Empirical Mean of β
^
6.00
5.98
5.96
0 2000 4000 6000 8000 10000
Number of Simulations
par(mar = c(5, 5, 1, 1)) # adjusted plot margins, otherwise the "hat" does not display
plot(cumsum(beta_0_hats) / (1:length(beta_0_hats)), type = "l", ylim = c(2.95, 3.05),
xlab = "Number of Simulations",
ylab = expression("Empirical Mean of " ~ hat(beta)[0]),
col = "dodgerblue")
3.04
0
3.02
Empirical Mean of β
^
3.00
2.98
2.96
0 2000 4000 6000 8000 10000
Number of Simulations
8.3 Standard Errors
So now we believe the two distributional results,
1 𝑥2̄
𝛽0̂ ∼ 𝑁 (𝛽0 , 𝜎2 ( + ))
𝑛 𝑆𝑥𝑥
𝜎2
𝛽1̂ ∼ 𝑁 (𝛽1 , ).
𝑆𝑥𝑥
Then by standardizing these results we find that
𝛽0̂ − 𝛽0
∼ 𝑁 (0, 1)
SD[𝛽 ̂ ]
0
and
𝛽1̂ − 𝛽1
∼ 𝑁 (0, 1)
SD[𝛽 ̂ ]
1
where
8.3. STANDARD ERRORS 135
1 𝑥2̄
SD[𝛽0̂ ] = 𝜎√ +
𝑛 𝑆𝑥𝑥
and
𝜎
SD[𝛽1̂ ] = .
√𝑆𝑥𝑥
Since we don’t know 𝜎 in practice, we will have to estimate it using 𝑠𝑒 , which we

plug into our existing expression for the standard deviations of our estimates.
These two new expressions are called standard errors which are the estimated
standard deviations of the sampling distributions.
1 𝑥2̄
SE[𝛽0̂ ] = 𝑠𝑒 √ +
𝑛 𝑆𝑥𝑥
𝑠𝑒
SE[𝛽1̂ ] =
√𝑆𝑥𝑥
Now if we divide by the standard error, instead of the standard deviation, we

obtain the following results which will allow us to make confidence intervals and
perform hypothesis testing.
𝛽0̂ − 𝛽0
∼ 𝑡𝑛−2
SE[𝛽 ̂ ]
0
𝛽1̂ − 𝛽1
∼ 𝑡𝑛−2
SE[𝛽 ̂ ]
1
To see this, first note that,
RSS (𝑛 − 2)𝑠2𝑒
= ∼ 𝜒2𝑛−2 .
𝜎2 𝜎2
Also recall that a random variable 𝑇 defined as,
𝑍
𝑇 = 2
√ 𝜒𝑑
𝑑
follows a 𝑡 distribution with 𝑑 degrees of freedom, where 𝜒2𝑑 is a 𝜒2 random

variable with 𝑑 degrees of freedom.
We write,
𝑇 ∼ 𝑡𝑑
to say that the random variable 𝑇 follows a 𝑡 distribution with 𝑑 degrees of

freedom.
Then we use the classic trick of “multiply by 1” and some rearranging to arrive
at
𝛽1̂ − 𝛽1 𝛽 ̂ − 𝛽1
= 1
SE[𝛽1̂ ] 𝑠𝑒 /√𝑆𝑥𝑥
𝛽 ̂ − 𝛽1 𝜎/√𝑆𝑥𝑥
= 1 ⋅
𝑠𝑒 /√𝑆𝑥𝑥 𝜎/√𝑆𝑥𝑥
𝛽 ̂ − 𝛽1 𝜎/√𝑆𝑥𝑥
= 1 ⋅
𝜎/√𝑆𝑥𝑥 𝑠𝑒 /√𝑆𝑥𝑥
𝛽 ̂ − 𝛽1 √ 𝑠2𝑒
= 1 /
𝜎/√𝑆𝑥𝑥 𝜎2
(𝑛−2)𝑠2
𝛽 ̂ − 𝛽1 √ 𝜎 2 𝑒 𝑍
= 1 / ∼ ∼ 𝑡𝑛−2
̂
SD[𝛽1 ] 𝑛−2 2
√ 𝜒𝑛−2
𝑛−2
where 𝑍 ∼ 𝑁 (0, 1).

Recall that a 𝑡 distribution is similar to a standard normal, but with heavier
tails. As the degrees of freedom increases, the 𝑡 distribution becomes more and
more like a standard normal. Below we plot a standard normal distribution
as well as two examples of a 𝑡 distribution with different degrees of freedom.
Notice how the 𝑡 distribution with the larger degrees of freedom is more similar
to the standard normal curve.
# define grid of x values

x = seq(-4, 4, length = 100)
# plot curve for standard normal

plot(x, dnorm(x), type = "l", lty = 1, lwd = 2,
xlab = "x", ylab = "Density", main = "Normal vs t Distributions")
# add curves for t distributions
lines(x, dt(x, df = 1), lty = 3, lwd = 2, col = "darkorange")
lines(x, dt(x, df = 10), lty = 2, lwd = 2, col = "dodgerblue")
# add legend
legend("topright", title = "Distributions",
8.4. CONFIDENCE INTERVALS FOR SLOPE AND INTERCEPT 137
legend = c("t, df = 1", "t, df = 10", "Standard Normal"),

lwd = 2, lty = c(3, 2, 1), col = c("darkorange", "dodgerblue", "black"))
Normal vs t Distributions
0.4
Distributions
t, df = 1
t, df = 10
0.3
Standard Normal
Density
0.2
0.1
0.0
-4 -2 0 2 4
8.4 Confidence Intervals for Slope and Intercept
Recall that confidence intervals for means often take the form:
EST ± CRIT ⋅ SE
or
EST ± MARGIN
where EST is an estimate for the parameter of interest, SE is the standard error
of the estimate, and MARGIN = CRIT ⋅ SE.
Then, for 𝛽0 and 𝛽1 we can create confidence intervals using
1 𝑥2̄
𝛽0̂ ± 𝑡𝛼/2,𝑛−2 ⋅ SE[𝛽0̂ ] 𝛽0̂ ± 𝑡𝛼/2,𝑛−2 ⋅ 𝑠𝑒 √ +
𝑛 𝑆𝑥𝑥
and
𝑠𝑒
𝛽1̂ ± 𝑡𝛼/2,𝑛−2 ⋅ SE[𝛽1̂ ] 𝛽1̂ ± 𝑡𝛼/2,𝑛−2 ⋅
√𝑆𝑥𝑥
where 𝑡𝛼/2,𝑛−2 is the critical value such that 𝑃 (𝑡𝑛−2 > 𝑡𝛼/2,𝑛−2 ) = 𝛼/2.
8.5 Hypothesis Tests

“We may speak of this hypothesis as the ‘null hypothesis,’ and it
should be noted that the null hypothesis is never proved or estab-
lished, but is possibly disproved, in the course of experimentation.”
— Ronald Aylmer Fisher
Recall that a test statistic (TS) for testing means often takes the form:
EST − HYP
TS =
SE
where EST is an estimate for the parameter of interest, HYP is a hypothesized
value of the parameter, and SE is the standard error of the estimate.
So, to test
𝐻0 ∶ 𝛽0 = 𝛽00 vs 𝐻1 ∶ 𝛽0 ≠ 𝛽00
we use the test statistic
𝛽0̂ − 𝛽00 𝛽0̂ − 𝛽00

𝑡= =
̂
SE[𝛽0 ] 𝑠𝑒 √ 𝑛1 + 𝑆𝑥̄
2
𝑥𝑥
which, under the null hypothesis, follows a 𝑡 distribution with 𝑛 − 2 degrees of

freedom. We use 𝛽00 to denote the hypothesized value of 𝛽0 .
Similarly, to test
𝐻0 ∶ 𝛽1 = 𝛽10 vs 𝐻1 ∶ 𝛽1 ≠ 𝛽10
we use the test statistic
𝛽1̂ − 𝛽10 𝛽 ̂ − 𝛽10

𝑡= = 1
which again, under the null hypothesis, follows a 𝑡 distribution with 𝑛−2 degrees
of freedom. We now use 𝛽10 to denote the hypothesized value of 𝛽1 .
8.6. CARS EXAMPLE 139
8.6 cars Example

We now return to the cars example from last chapter to illustrate these con-
cepts. We first fit the model using lm() then use summary() to view the results
in greater detail.

##
## Call:
##
## Residuals:
## -29.069 -9.525 -2.272 9.215 43.201
##
## Coefficients:
## (Intercept) -17.5791 6.7584 -2.601 0.0123 *
## speed 3.9324 0.4155 9.464 1.49e-12 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
8.6.1 Tests in R
We will now discuss the results displayed called Coefficients. First recall that
we can extract this information directly.
names(summary(stop_dist_model))
## [1] "call" "terms" "residuals" "coefficients"

## [5] "aliased" "sigma" "df" "r.squared"
## [9] "adj.r.squared" "fstatistic" "cov.unscaled"
summary(stop_dist_model)$coefficients

## (Intercept) -17.579095 6.7584402 -2.601058 1.231882e-02
## speed 3.932409 0.4155128 9.463990 1.489836e-12
The names() function tells us what information is available, and then we use
the $ operator and coefficients to extract the information we are interested
in. Two values here should be immediately familiar.
𝛽0̂ = −17.5790949
and
𝛽1̂ = 3.9324088
which are our estimates for the model parameters 𝛽0 and 𝛽1 .

Let’s now focus on the second row of output, which is relevant to 𝛽1 .
summary(stop_dist_model)$coefficients[2,]

## 3.932409e+00 4.155128e-01 9.463990e+00 1.489836e-12
Again, the first value, Estimate is
𝛽1̂ = 3.9324088.
The second value, Std. Error, is the standard error of 𝛽1̂ ,
𝑠𝑒
SE[𝛽1̂ ] = = 0.4155128.
√𝑆𝑥𝑥
The third value, t value, is the value of the test statistic for testing 𝐻0 ∶ 𝛽1 = 0
vs 𝐻1 ∶ 𝛽1 ≠ 0,
𝛽1̂ − 0 𝛽1̂ − 0
𝑡= = = 9.46399.
Lastly, Pr(>|t|), gives us the p-value of that test.
p-value = 1.4898365 × 10−12
Note here, we are specifically testing whether or not 𝛽1 = 0.

The first row of output reports the same values, but for 𝛽0 .
summary(stop_dist_model)$coefficients[1,]

## -17.57909489 6.75844017 -2.60105800 0.01231882
In summary, the following code stores the information of summary(stop_dist_model)$coefficients

in a new variable stop_dist_model_test_info, then extracts each element
into a new variable which describes the information it contains.
stop_dist_model_test_info = summary(stop_dist_model)$coefficients
beta_0_hat = stop_dist_model_test_info[1, 1] # Estimate

beta_0_hat_se = stop_dist_model_test_info[1, 2] # Std. Error
beta_0_hat_t = stop_dist_model_test_info[1, 3] # t value
beta_0_hat_pval = stop_dist_model_test_info[1, 4] # Pr(>|t|)
beta_1_hat = stop_dist_model_test_info[2, 1] # Estimate

beta_1_hat_se = stop_dist_model_test_info[2, 2] # Std. Error
beta_1_hat_t = stop_dist_model_test_info[2, 3] # t value
beta_1_hat_pval = stop_dist_model_test_info[2, 4] # Pr(>|t|)
We can then verify some equivalent expressions: the 𝑡 test statistic for 𝛽1̂ and
the two-sided p-value associated with that test statistic.
(beta_1_hat - 0) / beta_1_hat_se
## [1] 9.46399
beta_1_hat_t
## [1] 9.46399
2 * pt(abs(beta_1_hat_t), df = length(resid(stop_dist_model)) - 2, lower.tail = FALSE)
## [1] 1.489836e-12
beta_1_hat_pval
## [1] 1.489836e-12
8.6.2 Significance of Regression, t-Test
We pause to discuss the significance of regression test. First, note that

based on the above distributional results, we could test 𝛽0 and 𝛽1 against any
particular value, and perform both one and two-sided tests.
However, one very specific test,
𝐻0 ∶ 𝛽 1 = 0 vs 𝐻1 ∶ 𝛽1 ≠ 0
is used most often. Let’s think about this test in terms of the simple linear
regression model,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝜖𝑖 .
If we assume the null hypothesis is true, then 𝛽1 = 0 and we have the model,
𝑌𝑖 = 𝛽0 + 𝜖𝑖 .
In this model, the response does not depend on the predictor. So then we could
think of this test in the following way,
• Under 𝐻0 there is not a significant linear relationship between 𝑥 and 𝑦.

• Under 𝐻1 there is a significant linear relationship between 𝑥 and 𝑦.
For the cars example,
• Under 𝐻0 there is not a significant linear relationship between speed and

stopping distance.
• Under 𝐻1 there is a significant linear relationship between speed and
stopping distance.
Again, that test is seen in the output from summary(),
p-value = 1.4898365 × 10−12 .
With this extremely low p-value, we would reject the null hypothesis at any rea-
sonable 𝛼 level, say for example 𝛼 = 0.01. So we say there is a significant linear
relationship between speed and stopping distance. Notice that we emphasize
linear.
9
8
7
y
6
5
4
-1.0 -0.5 0.0 0.5 1.0
In this plot of simulated data, we see a clear relationship between 𝑥 and 𝑦,

however it is not a linear relationship. If we fit a line to this data, it is very flat.
The resulting test for 𝐻0 ∶ 𝛽1 = 0 vs 𝐻1 ∶ 𝛽1 ≠ 0 gives a large p-value, in this
case 0.7564548, so we would fail to reject and say that there is no significant
linear relationship between 𝑥 and 𝑦. We will see later how to fit a curve to this
data using a “linear” model, but for now, realize that testing 𝐻0 ∶ 𝛽1 = 0 vs
𝐻1 ∶ 𝛽1 ≠ 0 can only detect straight line relationships.
8.6.3 Confidence Intervals in R
Using R we can very easily obtain the confidence intervals for 𝛽0 and 𝛽1 .
confint(stop_dist_model, level = 0.99)
## 0.5 % 99.5 %
## (Intercept) -35.706610 0.5484205
## speed 2.817919 5.0468988
This automatically calculates 99% confidence intervals for both 𝛽0 and 𝛽1 , the
first row for 𝛽0 , the second row for 𝛽1 .
For the cars example when interpreting these intervals, we say, we are 99%
confident that for an increase in speed of 1 mile per hour, the average increase in
stopping distance is between 2.8179187 and 5.0468988 feet, which is the interval
for 𝛽1 .
Note that this 99% confidence interval does not contain the hypothesized value
of 0. Since it does not contain 0, it is equivalent to rejecting the test of 𝐻0 ∶
𝛽1 = 0 vs 𝐻1 ∶ 𝛽1 ≠ 0 at 𝛼 = 0.01, which we had seen previously.
You should be somewhat suspicious of the confidence interval for 𝛽0 , as it covers

negative values, which correspond to negative stopping distances. Technically
the interpretation would be that we are 99% confident that the average stopping
distance of a car traveling 0 miles per hour is between -35.7066103 and 0.5484205
feet, but we don’t really believe that, since we are actually certain that it would
be non-negative.
Note, we can extract specific values from this output a number of ways. This
code is not run, and instead, you should check how it relates to the output of
the code above.
confint(stop_dist_model, level = 0.99)[1,]

confint(stop_dist_model, level = 0.99)[1, 1]
confint(stop_dist_model, parm = "(Intercept)", level = 0.99)
confint(stop_dist_model, level = 0.99)[2,]
confint(stop_dist_model, parm = "speed", level = 0.99)
We can also verify that calculations that R is performing for the 𝛽1 interval.
# store estimate
beta_1_hat = coef(stop_dist_model)[2]
# store standard error

beta_1_hat_se = summary(stop_dist_model)$coefficients[2, 2]
# calculate critical value for two-sided 99% CI

crit = qt(0.995, df = length(resid(stop_dist_model)) - 2)
# est - margin, est + margin

c(beta_1_hat - crit * beta_1_hat_se, beta_1_hat + crit * beta_1_hat_se)
## speed speed
## 2.817919 5.046899
8.7. CONFIDENCE INTERVAL FOR MEAN RESPONSE 145
8.7 Confidence Interval for Mean Response

In addition to confidence intervals for 𝛽0 and 𝛽1 , there are two other common
interval estimates used with regression. The first is called a confidence inter-
val for the mean response. Often, we would like an interval estimate for the
mean, 𝐸[𝑌 ∣ 𝑋 = 𝑥] for a particular value of 𝑥.
In this situation we use 𝑦(𝑥)
̂ as our estimate of 𝐸[𝑌 ∣ 𝑋 = 𝑥]. We modify our
notation slightly to make it clear that the predicted value is a function of the 𝑥
value.
𝑦(𝑥)
̂ = 𝛽0̂ + 𝛽1̂ 𝑥
Recall that,
E[𝑌 ∣ 𝑋 = 𝑥] = 𝛽0 + 𝛽1 𝑥.
Thus, 𝑦(𝑥)
̂ is a good estimate since it is unbiased:
E[𝑦(𝑥)]
̂ = 𝛽0 + 𝛽1 𝑥.
We could then derive,
1 (𝑥 − 𝑥)̄ 2
Var[𝑦(𝑥)]
̂ = 𝜎2 ( + ).
𝑛 𝑆𝑥𝑥
Like the other estimates we have seen, 𝑦(𝑥)

̂ also follows a normal distribution.
Since 𝛽0̂ and 𝛽1̂ are linear combinations of normal random variables, 𝑦(𝑥)
̂ is as
well.
1 (𝑥 − 𝑥)̄ 2
𝑦(𝑥)
̂ ∼ 𝑁 (𝛽0 + 𝛽1 𝑥, 𝜎2 ( + ))
𝑛 𝑆𝑥𝑥
And lastly, since we need to estimate this variance, we arrive at the standard
error of our estimate,
1 (𝑥 − 𝑥)̄ 2
SE[𝑦(𝑥)]
̂ = 𝑠𝑒 √ + .
𝑛 𝑆𝑥𝑥
We can then use this to find the confidence interval for the mean response,
1 (𝑥 − 𝑥)̄ 2
𝑦(𝑥)
̂ ± 𝑡𝛼/2,𝑛−2 ⋅ 𝑠𝑒 √ +
𝑛 𝑆𝑥𝑥
To find confidence intervals for the mean response using R, we use the predict()
function. We give the function our fitted model as well as new data, stored as
a data frame. (This is important, so that R knows the name of the predictor
variable.) Here, we are finding the confidence interval for the mean stopping
distance when a car is travelling 5 miles per hour and when a car is travelling
21 miles per hour.
new_speeds = data.frame(speed = c(5, 21))

predict(stop_dist_model, newdata = new_speeds,
interval = c("confidence"), level = 0.99)
## fit lwr upr

## 1 2.082949 -10.89309 15.05898
## 2 65.001489 56.45836 73.54462
8.8 Prediction Interval for New Observations

Sometimes we would like an interval estimate for a new observation, 𝑌 , for a
particular value of 𝑥. This is very similar to an interval for the mean response,
E[𝑌 ∣ 𝑋 = 𝑥], but different in one very important way.
Our best guess for a new observation is still 𝑦(𝑥).
̂ The estimated mean is still
the best prediction we can make. The difference is in the amount of variability.
We know that observations will vary about the true regression line according
to a 𝑁 (0, 𝜎2 ) distribution. Because of this we add an extra factor of 𝜎2 to
our estimate’s variability in order to account for the variability of observations
about the regression line.
Var[𝑦(𝑥)
̂ + 𝜖] = Var[𝑦(𝑥)]
̂ + Var[𝜖]
1 (𝑥 − 𝑥)̄ 2
= 𝜎2 ( + ) + 𝜎2
𝑛 𝑆𝑥𝑥
1 (𝑥 − 𝑥)̄ 2
= 𝜎2 (1 + + )
𝑛 𝑆𝑥𝑥
1 (𝑥 − 𝑥)̄ 2
𝑦(𝑥)
̂ + 𝜖 ∼ 𝑁 (𝛽0 + 𝛽1 𝑥, 𝜎2 (1 + + ))
𝑛 𝑆𝑥𝑥
1 (𝑥 − 𝑥)̄ 2
SE[𝑦(𝑥)
̂ + 𝜖] = 𝑠𝑒 √1 + +
𝑛 𝑆𝑥𝑥
We can then find a prediction interval using,

8.9. CONFIDENCE AND PREDICTION BANDS 147
1 (𝑥 − 𝑥)̄ 2
𝑦(𝑥)
̂ ± 𝑡𝛼/2,𝑛−2 ⋅ 𝑠𝑒 √1 + + .
𝑛 𝑆𝑥𝑥
To calculate this for a set of points in R notice there is only a minor change in
syntax from finding a confidence interval for the mean response.
predict(stop_dist_model, newdata = new_speeds,

interval = c("prediction"), level = 0.99)
## fit lwr upr

## 1 2.082949 -41.16099 45.32689
## 2 65.001489 22.87494 107.12803
Also notice that these two intervals are wider than the corresponding confidence
intervals for the mean response.
8.9 Confidence and Prediction Bands

Often we will like to plot both confidence intervals for the mean response and
prediction intervals for all possible values of 𝑥. We calls these confidence and
prediction bands.
speed_grid = seq(min(cars$speed), max(cars$speed), by = 0.01)

dist_ci_band = predict(stop_dist_model,
newdata = data.frame(speed = speed_grid),
interval = "confidence", level = 0.99)
dist_pi_band = predict(stop_dist_model,
newdata = data.frame(speed = speed_grid),
interval = "prediction", level = 0.99)

pch = 20,
cex = 2,
col = "grey",
ylim = c(min(dist_pi_band), max(dist_pi_band)))
lines(speed_grid, dist_ci_band[,"lwr"], col = "dodgerblue", lwd = 3, lty = 2)

lines(speed_grid, dist_ci_band[,"upr"], col = "dodgerblue", lwd = 3, lty = 2)
lines(speed_grid, dist_pi_band[,"lwr"], col = "dodgerblue", lwd = 3, lty = 3)

lines(speed_grid, dist_pi_band[,"upr"], col = "dodgerblue", lwd = 3, lty = 3)
points(mean(cars$speed), mean(cars$dist), pch = "+", cex = 3)

100
50
+
0
-50
5 10 15 20 25
Some things to notice:
• We use the ylim argument to stretch the 𝑦-axis of the plot, since the bands
extend further than the points.
• We add a point at the point (𝑥,̄ 𝑦).
̄
– This is a point that the regression line will always pass through.
(Think about why.)
– This is the point where both the confidence and prediction bands are
the narrowest. Look at the standard errors of both to understand
why.
• The prediction bands (dotted blue) are less curved than the confidence
bands (dashed blue). This is a result of the extra factor of 𝜎2 added to
the variance at any value of 𝑥.
8.10. SIGNIFICANCE OF REGRESSION, F-TEST 149
8.10 Significance of Regression, F-Test

In the case of simple linear regression, the 𝑡 test for the significance of the
regression is equivalent to another test, the 𝐹 test for the significance of the
regression. This equivalence will only be true for simple linear regression, and in
the next chapter we will only use the 𝐹 test for the significance of the regression.
Recall from last chapter the decomposition of variance we saw before calculating
𝑅2 ,
𝑛 𝑛 𝑛
∑(𝑦𝑖 − 𝑦)̄ 2 = ∑(𝑦𝑖 − 𝑦𝑖̂ )2 + ∑(𝑦𝑖̂ − 𝑦)̄ 2 ,
𝑖=1 𝑖=1 𝑖=1
or, in short,
SST = SSE + SSReg.
To develop the 𝐹 test, we will arrange this information in an ANOVA table,
Degrees of Mean
Source Sum of Squares Freedom Square 𝐹
𝑛 2
Regression ∑𝑖=1 (𝑦𝑖̂− 𝑦)̄ 1 SSReg/1 MSReg/MSE
𝑛
Error ∑𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2 𝑛−2 SSE/(𝑛 − 2)
𝑛
Total ∑𝑖=1 (𝑦𝑖 − 𝑦)̄ 2 𝑛−1
ANOVA, or Analysis of Variance will be a concept we return to often in this

course. For now, we will focus on the results of the table, which is the 𝐹 statistic,
𝑛
∑𝑖=1 (𝑦𝑖̂ − 𝑦)̄ 2 /1
𝐹 = 𝑛 ∼ 𝐹1,𝑛−2
∑𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2 /(𝑛 − 2)
which follows an 𝐹 distribution with degrees of freedom 1 and 𝑛 − 2 under the

null hypothesis. An 𝐹 distribution is a continuous distribution which takes only
positive values and has two parameters, which are the two degrees of freedom.
Recall, in the significance of the regression test, 𝑌 does not depend on 𝑥 in the
null hypothesis.
𝐻0 ∶ 𝛽 1 = 0 𝑌 𝑖 = 𝛽 0 + 𝜖𝑖
While in the alternative hypothesis 𝑌 may depend on 𝑥.
𝐻1 ∶ 𝛽 1 ≠ 0 𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝜖𝑖
We can use the 𝐹 statistic to perform this test.
𝑛
∑𝑖=1 (𝑦𝑖̂ − 𝑦)̄ 2 /1
𝐹 = 𝑛
∑𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2 /(𝑛 − 2)
In particular, we will reject the null when the 𝐹 statistic is large, that is, when
there is a low probability that the observations could have come from the null
model by chance. We will let R calculate the p-value for us.
To perform the 𝐹 test in R you can look at the last row of the output from
summary() called F-statistic which gives the value of the test statistic, the
relevant degrees of freedom, as well as the p-value of the test.
##
## Call:
##
## Residuals:
## -29.069 -9.525 -2.272 9.215 43.201
##
## Coefficients:
## (Intercept) -17.5791 6.7584 -2.601 0.0123 *
## speed 3.9324 0.4155 9.464 1.49e-12 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
Additionally, you can use the anova() function to display the information in an
ANOVA table.
anova(stop_dist_model)
## Analysis of Variance Table

##
## Response: dist
## Df Sum Sq Mean Sq F value Pr(>F)
## speed 1 21186 21185.5 89.567 1.49e-12 ***
8.11. R MARKDOWN 151
## Residuals 48 11354 236.5

## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
This also gives a p-value for the test. You should notice that the p-value from
the 𝑡 test was the same. You might also notice that the value of the test statistic
for the 𝑡 test, 9.46399, can be squared to obtain the value of the 𝐹 statistic,
89.5671065.
Note that there is another equivalent way to do this in R, which we will return
to often to compare two models.
anova(lm(dist ~ 1, data = cars), lm(dist ~ speed, data = cars))

##
## Model 1: dist ~ 1
## Model 2: dist ~ speed
## Res.Df RSS Df Sum of Sq F Pr(>F)
## 1 49 32539
## 2 48 11354 1 21186 89.567 1.49e-12 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
The model statement lm(dist ~ 1, data = cars) applies the model 𝑌𝑖 =

𝛽0 + 𝜖𝑖 to the cars data. Note that 𝑦 ̂ = 𝑦 ̄ when 𝑌𝑖 = 𝛽0 + 𝜖𝑖 .
The model statement lm(dist ~ speed, data = cars) applies the model 𝑌𝑖 =
𝛽0 + 𝛽1 𝑥𝑖 + 𝜖𝑖 .
We can then think of this usage of anova() as directly comparing the two
models. (Notice we get the same p-value again.)
8.11 R Markdown
• slr-inf.Rmd

Chapter 9
Multiple Linear Regression
“Life is really simple, but we insist on making it complicated.”

— Confucius
• Construct and interpret linear regression models with more than one pre-
dictor.
• Understand how regression models are derived using matrices.
• Create interval estimates and perform hypothesis tests for multiple regres-
sion parameters.
• Formulate and interpret interval estimates for the mean response under
various conditions.
• Compare nested models using an ANOVA F-Test.
The last two chapters we saw how to fit a model that assumed a linear relation-
ship between a response variable and a single predictor variable. Specifically,
we defined the simple linear regression model,
However, it is rarely the case that a dataset will have a single predictor variable.
It is also rarely the case that a response variable will only depend on a single
variable. So in this chapter, we will extend our current linear model to allow a
response to depend on multiple predictors.
153
154 CHAPTER 9. MULTIPLE LINEAR REGRESSION
# read the data from the web

autompg = read.table(
"http://archive.ics.uci.edu/ml/machine-learning-databases/auto-mpg/auto-mpg.data",
quote = "\"",
comment.char = "",
stringsAsFactors = FALSE)
# give the dataframe headers
colnames(autompg) = c("mpg", "cyl", "disp", "hp", "wt", "acc", "year", "origin", "name"
# remove missing data, which is stored as "?"
autompg = subset(autompg, autompg$hp != "?")
# remove the plymouth reliant, as it causes some issues
autompg = subset(autompg, autompg$name != "plymouth reliant")
# give the dataset row names, based on the engine, year and name
rownames(autompg) = paste(autompg$cyl, "cylinder", autompg$year, autompg$name)
# remove the variable for name, as well as origin
autompg = subset(autompg, select = c("mpg", "cyl", "disp", "hp", "wt", "acc", "year"))
# change horsepower from character to numeric
autompg$hp = as.numeric(autompg$hp)
# check final structure of data
str(autompg)

## $ mpg : num 18 15 18 16 17 15 14 14 14 15 ...
## $ cyl : int 8 8 8 8 8 8 8 8 8 8 ...
## $ disp: num 307 350 318 304 302 429 454 440 455 390 ...
## $ hp : num 130 165 150 150 140 198 220 215 225 190 ...
## $ wt : num 3504 3693 3436 3433 3449 ...
## $ acc : num 12 11.5 11 12 10.5 10 9 8.5 10 8.5 ...
## $ year: int 70 70 70 70 70 70 70 70 70 70 ...
We will once again discuss a dataset with information about cars. This dataset,
which can be found at the UCI Machine Learning Repository contains a response
variable mpg which stores the city fuel efficiency of cars, as well as several pre-
dictor variables for the attributes of the vehicles. We load the data, and perform
some basic tidying before moving on to analysis.
For now we will focus on using two variables, wt and year, as predictor variables.
That is, we would like to model the fuel efficiency (mpg) of a car as a function
of its weight (wt) and model year (year). To do so, we will define the following
linear model,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + 𝜖𝑖 , 𝑖 = 1, 2, … , 𝑛
where 𝜖𝑖 ∼ 𝑁 (0, 𝜎2 ). In this notation we will define:

155
• 𝑥𝑖1 as the weight (wt) of the 𝑖th car.

• 𝑥𝑖2 as the model year (year) of the 𝑖th car.
The picture below will visualize what we would like to accomplish. The data
points (𝑥𝑖1 , 𝑥𝑖2 , 𝑦𝑖 ) now exist in 3-dimensional space, so instead of fitting a line
to the data, we will fit a plane. (We’ll soon move to higher dimensions, so this
will be the last example that is easy to visualize and think about this way.)
40
30
40
30 20
82
mp
20
80
g
10 78
0 76 10
r
yea
2000
74
3000
wt 4000 72
500070 0
How do we find such a plane? Well, we would like a plane that is as close as
possible to the data points. That is, we would like it to minimize the errors it
is making. How will we define these errors? Squared distance of course! So, we
would like to minimize
𝑛
𝑓(𝛽0 , 𝛽1 , 𝛽2 ) = ∑(𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 ))2
𝑖=1
with respect to 𝛽0 , 𝛽1 , and 𝛽2 . How do we do so? It is another straightforward

multivariate calculus problem. All we have done is add an extra variable since
we did this last time. So again, we take a derivative with respect to each of
𝛽0 , 𝛽1 , and 𝛽2 and set them equal to zero, then solve the resulting system of
equations. That is,
𝜕𝑓
=0
𝜕𝛽0
𝜕𝑓
=0
𝜕𝛽1
𝜕𝑓
=0
𝜕𝛽2
After doing so, we will once again obtain the normal equations.
𝑛 𝑛 𝑛
𝑛𝛽0 + 𝛽1 ∑ 𝑥𝑖1 + 𝛽2 ∑ 𝑥𝑖2 = ∑ 𝑦𝑖
𝑖=1 𝑖=1 𝑖=1
𝑛 𝑛 𝑛 𝑛
𝛽0 ∑ 𝑥𝑖1 + 𝛽1 ∑ 𝑥2𝑖1 + 𝛽2 ∑ 𝑥𝑖1 𝑥𝑖2 = ∑ 𝑥𝑖1 𝑦𝑖
𝑖=1 𝑖=1 𝑖=1 𝑖=1
𝑛 𝑛 𝑛 𝑛
𝛽0 ∑ 𝑥𝑖2 + 𝛽1 ∑ 𝑥𝑖1 𝑥𝑖2 + 𝛽2 ∑ 𝑥2𝑖2 = ∑ 𝑥𝑖2 𝑦𝑖
𝑖=1 𝑖=1 𝑖=1 𝑖=1
We now have three equations and three variables, which we could solve, or we
could simply let R solve for us.
mpg_model = lm(mpg ~ wt + year, data = autompg)

coef(mpg_model)
## (Intercept) wt year
## -14.637641945 -0.006634876 0.761401955
𝑦 ̂ = −14.6376419 + −0.0066349𝑥1 + 0.761402𝑥2
Here we have once again fit our model using lm(), however we have introduced
a new syntactical element. The formula mpg ~ wt + year now reads: “model
the response variable mpg as a linear function of wt and year.” That is, it will
estimate an intercept, as well as slope coefficients for wt and year. We then
extract these as we have done before using coef().
In the multiple linear regression setting, some of the interpretations of the co-
efficients change slightly.
9.1. MATRIX APPROACH TO REGRESSION 157
Here, 𝛽0̂ = −14.6376419 is our estimate for 𝛽0 , the mean miles per gallon for
a car that weighs 0 pounds and was built in 1900. We see our estimate here is
negative, which is a physical impossibility. However, this isn’t unexpected, as
we shouldn’t expect our model to be accurate for cars from 1900 which weigh 0
pounds. (Because they never existed!) This isn’t much of a change from SLR.
That is, 𝛽0 is still simply the mean when all of the predictors are 0.
The interpretation of the coefficients in front of our predictors is slightly different
than before. For example 𝛽1̂ = −0.0066349 is our estimate for 𝛽1 , the average
change in miles per gallon for an increase in weight (𝑥1 ) of one-pound for a
car of a certain model year, that is, for a fixed value of 𝑥2 . Note that this
coefficient is actually the same for any given value of 𝑥2 . Later, we will look at
models that allow for a different change in mean response for different values
of 𝑥2 . Also note that this estimate is negative, which we would expect since, in
general, fuel efficiency decreases for larger vehicles. Recall that in the multiple
linear regression setting, this interpretation is dependent on a fixed value for
𝑥2 , that is, “for a car of a certain model year.” It is possible that the indirect
relationship between fuel efficiency and weight does not hold when an additional
factor, say year, is included, and thus we could have the sign of our coefficient
flipped.
Lastly, 𝛽2̂ = 0.761402 is our estimate for 𝛽2 , the average change in miles per
gallon for a one-year increase in model year (𝑥2 ) for a car of a certain weight, that
is, for a fixed value of 𝑥1 . It is not surprising that the estimate is positive. We
expect that as time passes and the years march on, technology would improve
so that a car of a specific weight would get better mileage now as compared to
their predecessors. And yet, the coefficient could have been negative because
we are also including weight as variable, and not strictly as a fixed value.
9.1 Matrix Approach to Regression
In our above example we used two predictor variables, but it will only take a
little more work to allow for an arbitrary number of predictor variables and
derive their coefficient estimates. We can consider the model,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽𝑝−1 𝑥𝑖(𝑝−1) + 𝜖𝑖 , 𝑖 = 1, 2, … , 𝑛
where 𝜖𝑖 ∼ 𝑁 (0, 𝜎2 ). In this model, there are 𝑝 − 1 predictor variables,

𝑥1 , 𝑥2 , ⋯ , 𝑥𝑝−1 . There are a total of 𝑝 𝛽-parameters and a single parameter
𝜎2 for the variance of the errors. (It should be noted that almost as often,
authors will use 𝑝 as the number of predictors, making the total number of
𝛽 parameters 𝑝 + 1. This is always something you should be aware of when
reading about multiple regression. There is not a standard that is used most
often.)
If we were to stack together the 𝑛 linear equations that represent each 𝑌𝑖 into
a column vector, we get the following.
𝛽
𝑌1 1 𝑥11 𝑥12 ⋯ 𝑥1(𝑝−1) ⎡ 0 ⎤ 𝜖1
⎡ 𝑌 ⎤ ⎡1 𝑥 ⎤ 𝛽
21 𝑥22 ⋯ 𝑥2(𝑝−1) ⎢ 1 ⎥ ⎡ 𝜖2 ⎤
⎢ 2⎥ = ⎢ ⎥ 𝛽2 ⎥ + ⎢ ⎥
⎢ ⋮ ⎥ ⎢⋮ ⋮ ⋮ ⋮ ⎥⎢ ⎢ ⋮ ⎥
⎢⋮⎥
𝑌
⎣ 𝑛⎦ ⎣ 1 𝑥 𝑛1 𝑥𝑛2 ⋯ 𝑥𝑛(𝑝−1) ⎦ ⎣𝜖𝑛 ⎦
⎣𝛽𝑝−1 ⎦
𝑌 = 𝑋𝛽 + 𝜖
𝛽0
𝑌1 1 𝑥11 𝑥12 ⋯ 𝑥1(𝑝−1) ⎡ 𝛽 ⎤ 𝜖1
⎡𝑌 ⎤ ⎡1 𝑥 𝑥22 ⋯ 𝑥2(𝑝−1) ⎤ ⎢ 1 ⎥ ⎡𝜖 ⎤
21
𝑌 = ⎢ 2⎥ , 𝑋=⎢ ⎥, 𝛽 = ⎢ 𝛽2 ⎥ , 𝜖 = ⎢ 2⎥
⎢ ⋮ ⎥ ⎢⋮ ⋮ ⋮ ⋮ ⎥ ⎢⋮⎥
⎢ ⋮ ⎥
⎣𝑌𝑛 ⎦ ⎣1 𝑥𝑛1 𝑥𝑛2 ⋯ 𝑥𝑛(𝑝−1) ⎦ ⎣𝜖𝑛 ⎦
⎣𝛽𝑝−1 ⎦
So now with data,
𝑦1
⎡𝑦 ⎤
𝑦 = ⎢ 2⎥
⎢ ⋮ ⎥
⎣𝑦𝑛 ⎦
Just as before, we can estimate 𝛽 by minimizing,
𝑛
𝑓(𝛽0 , 𝛽1 , 𝛽2 , ⋯ , 𝛽𝑝−1 ) = ∑(𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽𝑝−1 𝑥𝑖(𝑝−1) ))2 ,
𝑖=1
which would require taking 𝑝 derivatives, which result in the following normal
equations.
𝑛 𝑛 𝑛 𝑛
𝑛 ∑𝑖=1 𝑥𝑖1 ∑𝑖=1 𝑥𝑖2 ⋯ ∑𝑖=1 𝑥𝑖(𝑝−1) 𝛽0 ∑𝑖=1 𝑦𝑖
⎡ 𝑛
∑𝑖=1 𝑥𝑖1
𝑛
∑𝑖=1 𝑥2𝑖1
𝑛
∑𝑖=1 𝑥𝑖1 𝑥𝑖2 ⋯ ∑𝑖=1 𝑥𝑖1 𝑥𝑖(𝑝−1) ⎤
𝑛 ⎡ 𝛽 ⎤ ⎡ ∑𝑛 𝑥 𝑦 ⎤
⎢ ⎥⎢ 1 ⎥ = ⎢ 𝑖=1 𝑖1 𝑖 ⎥
⎢ ⋮ ⋮ ⋮ ⋮ ⎥⎢ ⋮ ⎥ ⎢ ⋮ ⎥
𝑛 𝑛 𝑛 𝑛 𝑛
⎣∑𝑖=1 𝑥𝑖(𝑝−1) ∑𝑖=1 𝑥𝑖(𝑝−1) 𝑥𝑖1 ∑𝑖=1 𝑥𝑖(𝑝−1) 𝑥𝑖2 ⋯ ∑𝑖=1 𝑥2𝑖(𝑝−1) ⎦ ⎣𝛽𝑝−1 ⎦ ⎣∑𝑖=1 𝑥𝑖(𝑝−1) 𝑦𝑖 ⎦
The normal equations can be written much more succinctly in matrix notation,
𝑋 ⊤ 𝑋𝛽 = 𝑋 ⊤ 𝑦.
9.1. MATRIX APPROACH TO REGRESSION 159
We can then solve this expression by multiplying both sides by the inverse of
𝑋 ⊤ 𝑋, which exists, provided the columns of 𝑋 are linearly independent. Then
as always, we denote our solution with a hat.
−1
𝛽 ̂ = (𝑋 ⊤ 𝑋) 𝑋⊤𝑦
To verify that this is what R has done for us in the case of two predictors, we
create an 𝑋 matrix. Note that the first column is all 1s, and the remaining
columns contain the data.
n = nrow(autompg)
p = length(coef(mpg_model))
X = cbind(rep(1, n), autompg$wt, autompg$year)
y = autompg$mpg
(beta_hat = solve(t(X) %*% X) %*% t(X) %*% y)
## [,1]
## [1,] -14.637641945
## [2,] -0.006634876
## [3,] 0.761401955
coef(mpg_model)
## (Intercept) wt year
## -14.637641945 -0.006634876 0.761401955
−14.6376419
𝛽̂ = ⎡ ⎤
⎢ −0.0066349 ⎥
⎣ 0.761402 ⎦
In our new notation, the fitted values can be written
𝑦 ̂ = 𝑋 𝛽.̂
𝑦1̂
⎡𝑦 ̂ ⎤
𝑦̂ = ⎢ 2⎥
⎢ ⋮ ⎥
⎣𝑦𝑛̂ ⎦
Then, we can create a vector for the residual values,

𝑒1 𝑦1 𝑦1̂
⎡𝑒 ⎤ ⎡𝑦 ⎤ ⎡𝑦 ̂ ⎤
𝑒 = ⎢ 2⎥ = ⎢ 2⎥ − ⎢ 2⎥ .
⎢⋮⎥ ⎢ ⋮ ⎥ ⎢ ⋮ ⎥
⎣𝑒𝑛 ⎦ ⎣𝑦𝑛 ⎦ ⎣𝑦𝑛̂ ⎦
And lastly, we can update our estimate for 𝜎2 .
𝑛
∑𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2 𝑒⊤ 𝑒
𝑠2𝑒 = =
𝑛−𝑝 𝑛−𝑝
Recall, we like this estimate because it is unbiased, that is,
E[𝑠2𝑒 ] = 𝜎2
Note that the change from the SLR estimate to now is in the denominator.
Specifically we now divide by 𝑛 − 𝑝 instead of 𝑛 − 2. Or actually, we should note
that in the case of SLR, there are two 𝛽 parameters and thus 𝑝 = 2.
Also note that if we fit the model 𝑌𝑖 = 𝛽 + 𝜖𝑖 that 𝑦 ̂ = 𝑦 ̄ and 𝑝 = 1 and 𝑠2𝑒 would
become
𝑛
∑𝑖=1 (𝑦𝑖 − 𝑦)̄ 2
𝑠2𝑒 =
𝑛−1
which is likely the very first sample variance you saw in a mathematical statistics
class. The same reason for 𝑛−1 in this case, that we estimated one parameter, so
we lose one degree of freedom. Now, in general, we are estimating 𝑝 parameters,
the 𝛽 parameters, so we lose 𝑝 degrees of freedom.
Also, recall that most often we will be interested in 𝑠𝑒 , the residual standard
error as R calls it,
𝑛
∑𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2
𝑠𝑒 = √ .
𝑛−𝑝
In R, we could directly access 𝑠𝑒 for a fitted model, as we have seen before.
summary(mpg_model)$sigma
## [1] 3.431367
And we can now verify that our math above is indeed calculating the same
quantities.
9.2. SAMPLING DISTRIBUTION 161
y_hat = X %*% solve(t(X) %*% X) %*% t(X) %*% y

e = y - y_hat
sqrt(t(e) %*% e / (n - p))
## [,1]
## [1,] 3.431367
sqrt(sum((y - y_hat) ^ 2) / (n - p))
## [1] 3.431367
9.2 Sampling Distribution

As we can see in the output below, the results of calling summary() are similar
to SLR, but there are some differences, most obviously a new row for the added
predictor variable.
summary(mpg_model)
##
## Call:
## lm(formula = mpg ~ wt + year, data = autompg)
##
## Residuals:
## -8.852 -2.292 -0.100 2.039 14.325
##
## Coefficients:
## (Intercept) -1.464e+01 4.023e+00 -3.638 0.000312 ***
## wt -6.635e-03 2.149e-04 -30.881 < 2e-16 ***
## year 7.614e-01 4.973e-02 15.312 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## F-statistic: 815.6 on 2 and 387 DF, p-value: < 2.2e-16
To understand these differences in detail, we will need to first obtain the sam-
pling distribution of 𝛽.̂
The derivation of the sampling distribution of 𝛽 ̂ involves the multivariate normal

distribution. These brief notes from semesters past give a basic overview. These
are simply for your information, as we will not present the derivation in full here.
Our goal now is to obtain the distribution of the 𝛽 ̂ vector,
𝛽0̂
⎡ ̂ ⎤
⎢ 𝛽1 ⎥
𝛽 ̂ = ⎢ 𝛽2̂ ⎥
⎢ ⎥
⎢ ⋮ ⎥
̂ ⎦
⎣𝛽𝑝−1
Recall from last time that when discussing sampling distributions, we now con-
sider 𝛽 ̂ to be a random vector, thus we use 𝑌 instead of the data vector 𝑦.
−1
𝛽 ̂ = (𝑋 ⊤ 𝑋) 𝑋⊤𝑌
Then it is a consequence of the multivariate normal distribution that,
−1
𝛽 ̂ ∼ 𝑁 (𝛽, 𝜎2 (𝑋 ⊤ 𝑋) ) .
We then have
E[𝛽]̂ = 𝛽
and for any 𝛽𝑗̂ we have
E[𝛽𝑗̂ ] = 𝛽𝑗 .
We also have
−1
Var[𝛽]̂ = 𝜎2 (𝑋 ⊤ 𝑋)
and for any 𝛽𝑗̂ we have
Var[𝛽𝑗̂ ] = 𝜎2 𝐶𝑗𝑗
where
−1
𝐶 = (𝑋 ⊤ 𝑋)
and the elements of 𝐶 are denoted

𝐶00 𝐶01 𝐶02 ⋯ 𝐶0(𝑝−1)

⎡ 𝐶 𝐶11 𝐶12 ⋯ 𝐶1(𝑝−1) ⎤
⎢ 10 ⎥
𝐶 = ⎢ 𝐶20 𝐶21 𝐶22 ⋯ 𝐶2(𝑝−1) ⎥.
⎢ ⋮ ⋮ ⋮ ⋮ ⎥
⎣𝐶(𝑝−1)0 𝐶(𝑝−1)1 𝐶(𝑝−1)2 ⋯ 𝐶(𝑝−1)(𝑝−1) ⎦
Essentially, the diagonal elements correspond to the 𝛽 vector.

Then the standard error for the 𝛽 ̂ vector is given by
−1
SE[𝛽]̂ = 𝑠𝑒 √(𝑋 ⊤ 𝑋)
and for a particular 𝛽𝑗̂
SE[𝛽𝑗̂ ] = 𝑠𝑒 √𝐶𝑗𝑗 .
Lastly, each of the 𝛽𝑗̂ follows a normal distribution,
𝛽𝑗̂ ∼ 𝑁 (𝛽𝑗 , 𝜎2 𝐶𝑗𝑗 ) .
thus
𝛽𝑗̂ − 𝛽𝑗
∼ 𝑡𝑛−𝑝 .
𝑠𝑒 √𝐶𝑗𝑗
Now that we have the necessary distributional results, we can move on to per-
form tests and make interval estimates.
9.2.1 Single Parameter Tests
The first test we will see is a test for a single 𝛽𝑗 .
𝐻0 ∶ 𝛽 𝑗 = 0 vs 𝐻1 ∶ 𝛽𝑗 ≠ 0
Again, the test statistic takes the form
EST − HYP
TS = .
SE
In particular,
𝛽𝑗̂ − 𝛽𝑗 𝛽𝑗̂ − 0
𝑡= = ,
SE[𝛽𝑗̂ ] 𝑠𝑒 √𝐶𝑗𝑗
which, under the null hypothesis, follows a 𝑡 distribution with 𝑛 − 𝑝 degrees of

freedom.
Recall our model for mpg,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + 𝜖𝑖 , 𝑖 = 1, 2, … , 𝑛
• 𝑥𝑖1 as the weight (wt) of the 𝑖th car.

• 𝑥𝑖2 as the model year (year) of the 𝑖th car.
Then the test
𝐻0 ∶ 𝛽 1 = 0 vs 𝐻1 ∶ 𝛽1 ≠ 0
can be found in the summary() output, in particular:
summary(mpg_model)$coef

## (Intercept) -14.637641945 4.0233913563 -3.638135 3.118311e-04
## wt -0.006634876 0.0002148504 -30.881372 1.850466e-106
## year 0.761401955 0.0497265950 15.311765 1.036597e-41
The estimate (Estimate), standard error (Std. Error), test statistic (t

value), and p-value (Pr(>|t|)) for this test are displayed in the second row,
labeled wt. Remember that the p-value given here is specifically for a two-sided
test, where the hypothesized value is 0.
Also note in this case, by hypothesizing that 𝛽1 = 0 the null and alternative
essentially specify two different models:
• 𝐻0 : 𝑌 = 𝛽 0 + 𝛽 2 𝑥 2 + 𝜖
• 𝐻1 : 𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝜖
This is important. We are not simply testing whether or not there is a relation-
ship between weight and fuel efficiency. We are testing if there is a relationship
between weight and fuel efficiency, given that a term for year is in the model.
(Note, we dropped some indexing here, for readability.)
9.2.2 Confidence Intervals
Since 𝛽𝑗̂ is our estimate for 𝛽𝑗 and we have
E[𝛽𝑗̂ ] = 𝛽𝑗
as well as the standard error,
SE[𝛽𝑗̂ ] = 𝑠𝑒 √𝐶𝑗𝑗
and the sampling distribution of 𝛽𝑗̂ is Normal, then we can easily construct
confidence intervals for each of the 𝛽𝑗̂ .
𝛽𝑗̂ ± 𝑡𝛼/2,𝑛−𝑝 ⋅ 𝑠𝑒 √𝐶𝑗𝑗
We can find these in R using the same method as before. Now there will simply
be additional rows for the additional 𝛽.
confint(mpg_model, level = 0.99)
## 0.5 % 99.5 %
## (Intercept) -25.052563681 -4.222720208
## wt -0.007191036 -0.006078716
## year 0.632680051 0.890123859
9.2.3 Confidence Intervals for Mean Response
As we saw in SLR, we can create confidence intervals for the mean response,
that is, an interval estimate for E[𝑌 ∣ 𝑋 = 𝑥]. In SLR, the mean of 𝑌 was only
dependent on a single value 𝑥. Now, in multiple regression, E[𝑌 ∣ 𝑋 = 𝑥] is
dependent on the value of each of the predictors, so we define the vector 𝑥0 to
be,
1
⎡ 𝑥 ⎤
⎢ 01 ⎥
𝑥0 = ⎢ 𝑥02 ⎥ .
⎢ ⋮ ⎥
⎣𝑥0(𝑝−1) ⎦
Then our estimate of E[𝑌 ∣ 𝑋 = 𝑥0 ] for a set of values 𝑥0 is given by

̂ 0 ) = 𝑥⊤
𝑦(𝑥 ̂
0𝛽
= 𝛽0̂ + 𝛽1̂ 𝑥01 + 𝛽2̂ 𝑥02 + ⋯ + 𝛽𝑝−1

̂ 𝑥
0(𝑝−1) .
As with SLR, this is an unbiased estimate.
̂ 0 )] = 𝑥⊤
E[𝑦(𝑥 0𝛽
= 𝛽0 + 𝛽1 𝑥01 + 𝛽2 𝑥02 + ⋯ + 𝛽𝑝−1 𝑥0(𝑝−1)
To make an interval estimate, we will also need its standard error.
−1
̂ 0 )] = 𝑠𝑒 √𝑥⊤
SE[𝑦(𝑥 ⊤
0 (𝑋 𝑋) 𝑥0
Putting it all together, we obtain a confidence interval for the mean response.
−1
̂ 0 ) ± 𝑡𝛼/2,𝑛−𝑝 ⋅ 𝑠𝑒 √𝑥⊤
𝑦(𝑥 ⊤
0 (𝑋 𝑋) 𝑥0
The math has changed a bit, but the process in R remains almost identical.
Here, we create a data frame for two additional cars. One car that weighs 3500
pounds produced in 1976, as well as a second car that weighs 5000 pounds which
was produced in 1981.
new_cars = data.frame(wt = c(3500, 5000), year = c(76, 81))

new_cars
## wt year
## 1 3500 76
## 2 5000 81
We can then use the predict() function with interval = "confidence" to

obtain intervals for the mean fuel efficiency for both new cars. Again, it is
important to make the data passed to newdata a data frame, so that R knows
which values are for which variables.
predict(mpg_model, newdata = new_cars, interval = "confidence", level = 0.99)
## fit lwr upr

## 1 20.00684 19.4712 20.54248
## 2 13.86154 12.3341 15.38898
R then reports the estimate 𝑦(𝑥

̂ 0 ) (fit) for each, as well as the lower (lwr) and
upper (upr) bounds for the interval at a desired level (99%).
A word of caution here: one of these estimates is good while one is suspect.
new_cars$wt
## [1] 3500 5000
range(autompg$wt)
## [1] 1613 5140
Note that both of the weights of the new cars are within the range of observed
values.
new_cars$year
## [1] 76 81
range(autompg$year)
## [1] 70 82
As are the years of each of the new cars.
plot(year ~ wt, data = autompg, pch = 20, col = "dodgerblue", cex = 1.5)
points(new_cars, col = "darkorange", cex = 3, pch = "X")
82
80
78 X
X
year
76
74
72
70
1500 2000 2500 3000 3500 4000 4500 5000
wt
However, we have to consider weight and year together now. And based on the
above plot, one of the new cars is within the “blob” of observed values, while
the other, the car from 1981 weighing 5000 pounds, is noticeably outside of the
observed values. This is a hidden extrapolation which you should be aware of
when using multiple regression.
Shifting gears back to the new data pair that can be reasonably estimated, we
do a quick verification of some of the mathematics in R.
x0 = c(1, 3500, 76)

x0 %*% beta_hat
## [,1]
## [1,] 20.00684
1
𝑥0 = ⎡
⎢3500⎤
⎥
⎣ 76 ⎦
−14.6376419
𝛽̂ = ⎡ ⎤
⎢ −0.0066349 ⎥
⎣ 0.761402 ⎦
−14.6376419
̂ 0 ) = 𝑥⊤
𝑦(𝑥 ̂
0 𝛽 = [1 3500 76] ⎡ ⎤
⎢ −0.0066349 ⎥ = 20.0068411
⎣ 0.761402 ⎦
Also note that, using a particular value for 𝑥0 , we can essentially extract certain
𝛽𝑗̂ values.
beta_hat
## [,1]
## [1,] -14.637641945
## [2,] -0.006634876
## [3,] 0.761401955
x0 = c(0, 0, 1)
x0 %*% beta_hat
## [,1]
## [1,] 0.761402
With this in mind, confidence intervals for the individual 𝛽𝑗̂ are actually a special
case of a confidence interval for mean response.
9.2.4 Prediction Intervals
As with SLR, creating prediction intervals involves one slight change to the stan-
dard error to account for the fact that we are now considering an observation,
instead of a mean.
Here we use 𝑦(𝑥
̂ 0 ) to estimate 𝑌0 , a new observation of 𝑌 at the predictor vector
𝑥0 .
̂ 0 ) = 𝑥⊤
𝑦(𝑥 ̂
0𝛽
= 𝛽0̂ + 𝛽1̂ 𝑥01 + 𝛽2̂ 𝑥02 + ⋯ + 𝛽𝑝−1

̂ 𝑥
0(𝑝−1)
̂ 0 )] = 𝑥⊤
E[𝑦(𝑥 0𝛽
= 𝛽0 + 𝛽1 𝑥01 + 𝛽2 𝑥02 + ⋯ + 𝛽𝑝−1 𝑥0(𝑝−1)
As we did with SLR, we need to account for the additional variability of an

observation about its mean.
−1
̂ 0 ) + 𝜖] = 𝑠𝑒 √1 + 𝑥⊤
SE[𝑦(𝑥 ⊤
0 (𝑋 𝑋) 𝑥0
Then we arrive at our updated prediction interval for MLR.
−1
̂ 0 ) ± 𝑡𝛼/2,𝑛−𝑝 ⋅ 𝑠𝑒 √1 + 𝑥⊤
𝑦(𝑥 ⊤
0 (𝑋 𝑋) 𝑥0
new_cars
## wt year
## 1 3500 76
## 2 5000 81
predict(mpg_model, newdata = new_cars, interval = "prediction", level = 0.99)
## fit lwr upr

## 1 20.00684 11.108294 28.90539
## 2 13.86154 4.848751 22.87432
9.3 Significance of Regression
The decomposition of variation that we had seen in SLR still holds for MLR.
𝑛 𝑛 𝑛
∑(𝑦𝑖 − 𝑦)̄ 2 = ∑(𝑦𝑖 − 𝑦𝑖̂ )2 + ∑(𝑦𝑖̂ − 𝑦)̄ 2
𝑖=1 𝑖=1 𝑖=1
That is,
SST = SSE + SSReg.
This means that we can still calculate 𝑅2 in the same manner as before, which
R continues to do automatically.
summary(mpg_model)$r.squared
## [1] 0.8082355
9.3. SIGNIFICANCE OF REGRESSION 171
The interpretation changes slightly as compared to SLR. In this MLR case, we

say that 80.82% of the observed variation in miles per gallon is explained by the
linear relationship with the two predictor variables, weight and year.
In multiple regression, the significance of regression test is
𝐻0 ∶ 𝛽1 = 𝛽2 = ⋯ = 𝛽𝑝−1 = 0.
Here, we see that the null hypothesis sets all of the 𝛽𝑗 equal to 0, except the
intercept, 𝛽0 . We could then say that the null model, or “model under the null
hypothesis” is
𝑌𝑖 = 𝛽0 + 𝜖𝑖 .
This is a model where the regression is insignificant. None of the predictors

have a significant linear relationship with the response. Notationally, we will
denote the fitted values of this model as 𝑦0𝑖
̂ , which in this case happens to be:
𝑦0𝑖
̂ = 𝑦.̄
The alternative hypothesis here is that at least one of the 𝛽𝑗 from the null
hypothesis is not 0.
𝐻1 ∶ At least one of 𝛽𝑗 ≠ 0, 𝑗 = 1, 2, ⋯ , (𝑝 − 1)
We could then say that the full model, or “model under the alternative hypoth-
esis” is
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽(𝑝−1) 𝑥𝑖(𝑝−1) + 𝜖𝑖
This is a model where the regression is significant. At least one of the predic-
tors has a significant linear relationship with the response. There is some linear
relationship between 𝑦 and the predictors, 𝑥1 , 𝑥2 , … , 𝑥𝑝−1 .
We will denote the fitted values of this model as 𝑦1𝑖
̂ .
To develop the 𝐹 test for the significance of the regression, we will arrange the
variance decomposition into an ANOVA table.
Degrees of Mean
𝑛 2
Regression ∑𝑖=1 (𝑦1𝑖
̂ − 𝑦)̄ 𝑝−1 SSReg/(𝑝 − MSReg/MSE
1)
𝑛
Error ̂ )2
∑𝑖=1 (𝑦𝑖 − 𝑦1𝑖 𝑛−𝑝 SSE/(𝑛 − 𝑝)
Degrees of Mean
𝑛
Total ∑𝑖=1 (𝑦𝑖 − 𝑦)̄ 2 𝑛−1
In summary, the 𝐹 statistic is
𝑛
̂ − 𝑦)̄ 2 /(𝑝 − 1)
∑𝑖=1 (𝑦1𝑖
𝐹 = 𝑛 ,
̂ )2 /(𝑛 − 𝑝)
∑𝑖=1 (𝑦𝑖 − 𝑦1𝑖
and the p-value is calculated as
𝑃 (𝐹𝑝−1,𝑛−𝑝 > 𝐹 )
since we reject for large values of 𝐹 . A large value of the statistic corresponds to
a large portion of the variance being explained by the regression. Here 𝐹𝑝−1,𝑛−𝑝
represents a random variable which follows an 𝐹 distribution with 𝑝 − 1 and
𝑛 − 𝑝 degrees of freedom.
To perform this test in R, we first explicitly specify the two models in R and
save the results in different variables. We then use anova() to compare the
two models, giving anova() the null model first and the alternative (full) model
second. (Specifying the full model first will result in the same p-value, but some
nonsensical intermediate values.)
In this case,
• 𝐻0 : 𝑌 𝑖 = 𝛽 0 + 𝜖 𝑖
• 𝐻1 : 𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + 𝜖𝑖
That is, in the null model, we use neither of the predictors, whereas in the full
(alternative) model, at least one of the predictors is useful.
null_mpg_model = lm(mpg ~ 1, data = autompg)

full_mpg_model = lm(mpg ~ wt + year, data = autompg)
anova(null_mpg_model, full_mpg_model)

##
## Model 1: mpg ~ 1
## Model 2: mpg ~ wt + year
## 1 389 23761.7
## 2 387 4556.6 2 19205 815.55 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
9.3. SIGNIFICANCE OF REGRESSION 173
First, notice that R does not display the results in the same manner as the table
above. More important than the layout of the table are its contents. We see
that the value of the 𝐹 statistic is 815.55, and the p-value is extremely low, so
we reject the null hypothesis at any reasonable 𝛼 and say that the regression is
significant. At least one of wt or year has a useful linear relationship with mpg.
summary(mpg_model)
##
## Call:
## lm(formula = mpg ~ wt + year, data = autompg)
##
## Residuals:
## -8.852 -2.292 -0.100 2.039 14.325
##
## Coefficients:
## (Intercept) -1.464e+01 4.023e+00 -3.638 0.000312 ***
## wt -6.635e-03 2.149e-04 -30.881 < 2e-16 ***
## year 7.614e-01 4.973e-02 15.312 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
Notice that the value reported in the row for F-statistic is indeed the 𝐹 test
statistic for the significance of regression test, and additionally it reports the
two relevant degrees of freedom.
Also, note that none of the individual 𝑡-tests are equivalent to the 𝐹 -test as
they were in SLR. This equivalence only holds for SLR because the individual
test for 𝛽1 is the same as testing for all non-intercept parameters, since there is
only one.
We can also verify the sums of squares and degrees of freedom directly in R. You
should match these to the table from R and use this to match R’s output to the
written table above.
# SSReg
sum((fitted(full_mpg_model) - fitted(null_mpg_model)) ^ 2)
## [1] 19205.03
# SSE
sum(resid(full_mpg_model) ^ 2)
## [1] 4556.646
# SST
sum(resid(null_mpg_model) ^ 2)
## [1] 23761.67
# Degrees of Freedom: Regression

length(coef(full_mpg_model)) - length(coef(null_mpg_model))
## [1] 2
# Degrees of Freedom: Error

length(resid(full_mpg_model)) - length(coef(full_mpg_model))
## [1] 387
# Degrees of Freedom: Total

length(resid(null_mpg_model)) - length(coef(null_mpg_model))
## [1] 389
9.4 Nested Models

The significance of regression test is actually a special case of testing what we
will call nested models. More generally we can compare two models, where
one model is “nested” inside the other, meaning one model contains a subset of
the predictors from only the larger model.
Consider the following full model,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽(𝑝−1) 𝑥𝑖(𝑝−1) + 𝜖𝑖
This model has 𝑝 − 1 predictors, for a total of 𝑝 𝛽-parameters. We will denote

the fitted values of this model as 𝑦1𝑖
̂ .
Let the null model be
9.4. NESTED MODELS 175
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽(𝑞−1) 𝑥𝑖(𝑞−1) + 𝜖𝑖
where 𝑞 < 𝑝. This model has 𝑞 − 1 predictors, for a total of 𝑞 𝛽-parameters. We

will denote the fitted values of this model as 𝑦0𝑖
̂ .
The difference between these two models can be codified by the null hypothesis
of a test.
𝐻0 ∶ 𝛽𝑞 = 𝛽𝑞+1 = ⋯ = 𝛽𝑝−1 = 0.
Specifically, the 𝛽-parameters from the full model that are not in the null model
are zero. The resulting model, which is nested, is the null model.
We can then perform this test using an 𝐹 -test, which is the result of the following
ANOVA table.
Degrees of Mean
𝑛
Diff ∑𝑖=1 (𝑦1𝑖
̂ − 𝑦0𝑖̂ )2 𝑝−𝑞 SSD/(𝑝 − 𝑞) MSD/MSE
𝑛
Full ̂ )2
∑𝑖=1 (𝑦𝑖 − 𝑦1𝑖 𝑛−𝑝 SSE/(𝑛 − 𝑝)
𝑛
Null ̂ )2
∑𝑖=1 (𝑦𝑖 − 𝑦0𝑖 𝑛−𝑞
𝑛
∑𝑖=1 (𝑦1𝑖 ̂ )2 /(𝑝 − 𝑞)
̂ − 𝑦0𝑖
𝐹 = 𝑛 .
̂ )2 /(𝑛 − 𝑝)
∑𝑖=1 (𝑦𝑖 − 𝑦1𝑖
Notice that the row for “Diff” compares the sum of the squared differences of
the fitted values. The degrees of freedom is then the difference of the number
of 𝛽-parameters estimated between the two models.
For example, the autompg dataset has a number of additional variables that we
have yet to use.
names(autompg)
## [1] "mpg" "cyl" "disp" "hp" "wt" "acc" "year"
We’ll continue to use mpg as the response, but now we will consider two different
models.
• Full: mpg ~ wt + year + cyl + disp + hp + acc

• Null: mpg ~ wt + year
Note that these are nested models, as the null model contains a subset of the
predictors from the full model, and no additional predictors. Both models have
an intercept 𝛽0 as well as a coefficient in front of each of the predictors. We
could then write the null hypothesis for comparing these two models as,
𝐻0 ∶ 𝛽cyl = 𝛽disp = 𝛽hp = 𝛽acc = 0
The alternative is simply that at least one of the 𝛽𝑗 from the null is not 0.
To perform this test in R we first define both models, then give them to the
anova() commands.
null_mpg_model = lm(mpg ~ wt + year, data = autompg)

#full_mpg_model = lm(mpg ~ wt + year + cyl + disp + hp + acc, data = autompg)
full_mpg_model = lm(mpg ~ ., data = autompg)
anova(null_mpg_model, full_mpg_model)

##
## Model 1: mpg ~ wt + year
## Model 2: mpg ~ cyl + disp + hp + wt + acc + year
## 1 387 4556.6
## 2 383 4530.5 4 26.18 0.5533 0.6967
Here we have used the formula mpg ~ . to define to full model. This is the
same as the commented out line. Specifically, this is a common shortcut in R
which reads, “model mpg as the response with each of the remaining variables
in the data frame as predictors.”
Here we see that the value of the 𝐹 statistic is 0.553, and the p-value is very
large, so we fail to reject the null hypothesis at any reasonable 𝛼 and say that
none of cyl, disp, hp, and acc are significant with wt and year already in the
model.
Again, we verify the sums of squares and degrees of freedom directly in R. You
should match these to the table from R, and use this to match R’s output to the
written table above.
# SSDiff
sum((fitted(full_mpg_model) - fitted(null_mpg_model)) ^ 2)
## [1] 26.17981
9.5. SIMULATION 177
# SSE (For Full)

sum(resid(full_mpg_model) ^ 2)
## [1] 4530.466
# SST (For Null)

sum(resid(null_mpg_model) ^ 2)
## [1] 4556.646
# Degrees of Freedom: Diff

length(coef(full_mpg_model)) - length(coef(null_mpg_model))
## [1] 4
# Degrees of Freedom: Full

length(resid(full_mpg_model)) - length(coef(full_mpg_model))
## [1] 383
# Degrees of Freedom: Null

length(resid(null_mpg_model)) - length(coef(null_mpg_model))
## [1] 387
9.5 Simulation
Since we ignored the derivation of certain results, we will again use simulation to
convince ourselves of some of the above results. In particular, we will simulate
samples of size n = 100 from the model
𝑌𝑖 = 5 + −2𝑥𝑖1 + 6𝑥𝑖2 + 𝜖𝑖 , 𝑖 = 1, 2, … , 𝑛
where 𝜖𝑖 ∼ 𝑁 (0, 𝜎2 = 16). Here we have two predictors, so 𝑝 = 3.
set.seed(1337)
n = 100 # sample size
p = 3
beta_0 = 5
beta_1 = -2
beta_2 = 6
sigma = 4
As is the norm with regression, the 𝑥 values are considered fixed and known
quantities, so we will simulate those first, and they remain the same for the rest
of the simulation study. Also note we create an x0 which is all 1, which we need
to create our X matrix. If you look at the matrix formulation of regression, this
unit vector of all 1s is a “predictor” that puts the intercept into the model. We
also calculate the C matrix for later use.
x0 = rep(1, n)
x1 = sample(seq(1, 10, length = n))
x2 = sample(seq(1, 10, length = n))
X = cbind(x0, x1, x2)
C = solve(t(X) %*% X)
We then simulate the response according the model above. Lastly, we place the
two predictors and response into a data frame. Note that we do not place x0
in the data frame. This is a result of R adding an intercept by default.
eps = rnorm(n, mean = 0, sd = sigma)

y = beta_0 + beta_1 * x1 + beta_2 * x2 + eps
sim_data = data.frame(x1, x2, y)
Plotting this data and fitting the regression produces the following plot.
9.5. SIMULATION 179
60
60
50
40
40
y
20
30
0 20
10
2 10
8
4
6
6 x2 0
x1 4
8
2
10
We then calculate
−1
𝛽 ̂ = (𝑋 ⊤ 𝑋) 𝑋 ⊤ 𝑦.
(beta_hat = C %*% t(X) %*% y)
## [,1]
## x0 7.290735
## x1 -2.282176
## x2 5.843424
Notice that these values are the same as the coefficients found using lm() in R.
coef(lm(y ~ x1 + x2, data = sim_data))
## (Intercept) x1 x2
## 7.290735 -2.282176 5.843424
Also, these values are close to what we would expect.
c(beta_0, beta_1, beta_2)
## [1] 5 -2 6
We then calculated the fitted values in order to calculate 𝑠𝑒 , which we see is the
same as the sigma which is returned by summary().
y_hat = X %*% beta_hat

(s_e = sqrt(sum((y - y_hat) ^ 2) / (n - p)))
## [1] 4.294307
summary(lm(y ~ x1 + x2, data = sim_data))$sigma
## [1] 4.294307
So far so good. Everything checks out. Now we will finally simulate from this
model repeatedly in order to obtain an empirical distribution of 𝛽2̂ .
We expect 𝛽2̂ to follow a normal distribution,
𝛽2̂ ∼ 𝑁 (𝛽2 , 𝜎2 𝐶22 ) .
In this case,
𝛽2̂ ∼ 𝑁 (𝜇 = 6, 𝜎2 = 16 × 0.0014534 = 0.0232549) .
𝛽2̂ ∼ 𝑁 (𝜇 = 6, 𝜎2 = 0.0232549) .
Note that 𝐶22 corresponds to the element in the third row and third column
since 𝛽2 is the third parameter in the model and because R is indexed starting
at 1. However, we index the 𝐶 matrix starting at 0 to match the diagonal
elements to the corresponding 𝛽𝑗 .
9.5. SIMULATION 181
C[3, 3]
## [1] 0.00145343
C[2 + 1, 2 + 1]
## [1] 0.00145343
sigma ^ 2 * C[2 + 1, 2 + 1]
## [1] 0.02325487
We now perform the simulation a large number of times. Each time, we update
the y variable in the data frame, leaving the x variables the same. We then fit
a model, and store 𝛽2̂ .
num_sims = 10000
beta_hat_2 = rep(0, num_sims)
for(i in 1:num_sims) {
eps = rnorm(n, mean = 0 , sd = sigma)
sim_data$y = beta_0 * x0 + beta_1 * x1 + beta_2 * x2 + eps
fit = lm(y ~ x1 + x2, data = sim_data)
beta_hat_2[i] = coef(fit)[3]
}
We then see that the mean of the simulated values is close to the true value of
𝛽2 .
mean(beta_hat_2)
## [1] 5.999723
beta_2
## [1] 6
We also see that the variance of the simulated values is close to the true variance
of 𝛽2̂ .
Var[𝛽2̂ ] = 𝜎2 ⋅ 𝐶22 = 16 × 0.0014534 = 0.0232549

var(beta_hat_2)
## [1] 0.02343408
sigma ^ 2 * C[2 + 1, 2 + 1]
## [1] 0.02325487
The standard deviations found from the simulated data and the parent popu-
lation are also very close.
sd(beta_hat_2)
## [1] 0.1530819
sqrt(sigma ^ 2 * C[2 + 1, 2 + 1])
## [1] 0.1524955
Lastly, we plot a histogram of the simulated values, and overlay the true distri-
bution.
hist(beta_hat_2, prob = TRUE, breaks = 20,

curve(dnorm(x, mean = beta_2, sd = sqrt(sigma ^ 2 * C[2 + 1, 2 + 1])),
9.5. SIMULATION 183
2.5
2.0
1.5
Density
1.0
0.5
0.0
5.6 5.8 6.0 6.2 6.4

^
β 2
This looks good! The simulation-based histogram appears to be Normal

with mean 6 and spread of about 0.15 as you measure from center to in-
flection point. That matches really well with the sampling distribution of
𝛽2̂ ∼ 𝑁 (𝜇 = 6, 𝜎2 = 0.0232549).
One last check, we verify the 68 − 95 − 99.7 rule.
sd_bh2 = sqrt(sigma ^ 2 * C[2 + 1, 2 + 1])

# We expect these to be: 0.68, 0.95, 0.997
mean(beta_2 - 1 * sd_bh2 < beta_hat_2 & beta_hat_2 < beta_2 + 1 * sd_bh2)
## [1] 0.6807
## [1] 0.9529
## [1] 0.9967
9.6 R Markdown
• mlr.Rmd

Chapter 10
Model Building
“Statisticians, like artists, have the bad habit of falling in love with
their models.”
— George Box
Let’s take a step back and consider the process of finding a model for data at a
higher level. We are attempting to find a model for a response variable 𝑦 based
on a number of predictors 𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 .
Essentially, we are trying to discover the functional relationship between 𝑦 and
the predictors. In the previous chapter we were fitting models for a car’s fuel
efficiency (mpg) as a function of its attributes (wt, year, cyl, disp, hp, acc). We
also consider 𝑦 to be a function of some noise. Rarely if ever do we expect there
to be an exact functional relationship between the predictors and the response.
𝑦 = 𝑓(𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 ) + 𝜖
We can think of this as
response = signal + noise.
We could consider all sorts of complicated functions for 𝑓. You will likely en-
counter several ways of doing this in future machine learning courses. So far in
this course we have focused on (multiple) linear regression. That is
𝑦 = 𝑓(𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 ) + 𝜖
= 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + ⋯ + 𝛽𝑝−1 𝑥𝑝−1 + 𝜖
In the big picture of possible models that we could fit to this data, this is a
rather restrictive model. What do we mean by a restrictive model?
185
186 CHAPTER 10. MODEL BUILDING
10.1 Family, Form, and Fit
When modeling data, there are a number of choices that need to be made.
• What family of models will be considered?

• What form of the model will be used?
• How will the model be fit?
Let’s work backwards and discuss each of these.
10.1.1 Fit
Consider one of the simplest models we could fit to data, simple linear regression.
𝑦 = 𝑓(𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 ) + 𝜖 = 𝛽0 + 𝛽1 𝑥1 + 𝜖
So here, despite having multiple predictors, we chose to use only one. How is
this model fit? We will almost exclusively use the method of least squares, but
recall, we had seen alternative methods of fitting this model.
argmin max |𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖 )|

𝛽0 ,𝛽1
𝑛
argmin ∑ |𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖 )|
𝛽0 ,𝛽1 𝑖=1
𝑛
argmin ∑(𝑦𝑖 − (𝛽0 + 𝛽1 𝑥𝑖 ))2
𝛽0 ,𝛽1 𝑖=1
Any of these methods (we will always use the last, least squares) will obtain
estimates of the unknown parameters 𝛽0 and 𝛽1 . Since those are the only
unknowns of the specified model, we have then fit the model. The fitted model
is then
̂ ,𝑥 ,𝑥 ,…,𝑥 ) = 𝛽̂ + 𝛽̂ 𝑥
𝑦 ̂ = 𝑓(𝑥 1 2 3 𝑝−1 0 1 1
Note that, now we have dropped the term for the noise. We don’t make any
effort to model the noise, only the signal.
10.1. FAMILY, FORM, AND FIT 187
10.1.2 Form
What are the different forms a model could take? Currently, for the linear
models we have considered, the only method for altering the form of the model
is to control the predictors used. For example, one form of the multiple linear
regression model is simple linear regression.
𝑦 = 𝑓(𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 ) + 𝜖 = 𝛽0 + 𝛽1 𝑥1 + 𝜖
We could also consider a SLR model with a different predictor, thus altering
the form of the model.
𝑦 = 𝑓(𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 ) + 𝜖 = 𝛽0 + 𝛽2 𝑥2 + 𝜖
Often, we’ll use multiple predictors in our model. Very often, we will at least
try a model with all possible predictors.
𝑦 = 𝑓(𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 ) + 𝜖
= 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + ⋯ + 𝛽𝑝−1 𝑥𝑝−1 + 𝜖
We could also use some, but not all of the predictors.
𝑦 = 𝑓(𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 ) + 𝜖
= 𝛽0 + 𝛽1 𝑥1 + 𝛽3 𝑥3 + 𝛽5 𝑥5 + 𝜖
These forms are restrictive in two senses. First, they only allow for linear rela-
tionships between the response and the predictors. This seems like an obvious
restriction of linear models, but in fact, we will soon see how to use linear models
for non-linear relationships. (It will involve transforming variables.) Second,
how one variable affects the response is the same for any values of the other
predictors. Soon we will see how to create models where the effect of 𝑥1 can be
different for different values of 𝑥2 . We will discuss the concept of interaction.
10.1.3 Family
A family of models is a broader grouping of many possible forms of a model.

For example, above we saw several forms of models from the family of linear
models. We will only ever concern ourselves with linear models, which model a
response as a linear combination of predictors. There are certainly other families
of models.
For example, there are several families of non-parametric regression. Smoothing
is a broad family of models. As are regression trees.
In linear regression, we specified models with parameters, 𝛽𝑗 and fit the model
by finding the best values of these parameters. This is a parametric approach. A
non-parametric approach skips the step of specifying a model with parameters,
and is often described as more of an algorithm. Non-parametric models are
often used in machine learning.
Linear Regression Smoothing

10
10
5
5
Response
Response
0
0
-5
-5
0 2 4 6 8 10 0 2 4 6 8 10
Predictor Predictor
Here, SLR (parametric) is used on the left, while smoothing (non-parametric) is

used on the right. SLR finds the best slope and intercept. Smoothing produces
the fitted 𝑦 value at a particular 𝑥 value by considering the 𝑦 values of the data
in a neighborhood of the 𝑥 value considered. (Local smoothing.)
Why the focus on linear models? Two big reasons:
• Linear models are the go-to model. Linear models have been around for
a long time, and are computationally easy. A linear model may not be the
final model you use, but often, it should be the first model you try.
• The ideas behind linear models can be easily transferred to other modeling
techniques.
10.1.4 Assumed Model, Fitted Model
When searching for a model, we often need to make assumptions. These as-
sumptions are codified in the family and form of the model. For example
𝑦 = 𝛽0 + 𝛽1 𝑥1 + 𝛽3 𝑥3 + 𝛽5 𝑥5 + 𝜖
assumes that 𝑦 is a linear combination of 𝑥1 , 𝑥3 , and 𝑥5 as well as some noise.

This assumes that the effect of 𝑥1 on 𝑦 is 𝛽1 , which is the same for all values of
10.2. EXPLANATION VERSUS PREDICTION 189
𝑥3 and 𝑥5 . That is, we are using the family of linear models with a particular
form.
Suppose we then fit this model to some data and obtain the fitted model. For
example, in R we would use
fit = lm(y ~ x1 + x3 + x5, data = some_data)
This is R’s way of saying the family is linear and specifying the form from above.
An additive model with the specified predictors as well as an intercept. We then
obtain
𝑦 ̂ = 1.5 + 0.9𝑥1 + 1.1𝑥3 + 2.3𝑥5 .
This is our best guess for the function 𝑓 in
𝑦 = 𝑓(𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑝−1 ) + 𝜖
for the assumed family and form. Fitting a model only gives us the best fit
for the family and form that we specify. So the natural question is; how do we
choose the correct family and form? We’ll focus on form since we are focusing
on the family of linear models.
10.2 Explanation versus Prediction

What is the purpose of fitting a model to data? Usually it is to accomplish
one of two goals. We can use a model to explain the relationship between the
response and the predictors. Models can also be used to predict the response
based on the predictors. Often, a good model will do both, but we’ll discuss
both goals separately since the processes of finding models for explaining and
predicting have some differences.
For our purposes, since we are only considering linear models, searching for a
good model is essentially searching for a good form of a model.
10.2.1 Explanation
If the goal of a model is to explain the relationship between the response and
the predictors, we are looking for a model that is small and interpretable,
but still fits the data well. When discussing linear models, the size of a model
is essentially the number of 𝛽 parameters used.
Suppose we would like to find a model that explains fuel efficiency (mpg) based
on a car’s attributes (wt, year, cyl, disp, hp, acc). Perhaps we are a car
manufacturer trying to engineer a fuel-efficient vehicle. If this is the case, we

are interested in both which predictor variables are useful for explaining the
car’s fuel efficiency, as well as how those variables affect fuel efficiency. By
understanding this relationship, we can use this knowledge to our advantage
when designing a car.
To explain a relationship, we are interested in keeping models as small as pos-
sible, since smaller models are easy to interpret. The fewer predictors the fewer
considerations we need to make in our design process.
Note that linear models of any size are rather interpretable to begin with. Later
in your data analysis careers, you will see more complicated models that may fit
data better, but are much harder, if not impossible to interpret. These models
aren’t nearly as useful for explaining a relationship. This is another reason to
always attempt a linear model. If it fits as well as more complicated methods,
it will be the easiest to understand.
To find small and interpretable models, we will eventually use selection pro-
cedures, which search among many possible forms of a model. For now we
will do this in a more ad-hoc manner using inference techniques we have al-
ready encountered. To use inference as we have seen it, we need an additional
assumption in addition to the family and form of the model.
𝑦 = 𝛽0 + 𝛽1 𝑥1 + 𝛽3 𝑥3 + 𝛽5 𝑥5 + 𝜖
Our additional assumption is about the error term.
𝜖 ∼ 𝑁 (0, 𝜎2 )
This assumption that the errors are normally distributed with some common
variance is the key to all of the inference we have done so far. We will discuss
this in great detail later.
So with our inference tools (ANOVA and 𝑡-test) we have two potential strategies.
Start with a very small model (no predictors) and attempt to add predictors.
Or, start with a big model (all predictors) and attempt to remove predictors.
10.2.1.1 Correlation and Causation
A word of caution when using a model to explain a relationship. There are two
terms often used to describe a relationship between two variables: causation
and correlation. Correlation is often also referred to as association.
Just because two variables are correlated does not necessarily mean that one
causes the other. For example, consider modeling mpg as only a function of hp.
plot(mpg ~ hp, data = autompg, col = "dodgerblue", pch = 20, cex = 1.5)
40
30
mpg
20
10
50 100 150 200
hp
Does an increase in horsepower cause a drop in fuel efficiency? Or, perhaps

the causality is reversed and an increase in fuel efficiency causes a decrease in
horsepower. Or, perhaps there is a third variable that explains both!
The issue here is that we have observational data. With observational data,
we can only detect associations. To speak with confidence about causality, we
would need to run experiments. Often, this decision is made for us, before we
ever see data, so we can only modify our interpretation.
This is a concept that you should encounter often in your statistics education.
For some further reading, and some related fallacies, see: Wikipedia: Correla-
tion does not imply causation.
We’ll discuss this further when we discuss experimental design and traditional
ANOVA techniques. (All of which has recently been re-branded as A/B testing.)
10.2.2 Prediction
If the goal of a model is to predict the response, then the only consideration is
how well the model fits the data. For this, we will need a metric. In regression
problems, this is most often RMSE.
1 𝑛
RMSE(model, data) = √ ∑(𝑦 − 𝑦𝑖̂ )2
𝑛 𝑖=1 𝑖
where
• 𝑦𝑖 are the actual values of the response for the given data
• 𝑦𝑖̂ are the predicted values using the fitted model and the predictors from
the data
Correlation and causation are not an issue here. If a predictor is correlated

with the response, it is useful for prediction. For example, in elementary school
aged children their shoe size certainly doesn’t cause them to read at a higher
level, however we could very easily use shoe size to make a prediction about a
child’s reading ability. The larger their shoe size, the better they read. There’s
a lurking variable here though, their age! (Don’t send your kids to school with
size 14 shoes, it won’t make them read better!)
Also, since we are not performing inference, the extra assumption about the
errors is not needed. The only thing we care about is how close the fitted model
is to the data. Least squares is least squares. For a specified model, it will
find the values of the parameters which will minimize the squared error loss.
Your results might be largely uninterpretable and useless for inference, but for
prediction none of that matters.
Suppose instead of the manufacturer who would like to build a car, we are a
consumer who wishes to purchase a new car. However this particular car is so
new, it has not been rigorously tested, so we are unsure of what fuel efficiency to
expect. (And, as skeptics, we don’t trust what the manufacturer is telling us.)
In this case, we would like to use the model to help predict the fuel efficiency
of this car based on its attributes, which are the predictors of the model. The
smaller the errors the model makes, the more confident we are in its prediction.
10.2.2.1 Test-Train Split
The trouble with using RMSE to identify how well a model fits data is that
RMSE is always (equal or) lower for a larger model. This would suggest that
we should always use the largest model possible when looking for a model that
predicts well. The problem with this is the potential to overfit to the data. So,
we want a model that fits well, but does not overfit. To understand overfitting,
we need to think about applying a model to seen and unseen data.
Suppose we fit a model using all data available and we evaluate RMSE on this
fitted model and all of the seen data. We will call this data the training data,
and this RMSE the train RMSE.
Now, suppose we magically encounter some additional data. To truly assess

how well the model predicts, we should evaluate how well our model predicts
the response of this data. We will call this data the test data and this RMSE
the test RMSE.
• Train RMSE: model fit on seen data, evaluated on seen data

• Test RMSE: model fit on seen data, evaluated on unseen data
Below, we simulate some data and fit two models. We will call the solid blue
line the “simple” model. The dashed orange line will be called the “complex”
model, which was fit with methods we do not yet know.
Simulated Train Data Simulated Test Data

8
8
6
6
4
4
Response
Response
2
2
0
0
-2
-2
-4
-4
0 2 4 6 8 10 0 2 4 6 8 10
Predictor Predictor
The left panel shows the data that was used to fit the two models. Clearly the
“complex” model fits the data much better. The right panel shows additional
data that was simulated in the same manner as the original data. Here we see
that the “simple” model fits much better. The dashed orange line almost seems
random.
Model Train RMSE Test RMSE

Simple 1.71 1.45
Complex 1.41 2.07
The more “complex,” wiggly, model fits the training data much better as it has
a much lower train RMSE. However, we see that the “simple” model fits the test
data much better, with a much lower test RMSE. This means that the complex
model has overfit to the data, and we prefer the simple model. When choosing
a model for prediction, we prefer a model that predicts unseen data.
In practice, you can’t simply generate more data to evaluate your models. In-
stead we split existing data into data used to fit the model (train) and data
used to evaluate the model (test). Never fit a model with test data.
10.3 Summary
Models can be used to explain relationships and predict observations.
When using model to,
• explain; we prefer small and interpretable models.

• predict; we prefer models that make the smallest errors possible, without
overfitting.
Linear models can accomplish both these goals. Later, we will see that often
a linear model that accomplishes one of these goals, usually accomplishes the
other.
10.4 R Markdown
• model-building.Rmd

Chapter 11
Categorical Predictors and

Interactions
“The greatest value of a picture is when it forces us to notice what

we never expected to see.”
— John Tukey
• Include and interpret categorical variables in a linear regression model by

way of dummy variables.
• Understand the implications of using a model with a categorical variable
in two ways: levels serving as unique predictors versus levels serving as a
comparison to a baseline.
• Construct and interpret linear regression models with interaction terms.
• Identify categorical variables in a data set and convert them into factor
variables, if necessary, using R.
So far in each of our analyses, we have only used numeric variables as predictors.
We have also only used additive models, meaning the effect any predictor had
on the response was not dependent on the other predictors. In this chapter,
we will remove both of these restrictions. We will fit models with categorical
predictors, and use models that allow predictors to interact. The mathematics
of multiple regression will remain largely unchanging, however, we will pay close
attention to interpretation, as well as some difference in R usage.
195
196 CHAPTER 11. CATEGORICAL PREDICTORS AND INTERACTIONS
11.1 Dummy Variables

For this chapter, we will briefly use the built-in dataset mtcars before returning
to our autompg dataset that we created in the last chapter. The mtcars dataset
is somewhat smaller, so we’ll quickly take a look at the entire dataset.
mtcars
## mpg cyl disp hp drat wt qsec vs am gear carb

## Mazda RX4 21.0 6 160.0 110 3.90 2.620 16.46 0 1 4 4
## Mazda RX4 Wag 21.0 6 160.0 110 3.90 2.875 17.02 0 1 4 4
## Datsun 710 22.8 4 108.0 93 3.85 2.320 18.61 1 1 4 1
## Hornet 4 Drive 21.4 6 258.0 110 3.08 3.215 19.44 1 0 3 1
## Hornet Sportabout 18.7 8 360.0 175 3.15 3.440 17.02 0 0 3 2
## Valiant 18.1 6 225.0 105 2.76 3.460 20.22 1 0 3 1
## Duster 360 14.3 8 360.0 245 3.21 3.570 15.84 0 0 3 4
## Merc 240D 24.4 4 146.7 62 3.69 3.190 20.00 1 0 4 2
## Merc 230 22.8 4 140.8 95 3.92 3.150 22.90 1 0 4 2
## Merc 280 19.2 6 167.6 123 3.92 3.440 18.30 1 0 4 4
## Merc 280C 17.8 6 167.6 123 3.92 3.440 18.90 1 0 4 4
## Merc 450SE 16.4 8 275.8 180 3.07 4.070 17.40 0 0 3 3
## Merc 450SL 17.3 8 275.8 180 3.07 3.730 17.60 0 0 3 3
## Merc 450SLC 15.2 8 275.8 180 3.07 3.780 18.00 0 0 3 3
## Cadillac Fleetwood 10.4 8 472.0 205 2.93 5.250 17.98 0 0 3 4
## Lincoln Continental 10.4 8 460.0 215 3.00 5.424 17.82 0 0 3 4
## Chrysler Imperial 14.7 8 440.0 230 3.23 5.345 17.42 0 0 3 4
## Fiat 128 32.4 4 78.7 66 4.08 2.200 19.47 1 1 4 1
## Honda Civic 30.4 4 75.7 52 4.93 1.615 18.52 1 1 4 2
## Toyota Corolla 33.9 4 71.1 65 4.22 1.835 19.90 1 1 4 1
## Toyota Corona 21.5 4 120.1 97 3.70 2.465 20.01 1 0 3 1
## Dodge Challenger 15.5 8 318.0 150 2.76 3.520 16.87 0 0 3 2
## AMC Javelin 15.2 8 304.0 150 3.15 3.435 17.30 0 0 3 2
## Camaro Z28 13.3 8 350.0 245 3.73 3.840 15.41 0 0 3 4
## Pontiac Firebird 19.2 8 400.0 175 3.08 3.845 17.05 0 0 3 2
## Fiat X1-9 27.3 4 79.0 66 4.08 1.935 18.90 1 1 4 1
## Porsche 914-2 26.0 4 120.3 91 4.43 2.140 16.70 0 1 5 2
## Lotus Europa 30.4 4 95.1 113 3.77 1.513 16.90 1 1 5 2
## Ford Pantera L 15.8 8 351.0 264 4.22 3.170 14.50 0 1 5 4
## Ferrari Dino 19.7 6 145.0 175 3.62 2.770 15.50 0 1 5 6
## Maserati Bora 15.0 8 301.0 335 3.54 3.570 14.60 0 1 5 8
## Volvo 142E 21.4 4 121.0 109 4.11 2.780 18.60 1 1 4 2
We will be interested in three of the variables: mpg, hp, and am.
• mpg: fuel efficiency, in miles per gallon.

11.1. DUMMY VARIABLES 197
• hp: horsepower, in foot-pounds per second.

• am: transmission. Automatic or manual.
As we often do, we will start by plotting the data. We are interested in mpg as
the response variable, and hp as a predictor.
plot(mpg ~ hp, data = mtcars, cex = 2)

30
25
mpg
20
15
10
50 100 150 200 250 300
hp
Since we are also interested in the transmission type, we could also label the
points accordingly.
plot(mpg ~ hp, data = mtcars, col = am + 1, pch = am + 1, cex = 2)

legend("topright", c("Automatic", "Manual"), col = c(1, 2), pch = c(1, 2))
Automatic
30
25 Manual
mpg
20
15
10
50 100 150 200 250 300
hp
We used a common R “trick” when plotting this data. The am variable takes two
possible values; 0 for automatic transmission, and 1 for manual transmissions. R
can use numbers to represent colors, however the color for 0 is white. So we take
the am vector and add 1 to it. Then observations with automatic transmissions
are now represented by 1, which is black in R, and manual transmission are
represented by 2, which is red in R. (Note, we are only adding 1 inside the call
to plot(), we are not actually modifying the values stored in am.)
We now fit the SLR model
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝜖,
where 𝑌 is mpg and 𝑥1 is hp. For notational brevity, we drop the index 𝑖 for
observations.
mpg_hp_slr = lm(mpg ~ hp, data = mtcars)
We then re-plot the data and add the fitted line to the plot.

abline(mpg_hp_slr, lwd = 3, col = "grey")
Automatic
Manual
30
25
mpg
20
15
10
50 100 150 200 250 300
hp
We should notice a pattern here. The red, manual observations largely fall above
the line, while the black, automatic observations are mostly below the line. This
means our model underestimates the fuel efficiency of manual transmissions, and
overestimates the fuel efficiency of automatic transmissions. To correct for this,
we will add a predictor to our model, namely, am as 𝑥2 .
Our new model is
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝜖,
where 𝑥1 and 𝑌 remain the same, but now
1 manual transmission
𝑥2 = { .
0 automatic transmission
In this case, we call 𝑥2 a dummy variable. A dummy variable is somewhat

unfortunately named, as it is in no way “dumb.” In fact, it is actually somewhat
clever. A dummy variable is a numerical variable that is used in a regression
analysis to “code” for a binary categorical variable. Let’s see how this works.
First, note that am is already a dummy variable, since it uses the values 0 and
1 to represent automatic and manual transmissions. Often, a variable like am
would store the character values auto and man and we would either have to
convert these to 0 and 1, or, as we will see later, R will take care of creating
dummy variables for us.
So, to fit the above model, we do so like any other multiple regression model we
have seen before.
mpg_hp_add = lm(mpg ~ hp + am, data = mtcars)
Briefly checking the output, we see that R has estimated the three 𝛽 parameters.
mpg_hp_add
##
## Call:
## lm(formula = mpg ~ hp + am, data = mtcars)
##
## Coefficients:
## (Intercept) hp am
## 26.58491 -0.05889 5.27709
Since 𝑥2 can only take values 0 and 1, we can effectively write two different
models, one for manual and one for automatic transmissions.
For automatic transmissions, that is 𝑥2 = 0, we have,
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝜖.
Then for manual transmissions, that is 𝑥2 = 1, we have,
𝑌 = (𝛽0 + 𝛽2 ) + 𝛽1 𝑥1 + 𝜖.
Notice that these models share the same slope, 𝛽1 , but have different intercepts,
differing by 𝛽2 . So the change in mpg is the same for both models, but on average
mpg differs by 𝛽2 between the two transmission types.
We’ll now calculate the estimated slope and intercept of these two models so
that we can add them to a plot. Note that:
• 𝛽0̂ = coef(mpg_hp_add)[1] = 26.5849137

• 𝛽1̂ = coef(mpg_hp_add)[2] = -0.0588878
• 𝛽2̂ = coef(mpg_hp_add)[3] = 5.2770853
We can then combine these to calculate the estimated slope and intercepts.
int_auto = coef(mpg_hp_add)[1]
int_manu = coef(mpg_hp_add)[1] + coef(mpg_hp_add)[3]
slope_auto = coef(mpg_hp_add)[2]
slope_manu = coef(mpg_hp_add)[2]
Re-plotting the data, we use these slopes and intercepts to add the “two” fitted
models to the plot.

abline(int_auto, slope_auto, col = 1, lty = 1, lwd = 2) # add line for auto
abline(int_manu, slope_manu, col = 2, lty = 2, lwd = 2) # add line for manual
Automatic
Manual
30
25
mpg
20
15
10
50 100 150 200 250 300
hp
We notice right away that the points are no longer systematically incorrect. The
red, manual observations vary about the red line in no particular pattern without
underestimating the observations as before. The black, automatic points vary
about the black line, also without an obvious pattern.
They say a picture is worth a thousand words, but as a statistician, sometimes a
picture is worth an entire analysis. The above picture makes it plainly obvious
that 𝛽2 is significant, but let’s verify mathematically. Essentially we would like
to test:
𝐻0 ∶ 𝛽 2 = 0 vs 𝐻1 ∶ 𝛽2 ≠ 0.
This is nothing new. Again, the math is the same as the multiple regression
analyses we have seen before. We could perform either a 𝑡 or 𝐹 test here. The
only difference is a slight change in interpretation. We could think of this as
testing a model with a single line (𝐻0 ) against a model that allows two lines
(𝐻1 ).
To obtain the test statistic and p-value for the 𝑡-test, we would use
summary(mpg_hp_add)$coefficients["am",]

## 5.277085e+00 1.079541e+00 4.888270e+00 3.460318e-05
To do the same for the 𝐹 test, we would use
anova(mpg_hp_slr, mpg_hp_add)

##
## Model 1: mpg ~ hp
## Model 2: mpg ~ hp + am
## 1 30 447.67
## 2 29 245.44 1 202.24 23.895 3.46e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Notice that these are indeed testing the same thing, as the p-values are exactly
equal. (And the 𝐹 test statistic is the 𝑡 test statistic squared.)
Recapping some interpretations:
• 𝛽0̂ = 26.5849137 is the estimated average mpg for a car with an automatic
transmission and 0 hp.
• 𝛽0̂ + 𝛽2̂ = 31.8619991 is the estimated average mpg for a car with a manual
transmission and 0 hp.
• 𝛽2̂ = 5.2770853 is the estimated difference in average mpg for cars with
manual transmissions as compared to those with automatic transmission,
for any hp.
• 𝛽1̂ = −0.0588878 is the estimated change in average mpg for an increase

in one hp, for either transmission type.
We should take special notice of those last two. In the model,
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝜖,
11.2. INTERACTIONS 203
we see 𝛽1 is the average change in 𝑌 for an increase in 𝑥1 , no matter the value

of 𝑥2 . Also, 𝛽2 is always the difference in the average of 𝑌 for any value of 𝑥1 .
These are two restrictions we won’t always want, so we need a way to specify a
more flexible model.
Here we restricted ourselves to a single numerical predictor 𝑥1 and one dummy
variable 𝑥2 . However, the concept of a dummy variable can be used with larger
multiple regression models. We only use a single numerical predictor here for
ease of visualization since we can think of the “two lines” interpretation. But
in general, we can think of a dummy variable as creating “two models,” one for
each category of a binary categorical variable.
11.2 Interactions
To remove the “same slope” restriction, we will now discuss interaction. To
illustrate this concept, we will return to the autompg dataset we created in the
last chapter, with a few more modifications.
# read data frame from the web

quote = "\"",
comment.char = "",
# give the dataframe headers
colnames(autompg) = c("mpg", "cyl", "disp", "hp", "wt", "acc", "year", "origin", "name")
# remove missing data, which is stored as "?"
# remove the plymouth reliant, as it causes some issues
# give the dataset row names, based on the engine, year and name
# remove the variable for name
autompg = subset(autompg, select = c("mpg", "cyl", "disp", "hp", "wt", "acc", "year", "origin"))
# change horsepower from character to numeric
# create a dummy variable for foreign vs domestic cars. domestic = 1.
autompg$domestic = as.numeric(autompg$origin == 1)
# remove 3 and 5 cylinder cars (which are very rare.)
autompg = autompg[autompg$cyl != 5,]
# the following line would verify the remaining cylinder possibilities are 4, 6, 8
#unique(autompg$cyl)
# change cyl to a factor variable
autompg$cyl = as.factor(autompg$cyl)
str(autompg)

## $ mpg : num 18 15 18 16 17 15 14 14 14 15 ...
## $ cyl : Factor w/ 3 levels "4","6","8": 3 3 3 3 3 3 3 3 3 3 ...
## $ disp : num 307 350 318 304 302 429 454 440 455 390 ...
## $ hp : num 130 165 150 150 140 198 220 215 225 190 ...
## $ wt : num 3504 3693 3436 3433 3449 ...
## $ acc : num 12 11.5 11 12 10.5 10 9 8.5 10 8.5 ...
## $ year : int 70 70 70 70 70 70 70 70 70 70 ...
## $ origin : int 1 1 1 1 1 1 1 1 1 1 ...
## $ domestic: num 1 1 1 1 1 1 1 1 1 1 ...
We’ve removed cars with 3 and 5 cylinders , as well as created a new variable
domestic which indicates whether or not a car was built in the United States.
Removing the 3 and 5 cylinders is simply for ease of demonstration later in the
chapter and would not be done in practice. The new variable domestic takes
the value 1 if the car was built in the United States, and 0 otherwise, which
we will refer to as “foreign.” (We are arbitrarily using the United States as the
reference point here.) We have also made cyl and origin into factor variables,
which we will discuss later.
We’ll now be concerned with three variables: mpg, disp, and domestic. We will
use mpg as the response. We can fit a model,
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝜖,
where
• 𝑌 is mpg, the fuel efficiency in miles per gallon,

• 𝑥1 is disp, the displacement in cubic inches,
• 𝑥2 is domestic as described above, which is a dummy variable.
1 Domestic
𝑥2 = {
0 Foreign
We will fit this model, extract the slope and intercept for the “two lines,” plot
the data and add the lines.
mpg_disp_add = lm(mpg ~ disp + domestic, data = autompg)
int_for = coef(mpg_disp_add)[1]
int_dom = coef(mpg_disp_add)[1] + coef(mpg_disp_add)[3]
slope_for = coef(mpg_disp_add)[2]
slope_dom = coef(mpg_disp_add)[2]
plot(mpg ~ disp, data = autompg, col = domestic + 1, pch = domestic + 1)

abline(int_for, slope_for, col = 1, lty = 1, lwd = 2) # add line for foreign cars
abline(int_dom, slope_dom, col = 2, lty = 2, lwd = 2) # add line for domestic cars
legend("topright", c("Foreign", "Domestic"), pch = c(1, 2), col = c(1, 2))
Foreign
Domestic
40
30
mpg
20
10
100 200 300 400
disp
This is a model that allows for two parallel lines, meaning the mpg can be
different on average between foreign and domestic cars of the same engine dis-
placement, but the change in average mpg for an increase in displacement is the
same for both. We can see this model isn’t doing very well here. The red line fits
the red points fairly well, but the black line isn’t doing very well for the black
points, it should clearly have a more negative slope. Essentially, we would like
a model that allows for two different slopes.
Consider the following model,
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥1 𝑥2 + 𝜖,
where 𝑥1 , 𝑥2 , and 𝑌 are the same as before, but we have added a new inter-
action term 𝑥1 𝑥2 which multiplies 𝑥1 and 𝑥2 , so we also have an additional 𝛽
parameter 𝛽3 .
This model essentially creates two slopes and two intercepts, 𝛽2 being the dif-
ference in intercepts and 𝛽3 being the difference in slopes. To see this, we will
break down the model into the two “sub-models” for foreign and domestic cars.
For foreign cars, that is 𝑥2 = 0, we have
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝜖.
For domestic cars, that is 𝑥2 = 1, we have
𝑌 = (𝛽0 + 𝛽2 ) + (𝛽1 + 𝛽3 )𝑥1 + 𝜖.
These two models have both different slopes and intercepts.
• 𝛽0 is the average mpg for a foreign car with 0 disp.

• 𝛽1 is the change in average mpg for an increase of one disp, for foreign
cars.
• 𝛽0 + 𝛽2 is the average mpg for a domestic car with 0 disp.
• 𝛽1 + 𝛽3 is the change in average mpg for an increase of one disp, for
domestic cars.
How do we fit this model in R? There are a number of ways.

One method would be to simply create a new variable, then fit a model like any
other.
autompg$x3 = autompg$disp * autompg$domestic # THIS CODE NOT RUN!

do_not_do_this = lm(mpg ~ disp + domestic + x3, data = autompg) # THIS CODE NOT RUN!
You should only do this as a last resort. We greatly prefer not to have to modify
our data simply to fit a model. Instead, we can tell R we would like to use the
existing data with an interaction term, which it will create automatically when
we use the : operator.
mpg_disp_int = lm(mpg ~ disp + domestic + disp:domestic, data = autompg)
An alternative method, which will fit the exact same model as above would be
to use the * operator. This method automatically creates the interaction term,
as well as any “lower order terms,” which in this case are the first order terms
for disp and domestic
mpg_disp_int2 = lm(mpg ~ disp * domestic, data = autompg)
We can quickly verify that these are doing the same thing.
coef(mpg_disp_int)
## (Intercept) disp domestic disp:domestic

## 46.0548423 -0.1569239 -12.5754714 0.1025184
coef(mpg_disp_int2)

## 46.0548423 -0.1569239 -12.5754714 0.1025184
We see that both the variables, and their coefficient estimates are indeed the
same for both models.
summary(mpg_disp_int)
##
## Call:
## lm(formula = mpg ~ disp + domestic + disp:domestic, data = autompg)
##
## Residuals:
## -10.8332 -2.8956 -0.8332 2.2828 18.7749
##
## Coefficients:
## (Intercept) 46.05484 1.80582 25.504 < 2e-16 ***
## disp -0.15692 0.01668 -9.407 < 2e-16 ***
## domestic -12.57547 1.95644 -6.428 3.90e-10 ***
## disp:domestic 0.10252 0.01692 6.060 3.29e-09 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
We see that using summary() gives the usual output for a multiple regression
model. We pay close attention to the row for disp:domestic which tests,
𝐻0 ∶ 𝛽3 = 0.
In this case, testing for 𝛽3 = 0 is testing for two lines with parallel slopes
versus two lines with possibly different slopes. The disp:domestic line in the
summary() output uses a 𝑡-test to perform the test.
We could also use an ANOVA 𝐹 -test. The additive model without interaction
is our null model, and the interaction model is the alternative.
anova(mpg_disp_add, mpg_disp_int)

##
## Model 1: mpg ~ disp + domestic
## Model 2: mpg ~ disp + domestic + disp:domestic
## 1 380 7714.0
## 2 379 7032.6 1 681.36 36.719 3.294e-09 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Again we see this test has the same p-value as the 𝑡-test. Also the p-value is
extremely low, so between the two, we choose the interaction model.
int_for = coef(mpg_disp_int)[1]
int_dom = coef(mpg_disp_int)[1] + coef(mpg_disp_int)[3]
slope_for = coef(mpg_disp_int)[2]
slope_dom = coef(mpg_disp_int)[2] + coef(mpg_disp_int)[4]
Here we again calculate the slope and intercepts for the two lines for use in
plotting.
plot(mpg ~ disp, data = autompg, col = domestic + 1, pch = domestic + 1)

abline(int_for, slope_for, col = 1, lty = 1, lwd = 2) # line for foreign cars
abline(int_dom, slope_dom, col = 2, lty = 2, lwd = 2) # line for domestic cars
legend("topright", c("Foreign", "Domestic"), pch = c(1, 2), col = c(1, 2))
Foreign
Domestic
40
30
mpg
20
10
100 200 300 400
disp
We see that these lines fit the data much better, which matches the result of
our tests.
So far we have only seen interaction between a categorical variable (domestic)
and a numerical variable (disp). While this is easy to visualize, since it allows
for different slopes for two lines, it is not the only type of interaction we can use
in a model. We can also consider interactions between two numerical variables.
Consider the model,
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥1 𝑥2 + 𝜖,
where

• 𝑥2 is hp, the horsepower, in foot-pounds per second.
How does mpg change based on disp in this model? We can rearrange some
terms to see how.
𝑌 = 𝛽0 + (𝛽1 + 𝛽3 𝑥2 )𝑥1 + 𝛽2 𝑥2 + 𝜖
So, for a one unit increase in 𝑥1 (disp), the mean of 𝑌 (mpg) increases 𝛽1 +𝛽3 𝑥2 ,
which is a different value depending on the value of 𝑥2 (hp)!
Since we’re now working in three dimensions, this model can’t be easily justified
via visualizations like the previous example. Instead, we will have to rely on a
test.
mpg_disp_add_hp = lm(mpg ~ disp + hp, data = autompg)

mpg_disp_int_hp = lm(mpg ~ disp * hp, data = autompg)
summary(mpg_disp_int_hp)
##
## Call:
## lm(formula = mpg ~ disp * hp, data = autompg)
##
## Residuals:
## -10.7849 -2.3104 -0.5699 2.1453 17.9211
##
## Coefficients:
## (Intercept) 5.241e+01 1.523e+00 34.42 <2e-16 ***
## disp -1.002e-01 6.638e-03 -15.09 <2e-16 ***
## hp -2.198e-01 1.987e-02 -11.06 <2e-16 ***
## disp:hp 5.658e-04 5.165e-05 10.96 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
Using summary() we focus on the row for disp:hp which tests,
𝐻0 ∶ 𝛽3 = 0.
Again, we see a very low p-value so we reject the null (additive model) in favor
of the interaction model. Again, there is an equivalent 𝐹 -test.
anova(mpg_disp_add_hp, mpg_disp_int_hp)

##
## Model 1: mpg ~ disp + hp
## Model 2: mpg ~ disp * hp
## 1 380 7576.6
## 2 379 5754.2 1 1822.3 120.03 < 2.2e-16 ***

## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
We can take a closer look at the coefficients of our fitted interaction model.
coef(mpg_disp_int_hp)
## (Intercept) disp hp disp:hp

## 52.4081997848 -0.1001737655 -0.2198199720 0.0005658269
• 𝛽0̂ = 52.4081998 is the estimated average mpg for a car with 0 disp and 0
hp.
• 𝛽1̂ = −0.1001738 is the estimated change in average mpg for an increase
in 1 disp, for a car with 0 hp.
• 𝛽2̂ = −0.21982 is the estimated change in average mpg for an increase in
1 hp, for a car with 0 disp.
• 𝛽3̂ = 5.658269 × 10−4 is an estimate of the modification to the change
in average mpg for an increase in disp, for a car of a certain hp (or vice
versa).
That last coefficient needs further explanation. Recall the rearrangement we

made earlier
𝑌 = 𝛽0 + (𝛽1 + 𝛽3 𝑥2 )𝑥1 + 𝛽2 𝑥2 + 𝜖.
So, our estimate for 𝛽1 + 𝛽3 𝑥2 , is 𝛽1̂ + 𝛽3̂ 𝑥2 , which in this case is
−0.1001738 + 5.658269 × 10−4 𝑥2 .
This says that, for an increase of one disp we see an estimated change in
average mpg of −0.1001738 + 5.658269 × 10−4 𝑥2 . So how disp and mpg are
related, depends on the hp of the car.
So for a car with 50 hp, the estimated change in average mpg for an increase of
one disp is
−0.1001738 + 5.658269 × 10−4 ⋅ 50 = −0.0718824
And for a car with 350 hp, the estimated change in average mpg for an increase
of one disp is
−0.1001738 + 5.658269 × 10−4 ⋅ 350 = 0.0978657
Notice the sign changed!

11.3 Factor Variables

So far in this chapter, we have limited our use of categorical variables to binary
categorical variables. Specifically, we have limited ourselves to dummy variables
which take a value of 0 or 1 and represent a categorical variable numerically.
We will now discuss factor variables, which is a special way that R deals with
categorical variables. With factor variables, a human user can simply think
about the categories of a variable, and R will take care of the necessary dummy
variables without any 0/1 assignment being done by the user.
is.factor(autompg$domestic)
## [1] FALSE
Earlier when we used the domestic variable, it was not a factor variable. It was
simply a numerical variable that only took two possible values, 1 for domestic,
and 0 for foreign. Let’s create a new variable origin that stores the same
information, but in a different way.
autompg$origin[autompg$domestic == 1] = "domestic"
autompg$origin[autompg$domestic == 0] = "foreign"
head(autompg$origin)
## [1] "domestic" "domestic" "domestic" "domestic" "domestic" "domestic"
Now the origin variable stores "domestic" for domestic cars and "foreign"
for foreign cars.
is.factor(autompg$origin)
## [1] FALSE
However, this is simply a vector of character values. A vector of car models is

a character variable in R. A vector of Vehicle Identification Numbers (VINs) is
a character variable as well. But those don’t represent a short list of levels that
might influence a response variable. We will want to coerce this origin variable
to be something more: a factor variable.
autompg$origin = as.factor(autompg$origin)
Now when we check the structure of the autompg dataset, we see that origin
is a factor variable.
11.3. FACTOR VARIABLES 213
str(autompg)

## $ mpg : num 18 15 18 16 17 15 14 14 14 15 ...
## $ disp : num 307 350 318 304 302 429 454 440 455 390 ...
## $ hp : num 130 165 150 150 140 198 220 215 225 190 ...
## $ wt : num 3504 3693 3436 3433 3449 ...
## $ acc : num 12 11.5 11 12 10.5 10 9 8.5 10 8.5 ...
## $ year : int 70 70 70 70 70 70 70 70 70 70 ...
## $ origin : Factor w/ 2 levels "domestic","foreign": 1 1 1 1 1 1 1 1 1 1 ...
## $ domestic: num 1 1 1 1 1 1 1 1 1 1 ...
Factor variables have levels which are the possible values (categories) that the
variable may take, in this case foreign or domestic.
levels(autompg$origin)
## [1] "domestic" "foreign"
Recall that previously we have fit the model
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥1 𝑥2 + 𝜖,
where

• 𝑥2 is domestic, a dummy variable where 1 indicates a domestic car.
(mod_dummy = lm(mpg ~ disp * domestic, data = autompg))
##
## Call:
## lm(formula = mpg ~ disp * domestic, data = autompg)
##
## Coefficients:
## 46.0548 -0.1569 -12.5755 0.1025
So here we see that
𝛽0̂ + 𝛽2̂ = 46.0548423 + −12.5754714 = 33.4793709
is the estimated average mpg for a domestic car with 0 disp.

Now let’s try to do the same, but using our new factor variable.
(mod_factor = lm(mpg ~ disp * origin, data = autompg))
##
## Call:
## lm(formula = mpg ~ disp * origin, data = autompg)
##
## Coefficients:
## (Intercept) disp originforeign disp:originforeign
## 33.47937 -0.05441 12.57547 -0.10252
It seems that it doesn’t produce the same results. Right away we notice that
the intercept is different, as is the coefficient in front of disp. We also notice
that the remaining two coefficients are of the same magnitude as their respective
counterparts using the domestic variable, but with a different sign. Why is this
happening?
It turns out that by using a factor variable, R is automatically creating a dummy
variable for us. However, it is not the dummy variable that we had originally
used ourselves.
R is fitting the model
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥1 𝑥2 + 𝜖,
where

• 𝑥2 is a dummy variable created by R. It uses 1 to represent a foreign
car.
So now,
𝛽0̂ = 33.4793709
is the estimated average mpg for a domestic car with 0 disp, which is indeed
the same as before.
When R created 𝑥2 , the dummy variable, it used domestic cars as the reference
level, that is the default value of the factor variable. So when the dummy
variable is 0, the model represents this reference level, which is domestic. (R
makes this choice because domestic comes before foreign alphabetically.)
So the two models have different estimated coefficients, but due to the different
model representations, they are actually the same model.
11.3.1 Factors with More Than Two Levels
Let’s now consider a factor variable with more than two levels. In this dataset,
cyl is an example.
is.factor(autompg$cyl)
## [1] TRUE
levels(autompg$cyl)
## [1] "4" "6" "8"
Here the cyl variable has three possible levels: 4, 6, and 8. You may wonder,
why not simply use cyl as a numerical variable? You certainly could.
However, that would force the difference in average mpg between 4 and 6 cylin-
ders to be the same as the difference in average mpg between 6 and 8 cylinders.
That usually makes sense for a continuous variable, but not for a discrete vari-
able with so few possible values. In the case of this variable, there is no such
thing as a 7-cylinder engine or a 6.23-cylinder engine in personal vehicles. For
these reasons, we will simply consider cyl to be categorical. This is a decision
that will commonly need to be made with ordinal variables. Often, with a large
number of categories, the decision to treat them as numerical variables is appro-
priate because, otherwise, a large number of dummy variables are then needed
to represent these variables.
Let’s define three dummy variables related to the cyl factor variable.
1 4 cylinder
𝑣1 = {
0 not 4 cylinder
1 6 cylinder
𝑣2 = {
0 not 6 cylinder
1 8 cylinder
𝑣3 = {
0 not 8 cylinder
Now, let’s fit an additive model in R, using mpg as the response, and disp and
cyl as predictors. This should be a model that uses “three regression lines” to
model mpg, one for each of the possible cyl levels. They will all have the same
slope (since it is an additive model), but each will have its own intercept.
(mpg_disp_add_cyl = lm(mpg ~ disp + cyl, data = autompg))
##
## Call:
## lm(formula = mpg ~ disp + cyl, data = autompg)
##
## Coefficients:
## (Intercept) disp cyl6 cyl8
## 34.99929 -0.05217 -3.63325 -2.03603
The question is, what is the model that R has fit here? It has chosen to use the
model
𝑌 = 𝛽0 + 𝛽1 𝑥 + 𝛽2 𝑣2 + 𝛽3 𝑣3 + 𝜖,
where

• 𝑥 is disp, the displacement in cubic inches,
• 𝑣2 and 𝑣3 are the dummy variables defined above.
Why doesn’t R use 𝑣1 ? Essentially because it doesn’t need to. To create three
lines, it only needs two dummy variables since it is using a reference level, which
in this case is a 4 cylinder car. The three “sub models” are then:
• 4 Cylinder: 𝑌 = 𝛽0 + 𝛽1 𝑥 + 𝜖
• 6 Cylinder: 𝑌 = (𝛽0 + 𝛽2 ) + 𝛽1 𝑥 + 𝜖
• 8 Cylinder: 𝑌 = (𝛽0 + 𝛽3 ) + 𝛽1 𝑥 + 𝜖
Notice that they all have the same slope. However, using the two dummy
variables, we achieve the three intercepts.
• 𝛽0 is the average mpg for a 4 cylinder car with 0 disp.

• 𝛽0 + 𝛽2 is the average mpg for a 6 cylinder car with 0 disp.
• 𝛽0 + 𝛽3 is the average mpg for a 8 cylinder car with 0 disp.
So because 4 cylinder is the reference level, 𝛽0 is specific to 4 cylinders, but 𝛽2

and 𝛽3 are used to represent quantities relative to 4 cylinders.
As we have done before, we can extract these intercepts and slopes for the three
lines, and plot them accordingly.
int_4cyl = coef(mpg_disp_add_cyl)[1]
int_6cyl = coef(mpg_disp_add_cyl)[1] + coef(mpg_disp_add_cyl)[3]
int_8cyl = coef(mpg_disp_add_cyl)[1] + coef(mpg_disp_add_cyl)[4]
slope_all_cyl = coef(mpg_disp_add_cyl)[2]
plot_colors = c("Darkorange", "Darkgrey", "Dodgerblue")

plot(mpg ~ disp, data = autompg, col = plot_colors[cyl], pch = as.numeric(cyl))
abline(int_4cyl, slope_all_cyl, col = plot_colors[1], lty = 1, lwd = 2)
legend("topright", c("4 Cylinder", "6 Cylinder", "8 Cylinder"),
col = plot_colors, lty = c(1, 2, 3), pch = c(1, 2, 3))
4 Cylinder
6 Cylinder
40
8 Cylinder
30
mpg
20
10
100 200 300 400
disp
On this plot, we have
• 4 Cylinder: orange dots, solid orange line.

• 6 Cylinder: grey dots, dashed grey line.

• 8 Cylinder: blue dots, dotted blue line.
The odd result here is that we’re estimating that 8 cylinder cars have better
fuel efficiency than 6 cylinder cars at any displacement! The dotted blue line
is always above the dashed grey line. That doesn’t seem right. Maybe for very
large displacement engines that could be true, but that seems wrong for medium
to low displacement.
To attempt to fix this, we will try using an interaction model, that is, instead
of simply three intercepts and one slope, we will allow for three slopes. Again,
we’ll let R take the wheel (no pun intended), then figure out what model it has
applied.
(mpg_disp_int_cyl = lm(mpg ~ disp * cyl, data = autompg))
##
## Call:
## lm(formula = mpg ~ disp * cyl, data = autompg)
##
## Coefficients:
## (Intercept) disp cyl6 cyl8 disp:cyl6 disp:cyl8
## 43.59052 -0.13069 -13.20026 -20.85706 0.08299 0.10817
# could also use mpg ~ disp + cyl + disp:cyl
R has again chosen to use 4 cylinder cars as the reference level, but this also
now has an effect on the interaction terms. R has fit the model.
𝑌 = 𝛽0 + 𝛽1 𝑥 + 𝛽2 𝑣2 + 𝛽3 𝑣3 + 𝛾2 𝑥𝑣2 + 𝛾3 𝑥𝑣3 + 𝜖
We’re using 𝛾 like a 𝛽 parameter for simplicity, so that, for example 𝛽2 and 𝛾2
are both associated with 𝑣2 .
Now, the three “sub models” are:
• 4 Cylinder: 𝑌 = 𝛽0 + 𝛽1 𝑥 + 𝜖.
• 6 Cylinder: 𝑌 = (𝛽0 + 𝛽2 ) + (𝛽1 + 𝛾2 )𝑥 + 𝜖.
• 8 Cylinder: 𝑌 = (𝛽0 + 𝛽3 ) + (𝛽1 + 𝛾3 )𝑥 + 𝜖.
Interpreting some parameters and coefficients then:
• (𝛽0 + 𝛽2 ) is the average mpg of a 6 cylinder car with 0 disp

• (𝛽1̂ + 𝛾3̂ ) = −0.1306935+0.1081714 = −0.0225221 is the estimated change

in average mpg for an increase of one disp, for an 8 cylinder car.
So, as we have seen before 𝛽2 and 𝛽3 change the intercepts for 6 and 8 cylinder
cars relative to the reference level of 𝛽0 for 4 cylinder cars.
Now, similarly 𝛾2 and 𝛾3 change the slopes for 6 and 8 cylinder cars relative to
the reference level of 𝛽1 for 4 cylinder cars.
Once again, we extract the coefficients and plot the results.
int_4cyl = coef(mpg_disp_int_cyl)[1]
int_6cyl = coef(mpg_disp_int_cyl)[1] + coef(mpg_disp_int_cyl)[3]
int_8cyl = coef(mpg_disp_int_cyl)[1] + coef(mpg_disp_int_cyl)[4]
slope_4cyl = coef(mpg_disp_int_cyl)[2]
slope_6cyl = coef(mpg_disp_int_cyl)[2] + coef(mpg_disp_int_cyl)[5]
slope_8cyl = coef(mpg_disp_int_cyl)[2] + coef(mpg_disp_int_cyl)[6]
plot_colors = c("Darkorange", "Darkgrey", "Dodgerblue")

plot(mpg ~ disp, data = autompg, col = plot_colors[cyl], pch = as.numeric(cyl))
abline(int_4cyl, slope_4cyl, col = plot_colors[1], lty = 1, lwd = 2)
legend("topright", c("4 Cylinder", "6 Cylinder", "8 Cylinder"),
col = plot_colors, lty = c(1, 2, 3), pch = c(1, 2, 3))
4 Cylinder
6 Cylinder
40
8 Cylinder
30
mpg
20
10
100 200 300 400
disp
This looks much better! We can see that for medium displacement cars, 6
cylinder cars now perform better than 8 cylinder cars, which seems much more
reasonable than before.
To completely justify the interaction model (i.e., a unique slope for each cyl
level) compared to the additive model (single slope), we can perform an 𝐹 -test.
Notice first, that there is no 𝑡-test that will be able to do this since the difference
between the two models is not a single parameter.
We will test,
𝐻0 ∶ 𝛾 2 = 𝛾 3 = 0
which represents the parallel regression lines we saw before,
𝑌 = 𝛽0 + 𝛽1 𝑥 + 𝛽2 𝑣2 + 𝛽3 𝑣3 + 𝜖.
Again, this is a difference of two parameters, thus no 𝑡-test will be useful.
anova(mpg_disp_add_cyl, mpg_disp_int_cyl)

##
## Model 1: mpg ~ disp + cyl
## Model 2: mpg ~ disp * cyl
## 1 379 7299.5
## 2 377 6551.7 2 747.79 21.515 1.419e-09 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
As expected, we see a very low p-value, and thus reject the null. We prefer the
interaction model over the additive model.
Recapping a bit:
• Null Model: 𝑌 = 𝛽0 + 𝛽1 𝑥 + 𝛽2 𝑣2 + 𝛽3 𝑣3 + 𝜖
– Number of parameters: 𝑞 = 4
• Full Model: 𝑌 = 𝛽0 + 𝛽1 𝑥 + 𝛽2 𝑣2 + 𝛽3 𝑣3 + 𝛾2 𝑥𝑣2 + 𝛾3 𝑥𝑣3 + 𝜖
– Number of parameters: 𝑝 = 6
length(coef(mpg_disp_int_cyl)) - length(coef(mpg_disp_add_cyl))
11.4. PARAMETERIZATION 221
## [1] 2
We see there is a difference of two parameters, which is also displayed in the

resulting ANOVA table from R. Notice that the following two values also appear
on the ANOVA table.
nrow(autompg) - length(coef(mpg_disp_int_cyl))
## [1] 377
nrow(autompg) - length(coef(mpg_disp_add_cyl))
## [1] 379
11.4 Parameterization
So far we have been simply letting R decide how to create the dummy variables,
and thus R has been deciding the parameterization of the models. To illustrate
the ability to use alternative parameterizations, we will recreate the data, but
directly creating the dummy variables ourselves.
new_param_data = data.frame(
y = autompg$mpg,
x = autompg$disp,
v1 = 1 * as.numeric(autompg$cyl == 4),
v2 = 1 * as.numeric(autompg$cyl == 6),
v3 = 1 * as.numeric(autompg$cyl == 8))
head(new_param_data, 20)
## y x v1 v2 v3
## 1 18 307 0 0 1
## 2 15 350 0 0 1
## 3 18 318 0 0 1
## 4 16 304 0 0 1
## 5 17 302 0 0 1
## 6 15 429 0 0 1
## 7 14 454 0 0 1
## 8 14 440 0 0 1
## 9 14 455 0 0 1
## 10 15 390 0 0 1
## 11 15 383 0 0 1
## 12 14 340 0 0 1
## 13 15 400 0 0 1
## 14 14 455 0 0 1
## 15 24 113 1 0 0
## 16 22 198 0 1 0
## 17 18 199 0 1 0
## 18 21 200 0 1 0
## 19 27 97 1 0 0
## 20 26 97 1 0 0
Now,
• y is mpg
• x is disp, the displacement in cubic inches,
• v1, v2, and v3 are dummy variables as defined above.
First let’s try to fit an additive model using x as well as the three dummy
variables.
lm(y ~ x + v1 + v2 + v3, data = new_param_data)
##
## Call:
## lm(formula = y ~ x + v1 + v2 + v3, data = new_param_data)
##
## Coefficients:
## (Intercept) x v1 v2 v3
## 32.96326 -0.05217 2.03603 -1.59722 NA
What is happening here? Notice that R is essentially ignoring v3, but why?
Well, because R uses an intercept, it cannot also use v3. This is because
1 = 𝑣 1 + 𝑣2 + 𝑣3
which means that 1, 𝑣1 , 𝑣2 , and 𝑣3 are linearly dependent. This would make
the 𝑋 ⊤ 𝑋 matrix singular, but we need to be able to invert it to solve the
normal equations and obtain 𝛽.̂ With the intercept, v1, and v2, R can make the
necessary “three intercepts.” So, in this case v3 is the reference level.
If we remove the intercept, then we can directly obtain all “three intercepts”
without a reference level.
11.4. PARAMETERIZATION 223
lm(y ~ 0 + x + v1 + v2 + v3, data = new_param_data)
##
## Call:
## lm(formula = y ~ 0 + x + v1 + v2 + v3, data = new_param_data)
##
## Coefficients:
## x v1 v2 v3
## -0.05217 34.99929 31.36604 32.96326
Here, we are fitting the model
𝑌 = 𝜇1 𝑣1 + 𝜇2 𝑣2 + 𝜇3 𝑣3 + 𝛽𝑥 + 𝜖.
Thus we have:
• 4 Cylinder: 𝑌 = 𝜇1 + 𝛽𝑥 + 𝜖
We could also do something similar with the interaction model, and give each
line an intercept and slope, without the need for a reference level.
lm(y ~ 0 + v1 + v2 + v3 + x:v1 + x:v2 + x:v3, data = new_param_data)
##
## Call:
## lm(formula = y ~ 0 + v1 + v2 + v3 + x:v1 + x:v2 + x:v3, data = new_param_data)
##
## Coefficients:
## v1 v2 v3 v1:x v2:x v3:x
## 43.59052 30.39026 22.73346 -0.13069 -0.04770 -0.02252
𝑌 = 𝜇1 𝑣1 + 𝜇2 𝑣2 + 𝜇3 𝑣3 + 𝛽1 𝑥𝑣1 + 𝛽2 𝑥𝑣2 + 𝛽3 𝑥𝑣3 + 𝜖
• 4 Cylinder: 𝑌 = 𝜇1 + 𝛽1 𝑥 + 𝜖
Using the original data, we have (at least) three equivalent ways to specify the
interaction model with R.
lm(mpg ~ disp * cyl, data = autompg)
##
## Call:
## lm(formula = mpg ~ disp * cyl, data = autompg)
##
## Coefficients:
## (Intercept) disp cyl6 cyl8 disp:cyl6 disp:cyl8
## 43.59052 -0.13069 -13.20026 -20.85706 0.08299 0.10817
lm(mpg ~ 0 + cyl + disp : cyl, data = autompg)
##
## Call:
## lm(formula = mpg ~ 0 + cyl + disp:cyl, data = autompg)
##
## Coefficients:
## cyl4 cyl6 cyl8 cyl4:disp cyl6:disp cyl8:disp
## 43.59052 30.39026 22.73346 -0.13069 -0.04770 -0.02252
lm(mpg ~ 0 + disp + cyl + disp : cyl, data = autompg)
##
## Call:
## lm(formula = mpg ~ 0 + disp + cyl + disp:cyl, data = autompg)
##
## Coefficients:
## disp cyl4 cyl6 cyl8 disp:cyl6 disp:cyl8
## -0.13069 43.59052 30.39026 22.73346 0.08299 0.10817
They all fit the same model, importantly each using six parameters, but the
coefficients mean slightly different things in each. However, once they are in-
terpreted as slopes and intercepts for the “three lines” they will have the same
result.
Use ?all.equal to learn about the all.equal() function, and think about how
the following code verifies that the residuals of the two models are the same.
all.equal(fitted(lm(mpg ~ disp * cyl, data = autompg)),

fitted(lm(mpg ~ 0 + cyl + disp : cyl, data = autompg)))
## [1] TRUE
11.5. BUILDING LARGER MODELS 225
11.5 Building Larger Models

Now that we have seen how to incorporate categorical predictors as well as
interaction terms, we can start to build much larger, much more flexible models
which can potentially fit data better.
Let’s define a “big” model,
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥3 + 𝛽4 𝑥1 𝑥2 + 𝛽5 𝑥1 𝑥3 + 𝛽6 𝑥2 𝑥3 + 𝛽7 𝑥1 𝑥2 𝑥3 + 𝜖.
Here,
• 𝑌 is mpg.
• 𝑥1 is disp.
• 𝑥2 is hp.
• 𝑥3 is domestic, which is a dummy variable we defined, where 1 is a do-
mestic vehicle.
First thing to note here, we have included a new term 𝑥1 𝑥2 𝑥3 which is a three-
way interaction. Interaction terms can be larger and larger, up to the number
of predictors in the model.
Since we are using the three-way interaction term, we also use all possible two-
way interactions, as well as each of the first order (main effect) terms. This is
the concept of a hierarchy. Any time a “higher-order” term is in a model, the
related “lower-order” terms should also be included. Mathematically their inclu-
sion or exclusion is sometimes irrelevant, but from an interpretation standpoint,
it is best to follow the hierarchy rules.
Let’s do some rearrangement to obtain a “coefficient” in front of 𝑥1 .
𝑌 = 𝛽0 + 𝛽2 𝑥2 + 𝛽3 𝑥3 + 𝛽6 𝑥2 𝑥3 + (𝛽1 + 𝛽4 𝑥2 + 𝛽5 𝑥3 + 𝛽7 𝑥2 𝑥3 )𝑥1 + 𝜖.
Specifically, the “coefficient” in front of 𝑥1 is
(𝛽1 + 𝛽4 𝑥2 + 𝛽5 𝑥3 + 𝛽7 𝑥2 𝑥3 ).
Let’s discuss this “coefficient” to help us understand the idea of the flexibility
of a model. Recall that,
• 𝛽1 is the coefficient for a first order term,

• 𝛽4 and 𝛽5 are coefficients for two-way interactions,
• 𝛽7 is the coefficient for the three-way interaction.
If the two and three way interactions were not in the model, the whole “coeffi-
cient” would simply be
𝛽1 .
Thus, no matter the values of 𝑥2 and 𝑥3 , 𝛽1 would determine the relationship

between 𝑥1 (disp) and 𝑌 (mpg).
With the addition of the two-way interactions, now the “coefficient” would be
(𝛽1 + 𝛽4 𝑥2 + 𝛽5 𝑥3 ).
Now, changing 𝑥1 (disp) has a different effect on 𝑌 (mpg), depending on the

values of 𝑥2 and 𝑥3 .
Lastly, adding the three-way interaction gives the whole “coefficient”
(𝛽1 + 𝛽4 𝑥2 + 𝛽5 𝑥3 + 𝛽7 𝑥2 𝑥3 )
which is even more flexible. Now changing 𝑥1 (disp) has a different effect on 𝑌
(mpg), depending on the values of 𝑥2 and 𝑥3 , but in a more flexible way which
we can see with some more rearrangement. Now the “coefficient” in front of 𝑥3
in this “coefficient” is dependent on 𝑥2 .
(𝛽1 + 𝛽4 𝑥2 + (𝛽5 + 𝛽7 𝑥2 )𝑥3 )
It is so flexible, it is becoming hard to interpret!

Let’s fit this three-way interaction model in R.
big_model = lm(mpg ~ disp * hp * domestic, data = autompg)

summary(big_model)
##
## Call:
## lm(formula = mpg ~ disp * hp * domestic, data = autompg)
##
## Residuals:
## -11.9410 -2.2147 -0.4008 1.9430 18.4094
##
## Coefficients:
## (Intercept) 6.065e+01 6.600e+00 9.189 < 2e-16 ***
## disp -1.416e-01 6.344e-02 -2.232 0.0262 *
11.5. BUILDING LARGER MODELS 227
## hp -3.545e-01 8.123e-02 -4.364 1.65e-05 ***

## domestic -1.257e+01 7.064e+00 -1.780 0.0759 .
## disp:hp 1.369e-03 6.727e-04 2.035 0.0426 *
## disp:domestic 4.933e-02 6.400e-02 0.771 0.4414
## hp:domestic 1.852e-01 8.709e-02 2.126 0.0342 *
## disp:hp:domestic -9.163e-04 6.768e-04 -1.354 0.1766
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
Do we actually need this large of a model? Let’s first test for the necessity of
the three-way interaction term. That is,
𝐻0 ∶ 𝛽7 = 0.
So,
• Full Model: 𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥3 + 𝛽4 𝑥1 𝑥2 + 𝛽5 𝑥1 𝑥3 + 𝛽6 𝑥2 𝑥3 +
𝛽7 𝑥1 𝑥2 𝑥3 + 𝜖
• Null Model: 𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥3 + 𝛽4 𝑥1 𝑥2 + 𝛽5 𝑥1 𝑥3 + 𝛽6 𝑥2 𝑥3 + 𝜖
We fit the null model in R as two_way_int_mod, then use anova() to perform

an 𝐹 -test as usual.
two_way_int_mod = lm(mpg ~ disp * hp + disp * domestic + hp * domestic, data = autompg)

#two_way_int_mod = lm(mpg ~ (disp + hp + domestic) ^ 2, data = autompg)
anova(two_way_int_mod, big_model)

##
## Model 1: mpg ~ disp * hp + disp * domestic + hp * domestic
## Model 2: mpg ~ disp * hp * domestic
## 1 376 5673.2
## 2 375 5645.6 1 27.599 1.8332 0.1766
We see the p-value is somewhat large, so we would fail to reject. We prefer the
smaller, less flexible, null model, without the three-way interaction.
A quick note here: the full model does still “fit better.” Notice that it has a
smaller RMSE than the null model, which means the full model makes smaller
(squared) errors on average.
mean(resid(big_model) ^ 2)
## [1] 14.74053
mean(resid(two_way_int_mod) ^ 2)
## [1] 14.81259
However, it is not much smaller. We could even say that the difference is
insignificant. This is an idea we will return to later in greater detail.
Now that we have chosen the model without the three-way interaction, can we
go further? Do we need the two-way interactions? Let’s test
𝐻0 ∶ 𝛽4 = 𝛽5 = 𝛽6 = 0.
Remember we already chose 𝛽7 = 0, so,
• Full Model: 𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥3 + 𝛽4 𝑥1 𝑥2 + 𝛽5 𝑥1 𝑥3 + 𝛽6 𝑥2 𝑥3 + 𝜖
• Null Model: 𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥3 + 𝜖
We fit the null model in R as additive_mod, then use anova() to perform an

𝐹 -test as usual.
additive_mod = lm(mpg ~ disp + hp + domestic, data = autompg)

anova(additive_mod, two_way_int_mod)

##
## Model 1: mpg ~ disp + hp + domestic
## Model 2: mpg ~ disp * hp + disp * domestic + hp * domestic
## 1 379 7369.7
## 2 376 5673.2 3 1696.5 37.478 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Here the p-value is small, so we reject the null, and we prefer the full (alternative)
model. Of the models we have considered, our final preference is for
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽3 𝑥3 + 𝛽4 𝑥1 𝑥2 + 𝛽5 𝑥1 𝑥3 + 𝛽6 𝑥2 𝑥3 + 𝜖.
11.6 R Markdown
• cat-int.Rmd

Chapter 12
Analysis of Variance
Chapter Status: This chapter should be considered optional for a first reading
of this text. Its inclusion is mostly for the benefit of some courses that use the
text. Additionally, this chapter is currently somewhat underdeveloped compared
to the rest of the text. If you are interested in contributing, you can find several
lines marked “TODO” in the source. Pull requests encouraged!
“To find out what happens when you change something, it is neces-
sary to change it.”
— Box, Hunter, and Hunter, Statistics for Experimenters (1978)
Thus far, we have built models for numeric responses, when the predictors are all
numeric. We’ll take a minor detour to go back and consider models which only
have categorical predictors. A categorical predictor is a variable which takes
only a finite number of values, which are not ordered. For example a variable
which takes possible values red, blue, green is categorical. In the context
of using a categorical variable as a predictor, it would place observations into
different groups (categories).
We’ve also mostly been dealing with observational data. The methods in this
section are most useful in experimental settings, but still work with observational
data. (However, for determining causation, we require experiments.)
12.1 Experiments
The biggest difference between an observational study and an experiment is how

the predictor data is obtained. Is the experimenter in control?
231
232 CHAPTER 12. ANALYSIS OF VARIANCE
• In an observational study, both response and predictor data are obtained

via observation.
• In an experiment, the predictor data are values determined by the ex-
perimenter. The experiment is run and the response is observed.
In an experiment, the predictors, which are controlled by the experimenter, are

called factors. The possible values of these factors are called levels. Subjects
are randomly assigned to a level of each of the factors.
The design of experiments could be a course by itself. The Wikipedia article on
design of experiments gives a good overview. Originally, most of the methodol-
ogy was developed for agricultural applications by R. A. Fisher, but is still in
use today, now in a wide variety of application areas. Notably, these methods
have seen a resurgence as a part of “A/B Testing.”
12.2 Two-Sample t-Test

The simplest example of an experimental design is the setup for a two-sample
𝑡-test. There is a single factor variable with two levels which split the subjects
into two groups. Often, one level is considered the control, while the other is
the treatment. The subjects are randomly assigned to one of the two groups.
After being assigned to a group, each subject has some quantity measured,
which is the response variable.
Mathematically, we consider the model
𝑦𝑖𝑗 ∼ 𝑁 (𝜇𝑖 , 𝜎2 )
where 𝑖 = 1, 2 for the two groups and 𝑗 = 1, 2, … 𝑛𝑖 . Here 𝑛𝑖 is the number of

subjects in group 𝑖. So 𝑦13 would be the measurement for the third member of
the first group.
So measurements of subjects in group 1 follow a normal distribution with mean
𝜇1 .
𝑦1𝑗 ∼ 𝑁 (𝜇1 , 𝜎2 )
Then measurements of subjects in group 2 follow a normal distribution with

mean 𝜇2 .
𝑦2𝑗 ∼ 𝑁 (𝜇2 , 𝜎2 )
This model makes a number of assumptions. Specifically,

12.2. TWO-SAMPLE T-TEST 233
• The observations follow a normal distribution. The mean of each group

is different.
• Equal variance for each group.
• Independence. Which is believable if groups were randomly assigned.
Later, we will investigate the normal and equal variance assumptions. For now,
we will continue to assume they are reasonable.
The natural question to ask: Is there a difference between the two groups? The
specific question we’ll answer: Are the means of the two groups different?
Mathematically, that is
𝐻0 ∶ 𝜇 1 = 𝜇 2 vs 𝐻1 ∶ 𝜇1 ≠ 𝜇2
For the stated model and assuming the null hypothesis is true, the 𝑡 test statistic
would follow a 𝑡 distribution with degrees of freedom 𝑛1 + 𝑛2 − 2.
As an example, suppose we are interested in the effect of melatonin on sleep
duration. A researcher obtains a random sample of 20 adult males. Of these
subjects, 10 are randomly chosen for the control group, which will receive a
placebo. The remaining 10 will be given 5mg of melatonin before bed. The sleep
duration in hours of each subject is then measured. The researcher chooses a
significance level of 𝛼 = 0.10. Was sleep duration affected by the melatonin?
melatonin
## sleep group
## 1 8.145150 control
## 2 7.522362 treatment
## 3 6.935754 control
## 4 8.959435 treatment
## 5 6.985122 control
## 6 8.072651 treatment
## 7 8.313826 control
## 8 8.086409 treatment
## 9 8.922108 control
## 10 8.124743 treatment
## 11 8.065844 control
## 12 10.943974 treatment
## 13 4.833367 control
## 14 7.865453 treatment
## 15 6.340014 control
## 16 8.963140 treatment
## 17 6.158896 control
## 18 5.012253 treatment
## 19 3.571440 control
## 20 9.784136 treatment
Here, we would like to test,
𝐻0 ∶ 𝜇 𝐶 = 𝜇 𝑇 vs 𝐻1 ∶ 𝜇𝐶 ≠ 𝜇𝑇
To do so in R, we use the t.test() function, with the var.equal argument set

to TRUE.
t.test(sleep ~ group, data = melatonin, var.equal = TRUE)
##
##
## data: sleep by group
## t = -2.0854, df = 18, p-value = 0.05154
## alternative hypothesis: true difference in means between group control and group tre
## -3.02378261 0.01117547
## mean in group control mean in group treatment
## 6.827152 8.333456
At a significance level of 𝛼 = 0.10, we reject the null hypothesis. It seems that

the melatonin had a statistically significant effect. Be aware that statistical
significance is not always the same as scientific or practical significance. To
determine practical significance, we need to investigate the effect size in the
context of the situation. Here the effect size is the difference of the sample
means.
t.test(sleep ~ group, data = melatonin, var.equal = TRUE)$estimate
## mean in group control mean in group treatment

## 6.827152 8.333456
Here we see that the subjects in the melatonin group sleep an average of about
1.5 hours longer than the control group. An hour and a half of sleep is certainly
important!
With a big enough sample size, we could make an effect size of say, four minutes
statistically significant. Is it worth taking a pill every night to get an extra four
minutes of sleep? (Probably not.)
12.3. ONE-WAY ANOVA 235
boxplot(sleep ~ group, data = melatonin, col = 5:6)

10
8
sleep
6
4
control treatment
group
12.3 One-Way ANOVA

What if there are more than two groups? Consider the model
𝑦𝑖𝑗 = 𝜇 + 𝛼𝑖 + 𝑒𝑖𝑗
where
𝑔
∑ 𝛼𝑖 = 0
𝑖=1
and
𝑒𝑖𝑗 ∼ 𝑁 (0, 𝜎2 ).
Here,
• 𝑖 = 1, 2, … 𝑔 where 𝑔 is the number of groups.

• 𝑗 = 1, 2, … 𝑛𝑖 where 𝑛𝑖 is the number of observations in group 𝑖.
Then the total sample size is
𝑔
𝑁 = ∑ 𝑛𝑖
𝑖=1
Observations from group 𝑖 follow a normal distribution
𝑦𝑖𝑗 ∼ 𝑁 (𝜇𝑖 , 𝜎2 )
where the mean of each group is given by
𝜇𝑖 = 𝜇 + 𝛼𝑖 .
Here 𝛼𝑖 measures the effect of group 𝑖. It is the difference between the overall
mean and the mean of group 𝑖.
Essentially, the assumptions here are the same as the two-sample case, however
now, we simply have more groups.
Much like the two-sample case, we would again like to test if the means of the
groups are equal.
𝐻0 ∶ 𝜇 1 = 𝜇 2 = … 𝜇 𝑔 vs 𝐻1 ∶ Not all 𝜇𝑖 are equal.
Notice that the alternative simply indicates that some of the means are not
equal, not specifically which are not equal. More on that later.
Alternatively, we could write
𝐻0 ∶ 𝛼 1 = 𝛼 2 = … = 𝛼 𝑔 = 0 vs 𝐻1 ∶ Not all 𝛼𝑖 are 0.
This test is called Analysis of Variance (ANOVA). ANOVA compares the

variation due to specific sources (between groups) with the variation among
individuals who should be similar (within groups). In particular, ANOVA tests
whether several populations have the same mean by comparing how far apart
the sample means are with how much variation there is within the samples. We
use variability of means to test for equality of means, thus the use of variance
in the name for a test about means.
We’ll leave out most of the details about how the estimation is done, but we’ll
see later, that it is done via least squares. We’ll use R to obtain these estimates,
but they are actually rather simple. We only need to think about the sample
means of the groups.
• 𝑦𝑖̄ is the sample mean of group 𝑖.

• 𝑦 ̄ is the overall sample mean.
• 𝑠2𝑖 is the sample variance of group 𝑖.
We’ll then decompose the variance, as we’ve seen before in regression. The total
variation measures how much the observations vary about the overall sample
mean, ignoring the groups.
𝑔 𝑛𝑖
𝑆𝑆𝑇 = ∑ ∑(𝑦𝑖𝑗 − 𝑦)̄ 2
𝑖=1 𝑗=1
The variation between groups looks at how far the individual sample means
are from the overall sample mean.
𝑔 𝑛𝑖 𝑔
𝑆𝑆𝐵 = ∑ ∑(𝑦𝑖̄ − 𝑦)̄ = ∑ 𝑛𝑖 (𝑦𝑖̄ − 𝑦)̄ 2
2
𝑖=1 𝑗=1 𝑖=1
Lastly, the within group variation measures how far observations are from the
sample mean of its group.
𝑔 𝑛𝑖 𝑔
𝑆𝑆𝑊 = ∑ ∑(𝑦𝑖𝑗 − 𝑦𝑖̄ ) = ∑(𝑛𝑖 − 1)𝑠2𝑖
2
𝑖=1 𝑗=1 𝑖=1
This could also be thought of as the error sum of squares, where 𝑦𝑖𝑗 is an
observation and 𝑦𝑖̄ is its fitted (predicted) value from the model.
To develop the test statistic for ANOVA, we place this information into an
ANVOA table.
Source Sum of Squares Degrees of Freedom Mean Square 𝐹

Between SSB 𝑔−1 SSB / DFB MSB / MSW
Within SSW 𝑁 −𝑔 SSW / DFW
Total SST 𝑁 −1
We reject the null (equal means) when the 𝐹 statistic is large. This occurs when
the variation between groups is large compared to the variation within groups.
Under the null hypothesis, the distribution of the test statistic is 𝐹 with degrees
of freedom 𝑔 − 1 and 𝑁 − 𝑔.
Let’s see what this looks like in a few situations. In each of the following exam-
ples, we’ll consider sampling 20 observations (𝑛𝑖 = 20) from three populations
(groups; 𝑔 = 3).
First, consider 𝜇𝐴 = −5, 𝜇𝐵 = 0, 𝜇𝐶 = 5 with 𝜎 = 1.
Truth Observed Data
0.4
5
0.3
response
density
0.2
0
0.1
-5
0.0
-5 0 5 A B C
observations group
The left panel shows the three normal distributions we are sampling from. The
ticks along the 𝑥-axis show the randomly sampled observations. The right panel
re-displays only the sampled values in a boxplot. Note that the mid-line of the
boxes is usually the sample median. These boxplots have been modified to use
the sample mean.
Here the sample means vary a lot around the overall sample mean, which is the
solid grey line on the right panel. Within the groups there is variability, but it
is still obvious that the sample means are very different.
As a result, we obtain a large test statistic, thus small p-value.
• 𝐹 = 374.4469511
• p-value = 1.6349862 × 10−33
Now consider 𝜇𝐴 = 0, 𝜇𝐵 = 0, 𝜇𝐶 = 0 with 𝜎 = 1. That is, equal means for the

groups.
Truth Observed Data

0.4
3
0.3
2
response
1
density
0.2
0
0.1
-1
-2
0.0
-3 -2 -1 0 1 2 3 A B C
observations group
Here the sample means vary only a tiny bit around the overall sample mean.
Within the groups there is variability, this time much larger than the variability
of the sample means.
As a result, we obtain a small test statistic, thus large p-value.
• 𝐹 = 2.667892
• p-value = 0.0780579
The next two examples show different means, with different levels of noise.
Notice how these affect the test statistic and p-value.
• 𝜇𝐴 = −1, 𝜇𝐵 = 0, 𝜇𝐶 = 1, 𝜎 = 1
Truth Observed Data

0.4
3
2
0.3
1
response
density
0.2
0
-1
0.1
-2
-3
0.0
-4 -2 0 2 4 A B C
observations group
• 𝐹 = 16.4879492
• p-value = 2.2378806 × 10−6
Above, there isn’t obvious separation between the groups like the first example,
but it is still obvious the means are different. Below, there is more noise. Visu-
ally it is somewhat hard to tell, but the test still suggests a difference of means.
(At an 𝛼 of 0.05.)
• 𝜇𝐴 = −1, 𝜇𝐵 = 0, 𝜇𝐶 = 1, 𝜎 = 2
Truth Observed Data
6
0.4
4
0.3
2
response
density
0.2
0
-2
0.1
-4
0.0
-6
-6 -4 -2 0 2 4 6 A B C
observations group
• 𝐹 = 4.6256472
• p-value = 0.0137529
Let’s consider an example with real data. We’ll use the coagulation dataset
from the faraway package. Here four different diets (A, B, C, D) were adminis-
tered to a random sample of 24 animals. The subjects were randomly assigned
to one of the four diets. For each, their blood coagulation time was measured
in seconds.
Here we would like to test
𝐻0 ∶ 𝜇 𝐴 = 𝜇 𝐵 = 𝜇 𝐶 = 𝜇 𝐷
where, for example, 𝜇𝐴 is the mean blood coagulation time for an animal that
ate diet A.
library(faraway)
names(coagulation)
## [1] "coag" "diet"
plot(coag ~ diet, data = coagulation, col = 2:5)

70
65
coag
60
A B C D
diet
We first load the data and create the relevant boxplot. The plot alone suggests
a difference of means. The aov() function is used to obtain the relevant sums
of squares. Using the summary() function on the output from aov() creates the
desired ANOVA table (without the unneeded row for total).
coag_aov = aov(coag ~ diet, data = coagulation)

coag_aov
## Call:
## aov(formula = coag ~ diet, data = coagulation)
##
## Terms:
## diet Residuals
## Sum of Squares 228 112
## Deg. of Freedom 3 20
##
## Residual standard error: 2.366432
## Estimated effects may be unbalanced
summary(coag_aov)

## diet 3 228 76.0 13.57 4.66e-05 ***
## Residuals 20 112 5.6
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Were we to run this experiment, we would have pre-specified a significance

level. However, notice that the p-value of this test is incredibly low, so using
any reasonable significance level we would reject the null hypothesis. Thus we
believe the diets had an effect on blood coagulation time.
diets = data.frame(diet = unique(coagulation$diet))

data.frame(diets, coag = predict(coag_aov, diets))
## diet coag
## 1 A 61
## 2 B 66
## 3 C 68
## 4 D 61
Here, we’ve created a dataframe with a row for each diet. By predicting on this
dataframe, we obtain the sample means of each diet (group).
12.3.1 Factor Variables
When performing ANOVA in R, be sure the grouping variable is a factor variable.

If it is not, your result might not be ANOVA, but instead a linear regression
with the predictor variable considered numeric.
set.seed(42)
response = rnorm(15)
group = c(rep(1, 5), rep(2, 5), rep(3, 5))
bad = data.frame(response, group)
summary(aov(response ~ group, data = bad)) # wrong DF!

## group 1 0.017 0.0173 0.015 0.903
## Residuals 13 14.698 1.1306
good = data.frame(response, group = as.factor(group))

summary(aov(response ~ group, data = good))

## group 2 0.232 0.1158 0.096 0.909
## Residuals 12 14.484 1.2070
is.factor(bad$group) # 1, 2, and 3 are numbers.
## [1] FALSE
is.factor(good$group) # 1, 2, and 3 are labels.
## [1] TRUE
12.3.2 Some Simulation
Here we verify the distribution of the test statistic under the null hypothesis. We
simulate from a null model (equal variance) to obtain an empirical distribution
of the 𝐹 statistic. We add the curve for the expected distribution.
library(broom)
sim_anova = function(n = 10, mu_a = 0, mu_b = 0, mu_c = 0, mu_d = 0, sigma = 1, stat = TRUE) {
# create data from one-way ANOVA model with four groups of equal size
# response simulated from normal with group mean, shared variance
# group variable indicates group A, B, C or D
sim_data = data.frame(
response = c(rnorm(n = n, mean = mu_a, sd = sigma),
rnorm(n = n, mean = mu_b, sd = sigma),
rnorm(n = n, mean = mu_c, sd = sigma),
rnorm(n = n, mean = mu_d, sd = sigma)),
group = c(rep("A", times = n), rep("B", times = n),
rep("C", times = n), rep("D", times = n))
)
# obtain F-statistic and p-value for testing difference of means

# use lm instead of aov for better result formatting with glance
aov_results = lm(response ~ group, data = sim_data)
f_stat = glance(aov_results)$statistic
p_val = glance(aov_results)$p.value
# return f_stat if stat = TRUE, otherwise, p-value

ifelse(stat, f_stat, p_val)
f_stats = replicate(n = 5000, sim_anova(stat = TRUE))
hist(f_stats, breaks = 100, prob = TRUE, border = "dodgerblue", main = "Empirical Distr
curve(df(x, df1 = 4 - 1, df2 = 40 - 4), col = "darkorange", add = TRUE, lwd = 2)
Empirical Distribution of F
0.6
Density
0.4
0.2
0.0
0 2 4 6 8 10
f_stats
12.3.3 Power
Now that we’re performing experiments, getting more data means finding more
test subjects, running more lab tests, etc. In other words, it will cost more time
and money.
We’d like to design our experiment so that we have a good chance of detecting
an interesting effect size, without spending too much money. There’s no point in
running an experiment if there’s only a very low chance that it has a significant
result that you care about. (Remember, not all statistically significant results
have practical value.)
We’d like the ANOVA test to have high power for an alternative hypothesis
with a minimum desired effect size.
Power = 𝑃 (Reject 𝐻0 ∣ 𝐻0 False)
That is, for a true difference of means that we deem interesting, we want the
test to reject with high probability.
A number of things can affect the power of a test:
• Effect size. It is easier to detect larger effects.

• Noise level 𝜎. The less noise, the easier it is to detect signal (effect). We
don’t have much ability to control this, except maybe to measure more
accurately.
• Significance level 𝛼. Lower significance level makes rejecting more dif-
ficult. (But also allows for fewer false positives.)
• Sample size. Large samples mean easier to detect effects.
• Balanced design. An equal number of observations per group leads to
higher power.
The following simulations look at the effect of significance level, effect size, and
noise level on the power of an ANOVA 𝐹 -test. Homework will look into sample
size and balance.
p_vals = replicate(n = 1000, sim_anova(mu_a = -1, mu_b = 0, mu_c = 0, mu_d = 1,

sigma = 1.5, stat = FALSE))
mean(p_vals < 0.05)
## [1] 0.661
mean(p_vals < 0.01)
## [1] 0.384

mean(p_vals < 0.05)
## [1] 0.425
mean(p_vals < 0.01)
## [1] 0.19

mean(p_vals < 0.05)
## [1] 0.956
mean(p_vals < 0.01)
## [1] 0.837
12.4 Post Hoc Testing

Suppose we reject the null hypothesis from the ANOVA test for equal means.
That tells us that the means are different. But which means? All of them?
Some of them? The obvious strategy is to test all possible comparisons of two
means. We can do this easily in R.
with(coagulation, pairwise.t.test(coag, diet, p.adj = "none"))
##
## Pairwise comparisons using t tests with pooled SD
##
## data: coag and diet
##
## A B C
## B 0.00380 - -
## C 0.00018 0.15878 -
## D 1.00000 0.00086 2.3e-05
##
## P value adjustment method: none
# pairwise.t.test(coagulation$coag, coagulation$diet, p.adj = "none")
Notice the pairwise.t.test() function does not have a data argument. To

avoid using attach() or the $ operator, we introduce the with() function. The
commented line would perform the same operation.
Also note that we are using the argument p.adj = "none". What is this? An
adjustment (in this case not an adjustment) to the p-value of each test. Why
would we need to do this?
The adjustment is an attempt to correct for the multiple testing problem. (See
also: Relevant XKCD. ) Imagine that you knew ahead of time that you were
12.4. POST HOC TESTING 247
going to perform 100 𝑡-tests. Suppose you wish to do this with a false positive
rate of 𝛼 = 0.05. If we use this significance level for each test, for 100 tests,
we then expect 5 false positives. That means, with 100 tests, we’re almost
guaranteed to have at least one error.
What we’d really like is for the family-wise error rate (FWER) to be 0.05. If
we consider the 100 tests to be a single “experiment,” the FWER is the rate of
one or more false positives for the full experiment (100 tests). Consider it an
error rate for an entire procedure, instead of a single test.
With this in mind, one of the simplest adjustments we can make, is to increase
the p-values for each test, depending on the number of tests. In particular the
Bonferroni correction simply multiplies by the number of tests.
p-value-bonf = min(1, 𝑛𝑡𝑒𝑠𝑡𝑠 ⋅ p-value)
with(coagulation, pairwise.t.test(coag, diet, p.adj = "bonferroni"))
##
## Pairwise comparisons using t tests with pooled SD
##
## data: coag and diet
##
## A B C
## B 0.02282 - -
## C 0.00108 0.95266 -
## D 1.00000 0.00518 0.00014
##
## P value adjustment method: bonferroni
We see that these p-values are much higher than the unadjusted p-values, thus,
we are less likely to reject each test. As a result, the FWER is 0.05, instead of
an error rate of 0.05 for each test.
We can simulate the 100 test scenarios to illustrate this point.
get_p_val = function() {
# create data for two groups, equal mean

y = rnorm(20, mean = 0, sd = 1)
g = c(rep("A", 10), rep("B", 10))
# p-value of t-test when null is true

glance(t.test(y ~ g, var.equal = TRUE))$p.value
}
set.seed(1337)
# FWER with 100 tests

# desired rate = 0.05
# no adjustment
mean(replicate(1000, any(replicate(100, get_p_val()) < 0.05)))
## [1] 0.994
# FWER with 100 tests

# desired rate = 0.05
# bonferroni adjustment
mean(replicate(1000, any(p.adjust(replicate(100, get_p_val()), "bonferroni") < 0.05)))
## [1] 0.058
For the specific case of testing all two-way mean differences after an ANOVA
test, there are a number of potential methods for making an adjustment of this
type. The pros and cons of the potential methods are beyond the scope of
this course. We choose a method for its ease of use, and to a lesser extent, its
developer.
Tukey’s Honest Significance difference can be applied directly to an object which
was created using aov(). It will adjust the p-values of the pairwise comparisons
of the means to control the FWER, in this case, for 0.05. Notice it also gives
confidence intervals for the difference of the means.
TukeyHSD(coag_aov, conf.level = 0.95)
## Tukey multiple comparisons of means

## 95% family-wise confidence level
##
## Fit: aov(formula = coag ~ diet, data = coagulation)
##
## $diet
## diff lwr upr p adj
## B-A 5 0.7245544 9.275446 0.0183283
## C-A 7 2.7245544 11.275446 0.0009577
## D-A 0 -4.0560438 4.056044 1.0000000
## C-B 2 -1.8240748 5.824075 0.4766005
## D-B -5 -8.5770944 -1.422906 0.0044114
## D-C -7 -10.5770944 -3.422906 0.0001268
12.5. TWO-WAY ANOVA 249
Based on these results, we see no difference between A and D as well as B and

C. All other pairwise comparisons are significant. If you return to the original
boxplot, these results should not be surprising.
Also, nicely, we can easily produce a plot of these confidence intervals.
plot(TukeyHSD(coag_aov, conf.level = 0.95))
95% family-wise confidence level

B-A
C-A
D-A
C-B
D-B
D-C
-10 -5 0 5 10
Differences in mean levels of diet
The creator of this method, John Tukey, is an important figure in the history of
data science. He essentially predicted the rise of data science over 50 years ago.
For some retrospective thoughts on those 50 years, see this paper from David
Donoho.
12.5 Two-Way ANOVA

What if there is more than one factor variable? Why do we need to limit
ourselves to experiments with only one factor? We don’t! Consider the model
𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + (𝛼𝛽)𝑖𝑗 + 𝜖𝑖𝑗𝑘
where 𝜖𝑖𝑗𝑘 are 𝑁 (0, 𝜎2 ) random variables.

We add constraints
𝐼 𝐽
∑ 𝛼𝑖 = 0 ∑ 𝛽𝑗 = 0.
𝑖=1 𝑗=1
and
(𝛼𝛽)1𝑗 + (𝛼𝛽)2𝑗 + (𝛼𝛽)3𝑗 = 0(𝛼𝛽)𝑖1 + (𝛼𝛽)𝑖2 + (𝛼𝛽)𝑖3 + (𝛼𝛽)𝑖4 = 0
for any 𝑖 or 𝑗.
Here,
• 𝑖 = 1, 2, … 𝐼 where 𝐼 is the number of levels of factor 𝐴.

• 𝑗 = 1, 2, … 𝐽 where 𝐽 is the number of levels of factor 𝐵.
• 𝑘 = 1, 2, … 𝐾 where 𝐾 is the number of replicates per group.
Here, we can think of a group as a combination of a level from each of the

factors. So for example, one group will receive level 2 of factor 𝐴 and level 3
of factor 𝐵. The number of replicates is the number of subjects in each group.
Here 𝑦135 would be the measurement for the fifth member (replicate) of the
group for level 1 of factor 𝐴 and level 3 of factor 𝐵.
We call this setup an 𝐼 × 𝐽 factorial design with 𝐾 replicates. (Our current
notation only allows for equal replicates in each group. It isn’t difficult to
allow for different replicates for different groups, but we’ll proceed using equal
replicates per group, which if possible, is desirable.)
• 𝛼𝑖 measures the effect of level 𝑖 of factor 𝐴. We call these the main

effects of factor 𝐴.
• 𝛽𝑗 measures the effect of level 𝑗 of factor 𝐵. We call these the main
effects of factor 𝐵.
• (𝛼𝛽)𝑖𝑗 is a single parameter. We use 𝛼𝛽 to note that this parameter
measures the interaction between the two main effects.
Under this setup, there are a number of models that we can compare. Consider
a 2 × 2 factorial design. The following tables show the means for each of the
possible groups under each model.
Interaction Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + (𝛼𝛽)𝑖𝑗 + 𝜖𝑖𝑗𝑘
Factor B, Level 1 Factor B, Level 2

Factor A, Level 1 𝜇 + 𝛼1 + 𝛽1 + (𝛼𝛽)11 𝜇 + 𝛼1 + 𝛽2 + (𝛼𝛽)12
Factor A, Level 2 𝜇 + 𝛼2 + 𝛽1 + (𝛼𝛽)21 𝜇 + 𝛼2 + 𝛽2 + (𝛼𝛽)22
Additive Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + 𝜖𝑖𝑗𝑘

Factor A, Level 1 𝜇 + 𝛼 1 + 𝛽1 𝜇 + 𝛼 1 + 𝛽2
Factor A, Level 2 𝜇 + 𝛼 2 + 𝛽1 𝜇 + 𝛼 2 + 𝛽2
Factor B Only Model (One-Way): 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛽𝑗 + 𝜖𝑖𝑗𝑘

Factor A, Level 1 𝜇 + 𝛽1 𝜇 + 𝛽2
Factor A, Level 2 𝜇 + 𝛽1 𝜇 + 𝛽2
Factor A Only Model (One-Way): 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝜖𝑖𝑗𝑘

Factor A, Level 1 𝜇 + 𝛼1 𝜇 + 𝛼1
Factor A, Level 2 𝜇 + 𝛼2 𝜇 + 𝛼2
Null Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝜖𝑖𝑗𝑘

Factor A, Level 1 𝜇 𝜇
Factor A, Level 2 𝜇 𝜇
The question then, is which of these models should we use if we have two factors?
The most important question to consider is whether or not we should model the
interaction. Is the effect of Factor A the same for all levels of Factor B? In
the additive model, yes. In the interaction model, no. Both models would use
a different mean for each group, but in a very specific way in both cases.
Let’s discuss these comparisons by looking at some examples. We’ll first look at
the rats data from the faraway package. There are two factors here: poison
and treat. We use the levels() function to extract the levels of a factor
variable.
levels(rats$poison)
## [1] "I" "II" "III"

levels(rats$treat)
## [1] "A" "B" "C" "D"
Here, 48 rats were randomly assigned both one of three poisons and one of four
possible treatments. The experimenters then measured their survival time in
tens of hours. A total of 12 groups, each with 4 replicates.
Before running any tests, we should first look at the data. We will create
interaction plots, which will help us visualize the effect of one factor, as we
move through the levels of another factor.
par(mfrow = c(1, 2))

with(rats, interaction.plot(poison, treat, time, lwd = 2, col = 1:4))
with(rats, interaction.plot(treat, poison, time, lwd = 2, col = 1:3))
0.9
0.9
treat poison
0.8
0.8
B II
D I
0.7
0.7
C III
mean of time
mean of time
A
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
I II III A B C D
poison treat
If there is not interaction, thus an additive model, we would expect to see

parallel lines. That would mean, when we change the level of one factor, there
can be an effect on the response. However, the difference between the levels of
the other factor should still be the same.
The obvious indication of interaction would be lines that cross while heading in
different directions. Here we don’t see that, but the lines aren’t strictly parallel,
and there is some overlap on the right panel. However, is this interaction effect
significant?
Let’s fit each of the possible models, then investigate their estimates for each of
the group means.
rats_int = aov(time ~ poison * treat, data = rats) # interaction model

rats_add = aov(time ~ poison + treat, data = rats) # additive model
rats_pois = aov(time ~ poison , data = rats) # single factor model
rats_treat = aov(time ~ treat, data = rats) # single factor model
rats_null = aov(time ~ 1, data = rats) # null model
To get the estimates, we’ll create a table which we will predict on.
rats_table = expand.grid(poison = unique(rats$poison), treat = unique(rats$treat))

rats_table
## poison treat
## 1 I A
## 2 II A
## 3 III A
## 4 I B
## 5 II B
## 6 III B
## 7 I C
## 8 II C
## 9 III C
## 10 I D
## 11 II D
## 12 III D
matrix(paste0(rats_table$poison, "-", rats_table$treat) , 4, 3, byrow = TRUE)
## [,1] [,2] [,3]

## [1,] "I-A" "II-A" "III-A"
## [2,] "I-B" "II-B" "III-B"
## [3,] "I-C" "II-C" "III-C"
## [4,] "I-D" "II-D" "III-D"
Since we’ll be repeating ourselves a number of times, we write a function to

perform the prediction. Some housekeeping is done to keep the estimates in
order, and provide row and column names. Above, we’ve shown where each of
the estimates will be placed in the resulting matrix.
get_est_means = function(model, table) {

mat = matrix(predict(model, table), nrow = 4, ncol = 3, byrow = TRUE)
colnames(mat) = c("I", "II", "III")
rownames(mat) = c("A", "B", "C", "D")
mat
}
First, we obtain the estimates from the interaction model. Note that each cell
has a different value.
knitr::kable(get_est_means(model = rats_int, table = rats_table))
I II III
A 0.4125 0.3200 0.210
B 0.8800 0.8150 0.335
C 0.5675 0.3750 0.235
D 0.6100 0.6675 0.325
Next, we obtain the estimates from the additive model. Again, each cell has
a different value. We also see that these estimates are somewhat close to those
from the interaction model.
knitr::kable(get_est_means(model = rats_add, table = rats_table))
I II III
A 0.4522917 0.3791667 0.1110417
B 0.8147917 0.7416667 0.4735417
C 0.5306250 0.4575000 0.1893750
D 0.6722917 0.5991667 0.3310417
To understand the difference, let’s consider the effect of the treatments.
additive_means = get_est_means(model = rats_add, table = rats_table)

additive_means["A",] - additive_means["B",]
## I II III
## -0.3625 -0.3625 -0.3625
interaction_means = get_est_means(model = rats_int, table = rats_table)

interaction_means["A",] - interaction_means["B",]
## I II III
## -0.4675 -0.4950 -0.1250
This is the key difference between the interaction and additive models. The
difference between the effect of treatments A and B is the same for each poison
in the additive model. They are different in the interaction model.
The remaining three models are much simpler, having either only row or only
column effects. Or no effects in the case of the null model.
knitr::kable(get_est_means(model = rats_pois, table = rats_table))

I II III
A 0.6175 0.544375 0.27625
B 0.6175 0.544375 0.27625
C 0.6175 0.544375 0.27625
D 0.6175 0.544375 0.27625
knitr::kable(get_est_means(model = rats_treat, table = rats_table))
I II III
A 0.3141667 0.3141667 0.3141667
B 0.6766667 0.6766667 0.6766667
C 0.3925000 0.3925000 0.3925000
D 0.5341667 0.5341667 0.5341667
knitr::kable(get_est_means(model = rats_null, table = rats_table))
I II III
A 0.479375 0.479375 0.479375
B 0.479375 0.479375 0.479375
C 0.479375 0.479375 0.479375
D 0.479375 0.479375 0.479375
To perform the needed tests, we will need to create another ANOVA table.
(We’ll skip the details of the sums of squares calculations and simply let R take
care of them.)
Source Sum of Squares Degrees of Freedom Mean Square 𝐹

Factor A SSA 𝐼 −1 SSA / DFA MSA / MSE
Factor B SSB 𝐽 −1 SSB / DFB MSB / MSE
AB Interaction SSAB (𝐼 − 1)(𝐽 − 1) SSAB / DFAB MSAB / MSE
Error SSE 𝐼𝐽 (𝐾 − 1) SSE / DFE
Total SST 𝐼𝐽 𝐾 − 1
The row for AB Interaction tests:
𝐻0 ∶ All (𝛼𝛽)𝑖𝑗 = 0. vs 𝐻1 ∶ Not all (𝛼𝛽)𝑖𝑗 are 0.
• Null Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + 𝜖𝑖𝑗𝑘 . (Additive Model.)

• Alternative Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + (𝛼𝛽)𝑖𝑗 + 𝜖𝑖𝑗𝑘 . (Interaction Model.)
We reject the null when the 𝐹 statistic is large. Under the null hypothesis, the
distribution of the test statistic is 𝐹 with degrees of freedom (𝐼 − 1)(𝐽 − 1) and
𝐼𝐽 (𝐾 − 1).
The row for Factor B tests:
𝐻0 ∶ All 𝛽𝑗 = 0. vs 𝐻1 ∶ Not all 𝛽𝑗 are 0.
• Null Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝜖𝑖𝑗𝑘 . (Only Factor A Model.)

• Alternative Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + 𝜖𝑖𝑗𝑘 . (Additive Model.)
We reject the null when the 𝐹 statistic is large. Under the null hypothesis,
the distribution of the test statistic is 𝐹 with degrees of freedom 𝐽 − 1 and
𝐼𝐽 (𝐾 − 1).
The row for Factor A tests:
𝐻0 ∶ All 𝛼𝑖 = 0. vs 𝐻1 ∶ Not all 𝛼𝑖 are 0.
• Null Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛽𝑗 + 𝜖𝑖𝑗𝑘 . (Only Factor B Model.)

• Alternative Model: 𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + 𝜖𝑖𝑗𝑘 . (Additive Model.)
We reject the null when the 𝐹 statistic is large. Under the null hypothesis, the
distribution of the test statistic is 𝐹 with degrees of freedom 𝐼 −1 and 𝐼𝐽 (𝐾 −1).
These tests should be performed according to the model hierarchy. First con-
sider the test of interaction. If it is significant, we select the interaction model
and perform no further testing. If interaction is not significant, we then consider
the necessity of the individual factors of the additive model.
summary(aov(time ~ poison * treat, data = rats))

## poison 2 1.0330 0.5165 23.222 3.33e-07 ***
## treat 3 0.9212 0.3071 13.806 3.78e-06 ***
## poison:treat 6 0.2501 0.0417 1.874 0.112
## Residuals 36 0.8007 0.0222
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Using a significance level of 𝛼 = 0.05, we see that the interaction is not signif-
icant. Within the additive model, both factors are significant, so we select the
additive model.
Within the additive model, we could do further testing about the main effects.
Figure 12.1: Model Hierarchy

TukeyHSD(aov(time ~ poison + treat, data = rats))
## Tukey multiple comparisons of means

## 95% family-wise confidence level
##
## Fit: aov(formula = time ~ poison + treat, data = rats)
##
## $poison
## II-I -0.073125 -0.2089936 0.0627436 0.3989657
## III-I -0.341250 -0.4771186 -0.2053814 0.0000008
## III-II -0.268125 -0.4039936 -0.1322564 0.0000606
##
## $treat
## B-A 0.36250000 0.18976135 0.53523865 0.0000083
## C-A 0.07833333 -0.09440532 0.25107198 0.6221729
## D-A 0.22000000 0.04726135 0.39273865 0.0076661
## C-B -0.28416667 -0.45690532 -0.11142802 0.0004090
## D-B -0.14250000 -0.31523865 0.03023865 0.1380432
## D-C 0.14166667 -0.03107198 0.31440532 0.1416151
For an example with interaction, we investigate the warpbreaks dataset, a

default dataset in R.

with(warpbreaks, interaction.plot(wool, tension, breaks, lwd = 2, col = 2:4))
with(warpbreaks, interaction.plot(tension, wool, breaks, lwd = 2, col = 2:3))
45
45
tension wool
M A
40
40
L B
H
mean of breaks
mean of breaks
35
35
30
30
25
25
20
20
A B L M H
wool tension
Either plot makes it rather clear that the wool and tensions factors interact.
summary(aov(breaks ~ wool * tension, data = warpbreaks))

## wool 1 451 450.7 3.765 0.058213 .
## tension 2 2034 1017.1 8.498 0.000693 ***
## wool:tension 2 1003 501.4 4.189 0.021044 *
## Residuals 48 5745 119.7
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Using an 𝛼 of 0.05, the ANOVA test finds that the interaction is significant, so
we use the interaction model here.
12.6 R Markdown
• anova.Rmd

Chapter 13
Model Diagnostics
“Your assumptions are your windows on the world. Scrub them off
every once in a while, or the light won’t come in.”
— Isaac Asimov
• Understand the assumptions of a regression model.

• Assess regression model assumptions using visualizations and tests.
• Understand leverage, outliers, and influential points.
• Be able to identify unusual observations in regression models.
13.1 Model Assumptions

Recall the multiple linear regression model that we have defined.
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽𝑝−1 𝑥𝑖(𝑝−1) + 𝜖𝑖 , 𝑖 = 1, 2, … , 𝑛.
Using matrix notation, this model can be written much more succinctly as
𝑌 = 𝑋𝛽 + 𝜖.
Given data, we found the estimates for the 𝛽 parameters using
−1
𝛽 ̂ = (𝑋 ⊤ 𝑋) 𝑋 ⊤ 𝑦.
We then noted that these estimates had mean
261
262 CHAPTER 13. MODEL DIAGNOSTICS
E[𝛽]̂ = 𝛽,
and variance
−1
Var[𝛽]̂ = 𝜎2 (𝑋 ⊤ 𝑋) .
In particular, an individual parameter, say 𝛽𝑗̂ had a normal distribution
𝛽𝑗̂ ∼ 𝑁 (𝛽𝑗 , 𝜎2 𝐶𝑗𝑗 )
where 𝐶 was the matrix defined as
−1
𝐶 = (𝑋 ⊤ 𝑋) .
We then used this fact to define
𝛽𝑗̂ − 𝛽𝑗
∼ 𝑡𝑛−𝑝 ,
𝑠𝑒 √𝐶𝑗𝑗
which we used to perform hypothesis testing.

So far we have looked at various metrics such as RMSE, RSE, and 𝑅2 to deter-
mine how well our model fits our data. Each of these in some way considers the
expression
𝑛
∑(𝑦𝑖 − 𝑦𝑖̂ )2 .
𝑖=1
So, essentially each of these looks at how close the data points are to the model.
However is that all we care about?
• It could be that the errors are made in a systematic way, which means
that our model is misspecified. We may need additional interaction terms,
or polynomial terms which we will see later.
• It is also possible that at a particular set of predictor values, the errors are
very small, but at a different set of predictor values, the errors are large.
• Perhaps most of the errors are very small, but some are very large. This
would suggest that the errors do not follow a normal distribution.
13.2. CHECKING ASSUMPTIONS 263
Are these issues that we care about? If all we would like to do is predict, possibly
not, since we would only care about the size of our errors. However, if we would
like to perform inference, for example to determine if a particular predictor is
important, we care a great deal. All of the distributional results, such as a 𝑡-test
for a single predictor, are derived under the assumptions of our model.
Technically, the assumptions of the model are encoded directly in a model state-
ment such as,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽𝑝−1 𝑥𝑖(𝑝−1) + 𝜖𝑖
Often, the assumptions of linear regression, are stated as,
• Linearity: the response can be written as a linear combination of the

predictors. (With noise about this true linear relationship.)
• Independence: the errors are independent.
• Normality: the distribution of the errors should follow a normal distribu-
tion.
• Equal Variance: the error variance is the same at any set of predictor
values.
The linearity assumption is encoded as
𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽𝑝−1 𝑥𝑖(𝑝−1) ,
while the remaining three, are all encoded in
𝜖𝑖 ∼ 𝑁 (0, 𝜎2 ),
since the 𝜖𝑖 are 𝑖𝑖𝑑 normal random variables with constant variance.
If these assumptions are met, great! We can perform inference, and it is valid.
If these assumptions are not met, we can still “perform” a 𝑡-test using R, but
the results are not valid. The distributions of the parameter estimates will
not be what we expect. Hypothesis tests will then accept or reject incorrectly.
Essentially, garbage in, garbage out.
13.2 Checking Assumptions

We’ll now look at a number of tools for checking the assumptions of a linear
model. To test these tools, we’ll use data simulated from three models:
Model 1: 𝑌 = 3 + 5𝑥 + 𝜖, 𝜖 ∼ 𝑁 (0, 1)
Model 2: 𝑌 = 3 + 5𝑥 + 𝜖, 𝜖 ∼ 𝑁 (0, 𝑥2 )
Model 3: 𝑌 = 3 + 5𝑥2 + 𝜖, 𝜖 ∼ 𝑁 (0, 25)
sim_1 = function(sample_size = 500) {

x = runif(n = sample_size) * 5
y = 3 + 5 * x + rnorm(n = sample_size, mean = 0, sd = 1)
data.frame(x, y)
}

y = 3 + 5 * x + rnorm(n = sample_size, mean = 0, sd = x)
data.frame(x, y)
}

y = 3 + 5 * x ^ 2 + rnorm(n = sample_size, mean = 0, sd = 5)
data.frame(x, y)
}
13.2.1 Fitted versus Residuals Plot
Probably our most useful tool will be a Fitted versus Residuals Plot. It will
be useful for checking both the linearity and constant variance assumptions.
Data generated from Model 1 above should not show any signs of violating
assumptions, so we’ll use this to see what a good fitted versus residuals plot
should look like. First, we’ll simulate observations from this model.
set.seed(42)
sim_data_1 = sim_1()
head(sim_data_1)
## x y
## 1 4.574030 24.773995
## 2 4.685377 26.475936
## 3 1.430698 8.954993
## 4 4.152238 23.951210
## 5 3.208728 20.341344
## 6 2.595480 14.943525
We then fit the model and add the fitted line to a scatterplot.
plot(y ~ x, data = sim_data_1, col = "grey", pch = 20,

main = "Data from Model 1")
fit_1 = lm(y ~ x, data = sim_data_1)
abline(fit_1, col = "darkorange", lwd = 3)
Data from Model 1

30
25
20
15
y
10
5
0 1 2 3 4 5
We now plot a fitted versus residuals plot. Note, this is residuals on the 𝑦-axis
despite the ordering in the name. Sometimes you will see this called a residuals
versus fitted, or residuals versus predicted plot.
plot(fitted(fit_1), resid(fit_1), col = "grey", pch = 20,

xlab = "Fitted", ylab = "Residuals", main = "Data from Model 1")
Data from Model 1
3
2
1
Residuals
0
-1
-2
-3
5 10 15 20 25
Fitted
We should look for two things in this plot.
• At any fitted value, the mean of the residuals should be roughly 0. If this
is the case, the linearity assumption is valid. For this reason, we generally
add a horizontal line at 𝑦 = 0 to emphasize this point.
• At every fitted value, the spread of the residuals should be roughly the
same. If this is the case, the constant variance assumption is valid.
Here we see this is the case for both.
To get a better idea of how a fitted versus residuals plot can be useful, we will
simulate from models with violated assumptions.
Model 2 is an example of non-constant variance. In this case, the variance is

larger for larger values of the predictor variable 𝑥.
set.seed(42)
Data from Model 2

40
30
y
20
10
0 1 2 3 4 5
This actually is rather easy to see here by adding the fitted line to a scatterplot.
This is because we are only performing simple linear regression. With multiple
regression, a fitted versus residuals plot is a necessity, since adding a fitted
regression to a scatterplot isn’t exactly possible.

Data from Model 2
15
10
5
Residuals
0
-5
-10
-15
5 10 15 20 25
Fitted
On the fitted versus residuals plot, we see two things very clearly. For any fitted
value, the residuals seem roughly centered at 0. This is good! The linearity
assumption is not violated. However, we also see very clearly, that for larger
fitted values, the spread of the residuals is larger. This is bad! The constant
variance assumption is violated here.
Now we will demonstrate a model which does not meet the linearity assumption.
Model 3 is an example of a model where 𝑌 is not a linear combination of the
predictors. In this case the predictor is 𝑥, but the model uses 𝑥2 . (We’ll see later
that this is something that a “linear” model can deal with. The fix is simple,
just make 𝑥2 a predictor!)
set.seed(42)
Data from Model 3

120
80
y
60
40
20
0
0 1 2 3 4 5
Again, this is rather clear on the scatterplot, but again, we wouldn’t be able to
check this plot for multiple regression.

Data from Model 3
30
20
Residuals
10
0
-10
-20
-20 0 20 40 60 80 100
Fitted
This time on the fitted versus residuals plot, for any fitted value, the spread of
the residuals is about the same. However, they are not even close to centered
at zero! At small and large fitted values the model is underestimating, while at
medium fitted values, the model is overestimating. These are systematic errors,
not random noise. So the constant variance assumption is met, but the linearity
assumption is violated. The form of our model is simply wrong. We’re trying
to fit a line to a curve!
13.2.2 Breusch-Pagan Test
Constant variance is often called homoscedasticity. Conversely, non-constant

variance is called heteroscedasticity. We’ve seen how we can use a fitted
versus residuals plot to look for these attributes.
While a fitted versus residuals plot can give us an idea about homoscedasticity,
sometimes we would prefer a more formal test. There are many tests for constant
variance, but here we will present one, the Breusch-Pagan Test. The exact
details of the test will be omitted here, but importantly the null and alternative
can be considered to be,
• 𝐻0 : Homoscedasticity. The errors have constant variance about the true

model.
• 𝐻1 : Heteroscedasticity. The errors have non-constant variance about the
true model.
Isn’t that convenient? A test that will specifically test the constant variance
assumption.
The Breusch-Pagan Test can not be performed by default in R, however the
function bptest in the lmtest package implements the test.
#install.packages("lmtest")
library(lmtest)
Let’s try it on the three models we fit above. Recall,
• fit_1 had no violation of assumptions,

• fit_2 violated the constant variance assumption, but not linearity,
• fit_3 violated linearity, but not constant variance.
bptest(fit_1)
##
## studentized Breusch-Pagan test
##
## data: fit_1
## BP = 1.0234, df = 1, p-value = 0.3117
For fit_1 we see a large p-value, so we do not reject the null of homoscedasticity,
which is what we would expect.
bptest(fit_2)
##
##
## data: fit_2
## BP = 76.693, df = 1, p-value < 2.2e-16
For fit_2 we see a small p-value, so we reject the null of homoscedasticity.

The constant variance assumption is violated. This matches our findings with
a fitted versus residuals plot.
bptest(fit_3)
##
##
## data: fit_3
## BP = 0.33466, df = 1, p-value = 0.5629
Lastly, for fit_3 we again see a large p-value, so we do not reject the null of
homoscedasticity, which matches our findings with a fitted versus residuals plot.
13.2.3 Histograms
We have a number of tools for assessing the normality assumption. The most
obvious would be to make a histogram of the residuals. If it appears roughly
normal, then we’ll believe the errors could truly be normal.

hist(resid(fit_1),
xlab = "Residuals",
main = "Histogram of Residuals, fit_1",
col = "darkorange",
border = "dodgerblue",
breaks = 20)
hist(resid(fit_2),
xlab = "Residuals",
col = "darkorange",
breaks = 20)
hist(resid(fit_3),
xlab = "Residuals",
col = "darkorange",
breaks = 20)
Histogram of Residuals, fit_1 Histogram of Residuals, fit_2 Histogram of Residuals, fit_3

40
100
150
80
30
60
100
Frequency
Frequency
Frequency
20
40
50
10
20
0
-3 -2 -1 0 1 2 3 -15 -10 -5 0 5 10 15 -20 -10 0 10 20 30
Residuals Residuals Residuals
Above are histograms for each of the three regression models we have been
considering. Notice that the first, for fit_1 appears very normal. The third,
for fit_3, appears to be very non-normal. However fit_2 is not as clear. It
does have a rough bell shape, however, it also has a very sharp peak. For this
reason we will usually use more powerful tools such as Q-Q plots and the
Shapiro-Wilk test for assessing the normality of errors.
13.2.4 Q-Q Plots
Another visual method for assessing the normality of errors, which is more
powerful than a histogram, is a normal quantile-quantile plot, or Q-Q plot for
short.
In R these are very easy to make. The qqnorm() function plots the points, and
the qqline() function adds the necessary line. We create a Q-Q plot for the
residuals of fit_1 to check if the errors could truly be normally distributed.
qqnorm(resid(fit_1), main = "Normal Q-Q Plot, fit_1", col = "darkgrey")

qqline(resid(fit_1), col = "dodgerblue", lwd = 2)
Normal Q-Q Plot, fit_1

3
2
Sample Quantiles
1
0
-1
-2
-3
-3 -2 -1 0 1 2 3
Theoretical Quantiles
In short, if the points of the plot do not closely follow a straight line, this would
suggest that the data do not come from a normal distribution.
The calculations required to create the plot vary depending on the implementa-
tion, but essentially the 𝑦-axis is the sorted data (observed, or sample quantiles),
and the 𝑥-axis is the values we would expect if the data did come from a normal
distribution (theoretical quantiles).
The Wikipedia page for Normal probability plots gives details on how this is
implemented in R if you are interested.
Also, to get a better idea of how Q-Q plots work, here is a quick function which
creates a Q-Q plot:
qq_plot = function(e) {
n = length(e)
normal_quantiles = qnorm(((1:n - 0.5) / n))
# normal_quantiles = qnorm(((1:n) / (n + 1)))
# plot theoretical verus observed quantiles

plot(normal_quantiles, sort(e),
xlab = c("Theoretical Quantiles"),
ylab = c("Sample Quantiles"),
col = "darkgrey")
title("Normal Q-Q Plot")
# calculate line through the first and third quartiles

slope = (quantile(e, 0.75) - quantile(e, 0.25)) / (qnorm(0.75) - qnorm(0.25))
intercept = quantile(e, 0.25) - slope * qnorm(0.25)
# add to existing plot

abline(intercept, slope, lty = 2, lwd = 2, col = "dodgerblue")
}
We can then verify that it is essentially equivalent to using qqnorm() and

qqline() in R.
set.seed(420)
x = rnorm(100, mean = 0 , sd = 1)
qqnorm(x, col = "darkgrey")
qqline(x, lty = 2, lwd = 2, col = "dodgerblue")
qq_plot(x)
Normal Q-Q Plot Normal Q-Q Plot

2
2
1
1
Sample Quantiles
Sample Quantiles
0
0
-1
-1
-2
-2
-3
-3
-2 -1 0 1 2 -2 -1 0 1 2
Theoretical Quantiles Theoretical Quantiles
To get a better idea of what “close to the line” means, we perform a number of
simulations, and create Q-Q plots.
First we simulate data from a normal distribution with different sample sizes,
and each time create a Q-Q plot.

set.seed(420)
qq_plot(rnorm(10))
qq_plot(rnorm(25))
qq_plot(rnorm(100))
Normal Q-Q Plot Normal Q-Q Plot Normal Q-Q Plot

2.0
2
0.5
1.5
0.0
1
1.0
-0.5
Sample Quantiles
Sample Quantiles
Sample Quantiles
0
0.5
-1.0
-1.5
0.0
-1
-2.0
-0.5
-2
-2.5
-1.0
-3.0
-3
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 -2 -1 0 1 2 -2 -1 0 1 2
Theoretical Quantiles Theoretical Quantiles Theoretical Quantiles
Since this data is sampled from a normal distribution, these are all, by definition,
good Q-Q plots. The points are “close to the line” and we would conclude that
this data could have been sampled from a normal distribution. Notice in the first
plot, one point is somewhat far from the line, but just one point, in combination
with the small sample size, is not enough to make us worried. We see with the
large sample size, all of the points are rather close to the line.
Next, we simulate data from a 𝑡 distribution with a small degrees of freedom,

for different sample sizes.

set.seed(420)
qq_plot(rt(10, df = 4))

4
1.0
4
0.5
2
Sample Quantiles
Sample Quantiles
Sample Quantiles
0.0
0
-0.5
0
-2
-1.0
-2
-4
-1.5
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 -2 -1 0 1 2 -2 -1 0 1 2
Recall that as the degrees of freedom for a 𝑡 distribution become larger, the
distribution becomes more and more similar to a normal. Here, using 4 degrees
of freedom, we have a distribution that is somewhat normal, it is symmetrical
and roughly bell-shaped, however it has “fat tails.” This presents itself clearly in
the third panel. While many of the points are close to the line, at the edges, there
are large discrepancies. This indicates that the values are too small (negative) or
too large (positive) compared to what we would expect for a normal distribution.
So for the sample size of 100, we would conclude that the normality assumption
is violated. (If these were residuals of a model.) For sample sizes of 10 and 25
we may be suspicious, but not entirely confident. Reading Q-Q plots, is a bit
of an art, not completely a science.
Next, we simulate data from an exponential distribution.

set.seed(420)
qq_plot(rexp(10))
qq_plot(rexp(25))
qq_plot(rexp(100))

2.0
4
6
1.5
3
Sample Quantiles
Sample Quantiles
Sample Quantiles
4
1.0
2
3
2
0.5
1
1
0.0
0
0
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 -2 -1 0 1 2 -2 -1 0 1 2
This is a distribution that is not very similar to a normal, so in all three cases,
we see points that are far from the lines, so we would think that the normality
assumption is violated.
For a better understanding of which Q-Q plots are “good,” repeat the simula-
tions above a number of times (without setting the seed) and pay attention to
the differences between those that are simulated from normal, and those that
are not. Also consider different sample sizes and distribution parameters.
Returning to our three regressions, recall,
• fit_1 had no violation of assumptions,

• fit_2 violated the constant variance assumption, but not linearity,
• fit_3 violated linearity, but not constant variance.
We’ll now create a Q-Q plot for each to assess normality of errors.

3
2
Sample Quantiles
1
0
-1
-2
-3
-3 -2 -1 0 1 2 3
For fit_1, we have a near perfect Q-Q plot. We would believe the errors follow
a normal distribution.


15
10
Sample Quantiles
5
0
-5
-10
-15
-3 -2 -1 0 1 2 3
For fit_2, we have a suspect Q-Q plot. We would probably not believe the
errors follow a normal distribution.

30
20
Sample Quantiles
10
0
-10
-20
-3 -2 -1 0 1 2 3
Lastly, for fit_3, we again have a suspect Q-Q plot. We would probably not
believe the errors follow a normal distribution.
13.2.5 Shapiro-Wilk Test
Histograms and Q-Q Plots give a nice visual representation of the residuals
distribution, however if we are interested in formal testing, there are a number
of options available. A commonly used test is the Shapiro–Wilk test, which
is implemented in R.
set.seed(42)
shapiro.test(rnorm(25))
##
## Shapiro-Wilk normality test
##
## data: rnorm(25)
## W = 0.9499, p-value = 0.2495
shapiro.test(rexp(25))
##

##
## data: rexp(25)
## W = 0.71164, p-value = 1.05e-05
This gives us the value of the test statistic and its p-value. The null hypothesis
assumes the data were sampled from a normal distribution, thus a small p-value
indicates we believe there is only a small probability the data could have been
sampled from a normal distribution.
For details, see: Wikipedia: Shapiro–Wilk test.
In the above examples, we see we fail to reject for the data sampled from normal,
and reject on the non-normal data, for any reasonable 𝛼.
Returning again to fit_1, fit_2, and fit_3, we see the result of running
shapiro.test() on the residuals of each, returns a result for each that matches
decisions based on the Q-Q plots.
shapiro.test(resid(fit_1))
##
##
## data: resid(fit_1)
## W = 0.99858, p-value = 0.9622
##
##
## W = 0.93697, p-value = 1.056e-13
##
##
## W = 0.97643, p-value = 3.231e-07
13.3 Unusual Observations

In addition to checking the assumptions of regression, we also look for any
“unusual observations” in the data. Often a small number of data points can
have an extremely large influence on a regression, sometimes so much so that
the regression assumptions are violated as a result of these points.
The following three plots are inspired by an example from Linear Models with
R.

set.seed(42)
ex_data = data.frame(x = 1:10,
y = 10:1 + rnorm(n = 10))
ex_model = lm(y ~ x, data = ex_data)
# low leverage, large residual, small influence

point_1 = c(5.4, 11)
ex_data_1 = rbind(ex_data, point_1)
model_1 = lm(y ~ x, data = ex_data_1)
plot(y ~ x, data = ex_data_1, cex = 2, pch = 20, col = "grey",
main = "Low Leverage, Large Residual, Small Influence")
points(x = point_1[1], y = point_1[2], pch = 1, cex = 4, col = "black", lwd = 2)
abline(ex_model, col = "dodgerblue", lwd = 2)
abline(model_1, lty = 2, col = "darkorange", lwd = 2)
legend("bottomleft", c("Original Data", "Added Point"),
lty = c(1, 2), col = c("dodgerblue", "darkorange"))
# high leverage, small residual, small influence

point_2 = c(18, -5.7)
plot(y ~ x, data = ex_data_2, cex = 2, pch = 20, col = "grey",
main = "High Leverage, Small Residual, Small Influence")
# high leverage, large residual, large influence

point_3 = c(14, 5.1)
plot(y ~ x, data = ex_data_3, cex = 2, pch = 20, col = "grey", ylim = c(-3, 12),
main = "High Leverage, Large Residual, Large Influence")
13.3. UNUSUAL OBSERVATIONS 283

Low Leverage, Large Residual, Small Influence High Leverage, Small Residual, Small Influence High Leverage, Large Residual, Large Influence
10
10
10
8
5
y
y
6
0
4
0
2
-5
Original Data Original Data Original Data

Added Point Added Point Added Point
2 4 6 8 10 5 10 15 2 4 6 8 10 12 14
x x x
The blue solid line in each plot is a regression fit to the 10 original data points
stored in ex_data. The dashed orange line in each plot is the result of adding a
single point to the original data in ex_data. This additional point is indicated
by the circled point.
The slope of the regression for the original ten points, the solid blue line, is
given by:
coef(ex_model)[2]
## x
## -0.9696033
The added point in the first plot has a small effect on the slope, which becomes:
coef(model_1)[2]
## x
## -0.9749534
We will say that this point has low leverage, is an outlier due to its large residual,
but has small influence.
The added point in the second plot also has a small effect on the slope, which
is:
coef(model_2)[2]
## x
## -0.9507397
We will say that this point has high leverage, is not an outlier due to its small
residual, and has a very small influence.
Lastly, the added point in the third plot has a large effect on the slope, which
is now:
coef(model_3)[2]
## x
## -0.5892241
This added point is influential. It both has high leverage, and is an outlier due
to its large residual.
We’ve now mentioned three new concepts: leverage, outliers, and influential
points, each of which we will discuss in detail.
13.3.1 Leverage
A data point with high leverage, is a data point that could have a large influence
when fitting the model.
Recall that,
−1
𝛽 ̂ = (𝑋 ⊤ 𝑋) 𝑋 ⊤ 𝑦.
Thus,
−1
𝑦 ̂ = 𝑋 𝛽 ̂ = 𝑋 (𝑋 ⊤ 𝑋) 𝑋⊤𝑦
Now we define,
−1
𝐻 = 𝑋 (𝑋 ⊤ 𝑋) 𝑋⊤
which we will refer to as the hat matrix. The hat matrix is used to project
onto the subspace spanned by the columns of 𝑋. It is also simply known as a
projection matrix.
The hat matrix is a matrix that takes the original 𝑦 values, and adds a hat!
𝑦 ̂ = 𝐻𝑦
The diagonal elements of this matrix are called the leverages
𝐻𝑖𝑖 = ℎ𝑖 ,
where ℎ𝑖 is the leverage for the 𝑖th observation.

Large values of ℎ𝑖 indicate extreme values in 𝑋, which may influence regression.
Note that leverages only depend on 𝑋.
Here, 𝑝, the number of 𝛽s, is also the trace (and rank) of the hat matrix.
𝑛
∑ ℎ𝑖 = 𝑝
𝑖=1
What is a value of ℎ𝑖 that would be considered large? There is no exact answer

to this question. A common heuristic would be to compare each leverage to
two times the average leverage. A leverage larger than this is considered an
observation to be aware of. That is, if
ℎ𝑖 > 2ℎ̄
we say that observation 𝑖 has large leverage. Here,
𝑛
∑𝑖=1 ℎ𝑖 𝑝
ℎ̄ = = .
𝑛 𝑛
For simple linear regression, the leverage for each point is given by
1 (𝑥 − 𝑥)̄ 2
ℎ𝑖 = + 𝑖 .
𝑛 𝑆𝑥𝑥
This expression should be familiar. (Think back to inference for SLR.) It sug-
gests that the large leverages occur when 𝑥 values are far from their mean.
Recall that the regression goes through the point (𝑥,̄ 𝑦).
̄
There are multiple ways to find leverages in R.
lev_ex = data.frame(
x1 = c(0, 11, 11, 7, 4, 10, 5, 8),
x2 = c(1, 5, 4, 3, 1, 4, 4, 2),
y = c(11, 15, 13, 14, 0, 19, 16, 8))
plot(x2 ~ x1, data = lev_ex, cex = 2)

points(7, 3, pch = 20, col = "red", cex = 2)
5
4
x2
3
2
1
0 2 4 6 8 10
x1
Here we’ve created some multivariate data. Notice that we have plotted the 𝑥
values, not the 𝑦 values. The red point is (7, 3) which is the mean of x1 and the
mean of x2 respectively.
We could calculate the leverages using the expressions defined above. We first
create the 𝑋 matrix, then calculate 𝐻 as defined, and extract the diagonal
elements.
X = cbind(rep(1, 8), lev_ex$x1, lev_ex$x2)

H = X %*% solve(t(X) %*% X) %*% t(X)
diag(H)
## [1] 0.6000 0.3750 0.2875 0.1250 0.4000 0.2125 0.5875 0.4125
Notice here, we have two predictors, so the regression would have 3 𝛽 parameters,
so the sum of the diagonal elements is 3.
sum(diag(H))
## [1] 3
Alternatively, the method we will use more often, is to simply fit a regression,
then use the hatvalues() function, which returns the leverages.
lev_fit = lm(y ~ ., data = lev_ex)

hatvalues(lev_fit)
## 1 2 3 4 5 6 7 8
## 0.6000 0.3750 0.2875 0.1250 0.4000 0.2125 0.5875 0.4125
Again, note that here we have “used” the 𝑦 values to fit the regression, but R
still ignores them when calculating the leverages, as leverages only depend on
the 𝑥 values.
coef(lev_fit)
## 3.7 -0.7 4.4
Let’s see what happens to these coefficients when we modify the y value of the
point with the highest leverage.
which.max(hatvalues(lev_fit))
## 1
## 1
lev_ex[which.max(hatvalues(lev_fit)),]
## x1 x2 y
## 1 0 1 11
We see that the original y value is 11. We’ll create a copy of the data, and
modify this point to have a y value of 20.
lev_ex_1 = lev_ex
lev_ex_1$y[1] = 20
lm(y ~ ., data = lev_ex_1)
##
## Call:
## lm(formula = y ~ ., data = lev_ex_1)
##
## Coefficients:
## 8.875 -1.375 4.625
Notice the large changes in the coefficients. Also notice that each of the coef-
ficients has changed in some way. Note that the leverages of the points would
not have changed, as we have not modified any of the 𝑥 values.
Now let’s see what happens to these coefficients when we modify the y value of
the point with the lowest leverage.
which.min(hatvalues(lev_fit))
## 4
## 4
lev_ex[which.min(hatvalues(lev_fit)),]
## x1 x2 y
## 4 7 3 14
We see that the original y value is 14. We’ll again create a copy of the data,
and modify this point to have a y value of 30.
lev_ex_2 = lev_ex
lev_ex_2$y[4] = 30
lm(y ~ ., data = lev_ex_2)
##
## Call:
## lm(formula = y ~ ., data = lev_ex_2)
##
## Coefficients:
## 5.7 -0.7 4.4
This time despite a large change in the y value, there is only small change in
the coefficients. Also, only the intercept has changed!
mean(lev_ex$x1)
## [1] 7
mean(lev_ex$x2)
## [1] 3
lev_ex[4,]
## x1 x2 y
## 4 7 3 14
Notice that this point was the mean of both of the predictors.
Returning to our three plots, each with an added point, we can calculate the
leverages for each. Note that the 11th data point each time is the added data
point.
hatvalues(model_1)
## 1 2 3 4 5 6 7
## 0.33534597 0.23860732 0.16610842 0.11784927 0.09382988 0.09405024 0.11851036
## 8 9 10 11
## 0.16721022 0.24014985 0.33732922 0.09100926
hatvalues(model_2)
## 1 2 3 4 5 6 7
## 0.23238866 0.18663968 0.14979757 0.12186235 0.10283401 0.09271255 0.09149798
## 8 9 10 11
## 0.09919028 0.11578947 0.14129555 0.66599190
hatvalues(model_3)
## 1 2 3 4 5 6 7
## 0.27852761 0.21411043 0.16319018 0.12576687 0.10184049 0.09141104 0.09447853
## 8 9 10 11
## 0.11104294 0.14110429 0.18466258 0.49386503
Are any of these large?

hatvalues(model_1) > 2 * mean(hatvalues(model_1))
## 1 2 3 4 5 6 7 8 9 10 11
## FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
## 1 2 3 4 5 6 7 8 9 10 11
## FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE TRUE
## 1 2 3 4 5 6 7 8 9 10 11
## FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE TRUE
We see that in the second and third plots, the added point is a point of high
leverage. Recall that only in the third plot did that have an influence on the
regression. To understand why, we’ll need to discuss outliers.
13.3.2 Outliers
Outliers are points which do not fit the model well. They may or may not have
a large effect on the model. To identify outliers, we will look for observations
with large residuals.
Note,
𝑒 = 𝑦 − 𝑦 ̂ = 𝐼𝑦 − 𝐻𝑦 = (𝐼 − 𝐻)𝑦
Then, under the assumptions of linear regression,
Var(𝑒𝑖 ) = (1 − ℎ𝑖 )𝜎2
and thus estimating 𝜎2 with 𝑠2𝑒 gives
SE[𝑒𝑖 ] = 𝑠𝑒 √(1 − ℎ𝑖 ).
We can then look at the standardized residual for each observation, 𝑖 =

1, 2, … 𝑛,
𝑒𝑖 𝑎𝑝𝑝𝑟𝑜𝑥
𝑟𝑖 = ∼ 𝑁 (𝜇 = 0, 𝜎2 = 1)
𝑠𝑒 √1 − ℎ𝑖
when 𝑛 is large.
We can use this fact to identify “large” residuals. For example, standardized
residuals greater than 2 in magnitude should only happen approximately 5 per-
cent of the time.
Returning again to our three plots, each with an added point, we can calculate
the residuals and standardized residuals for each. Standardized residuals can
be obtained in R by using rstandard() where we would normally use resid().
resid(model_1)
## 1 2 3 4 5 6 7
## 0.4949887 -1.4657145 -0.5629345 -0.3182468 -0.5718877 -1.1073271 0.4852728
## 8 9 10 11
## -1.1459548 0.9420814 -1.1641029 4.4138254
rstandard(model_1)
## 1 2 3 4 5 6 7
## 0.3464701 -0.9585470 -0.3517802 -0.1933575 -0.3428264 -0.6638841 0.2949482
## 8 9 10 11
## -0.7165857 0.6167268 -0.8160389 2.6418234
rstandard(model_1)[abs(rstandard(model_1)) > 2]
## 11
## 2.641823
In the first plot, we see that the 11th point, the added point, is a large stan-
dardized residual.
resid(model_2)
## 1 2 3 4 5 6
## 1.03288292 -0.95203397 -0.07346766 0.14700626 -0.13084829 -0.69050140
## 7 8 9 10 11
## 0.87788484 -0.77755647 1.28626601 -0.84413207 0.12449986
rstandard(model_2)
## 1 2 3 4 5 6
## 1.41447023 -1.26655590 -0.09559792 0.18822094 -0.16574677 -0.86977220
## 7 8 9 10 11
## 1.10506546 -0.98294409 1.64121833 -1.09295417 0.25846620
## named numeric(0)
In the second plot, we see that there are no points with large standardized
residuals.
resid(model_3)
## 1 2 3 4 5 6
## 2.30296166 -0.04347087 0.47357980 0.33253808 -0.30683212 -1.22800087
## 7 8 9 10 11
## -0.02113027 -2.03808722 -0.33578039 -2.82769411 3.69191633
rstandard(model_3)
## 1 2 3 4 5 6
## 1.41302755 -0.02555591 0.26980722 0.18535382 -0.16873216 -0.67141143
## 7 8 9 10 11
## -0.01157256 -1.12656475 -0.18882474 -1.63206526 2.70453408
## 11
## 2.704534
In the last plot, we see that the 11th point, the added point, is a large stan-
dardized residual.
Recall that the added point in plots two and three were both high leverage, but
now only the point in plot three has a large residual. We will now combine this
information and discuss influence.
13.3.3 Influence
As we have now seen in the three plots, some outliers only change the regression
a small amount (plot one) and some outliers have a large effect on the regression
(plot three). Observations that fall into the latter category, points with (some
combination of) high leverage and large residual, we will call influential.
A common measure of influence is Cook’s Distance, which is defined as
1 2 ℎ𝑖
𝐷𝑖 = 𝑟 .
𝑝 𝑖 1 − ℎ𝑖
Notice that this is a function of both leverage and standardized residuals.

A Cook’s Distance is often considered large if
4
𝐷𝑖 >
𝑛
and an observation with a large Cook’s Distance is called influential. This is

again simply a heuristic, and not an exact rule.
The Cook’s distance for each point of a regression can be calculated using
cooks.distance() which is a default function in R. Let’s look for influential
points in the three plots we had been considering.
Low Leverage, Large Residual, Small Influence High Leverage, Small Residual, Small Influence High Leverage, Large Residual, Large Influence
10
10
10
8
5
y
y
6
0
4
0
2
-5
Original Data Original Data Original Data

Added Point Added Point Added Point
2 4 6 8 10 5 10 15 2 4 6 8 10 12 14
x x x
Recall that the circled points in each plot have different characteristics:
• Plot One: low leverage, large residual.

• Plot Two: high leverage, small residual.
• Plot Three: high leverage, large residual.
We’ll now directly check if each of these is influential.
cooks.distance(model_1)[11] > 4 / length(cooks.distance(model_1))
## 11
## FALSE

## 11
## FALSE
## 11
## TRUE
And, as expected, the added point in the third plot, with high leverage and a
large residual, is considered influential!
13.4 Data Analysis Examples
13.4.1 Good Diagnostics
Last chapter we fit an additive regression to the mtcars data with mpg as the
response and hp and am as predictors. Let’s perform some diagnostics on this
model.
First, fit the model as we did last chapter.
mpg_hp_add = lm(mpg ~ hp + am, data = mtcars)
plot(fitted(mpg_hp_add), resid(mpg_hp_add), col = "grey", pch = 20,

xlab = "Fitted", ylab = "Residual",
main = "mtcars: Fitted versus Residuals")
13.4. DATA ANALYSIS EXAMPLES 295
mtcars: Fitted versus Residuals

6
4
2
Residual
0
-2
-4
15 20 25
Fitted
The fitted versus residuals plot looks good. We don’t see any obvious pattern,
and the variance looks roughly constant. (Maybe a little larger for large fitted
values, but not enough to worry about.)
bptest(mpg_hp_add)
##
##
## data: mpg_hp_add
## BP = 7.5858, df = 2, p-value = 0.02253
The Breusch-Pagan test verifies this, at least for a small 𝛼 value.
qqnorm(resid(mpg_hp_add), col = "darkgrey")

qqline(resid(mpg_hp_add), col = "dodgerblue", lwd = 2)
Normal Q-Q Plot
6
4
Sample Quantiles
2
0
-2
-4
-2 -1 0 1 2
The Q-Q plot looks extremely good and the Shapiro-Wilk test agrees.
shapiro.test(resid(mpg_hp_add))
##
##
## data: resid(mpg_hp_add)
## W = 0.96485, p-value = 0.3706
sum(hatvalues(mpg_hp_add) > 2 * mean(hatvalues(mpg_hp_add)))
## [1] 2
We see that there are two points of large leverage.
sum(abs(rstandard(mpg_hp_add)) > 2)
## [1] 1
There is also one point with a large residual. Do these result in any points that
are considered influential?
cd_mpg_hp_add = cooks.distance(mpg_hp_add)
sum(cd_mpg_hp_add > 4 / length(cd_mpg_hp_add))
## [1] 2
large_cd_mpg = cd_mpg_hp_add > 4 / length(cd_mpg_hp_add)

cd_mpg_hp_add[large_cd_mpg]
## Toyota Corolla Maserati Bora

## 0.1772555 0.3447994
We find two influential points. Interestingly, they are very different cars.
coef(mpg_hp_add)
## 26.5849137 -0.0588878 5.2770853
Since the diagnostics looked good, there isn’t much need to worry about these
two points, but let’s see how much the coefficients change if we remove them.
mpg_hp_add_fix = lm(mpg ~ hp + am,

data = mtcars,
subset = cd_mpg_hp_add <= 4 / length(cd_mpg_hp_add))
coef(mpg_hp_add_fix)
## 27.22190933 -0.06286249 4.29765867
It seems there isn’t much of a change in the coefficients as a result of removing

the supposed influential points. Notice we did not create a new dataset to
accomplish this. We instead used the subset argument to lm(). Think about
what the code cd_mpg_hp_add <= 4 / length(cd_mpg_hp_add) does here.

plot(mpg_hp_add)
Residuals vs Fitted Normal Q-Q
6
Toyota Corolla Toyota Corolla
2
Lotus Europa Lotus Europa
Fiat 128 Fiat 128
Standardized residuals
4
1
Residuals
2
0
0
-2
-1
-4
15 20 25 -2 -1 0 1 2
Fitted values Theoretical Quantiles
Scale-Location Residuals vs Leverage

1.5
Toyota Corolla
Lotus Europa Toyota Corolla 1
2
Fiat 128
Fiat 128
0.5
Maserati Bora
1.0
1
0
0.5
-1
0.5
0.0
Cook's distance
-2
15 20 25 0.0 0.1 0.2 0.3 0.4
Fitted values Leverage
Notice that calling plot() on a variable which stores an object created by lm()
outputs four diagnostic plots by default. Use ?plot.lm to learn more. The first
two should already be familiar.
13.4.2 Suspect Diagnostics
Let’s consider the model big_model from last chapter which was fit to the
autompg dataset. It used mpg as the response, and considered many interaction
terms between the predictors disp, hp, and domestic.
str(autompg)

## $ mpg : num 18 15 18 16 17 15 14 14 14 15 ...

## $ disp : num 307 350 318 304 302 429 454 440 455 390 ...
## $ hp : num 130 165 150 150 140 198 220 215 225 190 ...
## $ wt : num 3504 3693 3436 3433 3449 ...
## $ acc : num 12 11.5 11 12 10.5 10 9 8.5 10 8.5 ...
## $ year : int 70 70 70 70 70 70 70 70 70 70 ...
## $ origin : int 1 1 1 1 1 1 1 1 1 1 ...
## $ domestic: num 1 1 1 1 1 1 1 1 1 1 ...
qqnorm(resid(big_model), col = "darkgrey")

qqline(resid(big_model), col = "dodgerblue", lwd = 2)
Normal Q-Q Plot

15
Sample Quantiles
10
5
0
-5
-10
-3 -2 -1 0 1 2 3
shapiro.test(resid(big_model))
##
##
## data: resid(big_model)
## W = 0.96161, p-value = 1.824e-08
Here both the Q-Q plot, and the Shapiro-Wilk test suggest that the normality
assumption is violated.
big_mod_cd = cooks.distance(big_model)
sum(big_mod_cd > 4 / length(big_mod_cd))
## [1] 31
Here, we find 31, so perhaps removing them will help!
big_model_fix = lm(mpg ~ disp * hp * domestic,

data = autompg,
subset = big_mod_cd < 4 / length(big_mod_cd))
qqnorm(resid(big_model_fix), col = "grey")
qqline(resid(big_model_fix), col = "dodgerblue", lwd = 2)
Normal Q-Q Plot

10
Sample Quantiles
5
0
-5
-3 -2 -1 0 1 2 3
shapiro.test(resid(big_model_fix))
##
##
## data: resid(big_model_fix)
## W = 0.99035, p-value = 0.02068
Removing these points results in a much better Q-Q plot, and now Shapiro-Wilk
fails to reject for a low 𝛼.
We’ve now seen that sometimes modifying the data can fix issues with regression.
However, next chapter, instead of modifying the data, we will modify the model
via transformations.
13.5 R Markdown
• diagnostics.Rmd

Chapter 14
Transformations
“Give me a lever long enough and a fulcrum on which to place it,

and I shall move the world.”
— Archimedes
Please note: some data currently used in this chapter was used, changed, and
passed around over the years in STAT 420 at UIUC. Its original sources, if they
exist, are at this time unknown to the author. As a result, they should only be
considered for use with STAT 420. Going forward they will likely be replaced
with alternative sourceable data that illustrates the same concepts. At the end
of this chapter you can find code seen in videos for Week 8 for STAT 420 in
the MCS-DS program. It is currently in the process of being merged into the
narrative of this chapter.
• Understand the concept of a variance stabilizing transformation.

• Use transformations of the response to improve regression models.
• Use polynomial terms as predictors to fit more flexible regression models.
Last chapter we checked the assumptions of regression models and looked at

ways to diagnose possible issues. This chapter we will use transformations of
both response and predictor variables in order to correct issues with model
diagnostics, and to also potentially simply make a model fit data better.
14.1 Response Transformation

Let’s look at some (fictional) salary data from the (fictional) company Initech.
We will try to model salary as a function of years of experience. The data
can be found in initech.csv.
303
304 CHAPTER 14. TRANSFORMATIONS
initech = read.csv("data/initech.csv")
plot(salary ~ years, data = initech, col = "grey", pch = 20, cex = 1.5,
main = "Salaries at Initech, By Seniority")
Salaries at Initech, By Seniority

250000
salary
150000
50000
5 10 15 20 25
years
We first fit a simple linear model.
initech_fit = lm(salary ~ years, data = initech)

summary(initech_fit)
##
## Call:
## lm(formula = salary ~ years, data = initech)
##
## Residuals:
## -57225 -18104 241 15589 91332
##
## Coefficients:
## (Intercept) 5302 5750 0.922 0.359
## years 8637 389 22.200 <2e-16 ***
## ---
14.1. RESPONSE TRANSFORMATION 305
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 27360 on 98 degrees of freedom
This model appears significant, but does it meet the model assumptions?
abline(initech_fit, col = "darkorange", lwd = 2)

250000
salary
150000
50000
5 10 15 20 25
years
Adding the fitted line to the plot, we see that the linear relationship appears
correct.
plot(fitted(initech_fit), resid(initech_fit), col = "grey", pch = 20,

xlab = "Fitted", ylab = "Residuals", main = "Fitted versus Residuals")
qqnorm(resid(initech_fit), main = "Normal Q-Q Plot", col = "darkgrey")

qqline(resid(initech_fit), col = "dodgerblue", lwd = 2)
Fitted versus Residuals Normal Q-Q Plot
50000
50000
Sample Quantiles
Residuals
0
-50000
-50000
50000 100000 150000 200000 -2 -1 0 1 2
Fitted Theoretical Quantiles
However, from the fitted versus residuals plot it appears there is non-constant
variance. Specifically, the variance increases as the fitted value increases.
14.1.1 Variance Stabilizing Transformations
Recall the fitted value is our estimate of the mean at a particular value of 𝑥.
Under our usual assumptions,
𝜖 ∼ 𝑁 (0, 𝜎2 )
and thus
Var[𝑌 |𝑋 = 𝑥] = 𝜎2
which is a constant value for any value of 𝑥.

However, here we see that the variance is a function of the mean,
Var[𝑌 ∣ 𝑋 = 𝑥] = ℎ(E[𝑌 ∣ 𝑋 = 𝑥]).
In this case, ℎ is some increasing function.

In order to correct for this, we would like to find some function of 𝑌 , 𝑔(𝑌 ) such
that,
Var[𝑔(𝑌 ) ∣ 𝑋 = 𝑥] = 𝑐
where 𝑐 is a constant that does not depend on the mean, E[𝑌 ∣ 𝑋 = 𝑥]. A
transformation that accomplishes this is called a variance stabilizing trans-
formation.
A common variance stabilizing transformation (VST) when we see increasing

variance in a fitted versus residuals plot is log(𝑌 ). Also, if the values of a variable
range over more than one order of magnitude and the variable is strictly positive,
then replacing the variable by its logarithm is likely to be helpful.
A reminder, that for our purposes, log and ln are both the natural log. R uses
log to mean the natural log, unless a different base is specified.
We will now use a model with a log transformed response for the Initech data,
log(𝑌𝑖 ) = 𝛽0 + 𝛽1 𝑥𝑖 + 𝜖𝑖 .
Note, if we re-scale the model from a log scale back to the original scale of the
data, we now have
𝑌𝑖 = exp(𝛽0 + 𝛽1 𝑥𝑖 ) ⋅ exp(𝜖𝑖 )
which has the errors entering the model in a multiplicative fashion.
Fitting this model in R requires only a minor modification to our formula spec-
ification.
initech_fit_log = lm(log(salary) ~ years, data = initech)
Note that while log(y) is considered the new response variable, we do not
actually create a new variable in R, but simply transform the variable inside the
model formula.
plot(log(salary) ~ years, data = initech, col = "grey", pch = 20, cex = 1.5,
abline(initech_fit_log, col = "darkorange", lwd = 2)
12.5
12.0
log(salary)
11.5
11.0
10.5
5 10 15 20 25
years
Plotting the data on the transformed log scale and adding the fitted line, the
relationship again appears linear, and we can already see that the variation
about the fitted line looks constant.
curve(exp(initech_fit_log$coef[1] + initech_fit_log$coef[2] * x),
from = 0, to = 30, add = TRUE, col = "darkorange", lwd = 2)

250000
salary
150000
50000
5 10 15 20 25
years
By plotting the data on the original scale, and adding the fitted regression, we
see an exponential relationship. However, this is still a linear model, since the
new transformed response, log(𝑦), is still a linear combination of the predictors.
plot(fitted(initech_fit_log), resid(initech_fit_log), col = "grey", pch = 20,

qqnorm(resid(initech_fit_log), main = "Normal Q-Q Plot", col = "darkgrey")

qqline(resid(initech_fit_log), col = "dodgerblue", lwd = 2)
0.4
0.4
0.2
0.2
Sample Quantiles
Residuals
0.0
0.0
-0.2
-0.2
-0.4
-0.4
-0.6
-0.6
10.5 11.0 11.5 12.0 12.5 -2 -1 0 1 2
The fitted versus residuals plot looks much better. It appears the constant
variance assumption is no longer violated.
Comparing the RMSE using the original and transformed response, we also
see that the log transformed model simply fits better, with a smaller average
squared error.
sqrt(mean(resid(initech_fit) ^ 2))
## [1] 27080.16
sqrt(mean(resid(initech_fit_log) ^ 2))
## [1] 0.1934907
But wait, that isn’t fair, this difference is simply due to the different scales being
used.
sqrt(mean((initech$salary - fitted(initech_fit)) ^ 2))
## [1] 27080.16
sqrt(mean((initech$salary - exp(fitted(initech_fit_log))) ^ 2))
## [1] 24280.36
Transforming the fitted values of the log model back to the data scale, we do
indeed see that it fits better!
summary(initech_fit_log)
##
## Call:
## lm(formula = log(salary) ~ years, data = initech)
##
## Residuals:
## -0.57022 -0.13560 0.03048 0.14157 0.41366
##
## Coefficients:
## (Intercept) 10.48381 0.04108 255.18 <2e-16 ***
## years 0.07888 0.00278 28.38 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
Again, the transformed response is a linear combination of the predictors,
log(𝑦(𝑥))
̂ = 𝛽0̂ + 𝛽1̂ 𝑥 = 10.484 + 0.079𝑥.
But now, if we re-scale the data from a log scale back to the original scale of
the data, we now have
𝑦(𝑥)
̂ = exp(𝛽0̂ ) exp(𝛽1̂ 𝑥) = exp(10.484) exp(0.079𝑥).
We see that for every one additional year of experience, average salary increases
exp(0.079) = 1.0822 times. We are now multiplying, not adding.
While using a log transform is possibly the most common response variable
transformation, many others exist. We will now consider a family of transfor-
mations and choose the best from among them, which includes the log transform.
14.1.2 Box-Cox Transformations
The Box-Cox method considers a family of transformations on strictly positive

response variables,
⎧ 𝑦𝜆 − 1
{ 𝜆≠0
𝑔𝜆 (𝑦) = 𝜆
⎨
{
⎩ log(𝑦) 𝜆=0
The 𝜆 parameter is chosen by numerically maximizing the log-likelihood,
𝑛
𝐿(𝜆) = − log(𝑅𝑆𝑆𝜆 /𝑛) + (𝜆 − 1) ∑ log(𝑦𝑖 ).
2
A 100(1 − 𝛼)% confidence interval for 𝜆 is,
1
{𝜆 ∶ 𝐿(𝜆) > 𝐿(𝜆)̂ − 𝜒21,𝛼 }
2
which R will plot for us to help quickly select an appropriate 𝜆 value. We often
choose a “nice” value from within the confidence interval, instead of the value
of 𝜆 that truly maximizes the likelihood.
library(MASS)
library(faraway)
Here we need the MASS package for the boxcox() function, and we will consider
a couple of datasets from the faraway package.
First we will use the savings dataset as an example of using the Box-Cox
method to justify the use of no transformation. We fit an additive multiple
regression model with sr as the response and each of the other variables as
predictors.
savings_model = lm(sr ~ ., data = savings)
We then use the boxcox() function to find the best transformation of the form
considered by the Box-Cox method.
boxcox(savings_model, plotit = TRUE)

-50
95%
-100
log-Likelihood
-150
-200
-2 -1 0 1 2
R automatically plots the log-Likelihood as a function of possible 𝜆 values. It in-

dicates both the value that maximizes the log-likelihood, as well as a confidence
interval for the 𝜆 value that maximizes the log-likelihood.
boxcox(savings_model, plotit = TRUE, lambda = seq(0.5, 1.5, by = 0.1))

-57.0
log-Likelihood
-58.0
95%
-59.0
-60.0
0.6 0.8 1.0 1.2 1.4
Note that we can specify a range of 𝜆 values to consider and thus be plotted. We
often specify a range that is more visually interesting. Here we see that 𝜆 = 1
is both in the confidence interval, and is extremely close to the maximum. This
suggests a transformation of the form
𝑦𝜆 − 1 𝑦1 − 1
= = 𝑦 − 1.
𝜆 1
This is essentially not a transformation. It would not change the variance or

make the model fit better. By subtracting 1 from every value, we would only
change the intercept of the model, and the resulting errors would be the same.
plot(fitted(savings_model), resid(savings_model), col = "dodgerblue",

pch = 20, cex = 1.5, xlab = "Fitted", ylab = "Residuals")
abline(h = 0, lty = 2, col = "darkorange", lwd = 2)
10
5
Residuals
0
-5
6 8 10 12 14 16
Fitted
Looking at a fitted versus residuals plot verifies that there likely are not any
issues with the assumptions of this model, which Breusch-Pagan and Shapiro-
Wilk tests verify.
library(lmtest)
bptest(savings_model)
##
##
## data: savings_model
## BP = 4.9852, df = 4, p-value = 0.2888
shapiro.test(resid(savings_model))
##
##
## data: resid(savings_model)
## W = 0.98698, p-value = 0.8524
Now we will use the gala dataset as an example of using the Box-Cox method
to justify a transformation other than log. We fit an additive multiple regres-
sion model with Species as the response and most of the other variables as
predictors.
gala_model = lm(Species ~ Area + Elevation + Nearest + Scruz + Adjacent, data = gala)
plot(fitted(gala_model), resid(gala_model), col = "dodgerblue",

100 150
Residuals
50
0
-100
0 100 200 300 400
Fitted
Even though there is not a lot of data for large fitted values, it still seems very
clear that the constant variance assumption is violated.
boxcox(gala_model, lambda = seq(-0.25, 0.75, by = 0.05), plotit = TRUE)

-58 -56 -54 -52 -50 -48 -46
95%
log-Likelihood
-0.2 0.0 0.2 0.4 0.6
Using the Box-Cox method, we see that 𝜆 = 0.3 is both in the confidence inter-
val, and is extremely close to the maximum, which suggests a transformation of
the form
𝑦𝜆 − 1 𝑦0.3 − 1
= .
𝜆 0.3
We then fit a model with this transformation applied to the response.
gala_model_cox = lm((((Species ^ 0.3) - 1) / 0.3) ~ Area + Elevation + Nearest + Scruz + Adjacent
plot(fitted(gala_model_cox), resid(gala_model_cox), col = "dodgerblue",

4
2
Residuals
0
-2
-4
5 10 15
Fitted
The resulting fitted versus residuals plot looks much better!
Lastly, we return to the initech data, and the initech_fit model we had
used earlier. Recall that this was the untransformed model, that we used a log
transform to fix.
boxcox(initech_fit)
14.2. PREDICTOR TRANSFORMATION 319
95%
-80
log-Likelihood
-120
-160
-200
-2 -1 0 1 2
Using the Box-Cox method, we see that 𝜆 = 0 is both in the interval, and
extremely close to the maximum, which suggests a transformation of the form
log(𝑦).
So the Box-Cox method justifies our previous choice of a log transform!
14.2 Predictor Transformation

In addition to transformation of the response variable, we can also consider
transformations of predictor variables. Sometimes these transformations can
help with violation of model assumptions, and other times they can be used to
simply fit a more flexible model.
str(autompg)

## $ mpg : num 18 15 18 16 17 15 14 14 14 15 ...
## $ disp : num 307 350 318 304 302 429 454 440 455 390 ...
## $ hp : num 130 165 150 150 140 198 220 215 225 190 ...
## $ wt : num 3504 3693 3436 3433 3449 ...
## $ acc : num 12 11.5 11 12 10.5 10 9 8.5 10 8.5 ...

## $ year : int 70 70 70 70 70 70 70 70 70 70 ...
## $ origin : int 1 1 1 1 1 1 1 1 1 1 ...
## $ domestic: num 1 1 1 1 1 1 1 1 1 1 ...
Recall the autompg dataset from the previous chapter. Here we will attempt to
model mpg as a function of hp.

mpg_hp = lm(mpg ~ hp, data = autompg)
abline(mpg_hp, col = "darkorange", lwd = 2)
plot(fitted(mpg_hp), resid(mpg_hp), col = "dodgerblue",
10 15
40
Residuals
5
30
mpg
0
20
-10
10
50 100 150 200 5 10 15 20 25 30
hp Fitted
We first attempt SLR, but we see a rather obvious pattern in the fitted versus
residuals plot, which includes increasing variance, so we attempt a log transform
of the response.

plot(log(mpg) ~ hp, data = autompg, col = "dodgerblue", pch = 20, cex = 1.5)
mpg_hp_log = lm(log(mpg) ~ hp, data = autompg)
abline(mpg_hp_log, col = "darkorange", lwd = 2)
plot(fitted(mpg_hp_log), resid(mpg_hp_log), col = "dodgerblue",
0.6
3.5
0.2
Residuals
log(mpg)
3.0
-0.2
2.5
-0.6
50 100 150 200 2.2 2.4 2.6 2.8 3.0 3.2 3.4
hp Fitted
After performing the log transform of the response, we still have some of the
same issues with the fitted versus response. Now, we will try also log transform-
ing the predictor.

plot(log(mpg) ~ log(hp), data = autompg, col = "dodgerblue", pch = 20, cex = 1.5)
mpg_hp_loglog = lm(log(mpg) ~ log(hp), data = autompg)
abline(mpg_hp_loglog, col = "darkorange", lwd = 2)
plot(fitted(mpg_hp_loglog), resid(mpg_hp_loglog), col = "dodgerblue",
0.6
3.5
0.2
Residuals
log(mpg)
3.0
-0.2
2.5
-0.6
4.0 4.5 5.0 5.5 2.4 2.6 2.8 3.0 3.2 3.4 3.6
log(hp) Fitted
Here, our fitted versus residuals plot looks good.

14.2.1 Polynomials
Another very common “transformation” of a predictor variable is the use of

polynomial transformations. They are extremely useful as they allow for more
flexible models, but do not change the units of the variables.
It should come as no surprise that sales of a product are related to the advertising
budget for the product, but there are diminishing returns. A company cannot
always expect linear returns based on an increased advertising budget.
Consider monthly data for the sales of Initech widgets, 𝑦, as a function of
Initech’s advertising expenditure for said widget, 𝑥, both in ten thousand dollars.
The data can be found in marketing.csv.
marketing = read.csv("data/marketing.csv")
plot(sales ~ advert, data = marketing,

xlab = "Advert Spending (in $10,000)", ylab = "Sales (in $10,000)",
pch = 20, cex = 2)
25
Sales (in $10,000)
20
15
10
2 4 6 8 10 12 14
Advert Spending (in $10,000)
We would like to fit the model,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝛽2 𝑥2𝑖 + 𝜖𝑖
where 𝜖𝑖 ∼ 𝑁 (0, 𝜎2 ) for 𝑖 = 1, 2, ⋯ 21.

The response 𝑦 is now a linear function of “two” variables which now allows
𝑦 to be a non-linear function of the original single predictor 𝑥. We consider
this a transformation, although we have actually in some sense added another
predictor.
Thus, our 𝑋 matrix is,
1 𝑥1 𝑥21
⎡ ⎤
⎢1 𝑥2 𝑥22 ⎥
⎢1 𝑥3 𝑥23 ⎥
⎢ ⎥
⎢⋮ ⋮ ⋮ ⎥
⎣1 𝑥𝑛 𝑥2𝑛 ⎦
We can then proceed to fit the model as we have in the past for multiple linear
regression.
−1
𝛽 ̂ = (𝑋 ⊤ 𝑋) 𝑋 ⊤ 𝑦.
Our estimates will have the usual properties. The mean is still
𝐸[𝛽]̂ = 𝛽,
and variance
−1
Var[𝛽]̂ = 𝜎2 (𝑋 ⊤ 𝑋) .
We also maintain the same distributional results
𝛽𝑗̂ ∼ 𝑁 (𝛽𝑗 , 𝜎2 𝐶𝑗𝑗 ) .
mark_mod = lm(sales ~ advert, data = marketing)

summary(mark_mod)
##
## Call:
## lm(formula = sales ~ advert, data = marketing)
##
## Residuals:
## -2.7845 -1.4762 -0.5103 1.2361 3.1869
##
## Coefficients:
## (Intercept) 9.4502 0.6806 13.88 2.13e-11 ***
## advert 1.1918 0.0937 12.72 9.65e-11 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
While the SLR model is significant, the fitted versus residuals plot would have
a very clear pattern.
mark_mod_poly2 = lm(sales ~ advert + I(advert ^ 2), data = marketing)

summary(mark_mod_poly2)
##
## Call:
## lm(formula = sales ~ advert + I(advert^2), data = marketing)
##
## Residuals:
## -1.9175 -0.8333 -0.1948 0.9292 2.1385
##
## Coefficients:
## (Intercept) 6.76161 0.67219 10.059 8.16e-09 ***
## advert 2.46231 0.24830 9.917 1.02e-08 ***
## I(advert^2) -0.08745 0.01658 -5.275 5.14e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## F-statistic: 209 on 2 and 18 DF, p-value: 3.486e-13
To add the second order term we need to use the I() function in the model
specification around our newly created predictor. We see that with the first
order term in the model, the quadratic term is also significant.
n = length(marketing$advert)
X = cbind(rep(1, n), marketing$advert, marketing$advert ^ 2)
t(X) %*% X
## [,1] [,2] [,3]

## [1,] 21.00 120.70 1107.95
## [2,] 120.70 1107.95 12385.86
## [3,] 1107.95 12385.86 151369.12
solve(t(X) %*% X) %*% t(X) %*% marketing$sales
## [,1]
## [1,] 6.76161045
## [2,] 2.46230964
## [3,] -0.08745394
Here we verify the parameter estimates were found as we would expect.

We could also add higher order terms, such as a third degree predictor. This is
easy to do. Our 𝑋 matrix simply becomes larger again.
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝛽2 𝑥2𝑖 + 𝛽3 𝑥3𝑖 + 𝜖𝑖
1 𝑥1 𝑥21 𝑥31
⎡ ⎤
⎢1 𝑥2 𝑥22 𝑥32 ⎥
⎢1 𝑥3 𝑥23 𝑥33 ⎥
⎢ ⎥
⎢⋮ ⋮ ⋮ ⋮ ⎥
⎣1 𝑥𝑛 𝑥2𝑛 𝑥3𝑛 ⎦
mark_mod_poly3 = lm(sales ~ advert + I(advert ^ 2) + I(advert ^ 3), data = marketing)

summary(mark_mod_poly3)
##
## Call:
## lm(formula = sales ~ advert + I(advert^2) + I(advert^3), data = marketing)
##
## Residuals:
## -1.44322 -0.61310 -0.01527 0.68131 1.22517
##
## Coefficients:
## (Intercept) 3.890070 0.761956 5.105 8.79e-05 ***
## advert 4.681864 0.501032 9.344 4.14e-08 ***
## I(advert^2) -0.455152 0.078977 -5.763 2.30e-05 ***
## I(advert^3) 0.016131 0.003429 4.704 0.000205 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
Now we see that with the first and second order terms in the model, the third
order term is also significant. But does this make sense practically? The fol-
lowing plot should gives hints as to why it doesn’t. (The model with the third
order term doesn’t have diminishing returns!)
plot(sales ~ advert, data = marketing,

xlab = "Advert Spending (in $10,000)", ylab = "Sales (in $10,000)",
pch = 20, cex = 2)
abline(mark_mod, lty = 2, col = "green", lwd = 2)
xplot = seq(0, 16, by = 0.01)
lines(xplot, predict(mark_mod_poly2, newdata = data.frame(advert = xplot)),
col = "blue", lwd = 2)
lines(xplot, predict(mark_mod_poly3, newdata = data.frame(advert = xplot)),
col = "red", lty = 3, lwd = 3)
25
Sales (in $10,000)
20
15
10
2 4 6 8 10 12 14
Advert Spending (in $10,000)
The previous plot was made using base graphics in R. The next plot was made
using the package ggplot2, an increasingly popular plotting method in R.
library(ggplot2)
ggplot(data = marketing, aes(x = advert, y = sales)) +
stat_smooth(method = "lm", se = FALSE, color = "green", formula = y ~ x) +
stat_smooth(method = "lm", se = FALSE, color = "blue", formula = y ~ x + I(x ^ 2)) +
stat_smooth(method = "lm", se = FALSE, color = "red", formula = y ~ x + I(x ^ 2)+ I(x ^ 3)) +
geom_point(colour = "black", size = 3)
25
20
sales
15
10
0 5 10 15
advert
Note we could fit a polynomial of an arbitrary order,
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝛽2 𝑥2𝑖 + ⋯ + 𝛽𝑝−1 𝑥𝑝−1

𝑖 + 𝜖𝑖
However, we should be careful about over-fitting, since with a polynomial of

degree one less than the number of observations, it is sometimes possible to fit
a model perfectly.
set.seed(1234)
x = seq(0, 10)
y = 3 + x + 4 * x ^ 2 + rnorm(11, 0, 20)
plot(x, y, ylim = c(-300, 400), cex = 2, pch = 20)
fit = lm(y ~ x + I(x ^ 2))
#summary(fit)
fit_perf = lm(y ~ x + I(x ^ 2) + I(x ^ 3) + I(x ^ 4) + I(x ^ 5) + I(x ^ 6)
+ I(x ^ 7) + I(x ^ 8) + I(x ^ 9) + I(x ^ 10))
summary(fit_perf)
##
## Call:
## lm(formula = y ~ x + I(x^2) + I(x^3) + I(x^4) + I(x^5) + I(x^6) +
## I(x^7) + I(x^8) + I(x^9) + I(x^10))
##
## Residuals:
## ALL 11 residuals are 0: no residual degrees of freedom!
##
## Coefficients:
## (Intercept) -2.114e+01 NaN NaN NaN
## x -1.918e+03 NaN NaN NaN
## I(x^2) 4.969e+03 NaN NaN NaN
## I(x^3) -4.932e+03 NaN NaN NaN
## I(x^4) 2.581e+03 NaN NaN NaN
## I(x^5) -8.035e+02 NaN NaN NaN
## I(x^6) 1.570e+02 NaN NaN NaN
## I(x^7) -1.947e+01 NaN NaN NaN
## I(x^8) 1.490e+00 NaN NaN NaN
## I(x^9) -6.424e-02 NaN NaN NaN
## I(x^10) 1.195e-03 NaN NaN NaN
##
## Residual standard error: NaN on 0 degrees of freedom
## Multiple R-squared: 1, Adjusted R-squared: NaN
## F-statistic: NaN on 10 and 0 DF, p-value: NA
xplot = seq(0, 10, by = 0.1)

lines(xplot, predict(fit, newdata = data.frame(x = xplot)),
col = "dodgerblue", lwd = 2, lty = 1)
lines(xplot, predict(fit_perf, newdata = data.frame(x = xplot)),
col = "darkorange", lwd = 2, lty = 2)
300
100
y
-100
-300
0 2 4 6 8 10
Notice in the summary, R could not calculate standard errors. This is a result
of being “out” of degrees of freedom. With 11 𝛽 parameters and 11 data points,
we use up all the degrees of freedom before we can estimate 𝜎.
In this example, the true relationship is quadratic, but the order 10 polynomial’s
fit is “perfect”. Next chapter we will focus on the trade-off between goodness of
fit (minimizing errors) and complexity of model.
Suppose you work for an automobile manufacturer which makes a large luxury
sedan. You would like to know how the car performs from a fuel efficiency
standpoint when it is driven at various speeds. Instead of testing the car at
every conceivable speed (which would be impossible) you create an experiment
where the car is driven at speeds of interest in increments of 5 miles per hour.
Our goal then, is to fit a model to this data in order to be able to predict fuel
efficiency when driving at certain speeds. The data from this example can be
found in fuel_econ.csv.
econ = read.csv("data/fuel_econ.csv")
In this example, we will be frequently looking at the fitted versus residuals plot,
so we should write a function to make our life easier, but this is left as an exercise
for homework.
We will also be adding fitted curves to scatterplots repeatedly, so smartly we
will write a function to do so.
plot_econ_curve = function(model) {
plot(mpg ~ mph, data = econ, xlab = "Speed (Miles per Hour)",
ylab = "Fuel Efficiency (Miles per Gallon)", col = "dodgerblue",
pch = 20, cex = 2)
xplot = seq(10, 75, by = 0.1)
lines(xplot, predict(model, newdata = data.frame(mph = xplot)),
}
So now we first fit a simple linear regression to this data.
fit1 = lm(mpg ~ mph, data = econ)

plot_econ_curve(fit1)
plot(fitted(fit1), resid(fit1), xlab = "Fitted", ylab = "Residuals",
col = "dodgerblue", pch = 20, cex = 2)
Fuel Efficiency (Miles per Gallon)
30
5
Residuals
25
0
20
-5
15
10 20 30 40 50 60 70 23.5 24.0 24.5 25.0 25.5
Speed (Miles per Hour) Fitted
Pretty clearly we can do better. Yes fuel efficiency does increase as speed in-
creases, but only up to a certain point.
We will now add polynomial terms until we fit a suitable fit.
fit2 = lm(mpg ~ mph + I(mph ^ 2), data = econ)

summary(fit2)
##
## Call:
## lm(formula = mpg ~ mph + I(mph^2), data = econ)
##
## Residuals:
## -2.8411 -0.9694 0.0017 1.0181 3.3900
##
## Coefficients:
## (Intercept) 2.4444505 1.4241091 1.716 0.0984 .
## mph 1.2716937 0.0757321 16.792 3.99e-15 ***
## I(mph^2) -0.0145014 0.0008719 -16.633 4.97e-15 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##

3
30
2
Residuals
1
25
0
-1
20
-2
15
-3
10 20 30 40 50 60 70 15 20 25 30
While this model clearly fits much better, and the second order term is sig-
nificant, we still see a pattern in the fitted versus residuals plot which suggests
higher order terms will help. Also, we would expect the curve to flatten as speed
increases or decreases, not go sharply downward as we see here.
fit3 = lm(mpg ~ mph + I(mph ^ 2) + I(mph ^ 3), data = econ)

summary(fit3)
##
## Call:
## lm(formula = mpg ~ mph + I(mph^2) + I(mph^3), data = econ)
##
## Residuals:
## -2.8112 -0.9677 0.0264 1.0345 3.3827
##
## Coefficients:
## (Intercept) 2.258e+00 2.768e+00 0.816 0.4227
## mph 1.291e+00 2.529e-01 5.103 3.2e-05 ***
## I(mph^2) -1.502e-02 6.604e-03 -2.274 0.0322 *
## I(mph^3) 4.066e-06 5.132e-05 0.079 0.9375
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##

3
30
2
Residuals
1
25
0
-1
20
-2
15
-3
10 20 30 40 50 60 70 15 20 25 30

Adding the third order term doesn’t seem to help at all. The fitted curve hardly
changes. This makes sense, since what we would like is for the curve to flatten
at the extremes. For this we will need an even degree polynomial term.
fit4 = lm(mpg ~ mph + I(mph ^ 2) + I(mph ^ 3) + I(mph ^ 4), data = econ)

summary(fit4)
##
## Call:
## lm(formula = mpg ~ mph + I(mph^2) + I(mph^3) + I(mph^4), data = econ)
##
## Residuals:
## -1.57410 -0.60308 0.04236 0.74481 1.93038
##
## Coefficients:
## (Intercept) 2.146e+01 2.965e+00 7.238 2.28e-07 ***
## mph -1.468e+00 3.913e-01 -3.751 0.00104 **
## I(mph^2) 1.081e-01 1.673e-02 6.463 1.35e-06 ***
## I(mph^3) -2.130e-03 2.844e-04 -7.488 1.31e-07 ***
## I(mph^4) 1.255e-05 1.665e-06 7.539 1.17e-07 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##

1.5
30
Residuals
0.5
25
-0.5
20
-1.5
15
10 20 30 40 50 60 70 15 20 25 30
Now we are making progress. The fourth order term is significant with the other
terms in the model. Also we are starting to see what we expected for low and
high speed. However, there still seems to be a bit of a pattern in the residuals,
so we will again try more higher order terms. We will add the fifth and sixth
together, since adding the fifth will be similar to adding the third.
fit6 = lm(mpg ~ mph + I(mph ^ 2) + I(mph ^ 3) + I(mph ^ 4) + I(mph ^ 5) + I(mph^6), dat

summary(fit6)
##
## Call:
## lm(formula = mpg ~ mph + I(mph^2) + I(mph^3) + I(mph^4) + I(mph^5) +
## I(mph^6), data = econ)
##
## Residuals:
## -1.1129 -0.5717 -0.1707 0.5026 1.5288
##
## Coefficients:
## (Intercept) -4.206e+00 1.204e+01 -0.349 0.7304
## mph 4.203e+00 2.553e+00 1.646 0.1146
## I(mph^2) -3.521e-01 2.012e-01 -1.750 0.0947 .
## I(mph^3) 1.579e-02 7.691e-03 2.053 0.0527 .
## I(mph^4) -3.473e-04 1.529e-04 -2.271 0.0338 *
## I(mph^5) 3.585e-06 1.518e-06 2.362 0.0279 *
## I(mph^6) -1.402e-08 5.941e-09 -2.360 0.0280 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## F-statistic: 186 on 6 and 21 DF, p-value: < 2.2e-16

0.0 0.5 1.0 1.5

30
Residuals
25
20
-1.0
15
10 20 30 40 50 60 70 15 20 25 30
Again the sixth order term is significant with the other terms in the model
and here we see less pattern in the residuals plot. Let’s now test which of the
previous two models we prefer. We will test
𝐻0 ∶ 𝛽5 = 𝛽6 = 0.
anova(fit4, fit6)

##
## Model 1: mpg ~ mph + I(mph^2) + I(mph^3) + I(mph^4)
## Model 2: mpg ~ mph + I(mph^2) + I(mph^3) + I(mph^4) + I(mph^5) + I(mph^6)
## 1 23 19.922
## 2 21 15.739 2 4.1828 2.7905 0.0842 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
So, this test does not reject the null hypothesis at a level of significance of
𝛼 = 0.05, however the p-value is still rather small, and the fitted versus residuals
plot is much better for the model with the sixth order term. This makes the
sixth order model a good choice. We could repeat this process one more time.
fit8 = lm(mpg ~ mph + I(mph ^ 2) + I(mph ^ 3) + I(mph ^ 4) + I(mph ^ 5)

+ I(mph ^ 6) + I(mph ^ 7) + I(mph ^ 8), data = econ)
summary(fit8)
##
## Call:
## I(mph^6) + I(mph^7) + I(mph^8), data = econ)
##
## Residuals:
## -1.21938 -0.50464 -0.09105 0.49029 1.45440
##
## Coefficients:
## (Intercept) -1.202e+01 7.045e+01 -0.171 0.866
## mph 6.021e+00 2.014e+01 0.299 0.768
## I(mph^2) -5.037e-01 2.313e+00 -0.218 0.830
## I(mph^3) 2.121e-02 1.408e-01 0.151 0.882
## I(mph^4) -4.008e-04 5.017e-03 -0.080 0.937
## I(mph^5) 1.789e-06 1.080e-04 0.017 0.987
## I(mph^6) 4.486e-08 1.381e-06 0.032 0.974
## I(mph^7) -6.456e-10 9.649e-09 -0.067 0.947
## I(mph^8) 2.530e-12 2.835e-11 0.089 0.930
##

0.0 0.5 1.0 1.5

30
Residuals
25
20
-1.0
15
10 20 30 40 50 60 70 15 20 25 30
summary(fit8)
##
## Call:
## I(mph^6) + I(mph^7) + I(mph^8), data = econ)
##
## Residuals:
## -1.21938 -0.50464 -0.09105 0.49029 1.45440
##
## Coefficients:
## (Intercept) -1.202e+01 7.045e+01 -0.171 0.866
## mph 6.021e+00 2.014e+01 0.299 0.768
## I(mph^2) -5.037e-01 2.313e+00 -0.218 0.830
## I(mph^3) 2.121e-02 1.408e-01 0.151 0.882
## I(mph^4) -4.008e-04 5.017e-03 -0.080 0.937
## I(mph^5) 1.789e-06 1.080e-04 0.017 0.987
## I(mph^6) 4.486e-08 1.381e-06 0.032 0.974
## I(mph^7) -6.456e-10 9.649e-09 -0.067 0.947
## I(mph^8) 2.530e-12 2.835e-11 0.089 0.930
##
anova(fit6, fit8)

##
## Model 1: mpg ~ mph + I(mph^2) + I(mph^3) + I(mph^4) + I(mph^5) + I(mph^6)
## Model 2: mpg ~ mph + I(mph^2) + I(mph^3) + I(mph^4) + I(mph^5) + I(mph^6) +
## I(mph^7) + I(mph^8)
## 1 21 15.739
## 2 19 15.506 2 0.2324 0.1424 0.8682
Here we would clearly stick with fit6. The eighth order term is not significant
with the other terms in the model and the F-test does not reject.
As an aside, be aware that there is a quicker way to specify a model with many
higher order terms.
fit6_alt = lm(mpg ~ poly(mph, 6), data = econ)

all.equal(fitted(fit6), fitted(fit6_alt))
## [1] TRUE
We first verify that this method produces the same fitted values. However, the
estimated coefficients are different.
coef(fit6)
## (Intercept) mph I(mph^2) I(mph^3) I(mph^4)

## -4.206224e+00 4.203382e+00 -3.521452e-01 1.579340e-02 -3.472665e-04
## I(mph^5) I(mph^6)
## 3.585201e-06 -1.401995e-08
coef(fit6_alt)
## (Intercept) poly(mph, 6)1 poly(mph, 6)2 poly(mph, 6)3 poly(mph, 6)4

## 24.40714286 4.16769628 -27.66685755 0.13446747 7.01671480
## poly(mph, 6)5 poly(mph, 6)6
## 0.09288754 -2.04307796
This is because poly() uses orthogonal polynomials, which solves an issue we

will discuss in the next chapter.
summary(fit6)
##
## Call:

## I(mph^6), data = econ)
##
## Residuals:
## -1.1129 -0.5717 -0.1707 0.5026 1.5288
##
## Coefficients:
## (Intercept) -4.206e+00 1.204e+01 -0.349 0.7304
## mph 4.203e+00 2.553e+00 1.646 0.1146
## I(mph^2) -3.521e-01 2.012e-01 -1.750 0.0947 .
## I(mph^3) 1.579e-02 7.691e-03 2.053 0.0527 .
## I(mph^4) -3.473e-04 1.529e-04 -2.271 0.0338 *
## I(mph^5) 3.585e-06 1.518e-06 2.362 0.0279 *
## I(mph^6) -1.402e-08 5.941e-09 -2.360 0.0280 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
summary(fit6_alt)
##
## Call:
## lm(formula = mpg ~ poly(mph, 6), data = econ)
##
## Residuals:
## -1.1129 -0.5717 -0.1707 0.5026 1.5288
##
## Coefficients:
## (Intercept) 24.40714 0.16360 149.184 < 2e-16 ***
## poly(mph, 6)1 4.16770 0.86571 4.814 9.31e-05 ***
## poly(mph, 6)2 -27.66686 0.86571 -31.958 < 2e-16 ***
## poly(mph, 6)3 0.13447 0.86571 0.155 0.878
## poly(mph, 6)4 7.01671 0.86571 8.105 6.68e-08 ***
## poly(mph, 6)5 0.09289 0.86571 0.107 0.916
## poly(mph, 6)6 -2.04308 0.86571 -2.360 0.028 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##

Notice though that the p-value for testing the degree 6 term is the same. Because
of this, for the most part we can use these interchangeably.
To use poly() to obtain the same results as using I() repeatedly, we would
need to set raw = TRUE.
fit6_alt2 = lm(mpg ~ poly(mph, 6, raw = TRUE), data = econ)

coef(fit6_alt2)
## (Intercept) poly(mph, 6, raw = TRUE)1 poly(mph, 6, raw = TRUE)2

## -4.206224e+00 4.203382e+00 -3.521452e-01
## poly(mph, 6, raw = TRUE)3 poly(mph, 6, raw = TRUE)4 poly(mph, 6, raw = TRUE)5
## 1.579340e-02 -3.472665e-04 3.585201e-06
## poly(mph, 6, raw = TRUE)6
## -1.401995e-08
We’ve now seen how to transform predictor and response variables. In this
chapter we have mostly focused on using this in the context of fixing SLR models.
However, these concepts can easily be used together with categorical variables
and interactions to build larger, more flexible models. In the next chapter, we
will discuss how to choose a good model from a collection of possible models.
Material below here is currently being merged into the content above.
Response Transformations
initech = read.csv("data/initech.csv")

250000
salary
150000
50000
5 10 15 20 25
years
initech_fit = lm(salary ~ years, data = initech)

summary(initech_fit)
##
## Call:
## lm(formula = salary ~ years, data = initech)
##
## Residuals:
## -57225 -18104 241 15589 91332
##
## Coefficients:
## (Intercept) 5302 5750 0.922 0.359
## years 8637 389 22.200 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 27360 on 98 degrees of freedom
abline(initech_fit, col = "darkorange", lwd = 2)

250000
salary
150000
50000
5 10 15 20 25
years
plot(fitted(initech_fit), resid(initech_fit), col = "grey", pch = 20,

qqnorm(resid(initech_fit), main = "Normal Q-Q Plot", col = "darkgrey")

qqline(resid(initech_fit), col = "dodgerblue", lwd = 2)

50000
50000
Sample Quantiles
Residuals
0
-50000
-50000
50000 100000 150000 200000 -2 -1 0 1 2
initech_fit_log = lm(log(salary) ~ years, data = initech)
log(𝑌𝑖 ) = 𝛽0 + 𝛽1 𝑥𝑖 + 𝜖𝑖
plot(log(salary) ~ years, data = initech, col = "grey", pch = 20, cex = 1.5,
abline(initech_fit_log, col = "darkorange", lwd = 2)

12.5
12.0
log(salary)
11.5
11.0
10.5
5 10 15 20 25
years
𝑌𝑖 = exp(𝛽0 + 𝛽1 𝑥𝑖 ) ⋅ exp(𝜖𝑖 )
curve(exp(initech_fit_log$coef[1] + initech_fit_log$coef[2] * x),
from = 0, to = 30, add = TRUE, col = "darkorange", lwd = 2)

250000
salary
150000
50000
5 10 15 20 25
years
plot(fitted(initech_fit_log), resid(initech_fit_log), col = "grey", pch = 20,

qqnorm(resid(initech_fit_log), main = "Normal Q-Q Plot", col = "darkgrey")

qqline(resid(initech_fit_log), col = "dodgerblue", lwd = 2)

0.4
0.4
0.2
0.2
Sample Quantiles
Residuals
0.0
0.0
-0.2
-0.2
-0.4
-0.4
-0.6
-0.6
10.5 11.0 11.5 12.0 12.5 -2 -1 0 1 2
sqrt(mean(resid(initech_fit) ^ 2))
## [1] 27080.16
sqrt(mean(resid(initech_fit_log) ^ 2))
## [1] 0.1934907
sqrt(mean((initech$salary - fitted(initech_fit)) ^ 2))
## [1] 27080.16
sqrt(mean((initech$salary - exp(fitted(initech_fit_log))) ^ 2))
## [1] 24280.36
Predictor Transformations
14.2.2 A Quadratic Model
sim_quad = function(sample_size = 500) {

y = 3 + 5 * x ^ 2 + rnorm(n = sample_size, mean = 0, sd = 5)
data.frame(x, y)
}
set.seed(314)
quad_data = sim_quad(sample_size = 200)
lin_fit = lm(y ~ x, data = quad_data)

summary(lin_fit)
##
## Call:
## lm(formula = y ~ x, data = quad_data)
##
## Residuals:
## -20.363 -7.550 -3.416 8.472 26.181
##
## Coefficients:
## (Intercept) -18.3271 1.5494 -11.83 <2e-16 ***
## x 24.8716 0.5343 46.55 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
plot(y ~ x, data = quad_data, col = "grey", pch = 20, cex = 1.5,

main = "Simulated Quadratic Data")
abline(lin_fit, col = "darkorange", lwd = 2)
Simulated Quadratic Data

120
80
60
y
40
20
0
0 1 2 3 4 5
plot(fitted(lin_fit), resid(lin_fit), col = "grey", pch = 20,

qqnorm(resid(lin_fit), main = "Normal Q-Q Plot", col = "darkgrey")

qqline(resid(lin_fit), col = "dodgerblue", lwd = 2)

20
20
Sample Quantiles
10
10
Residuals
0
-10
-10
-20
-20
-20 0 20 40 60 80 100 -3 -2 -1 0 1 2 3

𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝛽2 𝑥2𝑖 + 𝜖𝑖
quad_fit = lm(y ~ x + I(x^2), data = quad_data)

summary(quad_fit)
##
## Call:
## lm(formula = y ~ x + I(x^2), data = quad_data)
##
## Residuals:
## -11.4167 -3.0581 0.2297 3.1024 12.1256
##
## Coefficients:
## (Intercept) 3.0649 0.9577 3.200 0.0016 **
## x -0.5108 0.8637 -0.591 0.5549
## I(x^2) 5.0740 0.1667 30.433 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
plot(y ~ x, data = quad_data, col = "grey", pch = 20, cex = 1.5,

main = "Simulated Quadratic Data")
curve(quad_fit$coef[1] + quad_fit$coef[2] * x + quad_fit$coef[3] * x ^ 2,
from = -5, to = 30, add = TRUE, col = "darkorange", lwd = 2)
Simulated Quadratic Data

120
80
60
y
40
20
0
0 1 2 3 4 5
plot(fitted(quad_fit), resid(quad_fit), col = "grey", pch = 20,

qqnorm(resid(quad_fit), main = "Normal Q-Q Plot", col = "darkgrey")

qqline(resid(quad_fit), col = "dodgerblue", lwd = 2)

10
10
Sample Quantiles
5
5
Residuals
0
-5
-5
-10
-10
0 20 40 60 80 100 120 -3 -2 -1 0 1 2 3

14.2.3 Overfitting and Extrapolation
sim_for_perf = function() {
x = seq(0, 10)
y = 3 + x - 4 * x ^ 2 + rnorm(n = 11, mean = 0, sd = 25)
data.frame(x, y)
}
set.seed(1234)
data_for_perf = sim_for_perf()
fit_correct = lm(y ~ x + I(x ^ 2), data = data_for_perf)

fit_perfect = lm(y ~ x + I(x ^ 2) + I(x ^ 3) + I(x ^ 4) + I(x ^ 5) + I(x ^ 6) +
I(x ^ 7) + I(x ^ 8) + I(x ^ 9) + I(x ^ 10),
data = data_for_perf)
x_plot = seq(-5, 15, by = 0.1)

plot(y ~ x, data = data_for_perf, ylim = c(-450, 100), cex = 2, pch = 20)
lines(x_plot, predict(fit_correct, newdata = data.frame(x = x_plot)),
lines(x_plot, predict(fit_perfect, newdata = data.frame(x = x_plot)),
100
0
-100
y
-200
-300
-400
0 2 4 6 8 10
x
14.2.4 Comparing Polynomial Models
sim_higher = function(sample_size = 250) {

x = runif(n = sample_size, min = -1, max = 1) * 2
y = 3 + -6 * x ^ 2 + 1 * x ^ 4 + rnorm(n = sample_size, mean = 0, sd = 3)
data.frame(x, y)
}
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝛽2 𝑥2𝑖 + 𝜖𝑖
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝛽2 𝑥2𝑖 + 𝛽3 𝑥3𝑖 + 𝛽4 𝑥4𝑖 + 𝜖𝑖
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝛽2 𝑥2𝑖 + 𝛽3 𝑥3𝑖 + 𝛽4 𝑥4𝑖 + 𝛽5 𝑥5𝑖 + 𝛽6 𝑥6𝑖 + 𝜖𝑖
set.seed(42)
data_higher = sim_higher()
plot(y ~ x, data = data_higher, col = "grey", pch = 20, cex = 1.5,

main = "Simulated Quartic Data")
Simulated Quartic Data

10
5
0
y
-5
-10
-2 -1 0 1 2
x
fit_2 = lm(y ~ poly(x, 2), data = data_higher)

plot(y ~ x, data = data_higher, col = "grey", pch = 20, cex = 1.5,

main = "Simulated Quartic Data")
x_plot = seq(-5, 5, by = 0.05)
lines(x_plot, predict(fit_2, newdata = data.frame(x = x_plot)),
lines(x_plot, predict(fit_4, newdata = data.frame(x = x_plot)),
Simulated Quartic Data

10
5
0
y
-5
-10
-2 -1 0 1 2

qqnorm(resid(fit_2), main = "Normal Q-Q Plot", col = "darkgrey")


5
5
Sample Quantiles
Residuals
0
-5
-5
-8 -6 -4 -2 0 -3 -2 -1 0 1 2 3

qqnorm(resid(fit_4), main = "Normal Q-Q Plot", col = "darkgrey")


5
5
Sample Quantiles
Residuals
0
-5
-5
-6 -4 -2 0 2 -3 -2 -1 0 1 2 3
anova(fit_2, fit_4)

##
## Model 1: y ~ poly(x, 2)
## 1 247 2334.1
## 2 245 1912.6 2 421.51 26.997 2.536e-11 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
anova(fit_4, fit_6)

##
## 1 245 1912.6
## 2 243 1904.4 2 8.1889 0.5224 0.5937
14.2.5 poly() Function and Orthogonal Polynomials
𝑌𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝛽2 𝑥2𝑖 + 𝛽3 𝑥3𝑖 + 𝛽4 𝑥4𝑖 + 𝜖𝑖
fit_4a = lm(y ~ poly(x, degree = 4), data = data_higher)

fit_4b = lm(y ~ poly(x, degree = 4, raw = TRUE), data = data_higher)
fit_4c = lm(y ~ x + I(x^2) + I(x^3) + I(x^4), data = data_higher)
coef(fit_4a)
## (Intercept) poly(x, degree = 4)1 poly(x, degree = 4)2

## -1.980036 -2.053929 -49.344752
## poly(x, degree = 4)3 poly(x, degree = 4)4
## 0.669874 20.519759
coef(fit_4b)
## (Intercept) poly(x, degree = 4, raw = TRUE)1

## 2.9996256 -0.3880250
## poly(x, degree = 4, raw = TRUE)2 poly(x, degree = 4, raw = TRUE)3
## -6.1511166 0.1269046
## poly(x, degree = 4, raw = TRUE)4
## 1.0282139
coef(fit_4c)
## (Intercept) x I(x^2) I(x^3) I(x^4)

## 2.9996256 -0.3880250 -6.1511166 0.1269046 1.0282139
unname(coef(fit_4a))
## [1] -1.980036 -2.053929 -49.344752 0.669874 20.519759
unname(coef(fit_4b))
## [1] 2.9996256 -0.3880250 -6.1511166 0.1269046 1.0282139
unname(coef(fit_4c))
## [1] 2.9996256 -0.3880250 -6.1511166 0.1269046 1.0282139
all.equal(fitted(fit_4a),
fitted(fit_4b))
## [1] TRUE
all.equal(resid(fit_4a),
resid(fit_4b))
## [1] TRUE
summary(fit_4a)
##
## Call:
## lm(formula = y ~ poly(x, degree = 4), data = data_higher)
##
## Residuals:
## -7.6982 -2.0334 0.0042 1.9532 7.4626
##
## Coefficients:
## (Intercept) -1.9800 0.1767 -11.205 < 2e-16 ***
## poly(x, degree = 4)1 -2.0539 2.7940 -0.735 0.463

## poly(x, degree = 4)2 -49.3448 2.7940 -17.661 < 2e-16 ***
## poly(x, degree = 4)3 0.6699 2.7940 0.240 0.811
## poly(x, degree = 4)4 20.5198 2.7940 7.344 3.06e-12 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
summary(fit_4c)
##
## Call:
## lm(formula = y ~ x + I(x^2) + I(x^3) + I(x^4), data = data_higher)
##
## Residuals:
## -7.6982 -2.0334 0.0042 1.9532 7.4626
##
## Coefficients:
## (Intercept) 2.9996 0.3315 9.048 < 2e-16 ***
## x -0.3880 0.3828 -1.014 0.312
## I(x^2) -6.1511 0.5049 -12.183 < 2e-16 ***
## I(x^3) 0.1269 0.1456 0.871 0.384
## I(x^4) 1.0282 0.1400 7.344 3.06e-12 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
14.2.6 Inhibit Function
coef(lm(y ~ x + x ^ 2, data = quad_data))
## (Intercept) x
## -18.32715 24.87163
coef(lm(y ~ x + I(x ^ 2), data = quad_data))
## (Intercept) x I(x^2)
## 3.0649446 -0.5108131 5.0739805
coef(lm(y ~ x + x:x, data = quad_data))
## (Intercept) x
## -18.32715 24.87163
coef(lm(y ~ x * x, data = quad_data))
## (Intercept) x
## -18.32715 24.87163
coef(lm(y ~ x ^ 2, data = quad_data))
## (Intercept) x
## -18.32715 24.87163
coef(lm(y ~ x + x ^ 2, data = quad_data))
## (Intercept) x
## -18.32715 24.87163
coef(lm(y ~ I(x + x), data = quad_data))
## (Intercept) I(x + x)
## -18.32715 12.43582
coef(lm(y ~ x + x, data = quad_data))
## (Intercept) x
## -18.32715 24.87163
14.2.7 Data Example

pairs(autompg)
1.0 2.0 3.0 50 150 10 15 20 25 1.0 2.0 3.0
mpg
30
10
3.0
cyl
2.0
1.0
300
disp
100
150
hp
50
1500 3000 4500

wt
25
20
acc
15
10
82
78
year
74
70
3.0
origin
2.0
1.0
domestic 0.8
0.4
0.0
10 30 100 300 1500 3000 4500 70 74 78 82 0.0 0.4 0.8
mpg_hp = lm(mpg ~ hp, data = autompg)
abline(mpg_hp, col = "darkorange", lwd = 2)
plot(fitted(mpg_hp), resid(mpg_hp), col = "dodgerblue",

15
40
10
5
Residuals
30
mpg
0
20
-5
-10
10
50 100 150 200 5 10 15 20 25 30
hp Fitted
mpg_hp_log = lm(mpg ~ hp + I(hp ^ 2), data = autompg)
xplot = seq(min(autompg$hp), max(autompg$hp), by = 0.1)
lines(xplot, predict(mpg_hp_log, newdata = data.frame(hp = xplot)),

15
40
10
5
Residuals
30
mpg
0
-5
20
-10
10
-15
50 100 150 200 15 20 25 30 35
hp Fitted
mpg_hp_log = lm(log(mpg) ~ hp + I(hp ^ 2), data = autompg)
plot(log(mpg) ~ hp, data = autompg, col = "dodgerblue", pch = 20, cex = 1.5)
xplot = seq(min(autompg$hp), max(autompg$hp), by = 0.1)
lines(xplot, predict(mpg_hp_log, newdata = data.frame(hp = xplot)),

0.6
0.4
3.5
0.2
Residuals
log(mpg)
0.0
3.0
-0.6 -0.4 -0.2

2.5
50 100 150 200 2.6 2.8 3.0 3.2 3.4 3.6
hp Fitted
mpg_hp_loglog = lm(log(mpg) ~ log(hp), data = autompg)

plot(log(mpg) ~ log(hp), data = autompg, col = "dodgerblue", pch = 20, cex = 1.5)
abline(mpg_hp_loglog, col = "darkorange", lwd = 2)
plot(fitted(mpg_hp_loglog), resid(mpg_hp_loglog), col = "dodgerblue",

0.6
0.4
3.5
0.2
Residuals
log(mpg)
0.0
3.0
-0.2
-0.4
2.5
-0.6
4.0 4.5 5.0 5.5 2.4 2.6 2.8 3.0 3.2 3.4 3.6
log(hp) Fitted
qqnorm(resid(big_model), col = "darkgrey")

qqline(resid(big_model), col = "dodgerblue", lwd = 2)
Normal Q-Q Plot

15
Sample Quantiles
10
5
0
-5
-10
-3 -2 -1 0 1 2 3
bigger_model = lm(log(mpg) ~ disp * hp * domestic +

I(disp ^ 2) + I(hp ^ 2), data = autompg)
summary(bigger_model)
##
## Call:
## lm(formula = log(mpg) ~ disp * hp * domestic + I(disp^2) + I(hp^2),
## data = autompg)
##
## Residuals:
## -0.40381 -0.08635 -0.01040 0.09995 0.71365
##
## Coefficients:
## (Intercept) 4.276e+00 2.564e-01 16.674 <2e-16 ***
## disp -5.289e-03 2.565e-03 -2.062 0.0399 *
## hp -7.386e-03 3.309e-03 -2.232 0.0262 *
## domestic -2.496e-01 2.787e-01 -0.896 0.3710
## I(disp^2) 8.552e-06 4.141e-06 2.065 0.0396 *
## I(hp^2) -1.565e-05 1.679e-05 -0.932 0.3519
## disp:hp 2.685e-05 3.082e-05 0.871 0.3842
## disp:domestic -1.101e-03 2.526e-03 -0.436 0.6631
## hp:domestic 7.560e-03 3.689e-03 2.049 0.0411 *
## disp:hp:domestic -2.311e-05 2.662e-05 -0.868 0.3859
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
qqnorm(resid(bigger_model), col = "darkgrey")

qqline(resid(bigger_model), col = "dodgerblue", lwd = 2)
Normal Q-Q Plot

0.6
Sample Quantiles
0.4
0.2
0.0
-0.4 -0.2
-3 -2 -1 0 1 2 3
14.3 R Markdown
• transformations.Rmd

Chapter 15
Collinearity
“If I look confused it is because I am thinking.”

— Samuel Goldwyn
• Identify collinearity in regression.

• Understand the effect of collinearity on regression models.
15.1 Exact Collinearity

Let’s create a dataset where one of the predictors, 𝑥3 , is a linear combination
of the other predictors.
gen_exact_collin_data = function(num_samples = 100) {

x1 = rnorm(n = num_samples, mean = 80, sd = 10)
x2 = rnorm(n = num_samples, mean = 70, sd = 5)
x3 = 2 * x1 + 4 * x2 + 3
y = 3 + x1 + x2 + rnorm(n = num_samples, mean = 0, sd = 1)
data.frame(y, x1, x2, x3)
}
Notice that the way we are generating this data, the response 𝑦 only really
depends on 𝑥1 and 𝑥2 .
set.seed(42)
exact_collin_data = gen_exact_collin_data()
head(exact_collin_data)
365
366 CHAPTER 15. COLLINEARITY
## y x1 x2 x3
## 1 170.7135 93.70958 76.00483 494.4385
## 2 152.9106 74.35302 75.22376 452.6011
## 3 152.7866 83.63128 64.98396 430.1984
## 4 170.6306 86.32863 79.24241 492.6269
## 5 152.3320 84.04268 66.66613 437.7499
## 6 151.3155 78.93875 70.52757 442.9878
What happens when we attempt to fit a regression model in R using all of the
predictors?
exact_collin_fit = lm(y ~ x1 + x2 + x3, data = exact_collin_data)

summary(exact_collin_fit)
##
## Call:
## lm(formula = y ~ x1 + x2 + x3, data = exact_collin_data)
##
## Residuals:
## -2.57662 -0.66188 -0.08253 0.63706 2.52057
##
## Coefficients: (1 not defined because of singularities)
## (Intercept) 2.957336 1.735165 1.704 0.0915 .
## x1 0.985629 0.009788 100.702 <2e-16 ***
## x2 1.017059 0.022545 45.112 <2e-16 ***
## x3 NA NA NA NA
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
We see that R simply decides to exclude a variable. Why is this happening?
X = cbind(1, as.matrix(exact_collin_data[,-1]))
solve(t(X) %*% X)
𝑇 −1
If we attempt to find 𝛽 ̂ using (𝑋 𝑋) , we see that this is not possible, due
to the fact that the columns of 𝑋 are linearly dependent. The previous lines of
code were not run, because they produce an error!
When this happens, we say there is exact collinearity in the dataset.
15.1. EXACT COLLINEARITY 367
As a result of this issue, R essentially chose to fit the model y ~ x1 + x2.

However notice that two other models would accomplish exactly the same fit.
fit1 = lm(y ~ x1 + x2, data = exact_collin_data)

We see that the fitted values for each of the three models are exactly the same.
This is a result of 𝑥3 containing all of the information from 𝑥1 and 𝑥2 . As
long as one of 𝑥1 or 𝑥2 is included in the model, 𝑥3 can be used to recover the
information from the variable not included.
all.equal(fitted(fit1), fitted(fit2))
## [1] TRUE
all.equal(fitted(fit2), fitted(fit3))
## [1] TRUE
While their fitted values are all the same, their estimated coefficients are wildly
different. The sign of 𝑥2 is switched in two of the models! So only fit1 properly
explains the relationship between the variables, fit2 and fit3 still predict as
well as fit1, despite the coefficients having little to no meaning, a concept we
will return to later.
coef(fit1)
## 2.9573357 0.9856291 1.0170586
coef(fit2)
## 2.1945418 0.4770998 0.2542647
coef(fit3)
## 1.4788921 -0.9541995 0.4928145
15.2 Collinearity
Exact collinearity is an extreme example of collinearity, which occurs in mul-

tiple regression when predictor variables are highly correlated. Collinearity is
often called multicollinearity, since it is a phenomenon that really only occurs
during multiple regression.
Looking at the seatpos dataset from the faraway package, we will see an
example of this concept. The predictors in this dataset are various attributes
of car drivers, such as their height, weight and age. The response variable
hipcenter measures the “horizontal distance of the midpoint of the hips from a
fixed location in the car in mm.” Essentially, it measures the position of the seat
for a given driver. This is potentially useful information for car manufacturers
considering comfort and safety when designing vehicles.
We will attempt to fit a model that predicts hipcenter. Two predictor variables
are immediately interesting to us: HtShoes and Ht. We certainly expect a
person’s height to be highly correlated to their height when wearing shoes.
We’ll pay special attention to these two variables when fitting models.
library(faraway)
pairs(seatpos, col = "dodgerblue")
15.2. COLLINEARITY 369
100 200 300 150 180 26 32 38 30 36 42
20 40 60
Age
300
200
Weight
100
190
HtShoes
160
180
Ht
150
100
Seated
90
80
38
Arm
32
26
45
Thigh
35
42
Leg
36
30
-100
hipcenter
-250
20 40 60 160 190 80 90 100 35 45 -250 -100
round(cor(seatpos), 2)
## Age Weight HtShoes Ht Seated Arm Thigh Leg hipcenter

## Age 1.00 0.08 -0.08 -0.09 -0.17 0.36 0.09 -0.04 0.21
## Weight 0.08 1.00 0.83 0.83 0.78 0.70 0.57 0.78 -0.64
## HtShoes -0.08 0.83 1.00 1.00 0.93 0.75 0.72 0.91 -0.80
## Ht -0.09 0.83 1.00 1.00 0.93 0.75 0.73 0.91 -0.80
## Seated -0.17 0.78 0.93 0.93 1.00 0.63 0.61 0.81 -0.73
## Arm 0.36 0.70 0.75 0.75 0.63 1.00 0.67 0.75 -0.59
## Thigh 0.09 0.57 0.72 0.73 0.61 0.67 1.00 0.65 -0.59
## Leg -0.04 0.78 0.91 0.91 0.81 0.75 0.65 1.00 -0.79
## hipcenter 0.21 -0.64 -0.80 -0.80 -0.73 -0.59 -0.59 -0.79 1.00
After loading the faraway package, we do some quick checks of correlation

between the predictors. Visually, we can do this with the pairs() function,
which plots all possible scatterplots between pairs of variables in the dataset.
We can also do this numerically with the cor() function, which when applied to
a dataset, returns all pairwise correlations. Notice this is a symmetric matrix.
Recall that correlation measures strength and direction of the linear relationship
between two variables. The correlation between Ht and HtShoes is extremely
high. So high, that rounded to two decimal places, it appears to be 1!
Unlike exact collinearity, here we can still fit a model with all of the predictors,
but what effect does this have?
hip_model = lm(hipcenter ~ ., data = seatpos)

summary(hip_model)
##
## Call:
## lm(formula = hipcenter ~ ., data = seatpos)
##
## Residuals:
## -73.827 -22.833 -3.678 25.017 62.337
##
## Coefficients:
## (Intercept) 436.43213 166.57162 2.620 0.0138 *
## Age 0.77572 0.57033 1.360 0.1843
## Weight 0.02631 0.33097 0.080 0.9372
## HtShoes -2.69241 9.75304 -0.276 0.7845
## Ht 0.60134 10.12987 0.059 0.9531
## Seated 0.53375 3.76189 0.142 0.8882
## Arm -1.32807 3.90020 -0.341 0.7359
## Thigh -1.14312 2.66002 -0.430 0.6706
## Leg -6.43905 4.71386 -1.366 0.1824
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
One of the first things we should notice is that the 𝐹 -test for the regression tells
us that the regression is significant, however each individual predictor is not.
Another interesting result is the opposite signs of the coefficients for Ht and
HtShoes. This should seem rather counter-intuitive. Increasing Ht increases
hipcenter, but increasing HtShoes decreases hipcenter?
This happens as a result of the predictors being highly correlated. For example,
the HtShoe variable explains a large amount of the variation in Ht. When they
are both in the model, their effects on the response are lessened individually,
but together they still explain a large portion of the variation of hipcenter.
We define 𝑅𝑗2 to be the proportion of observed variation in the 𝑗-th predictor
explained by the other predictors. In other words 𝑅𝑗2 is the multiple R-Squared
for the regression of 𝑥𝑗 on each of the other predictors.
ht_shoes_model = lm(HtShoes ~ . - hipcenter, data = seatpos)

summary(ht_shoes_model)$r.squared
## [1] 0.9967472
Here we see that the other predictors explain 99.67% of the variation in HtShoe.
When fitting this model, we removed hipcenter since it is not a predictor.
15.2.1 Variance Inflation Factor.
Now note that the variance of 𝛽𝑗̂ can be written as
1 1
Var(𝛽𝑗̂ ) = 𝜎2 𝐶𝑗𝑗 = 𝜎2 ( )
1 − 𝑅𝑗2 𝑆𝑥𝑗 𝑥𝑗
where
𝑆𝑥𝑗 𝑥𝑗 = ∑(𝑥𝑖𝑗 − 𝑥𝑗̄ )2 .
This gives us a way to understand how collinearity affects our regression esti-
mates.
We will call,
1
1 − 𝑅𝑗2
the variance inflation factor. The variance inflation factor quantifies the
effect of collinearity on the variance of our regression estimates. When 𝑅𝑗2 is
large, that is close to 1, 𝑥𝑗 is well explained by the other predictors. With a
large 𝑅𝑗2 the variance inflation factor becomes large. This tells us that when 𝑥𝑗
is highly correlated with other predictors, our estimate of 𝛽𝑗 is highly variable.
The vif function from the faraway package calculates the VIFs for each of the
predictors of a model.
vif(hip_model)
## Age Weight HtShoes Ht Seated Arm Thigh

## 1.997931 3.647030 307.429378 333.137832 8.951054 4.496368 2.762886
## Leg
## 6.694291
In practice it is common to say that any VIF greater than 5 is cause for concern.
So in this example we see there is a huge multicollinearity issue as many of the
predictors have a VIF greater than 5.
Let’s further investigate how the presence of collinearity actually affects a model.
If we add a moderate amount of noise to the data, we see that the estimates of
the coefficients change drastically. This is a rather undesirable effect. Adding
random noise should not affect the coefficients of a model.
set.seed(1337)
noise = rnorm(n = nrow(seatpos), mean = 0, sd = 5)
hip_model_noise = lm(hipcenter + noise ~ ., data = seatpos)
Adding the noise had such a large effect, the sign of the coefficient for Ht has
changed.
coef(hip_model)
## (Intercept) Age Weight HtShoes Ht Seated

## 436.43212823 0.77571620 0.02631308 -2.69240774 0.60134458 0.53375170
## Arm Thigh Leg
## -1.32806864 -1.14311888 -6.43904627
coef(hip_model_noise)

## 415.32909380 0.76578240 0.01910958 -2.90377584 -0.12068122 2.03241638
## Arm Thigh Leg
## -1.02127944 -0.89034509 -5.61777220
This tells us that a model with collinearity is bad at explaining the relationship
between the response and the predictors. We cannot even be confident in the
direction of the relationship. However, does collinearity affect prediction?
plot(fitted(hip_model), fitted(hip_model_noise), col = "dodgerblue", pch = 20,

xlab = "Predicted, Without Noise", ylab = "Predicted, With Noise", cex = 1.5)
abline(a = 0, b = 1, col = "darkorange", lwd = 2)
-100
Predicted, With Noise
-150
-200
-250
-250 -200 -150 -100
Predicted, Without Noise
We see that by plotting the predicted values using both models against each
other, they are actually rather similar.
Let’s now look at a smaller model,
hip_model_small = lm(hipcenter ~ Age + Arm + Ht, data = seatpos)

summary(hip_model_small)
##
## Call:
## lm(formula = hipcenter ~ Age + Arm + Ht, data = seatpos)
##
## Residuals:
## -82.347 -24.745 -0.094 23.555 58.314
##
## Coefficients:
## (Intercept) 493.2491 101.0724 4.880 2.46e-05 ***
## Age 0.7988 0.5111 1.563 0.12735

## Arm -2.9385 3.5210 -0.835 0.40979
## Ht -3.4991 0.9954 -3.515 0.00127 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
vif(hip_model_small)
## Age Arm Ht
## 1.749943 3.996766 3.508693
Immediately we see that multicollinearity isn’t an issue here.
anova(hip_model_small, hip_model)

##
## Model 1: hipcenter ~ Age + Arm + Ht
## Model 2: hipcenter ~ Age + Weight + HtShoes + Ht + Seated + Arm + Thigh +
## Leg
## 1 34 44354
## 2 29 41262 5 3091.9 0.4346 0.8207
Also notice that using an 𝐹 -test to compare the two models, we would prefer
the smaller model.
We now investigate the effect of adding another variable to this smaller model.
Specifically we want to look at adding the variable HtShoes. So now our possible
predictors are HtShoes, Age, Arm, and Ht. Our response is still hipcenter.
To quantify this effect we will look at a variable added plot and a partial
correlation coefficient. For both of these, we will look at the residuals of two
models:
• Regressing the response (hipcenter) against all of the predictors except

the predictor of interest (HtShoes).
• Regressing the predictor of interest (HtShoes) against the other predictors
(Age, Arm, and Ht).
ht_shoes_model_small = lm(HtShoes ~ Age + Arm + Ht, data = seatpos)
So now, the residuals of hip_model_small give us the variation of hipcenter

that is unexplained by Age, Arm, and Ht. Similarly, the residuals of
ht_shoes_model_small give us the variation of HtShoes unexplained by Age,
Arm, and Ht.
The correlation of these two residuals gives us the partial correlation coeffi-
cient of HtShoes and hipcenter with the effects of Age, Arm, and Ht removed.
cor(resid(ht_shoes_model_small), resid(hip_model_small))
## [1] -0.03311061
Since this value is small, close to zero, it means that the variation of hipcenter
that is unexplained by Age, Arm, and Ht shows very little correlation with the
variation of HtShoes that is not explained by Age, Arm, and Ht. Thus adding
HtShoes to the model would likely be of little benefit.
Similarly a variable added plot visualizes these residuals against each other.
It is also helpful to regress the residuals of the response against the residuals of
the predictor and add the regression line to the plot.
plot(resid(hip_model_small) ~ resid(ht_shoes_model_small),
col = "dodgerblue", pch = 20,
xlab = "Residuals, Added Predictor",
ylab = "Residuals, Original Model",
main = "Variable Added Plot")
abline(h = 0, lty = 2)
abline(v = 0, lty = 2)
abline(lm(resid(hip_model_small) ~ resid(ht_shoes_model_small)),
col = "darkorange", lwd = 2)
Variable Added Plot
20 40 60
Residuals, Original Model
0
-40
-80
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
Residuals, Added Predictor
Here the variable added plot shows almost no linear relationship. This tells us
that adding HtShoes to the model would probably not be worthwhile. Since its
variation is largely explained by the other predictors, adding it to the model
will not do much to improve the model. However it will increase the variation
of the estimates and make the model much harder to interpret.
Had there been a strong linear relationship here, thus a large partial correlation
coefficient, it would likely have been useful to add the additional predictor to
the model.
This trade-off is mostly true in general. As a model gets more predictors, errors
will get smaller and its prediction will be better, but it will be harder to interpret.
This is why, if we are interested in explaining the relationship between the
predictors and the response, we often want a model that fits well, but with a
small number of predictors with little correlation.
Next chapter we will learn about methods to find models that both fit well,
but also have a small number of predictors. We will also discuss overfitting. Al-
though, adding additional predictors will always make errors smaller, sometimes
we will be “fitting the noise” and such a model will not generalize to additional
observations well.
15.3. SIMULATION 377
15.3 Simulation
Here we simulate example data with and without collinearity. We will note the
difference in the distribution of the estimates of the 𝛽 parameters, in particular
their variance. However, we will also notice the similarity in their 𝑀 𝑆𝐸.
We will use the model,
𝑌 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝜖
where 𝜖 ∼ 𝑁 (𝜇 = 0, 𝜎2 = 25) and the 𝛽 coefficients defined below.
set.seed(42)
beta_0 = 7
beta_1 = 3
beta_2 = 4
sigma = 5
We will use a sample size of 10, and 2500 simulations for both situations.
sample_size = 10
num_sim = 2500
We’ll first consider the situation with a collinearity issue, so we manually create
the two predictor variables.
x1 = c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
x2 = c(1, 2, 3, 4, 5, 7, 6, 10, 9, 8)
c(sd(x1), sd(x2))
## [1] 3.02765 3.02765
cor(x1, x2)
## [1] 0.9393939
Notice that they have extremely high correlation.
true_line_bad = beta_0 + beta_1 * x1 + beta_2 * x2

beta_hat_bad = matrix(0, num_sim, 2)
mse_bad = rep(0, num_sim)
We perform the simulation 2500 times, each time fitting a regression model, and
storing the estimated coefficients and the MSE.
for (s in 1:num_sim) {
y = true_line_bad + rnorm(n = sample_size, mean = 0, sd = sigma)
reg_out = lm(y ~ x1 + x2)
beta_hat_bad[s, ] = coef(reg_out)[-1]
mse_bad[s] = mean(resid(reg_out) ^ 2)
}
Now we move to the situation without a collinearity issue, so we again manually

create the two predictor variables.
z1 = c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
z2 = c(9, 2, 7, 4, 5, 6, 3, 8, 1, 10)
Notice that the standard deviations of each are the same as before, however,
now the correlation is extremely close to 0.
c(sd(z1), sd(z2))
## [1] 3.02765 3.02765
cor(z1, z2)
## [1] 0.03030303
true_line_good = beta_0 + beta_1 * z1 + beta_2 * z2

beta_hat_good = matrix(0, num_sim, 2)
mse_good = rep(0, num_sim)
We then perform simulations and store the same results.
for (s in 1:num_sim) {
y = true_line_good + rnorm(n = sample_size, mean = 0, sd = sigma)
reg_out = lm(y ~ z1 + z2)
beta_hat_good[s, ] = coef(reg_out)[-1]
mse_good[s] = mean(resid(reg_out) ^ 2)
}
We’ll now investigate the differences.


hist(beta_hat_bad[, 1],
col = "darkorange",
main = expression("Histogram of " *hat(beta)[1]* " with Collinearity"),
xlab = expression(hat(beta)[1]),
breaks = 20)
hist(beta_hat_good[, 1],
col = "darkorange",
main = expression("Histogram of " *hat(beta)[1]* " without Collinearity"),
breaks = 20)
^ with Collinearity
Histogram of β ^ without Collinearity
Histogram of β
1 1
300
300
250
200
Frequency
Frequency
200
150
100
100
50
0
-2 0 2 4 6 8 10 1 2 3 4 5
^
β ^
β
1 1
First, for 𝛽1 , which has a true value of 3, we see that both with and without
collinearity, the simulated values are centered near 3.
mean(beta_hat_bad[, 1])
## [1] 2.963325
mean(beta_hat_good[, 1])
## [1] 3.013414
The way the predictors were created, the 𝑆𝑥𝑗 𝑥𝑗 portion of the variance is the
same for the predictors in both cases, but the variance is still much larger in the
simulations performed with collinearity. The variance is so large in the collinear
case, that sometimes the estimated coefficient for 𝛽1 is negative!
sd(beta_hat_bad[, 1])
## [1] 1.633294
sd(beta_hat_good[, 1])
## [1] 0.5484684

hist(beta_hat_bad[, 2],
col = "darkorange",
main = expression("Histogram of " *hat(beta)[2]* " with Collinearity"),
breaks = 20)
hist(beta_hat_good[, 2],
col = "darkorange",
main = expression("Histogram of " *hat(beta)[2]* " without Collinearity"),
breaks = 20)
^ with Collinearity
Histogram of β ^ without Collinearity
Histogram of β
2 2
400
300
250
300
200
Frequency
Frequency
200
150
100
100
50
0
-2 0 2 4 6 8 10 2 3 4 5 6
^
β ^
β
2 2
We see the same issues with 𝛽2 . On average the estimates are correct, but the
variance is again much larger with collinearity.
mean(beta_hat_bad[, 2])
## [1] 4.025059
mean(beta_hat_good[, 2])
## [1] 4.004913
sd(beta_hat_bad[, 2])
## [1] 1.642592
sd(beta_hat_good[, 2])
## [1] 0.5470381

hist(mse_bad,
col = "darkorange",
main = "MSE, with Collinearity",
xlab = "MSE")
hist(mse_good,
col = "darkorange",
main = "MSE, without Collinearity",
xlab = "MSE")
MSE, with Collinearity MSE, without Collinearity

600
600
500
500
400
400
Frequency
Frequency
300
300
200
200
100
100
0
0 20 40 60 0 10 20 30 40 50 60 70
MSE MSE
Interestingly, in both cases, the MSE is roughly the same on average. Again,
this is because collinearity affects a model’s ability to explain, but not predict.
mean(mse_bad)
## [1] 17.7186
mean(mse_good)
## [1] 17.70513
15.4 R Markdown
• collinearity.Rmd

Chapter 16
Variable Selection and

Model Building
“Choose well. Your choice is brief, and yet endless.”

— Johann Wolfgang von Goethe
• Understand the trade-off between goodness-of-fit and model complexity.

• Use variable selection procedures to find a good model from a set of pos-
sible models.
• Understand the two uses of models: explanation and prediction.
Last chapter we saw how correlation between predictor variables can have un-
desirable effects on models. We used variance inflation factors to assess the
severity of the collinearity issues caused by these correlations. We also saw how
fitting a smaller model, leaving out some of the correlated predictors, results
in a model which no longer suffers from collinearity issues. But how should we
choose this smaller model?
This chapter, we will discuss several criteria and procedures for choosing a
“good” model from among a choice of many.
16.1 Quality Criterion

So far, we have seen criteria such as 𝑅2 and RMSE for assessing quality of fit.
However, both of these have a fatal flaw. By increasing the size of a model, that
is adding predictors, that can at worst not improve. It is impossible to add a
383
384 CHAPTER 16. VARIABLE SELECTION AND MODEL BUILDING
predictor to a model and make 𝑅2 or RMSE worse. That means, if we were to

use either of these to choose between models, we would always simply choose
the larger model. Eventually we would simply be fitting to noise.
This suggests that we need a quality criterion that takes into account the size
of the model, since our preference is for small models that still fit well. We are
willing to sacrifice a small amount of “goodness-of-fit” for obtaining a smaller
model. (Here we use “goodness-of-fit” to simply mean how far the data is from
the model, the smaller the errors the better. Often in statistics, goodness-of-
fit can have a more precise meaning.) We will look at three criteria that do
this explicitly: AIC, BIC, and Adjusted 𝑅2 . We will also look at one, Cross-
Validated RMSE, which implicitly considers the size of the model.
16.1.1 Akaike Information Criterion
The first criteria we will discuss is the Akaike Information Criterion, or AIC
for short. (Note that, when Akaike first introduced this metric, it was simply
called An Information Criterion. The A has changed meaning over the years.)
Recall, the maximized log-likelihood of a regression model can be written as
𝑛 𝑛 RSS 𝑛
log 𝐿(𝛽,̂ 𝜎̂ 2 ) = − log(2𝜋) − log ( )− ,
2 2 𝑛 2
𝑛
where RSS = ∑𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2 and 𝛽 ̂ and 𝜎̂ 2 were chosen to maximize the likeli-
hood.
Then we can define AIC as
RSS
AIC = −2 log 𝐿(𝛽,̂ 𝜎̂ 2 ) + 2𝑝 = 𝑛 + 𝑛 log(2𝜋) + 𝑛 log ( ) + 2𝑝,
𝑛
which is a measure of quality of the model. The smaller the AIC, the better.
To see why, let’s talk about the two main components of AIC, the likelihood
(which measures “goodness-of-fit”) and the penalty (which is a function of the
size of the model).
The likelihood portion of AIC is given by
RSS
−2 log 𝐿(𝛽,̂ 𝜎̂ 2 ) = 𝑛 + 𝑛 log(2𝜋) + 𝑛 log ( ).
𝑛
For the sake of comparing models, the only term here that will change is
𝑛 log ( RSS
𝑛 ), which is a function of RSS. The
𝑛 + 𝑛 log(2𝜋)
16.1. QUALITY CRITERION 385
terms will be constant across all models applied to the same data. So, when a
model fits well, that is, has a low RSS, then this likelihood component will be
small.
Similarly, we can discuss the penalty component of AIC which is,
2𝑝,
where 𝑝 is the number of 𝛽 parameters in the model. We call this a penalty,

because it is large when 𝑝 is large, but we are seeking to find a small AIC
Thus, a good model, that is one with a small AIC, will have a good balance
between fitting well, and using a small number of parameters. For comparing
models
RSS
AIC = 𝑛 log ( ) + 2𝑝
𝑛
is a sufficient expression, as 𝑛 + 𝑛 log(2𝜋) is the same across all models for any
particular dataset.
16.1.2 Bayesian Information Criterion
The Bayesian Information Criterion, or BIC, is similar to AIC, but has a larger
penalty. BIC also quantifies the trade-off between a model which fits well and
the number of model parameters, however for a reasonable sample size, generally
picks a smaller model than AIC. Again, for model selection use the model with
the smallest BIC.
RSS
BIC = −2 log 𝐿(𝛽,̂ 𝜎̂ 2 ) + log(𝑛)𝑝 = 𝑛 + 𝑛 log(2𝜋) + 𝑛 log ( ) + log(𝑛)𝑝.
𝑛
Notice that the AIC penalty was
2𝑝,
whereas for BIC, the penalty is
log(𝑛)𝑝.
So, for any dataset where 𝑙𝑜𝑔(𝑛) > 2 the BIC penalty will be larger than the
AIC penalty, thus BIC will likely prefer a smaller model.
Note that sometimes the penalty is considered a general expression of the form
𝑘 ⋅ 𝑝.
Then, for AIC 𝑘 = 2, and for BIC 𝑘 = log(𝑛).

For comparing models
RSS
BIC = 𝑛 log ( ) + log(𝑛)𝑝
𝑛
is again a sufficient expression, as 𝑛 + 𝑛 log(2𝜋) is the same across all models

for any particular dataset.
16.1.3 Adjusted R-Squared
Recall,
𝑛
SSE ∑ (𝑦𝑖 − 𝑦𝑖̂ )2
𝑅2 = 1 − = 1 − 𝑖=1
𝑛 .
SST ∑𝑖=1 (𝑦𝑖 − 𝑦)̄ 2
We now define
SSE/(𝑛 − 𝑝) 𝑛−1
𝑅𝑎2 = 1 − =1−( ) (1 − 𝑅2 )
SST/(𝑛 − 1) 𝑛−𝑝
which we call the Adjusted 𝑅2 .

Unlike 𝑅2 which can never become smaller with added predictors, Adjusted
𝑅2 effectively penalizes for additional predictors, and can decrease with added
predictors. Like 𝑅2 , larger is still better.
16.1.4 Cross-Validated RMSE
Each of the previous three metrics explicitly used 𝑝, the number of parameters,
in their calculations. Thus, they all explicitly limit the size of models chosen
when used to compare models.
We’ll now briefly introduce overfitting and cross-validation.
make_poly_data = function(sample_size = 11) {

x = seq(0, 10)
y = 3 + x + 4 * x ^ 2 + rnorm(n = sample_size, mean = 0, sd = 20)
data.frame(x, y)
}
set.seed(1234)
poly_data = make_poly_data()
Here we have generated data where the mean of 𝑌 is a quadratic function of a

single predictor 𝑥, specifically,
𝑌 = 3 + 𝑥 + 4𝑥2 + 𝜖.
We’ll now fit two models to this data, one which has the correct form, quadratic,
and one that is large, which includes terms up to and including an eighth degree.
fit_quad = lm(y ~ poly(x, degree = 2), data = poly_data)

fit_big = lm(y ~ poly(x, degree = 8), data = poly_data)
We then plot the data and the results of the two models.
plot(y ~ x, data = poly_data, ylim = c(-100, 400), cex = 2, pch = 20)

xplot = seq(0, 10, by = 0.1)
lines(xplot, predict(fit_quad, newdata = data.frame(x = xplot)),
lines(xplot, predict(fit_big, newdata = data.frame(x = xplot)),
400
300
200
y
100
0
-100
0 2 4 6 8 10
x
We can see that the solid blue curve models this data rather nicely. The dashed
orange curve fits the points better, making smaller errors, however it is unlikely
that it is correctly modeling the true relationship between 𝑥 and 𝑦. It is fitting
the random noise. This is an example of overfitting.
We see that the larger model indeed has a lower RMSE.
sqrt(mean(resid(fit_quad) ^ 2))
## [1] 17.61812
sqrt(mean(resid(fit_big) ^ 2))
## [1] 10.4197
To correct for this, we will introduce cross-validation. We define the leave-one-

out cross-validated RMSE to be
1 𝑛 2
RMSELOOCV = √ ∑𝑒 .
𝑛 𝑖=1 [𝑖]
The 𝑒[𝑖] are the residual for the 𝑖th observation, when that observation is not
used to fit the model.
𝑒[𝑖] = 𝑦𝑖 − 𝑦[𝑖]
̂
That is, the fitted value is calculated as
̂ = 𝑥⊤
𝑦[𝑖] ̂
𝑖 𝛽[𝑖]
̂ are the estimated coefficients when the 𝑖th observation is removed

where 𝛽[𝑖]
from the dataset.
In general, to perform this calculation, we would be required to fit the model 𝑛
times, once with each possible observation removed. However, for leave-one-out
cross-validation and linear models, the equation can be rewritten as
√
√1 𝑛 2
𝑒𝑖
RMSELOOCV = √ ∑ ( ) ,
𝑛 1 − ℎ𝑖
⎷ 𝑖=1
where ℎ𝑖 are the leverages and 𝑒𝑖 are the usual residuals. This is great, because
now we can obtain the LOOCV RMSE by fitting only one model! In practice 5
or 10 fold cross-validation are much more popular. For example, in 5-fold cross-
validation, the model is fit 5 times, each time leaving out a fifth of the data, then
predicting on those values. We’ll leave in-depth examination of cross-validation
to a machine learning course, and simply use LOOCV here.
Let’s calculate LOOCV RMSE for both models, then discuss why we want to
do so. We first write a function which calculates the LOOCV RMSE as defined
using the shortcut formula for linear models.
calc_loocv_rmse = function(model) {
sqrt(mean((resid(model) / (1 - hatvalues(model))) ^ 2))
}
Then calculate the metric for both models.
calc_loocv_rmse(fit_quad)
## [1] 23.57189
calc_loocv_rmse(fit_big)
## [1] 1334.357
Now we see that the quadratic model has a much smaller LOOCV RMSE, so we
would prefer this quadratic model. This is because the large model has severely
over-fit the data. By leaving a single data point out and fitting the large model,
the resulting fit is much different than the fit using all of the data. For example,
let’s leave out the third data point and fit both models, then plot the result.
fit_quad_removed = lm(y ~ poly(x, degree = 2), data = poly_data[-3, ])

fit_big_removed = lm(y ~ poly(x, degree = 8), data = poly_data[-3, ])
plot(y ~ x, data = poly_data, ylim = c(-100, 400), cex = 2, pch = 20)

xplot = seq(0, 10, by = 0.1)
lines(xplot, predict(fit_quad_removed, newdata = data.frame(x = xplot)),
lines(xplot, predict(fit_big_removed, newdata = data.frame(x = xplot)),
400
300
200
y
100
0
-100
0 2 4 6 8 10
We see that on average, the solid blue line for the quadratic model has similar
errors as before. It has changed very slightly. However, the dashed orange line
for the large model, has a huge error at the point that was removed and is much
different than the previous fit.
This is the purpose of cross-validation. By assessing how the model fits points
that were not used to perform the regression, we get an idea of how well the
model will work for future observations. It assesses how well the model works
in general, not simply on the observed data.
16.2 Selection Procedures

We’ve now seen a number of model quality criteria, but now we need to address
which models to consider. Model selection involves both a quality criterion, plus
a search procedure.
library(faraway)
hipcenter_mod = lm(hipcenter ~ ., data = seatpos)
coef(hipcenter_mod)

## 436.43212823 0.77571620 0.02631308 -2.69240774 0.60134458 0.53375170
## Arm Thigh Leg
## -1.32806864 -1.14311888 -6.43904627
16.2. SELECTION PROCEDURES 391
Let’s return to the seatpos data from the faraway package. Now, let’s consider
only models with first order terms, thus no interactions and no polynomials.
There are eight predictors in this model. So if we consider all possible models,
ranging from using 0 predictors, to all eight predictors, there are
𝑝−1
𝑝−1
∑( ) = 2𝑝−1 = 28 = 256
𝑘=0
𝑘
possible models.
If we had 10 or more predictors, we would already be considering over 1000
models! For this reason, we often search through possible models in an intelligent
way, bypassing some models that are unlikely to be considered good. We will
consider three search procedures: backwards, forwards, and stepwise.
16.2.1 Backward Search
Backward selection procedures start with all possible predictors in the model,
then consider how deleting a single predictor will affect a chosen metric. Let’s
try this on the seatpos data. We will use the step() function in R which by
default uses AIC as its metric of choice.
hipcenter_mod_back_aic = step(hipcenter_mod, direction = "backward")
## Start: AIC=283.62
## hipcenter ~ Age + Weight + HtShoes + Ht + Seated + Arm + Thigh +
## Leg
##
## Df Sum of Sq RSS AIC
## - Ht 1 5.01 41267 281.63
## - Weight 1 8.99 41271 281.63
## - Seated 1 28.64 41290 281.65
## - HtShoes 1 108.43 41370 281.72
## - Arm 1 164.97 41427 281.78
## - Thigh 1 262.76 41525 281.87
## <none> 41262 283.62
## - Age 1 2632.12 43894 283.97
## - Leg 1 2654.85 43917 283.99
##
## Step: AIC=281.63
## hipcenter ~ Age + Weight + HtShoes + Seated + Arm + Thigh + Leg
##
## - Weight 1 11.10 41278 279.64
## - Seated 1 30.52 41297 279.66

## - Arm 1 160.50 41427 279.78
## - Thigh 1 269.08 41536 279.88
## - HtShoes 1 971.84 42239 280.51
## <none> 41267 281.63
## - Leg 1 2664.65 43931 282.01
## - Age 1 2808.52 44075 282.13
##
## Step: AIC=279.64
## hipcenter ~ Age + HtShoes + Seated + Arm + Thigh + Leg
##
## - Seated 1 35.10 41313 277.67
## - Arm 1 156.47 41434 277.78
## - Thigh 1 285.16 41563 277.90
## - HtShoes 1 975.48 42253 278.53
## <none> 41278 279.64
## - Leg 1 2661.39 43939 280.01
## - Age 1 3011.86 44290 280.31
##
## Step: AIC=277.67
## hipcenter ~ Age + HtShoes + Arm + Thigh + Leg
##
## - Arm 1 172.02 41485 275.83
## - Thigh 1 344.61 41658 275.99
## - HtShoes 1 1853.43 43166 277.34
## <none> 41313 277.67
## - Leg 1 2871.07 44184 278.22
## - Age 1 2976.77 44290 278.31
##
## Step: AIC=275.83
## hipcenter ~ Age + HtShoes + Thigh + Leg
##
## - Thigh 1 472.8 41958 274.26
## <none> 41485 275.83
## - HtShoes 1 2340.7 43826 275.92
## - Age 1 3501.0 44986 276.91
## - Leg 1 3591.7 45077 276.98
##
## Step: AIC=274.26
## hipcenter ~ Age + HtShoes + Leg
##
## <none> 41958 274.26
## - Age 1 3108.8 45067 274.98

## - Leg 1 3476.3 45434 275.28
## - HtShoes 1 4218.6 46176 275.90
We start with the model hipcenter ~ ., which is otherwise known as

hipcenter ~ Age + Weight + HtShoes + Ht + Seated + Arm + Thigh +
Leg. R will then repeatedly attempt to delete a predictor until it stops, or
reaches the model hipcenter ~ 1, which contains no predictors.
At each “step,” R reports the current model, its AIC, and the possible steps
with their RSS and more importantly AIC.
In this example, at the first step, the current model is hipcenter ~ Age +
Weight + HtShoes + Ht + Seated + Arm + Thigh + Leg which has an AIC
of 283.62. Note that when R is calculating this value, it is using extractAIC(),
which uses the expression
RSS
AIC = 𝑛 log ( ) + 2𝑝,
𝑛
which we quickly verify.
extractAIC(hipcenter_mod) # returns both p and AIC
## [1] 9.000 283.624
n = length(resid(hipcenter_mod))
(p = length(coef(hipcenter_mod)))
## [1] 9
n * log(mean(resid(hipcenter_mod) ^ 2)) + 2 * p
## [1] 283.624
Returning to the first step, R then gives us a row which shows the effect of
deleting each of the current predictors. The - signs at the beginning of each
row indicate we are considering removing a predictor. There is also a row with
<none> which is a row for keeping the current model. Notice that this row has
the smallest RSS, as it is the largest model.
We see that every row above <none> has a smaller AIC than the row for <none>
with the one at the top, Ht, giving the lowest AIC. Thus we remove Ht from
the model, and continue the process.
Notice, in the second step, we start with the model hipcenter ~ Age + Weight
+ HtShoes + Seated + Arm + Thigh + Leg and the variable Ht is no longer
considered.
We continue the process until we reach the model hipcenter ~ Age + HtShoes
+ Leg. At this step, the row for <none> tops the list, as removing any addi-
tional variable will not improve the AIC This is the model which is stored in
hipcenter_mod_back_aic.
coef(hipcenter_mod_back_aic)
## (Intercept) Age HtShoes Leg

## 456.2136538 0.5998327 -2.3022555 -6.8297461
We could also search through the possible models in a backwards fashion using
BIC. To do so, we again use the step() function, but now specify k = log(n),
where n stores the number of observations in the data.
hipcenter_mod_back_bic = step(hipcenter_mod, direction = "backward", k = log(n))
## hipcenter ~ Age + Weight + HtShoes + Ht + Seated + Arm + Thigh +
## Leg
##
## - Ht 1 5.01 41267 294.73
## - Weight 1 8.99 41271 294.73
## - Seated 1 28.64 41290 294.75
## - HtShoes 1 108.43 41370 294.82
## - Arm 1 164.97 41427 294.88
## - Thigh 1 262.76 41525 294.97
## - Age 1 2632.12 43894 297.07
## - Leg 1 2654.85 43917 297.09
## <none> 41262 298.36
##
## Step: AIC=294.73
## hipcenter ~ Age + Weight + HtShoes + Seated + Arm + Thigh + Leg
##
## - Weight 1 11.10 41278 291.10
## - Seated 1 30.52 41297 291.12
## - Arm 1 160.50 41427 291.24
## - Thigh 1 269.08 41536 291.34
## - HtShoes 1 971.84 42239 291.98
## - Leg 1 2664.65 43931 293.47

## - Age 1 2808.52 44075 293.59
## <none> 41267 294.73
##
## Step: AIC=291.1
## hipcenter ~ Age + HtShoes + Seated + Arm + Thigh + Leg
##
## - Seated 1 35.10 41313 287.50
## - Arm 1 156.47 41434 287.61
## - Thigh 1 285.16 41563 287.73
## - HtShoes 1 975.48 42253 288.35
## - Leg 1 2661.39 43939 289.84
## - Age 1 3011.86 44290 290.14
## <none> 41278 291.10
##
## Step: AIC=287.5
## hipcenter ~ Age + HtShoes + Arm + Thigh + Leg
##
## - Arm 1 172.02 41485 284.02
## - Thigh 1 344.61 41658 284.18
## - HtShoes 1 1853.43 43166 285.53
## - Leg 1 2871.07 44184 286.41
## - Age 1 2976.77 44290 286.50
## <none> 41313 287.50
##
## Step: AIC=284.02
## hipcenter ~ Age + HtShoes + Thigh + Leg
##
## - Thigh 1 472.8 41958 280.81
## - HtShoes 1 2340.7 43826 282.46
## - Age 1 3501.0 44986 283.46
## - Leg 1 3591.7 45077 283.54
## <none> 41485 284.02
##
## Step: AIC=280.81
## hipcenter ~ Age + HtShoes + Leg
##
## - Age 1 3108.8 45067 279.89
## - Leg 1 3476.3 45434 280.20
## <none> 41958 280.81
## - HtShoes 1 4218.6 46176 280.81
##
## Step: AIC=279.89
## hipcenter ~ HtShoes + Leg
##
## - Leg 1 3038.8 48105 278.73
## <none> 45067 279.89
## - HtShoes 1 5004.4 50071 280.25
##
## Step: AIC=278.73
## hipcenter ~ HtShoes
##
## <none> 48105 278.73
## - HtShoes 1 83534 131639 313.35
The procedure is exactly the same, except at each step we look to improve the
BIC, which R still labels AIC in the output.
The variable hipcenter_mod_back_bic stores the model chosen by this proce-
dure.
coef(hipcenter_mod_back_bic)
## (Intercept) HtShoes
## 565.592659 -4.262091
We note that this model is smaller, has fewer predictors, than the model chosen
by AIC, which is what we would expect. Also note that while both models are
different, neither uses both Ht and HtShoes which are extremely correlated.
We can use information from the summary() function to compare their Adjusted
𝑅2 values. Note that either selected model performs better than the original
full model.
summary(hipcenter_mod)$adj.r.squared
## [1] 0.6000855
summary(hipcenter_mod_back_aic)$adj.r.squared
## [1] 0.6531427
summary(hipcenter_mod_back_bic)$adj.r.squared
## [1] 0.6244149
We can also calculate the LOOCV RMSE for both selected models, as well as
the full model.
calc_loocv_rmse(hipcenter_mod)
## [1] 44.44564
calc_loocv_rmse(hipcenter_mod_back_aic)
## [1] 37.58473
calc_loocv_rmse(hipcenter_mod_back_bic)
## [1] 37.40564
We see that we would prefer the model chosen via BIC if using LOOCV RMSE
as our metric.
16.2.2 Forward Search
Forward selection is the exact opposite of backwards selection. Here we tell

R to start with a model using no predictors, that is hipcenter ~ 1, then
at each step R will attempt to add a predictor until it finds a good model
or reaches hipcenter ~ Age + Weight + HtShoes + Ht + Seated + Arm +
Thigh + Leg.
hipcenter_mod_start = lm(hipcenter ~ 1, data = seatpos)

hipcenter_mod_forw_aic = step(
hipcenter_mod_start,
scope = hipcenter ~ Age + Weight + HtShoes + Ht + Seated + Arm + Thigh + Leg,
direction = "forward")
## hipcenter ~ 1
##
## + Ht 1 84023 47616 275.07

## + HtShoes 1 83534 48105 275.45
## + Leg 1 81568 50071 276.98
## + Seated 1 70392 61247 284.63
## + Weight 1 53975 77664 293.66
## + Thigh 1 46010 85629 297.37
## + Arm 1 45065 86574 297.78
## <none> 131639 311.71
## + Age 1 5541 126098 312.07
##
## Step: AIC=275.07
## hipcenter ~ Ht
##
## + Leg 1 2781.10 44835 274.78
## <none> 47616 275.07
## + Age 1 2353.51 45262 275.14
## + Weight 1 195.86 47420 276.91
## + Seated 1 101.56 47514 276.99
## + Arm 1 75.78 47540 277.01
## + HtShoes 1 25.76 47590 277.05
## + Thigh 1 4.63 47611 277.06
##
## Step: AIC=274.78
## hipcenter ~ Ht + Leg
##
## + Age 1 2896.60 41938 274.24
## <none> 44835 274.78
## + Arm 1 522.72 44312 276.33
## + Weight 1 445.10 44390 276.40
## + HtShoes 1 34.11 44801 276.75
## + Thigh 1 32.96 44802 276.75
## + Seated 1 1.12 44834 276.78
##
## Step: AIC=274.24
## hipcenter ~ Ht + Leg + Age
##
## <none> 41938 274.24
## + Thigh 1 372.71 41565 275.90
## + Arm 1 257.09 41681 276.01
## + Seated 1 121.26 41817 276.13
## + Weight 1 46.83 41891 276.20
## + HtShoes 1 13.38 41925 276.23
Again, by default R uses AIC as its quality metric when using the step()
function. Also note that now the rows begin with a + which indicates addition
of predictors to the current model from any step.
hipcenter_mod_forw_bic = step(
direction = "forward", k = log(n))
## hipcenter ~ 1
##
## + Ht 1 84023 47616 278.34
## + HtShoes 1 83534 48105 278.73
## + Leg 1 81568 50071 280.25
## + Seated 1 70392 61247 287.91
## + Weight 1 53975 77664 296.93
## + Thigh 1 46010 85629 300.64
## + Arm 1 45065 86574 301.06
## <none> 131639 313.35
## + Age 1 5541 126098 315.35
##
## Step: AIC=278.34
## hipcenter ~ Ht
##
## <none> 47616 278.34
## + Leg 1 2781.10 44835 279.69
## + Age 1 2353.51 45262 280.05
## + Weight 1 195.86 47420 281.82
## + Seated 1 101.56 47514 281.90
## + Arm 1 75.78 47540 281.92
## + HtShoes 1 25.76 47590 281.96
## + Thigh 1 4.63 47611 281.98
We can make the same modification as last time to instead use BIC with forward
selection.
## [1] 0.6000855
summary(hipcenter_mod_forw_aic)$adj.r.squared
## [1] 0.6533055
summary(hipcenter_mod_forw_bic)$adj.r.squared
## [1] 0.6282374
We can compare the two selected models’ Adjusted 𝑅2 as well as their LOOCV
RMSE The results are very similar to those using backwards selection, although
the models are not exactly the same.
## [1] 44.44564
calc_loocv_rmse(hipcenter_mod_forw_aic)
## [1] 37.62516
calc_loocv_rmse(hipcenter_mod_forw_bic)
## [1] 37.2511
16.2.3 Stepwise Search
Stepwise search checks going both backwards and forwards at every step. It
considers the addition of any variable not currently in the model, as well as the
removal of any variable currently in the model.
Here we perform stepwise search using AIC as our metric. We start with
the model hipcenter ~ 1 and search up to hipcenter ~ Age + Weight +
HtShoes + Ht + Seated + Arm + Thigh + Leg. Notice that at many of the
steps, some row begin with -, while others begin with +.
hipcenter_mod_both_aic = step(
direction = "both")
## hipcenter ~ 1
##
## + Ht 1 84023 47616 275.07
## + HtShoes 1 83534 48105 275.45
## + Leg 1 81568 50071 276.98
## + Seated 1 70392 61247 284.63
## + Weight 1 53975 77664 293.66
## + Thigh 1 46010 85629 297.37
## + Arm 1 45065 86574 297.78
## <none> 131639 311.71
## + Age 1 5541 126098 312.07
##
## Step: AIC=275.07
## hipcenter ~ Ht
##
## + Leg 1 2781 44835 274.78
## <none> 47616 275.07
## + Age 1 2354 45262 275.14
## + Weight 1 196 47420 276.91
## + Seated 1 102 47514 276.99
## + Arm 1 76 47540 277.01
## + HtShoes 1 26 47590 277.05
## + Thigh 1 5 47611 277.06
## - Ht 1 84023 131639 311.71
##
## Step: AIC=274.78
## hipcenter ~ Ht + Leg
##
## + Age 1 2896.6 41938 274.24
## <none> 44835 274.78
## - Leg 1 2781.1 47616 275.07
## + Arm 1 522.7 44312 276.33
## + Weight 1 445.1 44390 276.40
## + HtShoes 1 34.1 44801 276.75
## + Thigh 1 33.0 44802 276.75
## + Seated 1 1.1 44834 276.78
## - Ht 1 5236.3 50071 276.98
##
## Step: AIC=274.24
## hipcenter ~ Ht + Leg + Age
##
## <none> 41938 274.24

## - Age 1 2896.6 44835 274.78
## - Leg 1 3324.2 45262 275.14
## - Ht 1 4238.3 46176 275.90
## + Thigh 1 372.7 41565 275.90
## + Arm 1 257.1 41681 276.01
## + Seated 1 121.3 41817 276.13
## + Weight 1 46.8 41891 276.20
## + HtShoes 1 13.4 41925 276.23
We could again instead use BIC as our metric.
hipcenter_mod_both_bic = step(
direction = "both", k = log(n))
## hipcenter ~ 1
##
## + Ht 1 84023 47616 278.34
## + HtShoes 1 83534 48105 278.73
## + Leg 1 81568 50071 280.25
## + Seated 1 70392 61247 287.91
## + Weight 1 53975 77664 296.93
## + Thigh 1 46010 85629 300.64
## + Arm 1 45065 86574 301.06
## <none> 131639 313.35
## + Age 1 5541 126098 315.35
##
## Step: AIC=278.34
## hipcenter ~ Ht
##
## <none> 47616 278.34
## + Leg 1 2781 44835 279.69
## + Age 1 2354 45262 280.05
## + Weight 1 196 47420 281.82
## + Seated 1 102 47514 281.90
## + Arm 1 76 47540 281.92
## + HtShoes 1 26 47590 281.96
## + Thigh 1 5 47611 281.98
## - Ht 1 84023 131639 313.35
Adjusted 𝑅2 and LOOCV RMSE comparisons are similar to backwards and

forwards, which is not at all surprising, as some of the models selected are the
same as before.
## [1] 0.6000855
summary(hipcenter_mod_both_aic)$adj.r.squared
## [1] 0.6533055
summary(hipcenter_mod_both_bic)$adj.r.squared
## [1] 0.6282374
## [1] 44.44564
calc_loocv_rmse(hipcenter_mod_both_aic)
## [1] 37.62516
calc_loocv_rmse(hipcenter_mod_both_bic)
## [1] 37.2511
16.2.4 Exhaustive Search
Backward, forward, and stepwise search are all useful, but do have an obvious
issue. By not checking every possible model, sometimes they will miss the best
possible model. With an extremely large number of predictors, sometimes this is
necessary since checking every possible model would be rather time consuming,
even with current computers.
However, with a reasonably sized dataset, it isn’t too difficult to check all pos-
sible models. To do so, we will use the regsubsets() function in the R package
leaps.
library(leaps)
all_hipcenter_mod = summary(regsubsets(hipcenter ~ ., data = seatpos))
A few points about this line of code. First, note that we immediately
use summary() and store those results. That is simply the intended use
of regsubsets(). Second, inside of regsubsets() we specify the model
hipcenter ~ .. This will be the largest model considered, that is the model
using all first-order predictors, and R will check all possible subsets. Third,
regsubsets() has an argument of nvmax to set the maximum size of subsets to
examine with 8 as default. You need to change it if you investigate the subsets
with more than 8 predictors, although increasing nvmax will significantly
increase computation time.
We’ll now look at the information stored in all_hipcenter_mod.
all_hipcenter_mod$which
## (Intercept) Age Weight HtShoes Ht Seated Arm Thigh Leg

## 1 TRUE FALSE FALSE FALSE TRUE FALSE FALSE FALSE FALSE
## 2 TRUE FALSE FALSE FALSE TRUE FALSE FALSE FALSE TRUE
## 3 TRUE TRUE FALSE FALSE TRUE FALSE FALSE FALSE TRUE
## 4 TRUE TRUE FALSE TRUE FALSE FALSE FALSE TRUE TRUE
## 5 TRUE TRUE FALSE TRUE FALSE FALSE TRUE TRUE TRUE
## 6 TRUE TRUE FALSE TRUE FALSE TRUE TRUE TRUE TRUE
## 7 TRUE TRUE TRUE TRUE FALSE TRUE TRUE TRUE TRUE
## 8 TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
Using $which gives us the best model, according to RSS, for a model of each
possible size, in this case ranging from one to eight predictors. For example the
best model with four predictors (𝑝 = 5) would use Age, HtShoes, Thigh, and
Leg.
all_hipcenter_mod$rss
## [1] 47615.79 44834.69 41938.09 41485.01 41313.00 41277.90 41266.80 41261.78
We can obtain the RSS for each of these models using $rss. Notice that these
are decreasing since the models range from small to large.
Now that we have the RSS for each of these models, it is rather easy to obtain
AIC, BIC, and Adjusted 𝑅2 since they are all a function of RSS Also, since we
have the models with the best RSS for each size, they will result in the models
with the best AIC, BIC, and Adjusted 𝑅2 for each size. Then by picking from
those, we can find the overall best AIC, BIC, and Adjusted 𝑅2 .
Conveniently, Adjusted 𝑅2 is automatically calculated.
all_hipcenter_mod$adjr2
## [1] 0.6282374 0.6399496 0.6533055 0.6466586 0.6371276 0.6257403 0.6133690

## [8] 0.6000855
To find which model has the highest Adjusted 𝑅2 we can use the which.max()
function.
(best_r2_ind = which.max(all_hipcenter_mod$adjr2))
## [1] 3
We can then extract the predictors of that model.
all_hipcenter_mod$which[best_r2_ind, ]

## TRUE TRUE FALSE FALSE TRUE FALSE
## Arm Thigh Leg
## FALSE FALSE TRUE
We’ll now calculate AIC and BIC for each of the models with the best RSS. To
do so, we will need both 𝑛 and the 𝑝 for the largest possible model.
p = length(coef(hipcenter_mod))
We’ll use the form of AIC which leaves out the constant term that is equal
across all models.
RSS
AIC = 𝑛 log ( ) + 2𝑝.
𝑛
Since we have the RSS of each model stored, this is easy to calculate.
hipcenter_mod_aic = n * log(all_hipcenter_mod$rss / n) + 2 * (2:p)
We can then extract the predictors of the model with the best AIC.
best_aic_ind = which.min(hipcenter_mod_aic)
all_hipcenter_mod$which[best_aic_ind,]

## TRUE TRUE FALSE FALSE TRUE FALSE
## Arm Thigh Leg
## FALSE FALSE TRUE
Let’s fit this model so we can compare to our previously chosen models using
AIC and search procedures.
hipcenter_mod_best_aic = lm(hipcenter ~ Age + Ht + Leg, data = seatpos)
The extractAIC() function will calculate the AIC defined above for a fitted
model.
extractAIC(hipcenter_mod_best_aic)
## [1] 4.0000 274.2418
extractAIC(hipcenter_mod_back_aic)
## [1] 4.0000 274.2597
extractAIC(hipcenter_mod_forw_aic)
## [1] 4.0000 274.2418
extractAIC(hipcenter_mod_both_aic)
## [1] 4.0000 274.2418
We see that two of the models chosen by search procedures have the best possible
AIC, as they are the same model. This is however never guaranteed. We see
that the model chosen using backwards selection does not achieve the smallest
possible AIC.
plot(hipcenter_mod_aic ~ I(2:p), ylab = "AIC", xlab = "p, number of parameters",

pch = 20, col = "dodgerblue", type = "b", cex = 2,
main = "AIC vs Model Complexity")
AIC vs Model Complexity

282
280
AIC
278
276
274
2 3 4 5 6 7 8 9
p, number of parameters
We could easily repeat this process for BIC.
RSS
BIC = 𝑛 log ( ) + log(𝑛)𝑝.
𝑛
hipcenter_mod_bic = n * log(all_hipcenter_mod$rss / n) + log(n) * (2:p)
which.min(hipcenter_mod_bic)
## [1] 1
all_hipcenter_mod$which[1,]

## TRUE FALSE FALSE FALSE TRUE FALSE
## Arm Thigh Leg
## FALSE FALSE FALSE
hipcenter_mod_best_bic = lm(hipcenter ~ Ht, data = seatpos)

extractAIC(hipcenter_mod_best_bic, k = log(n))
## [1] 2.0000 278.3418
extractAIC(hipcenter_mod_back_bic, k = log(n))
## [1] 2.0000 278.7306
extractAIC(hipcenter_mod_forw_bic, k = log(n))
## [1] 2.0000 278.3418
extractAIC(hipcenter_mod_both_bic, k = log(n))
## [1] 2.0000 278.3418
16.3 Higher Order Terms

So far we have only allowed first-order terms in our models. Let’s return to the
autompg dataset to explore higher-order terms.
quote = "\"",
comment.char = "",
colnames(autompg) = c("mpg", "cyl", "disp", "hp", "wt", "acc",
"year", "origin", "name")
autompg$domestic = as.numeric(autompg$origin == 1)
autompg$cyl = as.factor(autompg$cyl)
autompg$domestic = as.factor(autompg$domestic)
autompg = subset(autompg, select = c("mpg", "cyl", "disp", "hp",
"wt", "acc", "year", "domestic"))
16.3. HIGHER ORDER TERMS 409
str(autompg)

## $ mpg : num 18 15 18 16 17 15 14 14 14 15 ...
## $ disp : num 307 350 318 304 302 429 454 440 455 390 ...
## $ hp : num 130 165 150 150 140 198 220 215 225 190 ...
## $ wt : num 3504 3693 3436 3433 3449 ...
## $ acc : num 12 11.5 11 12 10.5 10 9 8.5 10 8.5 ...
## $ year : int 70 70 70 70 70 70 70 70 70 70 ...
## $ domestic: Factor w/ 2 levels "0","1": 2 2 2 2 2 2 2 2 2 2 ...
Recall that we have two factor variables, cyl and domestic. The cyl variable
has three levels, while the domestic variable has only two. Thus the cyl variable
will be coded using two dummy variables, while the domestic variable will only
need one. We will pay attention to this later.
pairs(autompg, col = "dodgerblue")

1.0 2.0 3.0 50 150 10 20 1.0 1.4 1.8
mpg
30
10
3.0
cyl
2.0
1.0
300
disp
100
150
hp
50
3500
wt
1500
20
acc
10
70 74 78 82
year
1.8
domestic
1.4
1.0
10 30 100 300 1500 3500 70 74 78 82
We’ll use the pairs() plot to determine which variables may benefit from a
quadratic relationship with the response. We’ll also consider all possible two-
way interactions. We won’t consider any three-order or higher. For example, we
won’t consider the interaction between first-order terms and the added quadratic
terms.
So now, we’ll fit this rather large model. We’ll use a log-transformed response.
Notice that log(mpg) ~ . ^ 2 will automatically consider all first-order terms,
as well as all two-way interactions. We use I(var_name ^ 2) to add quadratic
terms for some variables. This generally works better than using poly() when
performing variable selection.
autompg_big_mod = lm(
log(mpg) ~ . ^ 2 + I(disp ^ 2) + I(hp ^ 2) + I(wt ^ 2) + I(acc ^ 2),
data = autompg)
16.3. HIGHER ORDER TERMS 411
We think it is rather unlikely that we truly need all of these terms. There are
quite a few!
length(coef(autompg_big_mod))
## [1] 40
We’ll try backwards search with both AIC and BIC to attempt to find a smaller,
more reasonable model.
autompg_mod_back_aic = step(autompg_big_mod, direction = "backward", trace = 0)
Notice that we used trace = 0 in the function call. This suppresses the output
for each step, and simply stores the chosen model. This is useful, as this code
would otherwise create a large amount of output. If we had viewed the output,
which you can try on your own by removing trace = 0, we would see that R
only considers the cyl variable as a single variable, despite the fact that it is
coded using two dummy variables. So removing cyl would actually remove two
parameters from the resulting model.
You should also notice that R respects hierarchy when attempting to remove
variables. That is, for example, R will not consider removing hp if hp:disp or
I(hp ^ 2) are currently in the model.
We also use BIC.
n = length(resid(autompg_big_mod))
autompg_mod_back_bic = step(autompg_big_mod, direction = "backward",
k = log(n), trace = 0)
Looking at the coefficients of the two chosen models, we see they are still rather
large.
coef(autompg_mod_back_aic)
## (Intercept) cyl6 cyl8 disp hp

## 3.671884e+00 -1.602563e-01 -8.581644e-01 -9.371971e-03 2.293534e-02
## wt acc year domestic1 I(hp^2)
## -3.064497e-04 -1.393888e-01 -1.966361e-03 9.369324e-01 -1.497669e-05
## cyl6:acc cyl8:acc disp:wt disp:year hp:acc
## 7.220298e-03 5.041915e-02 5.797816e-07 9.493770e-05 -5.062295e-04
## hp:year acc:year acc:domestic1 year:domestic1
## -1.838985e-04 2.345625e-03 -2.372468e-02 -7.332725e-03
coef(autompg_mod_back_bic)

## 4.657847e+00 -1.086165e-01 -7.611631e-01 -1.609316e-03 2.621266e-03
## wt acc year domestic1 cyl6:acc
## -2.635972e-04 -1.670601e-01 -1.045646e-02 3.341579e-01 4.315493e-03
## cyl8:acc disp:wt hp:acc acc:year acc:domestic1
## 4.610095e-02 4.102804e-07 -3.386261e-04 2.500137e-03 -2.193294e-02
However, they are much smaller than the original full model. Also notice that
the resulting models respect hierarchy.
length(coef(autompg_big_mod))
## [1] 40
length(coef(autompg_mod_back_aic))
## [1] 19
length(coef(autompg_mod_back_bic))
## [1] 15
Calculating the LOOCV RMSE for each, we see that the model chosen using
BIC performs the best. That means that it is both the best model for pre-
diction, since it achieves the best LOOCV RMSE, but also the best model for
explanation, as it is also the smallest.
calc_loocv_rmse(autompg_big_mod)
## [1] 0.1112024
calc_loocv_rmse(autompg_mod_back_aic)
## [1] 0.1032888
calc_loocv_rmse(autompg_mod_back_bic)
## [1] 0.103134
16.4 Explanation versus Prediction
Throughout this chapter, we have attempted to find reasonably “small” models,

which are good at explaining the relationship between the response and the
predictors, that also have small errors which are thus good for making predic-
tions.
We’ll further discuss the model autompg_mod_back_bic to better explain the
difference between using models for explaining and predicting. This is the model
fit to the autompg data that was chosen using Backwards Search and BIC, which
obtained the lowest LOOCV RMSE of the models we considered.
autompg_mod_back_bic
##
## Call:
## lm(formula = log(mpg) ~ cyl + disp + hp + wt + acc + year + domestic +
## cyl:acc + disp:wt + hp:acc + acc:year + acc:domestic, data = autompg)
##
## Coefficients:
## 4.658e+00 -1.086e-01 -7.612e-01 -1.609e-03 2.621e-03
## wt acc year domestic1 cyl6:acc
## -2.636e-04 -1.671e-01 -1.046e-02 3.342e-01 4.315e-03
## cyl8:acc disp:wt hp:acc acc:year acc:domestic1
## 4.610e-02 4.103e-07 -3.386e-04 2.500e-03 -2.193e-02
Notice this is a somewhat “large” model, which uses 15 parameters, including

several interaction terms. Do we care that this is a “large” model? The answer
is, it depends.
16.4.1 Explanation
Suppose we would like to use this model for explanation. Perhaps we are a car
manufacturer trying to engineer a fuel efficient vehicle. If this is the case, we are
interested in both what predictor variables are useful for explaining the car’s fuel
efficiency, as well as how those variables affect fuel efficiency. By understanding
this relationship, we can use this knowledge to our advantage when designing a
car.
To explain a relationship, we are interested in keeping models as small as pos-
sible, since smaller models are easy to interpret. The fewer predictors the fewer
considerations we need to make in our design process. Also the fewer inter-
actions and polynomial terms, the easier it is to interpret any one parameter,
since the parameter interpretations are conditional on which parameters are in

the model.
Note that linear models are rather interpretable to begin with. Later in your
data analysis careers, you will see more complicated models that may fit data
better, but are much harder, if not impossible to interpret. These models aren’t
very useful for explaining a relationship.
To find small and interpretable models, we would use selection criteria that
explicitly penalize larger models, such as AIC and BIC. In this case we still
obtained a somewhat large model, but much smaller than the model we used to
start the selection process.
16.4.1.1 Correlation and Causation
A word of caution when using a model to explain a relationship. There are two
terms often used to describe a relationship between two variables: causation
and correlation. Correlation is often also referred to as association.
Just because two variables are correlated does not necessarily mean that one
causes the other. For example, considering modeling mpg as only a function of
hp.
40
30
mpg
20
10
50 100 150 200
hp
Does an increase in horsepower cause a drop in fuel efficiency? Or, perhaps

the causality is reversed and an increase in fuel efficiency causes a decrease in
horsepower. Or, perhaps there is a third variable that explains both!
The issue here is that we have observational data. With observational data,
we can only detect associations. To speak with confidence about causality, we
would need to run experiments.
This is a concept that you should encounter often in your statistics education.
For some further reading, and some related fallacies, see: Wikipedia: Correla-
tion does not imply causation.
16.4.2 Prediction
Suppose now instead of the manufacturer who would like to build a car, we are
a consumer who wishes to purchase a new car. However this particular car is so
new, it has not been rigorously tested, so we are unsure of what fuel efficiency
to expect. (And, as skeptics, we don’t trust what the manufacturer is telling
us.)
In this case, we would like to use the model to help predict the fuel efficiency
of this car based on its attributes, which are the predictors of the model. The
smaller the errors the model makes, the more confident we are in its prediction.
Thus, to find models for prediction, we would use selection criteria that implicitly
penalize larger models, such as LOOCV RMSE. So long as the model does not
over-fit, we do not actually care how large the model becomes. Explaining the
relationship between the variables is not our goal here, we simply want to know
what kind of fuel efficiency we should expect!
If we only care about prediction, we don’t need to worry about correlation vs

causation, and we don’t need to worry about model assumptions.
If a variable is correlated with the response, it doesn’t actually matter if it causes

an effect on the response, it can still be useful for prediction. For example, in
elementary school aged children their shoe size certainly doesn’t cause them to
read at a higher level, however we could very easily use shoe size to make a
prediction about a child’s reading ability. The larger their shoe size, the better
they read. There’s a lurking variable here though, their age! (Don’t send your
kids to school with size 14 shoes, it won’t make them read better!)
We also don’t care about model assumptions. Least squares is least squares. For
a specified model, it will find the values of the parameters which will minimize
the squared error loss. Your results might be largely uninterpretable and useless
for inference, but for prediction none of that matters.
16.5 R Markdown
• selection.Rmd

Chapter 17
Logistic Regression
Note to current readers: This chapter is slightly less tested than previous
chapters. Please do not hesitate to report any errors, or suggest sections that
need better explanation! Also, as a result, this material is more likely to receive
edits.
• Understand how generalized linear models are a generalization of ordinary

linear models.
• Use logistic regression to model a binary response.
• Apply concepts learned for ordinary linear models to logistic regression.
• Use logistic regression to perform classification.
So far we have only considered models for numeric response variables. What
about response variables that only take integer values? What about a response
variable that is categorical? Can we use linear models in these situations? Yes!
The model that we have been using, which we will call ordinary linear regression,
is actually a specific case of the more general, generalized linear model. (Aren’t
statisticians great at naming things?)
17.1 Generalized Linear Models

So far, we’ve had response variables that, conditioned on the predictors, were
modeled using a normal distribution with a mean that is some linear combi-
nation of the predictors. This linear combination is what made a linear model
“linear.”
𝑌 ∣ X = x ∼ 𝑁 (𝛽0 + 𝛽1 𝑥1 + … + 𝛽𝑝−1 𝑥𝑝−1 , 𝜎2 )
417
418 CHAPTER 17. LOGISTIC REGRESSION
Now we’ll allow for two modifications of this situation, which will let us use linear
models in many more situations. Instead of using a normal distribution for the
response conditioned on the predictors, we’ll allow for other distributions. Also,
instead of the conditional mean being a linear combination of the predictors, it
can be some function of a linear combination of the predictors.
In general, a generalized linear model has three parts:
• A distribution of the response conditioned on the predictors. (Techni-

cally this distribution needs to be from the exponential family of distri-
butions.)
• A linear combination of the 𝑝 − 1 predictors, 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + … +
𝛽𝑝−1 𝑥𝑝−1 , which we write as 𝜂(x). That is,
𝜂(x) = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + … + 𝛽𝑝−1 𝑥𝑝−1
• A link function, 𝑔(), that defines how 𝜂(x), the linear combination of
the predictors, is related to the mean of the response conditioned on the
predictors, E[𝑌 ∣ X = x].
𝜂(x) = 𝑔 (E[𝑌 ∣ X = x]) .
The following table summarizes three examples of a generalized linear model:
Linear Poisson
Regression Regression Logistic Regression
2
𝑌 ∣X=x 𝑁 (𝜇(x), 𝜎 ) Pois(𝜆(x)) Bern(𝑝(x))
Distribution Normal Poisson Bernoulli
Name (Binomial)
E[𝑌 ∣ X = x] 𝜇(x) 𝜆(x) 𝑝(x)
Support Real: (−∞, ∞) Integer: 0, 1, 2, … Integer: 0, 1
Usage Numeric Data Count (Integer) Binary (Class )
Data Data
Link Name Identity Log Logit
𝑝(x)
Link 𝜂(x) = 𝜇(x) 𝜂(x) = log(𝜆(x)) 𝜂(x) = log ( 1−𝑝(x) )
Function
𝑒𝜂(x)
Mean 𝜇(x) = 𝜂(x) 𝜆(x) = 𝑒𝜂(x) 𝑝(x) = 1+𝑒𝜂(x)
=
Function 1
1+𝑒−𝜂(x)
Like ordinary linear regression, we will seek to “fit” the model by estimating
the 𝛽 parameters. To do so, we will use the method of maximum likelihood.
17.2. BINARY RESPONSE 419
Note that a Bernoulli distribution is a specific case of a binomial distribution

where the 𝑛 parameter of a binomial is 1. Binomial regression is also possible,
but we’ll focus on the much more popular Bernoulli case.
So, in general, GLMs relate the mean of the response to a linear combination
of the predictors, 𝜂(x), through the use of a link function, 𝑔(). That is,
𝜂(x) = 𝑔 (E[𝑌 ∣ X = x]) .
The mean is then
E[𝑌 ∣ X = x] = 𝑔−1 (𝜂(x)).
17.2 Binary Response

To illustrate the use of a GLM we’ll focus on the case of binary responses variable
coded using 0 and 1. In practice, these 0 and 1s will code for two classes such
as yes/no, cat/dog, sick/healthy, etc.
1 yes
𝑌 ={
0 no
First, we define some notation that we will use throughout.
𝑝(x) = 𝑃 [𝑌 = 1 ∣ X = x]
With a binary (Bernoulli) response, we’ll mostly focus on the case when 𝑌 = 1,
since with only two possibilities, it is trivial to obtain probabilities when 𝑌 = 0.
𝑃 [𝑌 = 0 ∣ X = x] + 𝑃 [𝑌 = 1 ∣ X = x] = 1
𝑃 [𝑌 = 0 ∣ X = x] = 1 − 𝑝(x)
We now define the logistic regression model.
𝑝(x)
log ( ) = 𝛽0 + 𝛽1 𝑥1 + … + 𝛽𝑝−1 𝑥𝑝−1
1 − 𝑝(x)
Immediately we notice some similarities to ordinary linear regression, in partic-

ular, the right-hand side. This is our usual linear combination of the predictors.
We have our usual 𝑝 − 1 predictors for a total of 𝑝 𝛽 parameters. (Note, many
more machine learning focused texts will use 𝑝 as the number of predictors.
This is an arbitrary choice, but you should be aware of it.)
The left-hand side is called the log odds, which is the log of the odds. The
odds are the probability for a positive event (𝑌 = 1) divided by the probability
of a negative event (𝑌 = 0). So when the odds are 1, the two events have equal
probability. Odds greater than 1 favor a positive event. The opposite is true
when the odds are less than 1.
𝑝(x) 𝑃 [𝑌 = 1 ∣ X = x]
=
1 − 𝑝(x) 𝑃 [𝑌 = 0 ∣ X = x]
Essentially, the log odds are the logit transform applied to 𝑝(x).
𝜉
logit(𝜉) = log ( )
1−𝜉
It will also be useful to define the inverse logit, otherwise known as the “logistic”
or sigmoid function.
−1 𝑒𝜉 1
logit (𝜉) = =
1 + 𝑒𝜉 1 + 𝑒−𝜉
Note that for 𝑥 ∈ (−∞, ∞)), this function outputs values between 0 and 1.
Students often ask, where is the error term? The answer is that it’s something
that is specific to the normal model. First notice that the model with the error
term,
𝑌 = 𝛽0 + 𝛽1 𝑥1 + … + 𝛽𝑞 𝑥𝑞 + 𝜖, 𝜖 ∼ 𝑁 (0, 𝜎2 )
can instead be written as
𝑌 ∣ X = x ∼ 𝑁 (𝛽0 + 𝛽1 𝑥1 + … + 𝛽𝑞 𝑥𝑞 , 𝜎2 ).
While our main focus is on estimating the mean, 𝛽0 + 𝛽1 𝑥1 + … + 𝛽𝑞 𝑥𝑞 , there

is also another parameter, 𝜎2 which needs to be estimated. This is the result of
the normal distribution having two parameters.
With logistic regression, which uses the Bernoulli distribution, we only need to
estimate the Bernoulli distribution’s single parameter 𝑝(x), which happens to
be its mean.
𝑝(x)
log ( ) = 𝛽 0 + 𝛽 1 𝑥1 + … + 𝛽 𝑞 𝑥𝑞
1 − 𝑝(x)
So even though we introduced ordinary linear regression first, in some ways,

logistic regression is actually simpler.
Note that applying the inverse logit transformation allow us to obtain an ex-
pression for 𝑝(x).
𝑒𝛽0 +𝛽1 𝑥1 +⋯+𝛽𝑝−1 𝑥(𝑝−1)

𝑝(x) = 𝑃 [𝑌 = 1 ∣ X = x] =
1 + 𝑒𝛽0 +𝛽1 𝑥1 +⋯+𝛽𝑝−1 𝑥(𝑝−1)
17.2.1 Fitting Logistic Regression
With 𝑛 observations, we write the model indexed with 𝑖 to note that it is being
applied to each observation.
𝑝(xi )
log ( ) = 𝛽0 + 𝛽1 𝑥𝑖1 + ⋯ + 𝛽𝑝−1 𝑥𝑖(𝑝−1)
1 − 𝑝(xi )
We can apply the inverse logit transformation to obtain 𝑃 [𝑌𝑖 = 1 ∣ Xi = xi ] for

each observation. Since these are probabilities, it’s good that we used a function
that returns values between 0 and 1.
𝑒𝛽0 +𝛽1 𝑥𝑖1 +⋯+𝛽𝑝−1 𝑥𝑖(𝑝−1)

𝑝(xi ) = 𝑃 [𝑌𝑖 = 1 ∣ Xi = xi ] =
1 + 𝑒𝛽0 +𝛽1 𝑥𝑖1 +⋯+𝛽𝑝−1 𝑥𝑖(𝑝−1)
1
1 − 𝑝(xi ) = 𝑃 [𝑌𝑖 = 0 ∣ X = xi ] =
1 + 𝑒𝛽0 +𝛽1 𝑥𝑖1 +⋯+𝛽𝑝−1 𝑥𝑖(𝑝−1)
To “fit” this model, that is estimate the 𝛽 parameters, we will use maximum
likelihood.
𝛽 = [𝛽0 , 𝛽1 , 𝛽2 , 𝛽3 , … , 𝛽𝑝−1 ]
We first write the likelihood given the observed data.
𝑛
𝐿(𝛽) = ∏ 𝑃 [𝑌𝑖 = 𝑦𝑖 ∣ Xi = xi ]
𝑖=1
This is already technically a function of the 𝛽 parameters, but we’ll do some

rearrangement to make this more explicit.
𝑛
𝐿(𝛽) = ∏ 𝑝(xi )𝑦𝑖 (1 − 𝑝(xi ))(1−𝑦𝑖 )
𝑖=1
𝑛 𝑛
𝐿(𝛽) = ∏ 𝑝(xi ) ∏ (1 − 𝑝(xj ))
𝑖∶𝑦𝑖 =1 𝑗∶𝑦𝑗 =0
𝑒𝛽0 +𝛽1 𝑥𝑖1 +⋯+𝛽𝑝−1 𝑥𝑖(𝑝−1) 1

𝐿(𝛽) = ∏ 𝛽0 +𝛽1 𝑥𝑖1 +⋯+𝛽𝑝−1 𝑥𝑖(𝑝−1)
∏ 𝛽0 +𝛽1 𝑥𝑗1 +⋯+𝛽𝑝−1 𝑥𝑗(𝑝−1)
𝑖∶𝑦𝑖 =1 1 + 𝑒 𝑗∶𝑦𝑗 =0 1 + 𝑒
Unfortunately, unlike ordinary linear regression, there is no analytical solution

for this maximization problem. Instead, it will need to be solved numerically.
Fortunately, R will take care of this for us using an iteratively reweighted least
squares algorithm. (We’ll leave the details for a machine learning or optimization
course, which would likely also discuss alternative optimization strategies.)
17.2.2 Fitting Issues
We should note that, if there exists some 𝛽 ∗ such that
xi ⊤ 𝛽 ∗ > 0 ⟹ 𝑦𝑖 = 1
and
xi ⊤ 𝛽 ∗ < 0 ⟹ 𝑦𝑖 = 0
for all observations, then the MLE is not unique. Such data is said to be
separable.
This, and similar numeric issues related to estimated probabilities near 0 or 1,
will return a warning in R:
## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
When this happens, the model is still “fit,” but there are consequences, namely,
the estimated coefficients are highly suspect. This is an issue when then trying
to interpret the model. When this happens, the model will often still be useful
for creating a classifier, which will be discussed later. However, it is still subject
to the usual evaluations for classifiers to determine how well it is performing.
For details, see Modern Applied Statistics with S-PLUS, Chapter 7.
17.2.3 Simulation Examples

sim_logistic_data = function(sample_size = 25, beta_0 = -2, beta_1 = 3) {

x = rnorm(n = sample_size)
eta = beta_0 + beta_1 * x
p = 1 / (1 + exp(-eta))
y = rbinom(n = sample_size, size = 1, prob = p)
data.frame(y, x)
}
You might think, why not simply use ordinary linear regression? Even with a
binary response, our goal is still to model (some function of) E[𝑌 ∣ X = x].
However, with a binary response coded as 0 and 1, E[𝑌 ∣ X = x] = 𝑃 [𝑌 = 1 ∣
X = x] since
E[𝑌 ∣ X = x] = 1 ⋅ 𝑃 [𝑌 = 1 ∣ X = x] + 0 ⋅ 𝑃 [𝑌 = 0 ∣ X = x]
= 𝑃 [𝑌 = 1 ∣ X = x]
Then why can’t we just use ordinary linear regression to estimate E[𝑌 ∣ X = x],
and thus 𝑃 [𝑌 = 1 ∣ X = x]?
To investigate, let’s simulate data from the following model:
𝑝(x)
log ( ) = −2 + 3𝑥
1 − 𝑝(x)
Another way to write this, which better matches the function we’re using to
simulate the data:
𝑌𝑖 ∣ Xi = xi ∼ Bern(𝑝𝑖 )
1
𝑝𝑖 = 𝑝(xi ) =
1 + 𝑒−𝜂(xi )
𝜂(xi ) = −2 + 3𝑥𝑖
set.seed(1)
example_data = sim_logistic_data()
head(example_data)
## y x
## 1 0 -0.6264538
## 2 1 0.1836433
## 3 0 -0.8356286
## 4 1 1.5952808
## 5 0 0.3295078
## 6 0 -0.8204684
After simulating a dataset, we’ll then fit both ordinary linear regression and
logistic regression. Notice that currently the responses variable y is a numeric
variable that only takes values 0 and 1. Later we’ll see that we can also fit
logistic regression when the response is a factor variable with only two levels.
(Generally, having a factor response is preferred, but having a dummy response
allows us to make the comparison to using ordinary linear regression.)
# ordinary linear regression

fit_lm = lm(y ~ x, data = example_data)
# logistic regression
fit_glm = glm(y ~ x, data = example_data, family = binomial)
Notice that the syntax is extremely similar. What’s changed?
• lm() has become glm()

• We’ve added family = binomial argument
In a lot of ways, lm() is just a more specific version of glm(). For example
glm(y ~ x, data = example_data)
would actually fit the ordinary linear regression that we have seen in the past.
By default, glm() uses family = gaussian argument. That is, we’re fitting
a GLM with a normally distributed response and the identity function as the
link.
The family argument to glm() actually specifies both the distribution and the
link function. If not made explicit, the link function is chosen to be the canon-
ical link function, which is essentially the most mathematical convenient link
function. See ?glm and ?family for details. For example, the following code
explicitly specifies the link function which was previously used by default.
# more detailed call to glm for logistic regression

fit_glm = glm(y ~ x, data = example_data, family = binomial(link = "logit"))
Making predictions with an object of type glm is slightly different than making
predictions after fitting with lm(). In the case of logistic regression, with family
= binomial, we have:
type Returned
"link" [default] 𝜂(x)
̂ = log ( 1−𝑝(x)
̂
𝑝(x)
̂ )
𝑒𝜂(x)
̂
1
"response" 𝑝(x)
̂ = 1+𝑒𝜂(x)
̂ = 1+𝑒−𝜂(x)
̂
That is, type = "link" will get you the log odds, while type = "response"
will return the estimated mean, in this case, 𝑃 [𝑌 = 1 ∣ X = x] for each
observation.
predict(fit_glm, data.frame(x = 1.2)) # type = "link" as default
## 1
## 2.076166
predict(fit_glm, data.frame(x = 1.2), type = "link")
## 1
## 2.076166
predict(fit_glm, data.frame(x = 1.2), type = "response")
## 1
## 0.888565
plot(y ~ x, data = example_data,

pch = 20, ylab = "Estimated Probability",
main = "Ordinary vs Logistic Regression")
grid()
abline(fit_lm, col = "darkorange")
curve(predict(fit_glm, data.frame(x), type = "response"),
add = TRUE, col = "dodgerblue", lty = 2)
legend("topleft", c("Ordinary", "Logistic", "Data"), lty = c(1, 2, 0),
pch = c(NA, NA, 20), lwd = 2, col = c("darkorange", "dodgerblue", "black"))
Ordinary vs Logistic Regression
1.0
Ordinary
0.8 Logistic
Data
Estimated Probability
0.6
0.4
0.2
0.0
-2 -1 0 1
Since we only have a single predictor variable, we are able to graphically show
this situation. First, note that the data, is plotted using black dots. The
response y only takes values 0 and 1.
Next, we need to discuss the two added lines to the plot. The first, the solid
orange line, is the fitted ordinary linear regression.
The dashed blue curve is the estimated logistic regression. It is helpful to realize
that we are not plotting an estimate of 𝑌 for either. (Sometimes it might seem
that way with ordinary linear regression, but that isn’t what is happening.) For
̂
both, we are plotting E[𝑌 ∣ X = x], the estimated mean, which for a binary
response happens to be an estimate of 𝑃 [𝑌 = 1 ∣ X = x].
We immediately see why ordinary linear regression is not a good idea. While it
is estimating the mean, we see that it produces estimates that are less than 0!
(And in other situations could produce estimates greater than 1!) If the mean
is a probability, we don’t want probabilities less than 0 or greater than 1.
Enter logistic regression. Since the output of the inverse logit function is re-
stricted to be between 0 and 1, our estimates make much more sense as prob-
abilities. Let’s look at our estimated coefficients. (With a lot of rounding, for
simplicity.)
round(coef(fit_glm), 1)
## (Intercept) x
## -2.3 3.7
Our estimated model is then:
𝑝(x)
̂
log ( ) = −2.3 + 3.7𝑥
1 − 𝑝(x)
̂
Because we’re not directly estimating the mean, but instead a function of the
mean, we need to be careful with our interpretation of 𝛽1̂ = 3.7. This means
that, for a one unit increase in 𝑥, the log odds change (in this case increase) by
3.7. Also, since 𝛽1̂ is positive, as we increase 𝑥 we also increase 𝑝(x).
̂ To see
how much, we have to consider the inverse logistic function.
For example, we have:
𝑒−2.3+3.7⋅(−0.5)
𝑃 ̂ [𝑌 = 1 ∣ 𝑋 = −0.5] = ≈ 0.016
1 + 𝑒−2.3+3.7⋅(−0.5)
𝑒−2.3+3.7⋅(0)
𝑃 ̂ [𝑌 = 1 ∣ 𝑋 = 0] = ≈ 0.09112296
1 + 𝑒−2.3+3.7⋅(0)
𝑒−2.3+3.7⋅(1)
𝑃 ̂ [𝑌 = 1 ∣ 𝑋 = 1] = ≈ 0.8021839
1 + 𝑒−2.3+3.7⋅(1)
Now that we know we should use logistic regression, and not ordinary linear
regression, let’s consider another example. This time, let’s consider the model
𝑝(x)
log ( ) = 1 + −4𝑥.
1 − 𝑝(x)
Again, we could re-write this to better match the function we’re using to simu-
late the data:
1
𝑝𝑖 = 𝑝(xi ) =
1 + 𝑒−𝜂(xi )
𝜂(xi ) = 1 + −4𝑥𝑖
In this model, as 𝑥 increases, the log odds decrease.
set.seed(1)
example_data = sim_logistic_data(sample_size = 50, beta_0 = 1, beta_1 = -4)
We again simulate some observations form this model, then fit logistic regression.
fit_glm = glm(y ~ x, data = example_data, family = binomial)

main = "Logistic Regression, Decreasing Probability")
grid()
curve(boot::inv.logit(1 - 4 * x), add = TRUE, col = "darkorange", lty = 1)
legend("bottomleft", c("True Probability", "Estimated Probability", "Data"), lty = c(1,
Logistic Regression, Decreasing Probability

1.0
0.8
0.6
0.4
0.2
True Probability
Data
0.0
-2 -1 0 1
We see that this time, as 𝑥 increases, 𝑝(x)

̂ decreases.
Now let’s look at an example where the estimated probability doesn’t always
simply increase or decrease. Much like ordinary linear regression, the linear
combination of predictors can contain transformations of predictors (in this
case a quadratic term) and interactions.
sim_quadratic_logistic_data = function(sample_size = 25) {

x = rnorm(n = sample_size)
eta = -1.5 + 0.5 * x + x ^ 2
p = 1 / (1 + exp(-eta))
y = rbinom(n = sample_size, size = 1, prob = p)

data.frame(y, x)
}
𝑝(x)
log ( ) = −1.5 + 0.5𝑥 + 𝑥2 .
1 − 𝑝(x)
Again, we could re-write this to better match the function we’re using to simu-
late the data:
1
𝑝𝑖 = 𝑝(xi ) =
1 + 𝑒−𝜂(xi )
𝜂(xi ) = −1.5 + 0.5𝑥𝑖 + 𝑥2𝑖
set.seed(42)
example_data = sim_quadratic_logistic_data(sample_size = 50)
fit_glm = glm(y ~ x + I(x^2), data = example_data, family = binomial)

main = "Logistic Regression, Quadratic Relationship")
grid()
curve(boot::inv.logit(-1.5 + 0.5 * x + x ^ 2),
add = TRUE, col = "darkorange", lty = 1)
legend("bottomleft", c("True Probability", "Estimated Probability", "Data"), lty = c(1, 2, 0),
Logistic Regression, Quadratic Relationship
1.0
0.8
0.6
0.4
0.2
True Probability
Data
0.0
-2 -1 0 1 2
17.3 Working with Logistic Regression
While the logistic regression model isn’t exactly the same as the ordinary linear
regression model, because they both use a linear combination of the predictors
𝜂(x) = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + … + 𝛽𝑝−1 𝑥𝑝−1
working with logistic regression is very similar. Many of the things we did with
ordinary linear regression can be done with logistic regression in a very similar
fashion. For example,
• Testing for a single 𝛽 parameter

• Testing for a set of 𝛽 parameters
• Formula specification in R
• Interpreting parameters and estimates
• Confidence intervals for parameters
• Confidence intervals for mean response
• Variable selection
After some introduction to the new tests, we’ll demonstrate each of these using
an example.
17.3. WORKING WITH LOGISTIC REGRESSION 431
17.3.1 Testing with GLMs
Like ordinary linear regression, we’ll want to be able to perform hypothesis

testing. We’ll again want both single parameter, and multiple parameter tests.
17.3.2 Wald Test
In ordinary linear regression, we performed the test of
𝐻0 ∶ 𝛽 𝑗 = 0 vs 𝐻1 ∶ 𝛽𝑗 ≠ 0
using a 𝑡-test.
For the logistic regression model,
𝑝(x)
log ( ) = 𝛽0 + 𝛽1 𝑥1 + … + 𝛽𝑝−1 𝑥𝑝−1
1 − 𝑝(x)
we can again perform a test of
𝐻0 ∶ 𝛽 𝑗 = 0 vs 𝐻1 ∶ 𝛽𝑗 ≠ 0
however, the test statistic and its distribution are no longer 𝑡. We see that the
test statistic takes the same form
𝛽𝑗̂ − 𝛽𝑗 approx
𝑧= ∼ 𝑁 (0, 1)
SE[𝛽 ̂ ]𝑗
but now we are performing a 𝑧-test, as the test statistic is approximated by a

standard normal distribution, provided we have a large enough sample. (The
𝑡-test for ordinary linear regression, assuming the assumptions were correct, had
an exact distribution for any sample size.)
We’ll skip some of the exact details of the calculations, as R will obtain the
standard error for us. The use of this test will be extremely similar to the 𝑡-test
for ordinary linear regression. Essentially the only thing that changes is the
distribution of the test statistic.
17.3.3 Likelihood-Ratio Test
Consider the following full model,
𝑝(xi )
log ( ) = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽(𝑝−1) 𝑥𝑖(𝑝−1)
1 − 𝑝(xi )
This model has 𝑝 − 1 predictors, for a total of 𝑝 𝛽-parameters. We will denote

̂
the MLE of these 𝛽-parameters as 𝛽Full
Now consider a null (or reduced) model,
𝑝(xi )
log ( ) = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽(𝑞−1) 𝑥𝑖(𝑞−1)
1 − 𝑝(xi )
where 𝑞 < 𝑝. This model has 𝑞 − 1 predictors, for a total of 𝑞 𝛽-parameters. We

̂
will denote the MLE of these 𝛽-parameters as 𝛽Null
The difference between these two models can be codified by the null hypothesis
of a test.
𝐻0 ∶ 𝛽𝑞 = 𝛽𝑞+1 = ⋯ = 𝛽𝑝−1 = 0.
This implies that the reduced model is nested inside the full model.
We then define a test statistic, 𝐷,
̂ )
𝐿(𝛽Null ̂ )
𝐿(𝛽Full
𝐷 = −2 log ( ) = 2 log ( ̂ ) − ℓ(𝛽 ̂ ))
) = 2 (ℓ(𝛽Full Null
𝐿(𝛽 ̂ )
Full 𝐿(𝛽 ̂ ) Null
where 𝐿 denotes a likelihood and ℓ denotes a log-likelihood. For a large enough

sample, this test statistic has an approximate Chi-square distribution
approx
𝐷 ∼ 𝜒2𝑘
where 𝑘 = 𝑝 − 𝑞, the difference in number of parameters of the two models.

This test, which we will call the Likelihood-Ratio Test, will be the ana-
logue to the ANOVA 𝐹 -test for logistic regression. Interestingly, to perform the
Likelihood-Ratio Test, we’ll actually again use the anova() function in R!.
The Likelihood-Ratio Test is actually a rather general test, however, here we
have presented a specific application to nested logistic regression models.
17.3.4 SAheart Example
To illustrate the use of logistic regression, we will use the SAheart dataset from
the bestglm package.
# "leaps" is required for "bestglm"

if (!require("leaps")) install.packages("leaps")
if (!require("bestglm")) install.packages("bestglm")
library(leaps)
library(bestglm)
data("SAheart")
sbp tobacco ldl adiposity famhist typea obesity alcohol age chd
160 12.00 5.73 23.11 Present 49 25.30 97.20 52 1
144 0.01 4.41 28.61 Absent 55 28.87 2.06 63 1
118 0.08 3.48 32.28 Present 52 29.14 3.81 46 0
170 7.50 6.41 38.03 Present 51 31.99 24.26 58 1
134 13.60 3.50 27.78 Present 60 25.99 57.34 49 1
132 6.20 6.47 36.21 Present 62 30.77 14.14 45 0
This data comes from a retrospective sample of males in a heart-disease high-

risk region of the Western Cape, South Africa. The chd variable, which we will
use as a response, indicates whether or not coronary heart disease is present
in an individual. Note that this is coded as a numeric 0 / 1 variable. Using
this as a response with glm() it is important to indicate family = binomial,
otherwise ordinary linear regression will be fit. Later, we will see the use of a
factor variable response, which is actually preferred, as you cannot accidentally
fit ordinary linear regression.
The predictors are various measurements for each individual, many related to
heart health. For example sbp, systolic blood pressure, and ldl, low density
lipoprotein cholesterol. For full details, use ?SAheart.
We’ll begin by attempting to model the probability of coronary heart disease

based on low density lipoprotein cholesterol. That is, we will fit the model
𝑃 [chd = 1]
log ( ) = 𝛽0 + 𝛽ldl 𝑥ldl
1 − 𝑃 [chd = 1]
chd_mod_ldl = glm(chd ~ ldl, data = SAheart, family = binomial)

plot(jitter(chd, factor = 0.1) ~ ldl, data = SAheart, pch = 20,
ylab = "Probability of CHD", xlab = "Low Density Lipoprotein Cholesterol")
grid()
curve(predict(chd_mod_ldl, data.frame(ldl = x), type = "response"),
1.0
0.8
Probability of CHD
0.6
0.4
0.2
0.0
2 4 6 8 10 12 14
Low Density Lipoprotein Cholesterol
As before, we plot the data in addition to the estimated probabilities. Note that
we have “jittered” the data to make it easier to visualize, but the data do only
take values 0 and 1.
As we would expect, this plot indicates that as ldl increases, so does the prob-
ability of chd.
coef(summary(chd_mod_ldl))
## Estimate Std. Error z value Pr(>|z|)

## (Intercept) -1.9686681 0.27307908 -7.209150 5.630207e-13
## ldl 0.2746613 0.05163983 5.318787 1.044615e-07
To perform the test
𝐻0 ∶ 𝛽ldl = 0
we use the summary() function as we have done so many times before. Like the
𝑡-test for ordinary linear regression, this returns the estimate of the parameter,
its standard error, the relevant test statistic (𝑧), and its p-value. Here we have
an incredibly low p-value, so we reject the null hypothesis. The ldl variable
appears to be a significant predictor.
When fitting logistic regression, we can use the same formula syntax as ordinary
linear regression. So, to fit an additive model using all available predictors, we
use:
chd_mod_additive = glm(chd ~ ., data = SAheart, family = binomial)
We can then use the likelihood-ratio test to compare the two models. Specifi-
cally, we are testing
𝐻0 ∶ 𝛽sbp = 𝛽tobacco = 𝛽adiposity = 𝛽famhist = 𝛽typea = 𝛽obesity = 𝛽alcohol = 𝛽age = 0
We could manually calculate the test statistic,
-2 * as.numeric(logLik(chd_mod_ldl) - logLik(chd_mod_additive))
## [1] 92.13879
Or we could utilize the anova() function. By specifying test = "LRT", R will

use the likelihood-ratio test to compare the two models.
anova(chd_mod_ldl, chd_mod_additive, test = "LRT")
## Analysis of Deviance Table

##
## Model 1: chd ~ ldl
## Model 2: chd ~ sbp + tobacco + ldl + adiposity + famhist + typea + obesity +
## alcohol + age
## Resid. Df Resid. Dev Df Deviance Pr(>Chi)
## 1 460 564.28
## 2 452 472.14 8 92.139 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
We see that the test statistic that we had just calculated appears in the output.
The very small p-value suggests that we prefer the larger model.
While we prefer the additive model compared to the model with only a single
predictor, do we actually need all of the predictors in the additive model? To
select a subset of predictors, we can use a stepwise procedure as we did with
ordinary linear regression. Recall that AIC and BIC were defined in terms
of likelihoods. Here we demonstrate using AIC with a backwards selection
procedure.
chd_mod_selected = step(chd_mod_additive, trace = 0)

coef(chd_mod_selected)
## (Intercept) tobacco ldl famhistPresent typea

## -6.44644451 0.08037533 0.16199164 0.90817526 0.03711521
## age
## 0.05046038
We could again compare this model to the additive models.
𝐻0 ∶ 𝛽sbp = 𝛽adiposity = 𝛽obesity = 𝛽alcohol = 0
anova(chd_mod_selected, chd_mod_additive, test = "LRT")
## Analysis of Deviance Table

##
## Model 1: chd ~ tobacco + ldl + famhist + typea + age
## Model 2: chd ~ sbp + tobacco + ldl + adiposity + famhist + typea + obesity +
## alcohol + age
## Resid. Df Resid. Dev Df Deviance Pr(>Chi)
## 1 456 475.69
## 2 452 472.14 4 3.5455 0.471
Here it seems that we would prefer the selected model.
17.3.5 Confidence Intervals
We can create confidence intervals for the 𝛽 parameters using the confint()
function as we did with ordinary linear regression.
confint(chd_mod_selected, level = 0.99)
## Waiting for profiling to be done...
## 0.5 % 99.5 %
## (Intercept) -8.941825274 -4.18278990
## tobacco 0.015704975 0.14986616
## ldl 0.022923610 0.30784590
## famhistPresent 0.330033483 1.49603366
## typea 0.006408724 0.06932612
## age 0.024847330 0.07764277
Note that we could create intervals by rearranging the results of the Wald test
to obtain the Wald confidence interval. This would be given by
𝛽𝑗̂ ± 𝑧𝛼/2 ⋅ SE[𝛽𝑗̂ ].
However, R is using a slightly different approach based on a concept called the

profile likelihood. (The details of which we will omit.) Ultimately the intervals
reported will be similar, but the method used by R is more common in practice,
probably at least partially because it is the default approach in R. Check to
see how intervals using the formula above compare to those from the output of
confint(). (Or, note that using confint.default() will return the results of
calculating the Wald confidence interval.)
17.3.6 Confidence Intervals for Mean Response
Confidence intervals for the mean response require some additional thought.
With a “large enough” sample, we have
𝜂(x)
̂ − 𝜂(x) approx
∼ 𝑁 (0, 1)
SE[𝜂(x)]
̂
Then we can create an approximate (1 − 𝛼)% confidence intervals for 𝜂(x) using
𝜂(x)
̂ ± 𝑧𝛼/2 ⋅ SE[𝜂(x)]
̂
where 𝑧𝛼/2 is the critical value such that 𝑃 (𝑍 > 𝑧𝛼/2 ) = 𝛼/2.
This isn’t a particularly interesting interval. Instead, what we really want is an
interval for the mean response, 𝑝(x). To obtain an interval for 𝑝(x), we simply
apply the inverse logit transform to the endpoints of the interval for 𝜂.
−1 −1
(logit (𝜂(x)
̂ − 𝑧𝛼/2 ⋅ SE[𝜂(x)]),
̂ logit (𝜂(x)
̂ + 𝑧𝛼/2 ⋅ SE[𝜂(x)]))
̂
To demonstrate creating these intervals, we’ll consider a new observation.
new_obs = data.frame(
sbp = 148.0,
tobacco = 5,
ldl = 12,
adiposity = 31.23,
famhist = "Present",
typea = 47,
obesity = 28.50,
alcohol = 23.89,
age = 60
)
Fist, we’ll use the predict() function to obtain 𝜂(x)

̂ for this observation.
eta_hat = predict(chd_mod_selected, new_obs, se.fit = TRUE, type = "link")

eta_hat
## $fit
## 1
## 1.579545
##
## $se.fit
## [1] 0.4114796
##
## $residual.scale
## [1] 1
By setting se.fit = TRUE, R also computes SE[𝜂(x)].

̂ Note that we used type
= "link", but this is actually a default value. We added it here to stress that
the output from predict() will be the value of the link function.
z_crit = qnorm(0.975)
round(z_crit, 2)
## [1] 1.96
After obtaining the correct critical value, we can easily create a 95% confidence
interval for 𝜂(x).
eta_hat$fit + c(-1, 1) * z_crit * eta_hat$se.fit
## [1] 0.7730599 2.3860302
Now we simply need to apply the correct transformation to make this a confi-
dence interval for 𝑝(x), the probability of coronary heart disease for this observa-
tion. Note that the boot package contains functions logit() and inv.logit()
which are the logit and inverse logit transformations, respectively.
boot::inv.logit(eta_hat$fit + c(-1, 1) * z_crit * eta_hat$se.fit)
## [1] 0.6841824 0.9157558
Notice, as we would expect, the bounds of this interval are both between 0 and
1. Also, since both bounds of the interval for 𝜂(x) are positive, both bounds of
the interval for 𝑝(x) are greater than 0.5.
17.3.7 Formula Syntax
Without really thinking about it, we’ve been using our previous knowledge of
R’s model formula syntax to fit logistic regression.
17.3.7.1 Interactions
Let’s add an interaction between LDL and family history for the model we
selected.
chd_mod_interaction = glm(chd ~ alcohol + ldl + famhist + typea + age + ldl:famhist,

data = SAheart, family = binomial)
summary(chd_mod_interaction)
##
## Call:
## glm(formula = chd ~ alcohol + ldl + famhist + typea + age + ldl:famhist,
## family = binomial, data = SAheart)
##
## Deviance Residuals:
## -1.9082 -0.8308 -0.4550 0.9286 2.5152
##
## Coefficients:
## (Intercept) -6.043472 0.937186 -6.449 1.13e-10 ***
## alcohol 0.003800 0.004332 0.877 0.38033
## ldl 0.035593 0.071448 0.498 0.61837
## famhistPresent -0.733836 0.618131 -1.187 0.23515
## typea 0.036253 0.012172 2.978 0.00290 **
## age 0.062416 0.009723 6.419 1.37e-10 ***
## ldl:famhistPresent 0.314311 0.114922 2.735 0.00624 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 596.11 on 461 degrees of freedom
## Residual deviance: 477.46 on 455 degrees of freedom
## AIC: 491.46
##
## Number of Fisher Scoring iterations: 5
Based on the 𝑧-test seen in the above summary, this interaction is significant.
The effect of LDL on the probability of CHD is different depending on family
history.
17.3.7.2 Polynomial Terms
Let’s take the previous model, and now add a polynomial term.
chd_mod_int_quad = glm(chd ~ alcohol + ldl + famhist + typea + age + ldl:famhist + I(ld

data = SAheart, family = binomial)
summary(chd_mod_int_quad)
##
## Call:
## glm(formula = chd ~ alcohol + ldl + famhist + typea + age + ldl:famhist +
## I(ldl^2), family = binomial, data = SAheart)
##
## Deviance Residuals:
## -1.8953 -0.8311 -0.4556 0.9276 2.5204
##
## Coefficients:
## (Intercept) -6.096747 1.065952 -5.720 1.07e-08 ***
## alcohol 0.003842 0.004350 0.883 0.37716
## ldl 0.056876 0.214420 0.265 0.79081
## famhistPresent -0.723769 0.625167 -1.158 0.24698
## typea 0.036248 0.012171 2.978 0.00290 **
## age 0.062299 0.009788 6.365 1.95e-10 ***
## I(ldl^2) -0.001587 0.015076 -0.105 0.91617
## ldl:famhistPresent 0.311615 0.117559 2.651 0.00803 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 596.11 on 461 degrees of freedom
## Residual deviance: 477.45 on 454 degrees of freedom
## AIC: 493.45
##
## Number of Fisher Scoring iterations: 5
Unsurprisingly, since this additional transformed variable wasn’t intelligently

chosen, it is not significant. However, this does allow us to stress the fact that
the syntax notation that we had been using with lm() works basically exactly
the same for glm(), however now we understand that this is specifying the linear
combination of predictions, 𝜂(x).
That is, the above fits the model
𝑝(x)
log ( ) = 𝛽0 +𝛽1 𝑥alcohol +𝛽2 𝑥ldl +𝛽3 𝑥famhist +𝛽4 𝑥typea +𝛽5 𝑥age +𝛽6 𝑥ldl 𝑥famhist +𝛽7 𝑥2ldl
1 − 𝑝(x)
You may have realized this before we actually explicitly wrote it down!
17.3.8 Deviance
You have probably noticed that the output from summary() is also very similar
to that of ordinary linear regression. One difference, is the “deviance” being
reported. The Null deviance is the deviance for the null model, that is, a
model with no predictors. The Residual deviance is the deviance for the
model that was fit.
Deviance compares the model to a saturated model. (Without repeated obser-
vations, a saturated model is a model that fits perfectly, using a parameter for
each observation.) Essentially, deviance is a generalized residual sum of squares
for GLMs. Like RSS, deviance decreases as the model complexity increases.
deviance(chd_mod_ldl)
## [1] 564.2788
deviance(chd_mod_selected)
## [1] 475.6856
deviance(chd_mod_additive)
## [1] 472.14
Note that these are nested, and we see that deviance does decrease as the model
size becomes larger. So while a lower deviance is better, if the model becomes
too big, it may be overfitting. Note that R also outputs AIC in the summary,
which will penalize according to model size, to prevent overfitting.
17.4 Classification
So far we’ve mostly used logistic regression to estimate class probabilities. The
somewhat obvious next step is to use these probabilities to make “predictions,”
which in this context, we would call classifications. Based on the values of the
predictors, should an observation be classified as 𝑌 = 1 or as 𝑌 = 0?
Suppose we didn’t need to estimate probabilities from data, and instead, we
actually knew both
𝑝(x) = 𝑃 [𝑌 = 1 ∣ X = x]
and
1 − 𝑝(x) = 𝑃 [𝑌 = 0 ∣ X = x].
With this information, classifying observations based on the values of the pre-
dictors is actually extremely easy. Simply classify an observation to the class
(0 or 1) with the larger probability. In general, this result is called the Bayes
Classifier,
𝐶 𝐵 (x) = argmax 𝑃 [𝑌 = 𝑘 ∣ X = x].

𝑘
For a binary response, that is,
̂ 1 p(x) > 0.5

𝐶(x) ={
0 p(x) ≤ 0.5
Simply put, the Bayes classifier (not to be confused with the Naive Bayes Clas-
sifier) minimizes the probability of misclassification by classifying each obser-
vation to the class with the highest probability. Unfortunately, in practice,
we won’t know the necessary probabilities to directly use the Bayes classifier.
17.4. CLASSIFICATION 443
Instead we’ll have to use estimated probabilities. So to create a classifier that

seeks to minimize misclassifications, we would use,
̂
𝐶(x) = argmax 𝑃 ̂ [𝑌 = 𝑘 ∣ X = x].
𝑘
In the case of a binary response since 𝑝(x)

̂ = 1 − 𝑝(x),
̂ this becomes
̂ 1 p(x)
̂ > 0.5
𝐶(x) ={
0 p(x)
̂ ≤ 0.5
Using this simple classification rule, we can turn logistic regression into a clas-
sifier. To use logistic regression for classification, we first use logistic regression
to obtain estimated probabilities, 𝑝(x),
̂ then use these in conjunction with the
above classification rule.
Logistic regression is just one of many ways that these probabilities could be
estimated. In a course completely focused on machine learning, you’ll learn
many additional ways to do this, as well as methods to directly make classifica-
tions without needing to first estimate probabilities. But since we had already
introduced logistic regression, it makes sense to discuss it in the context of
classification.
17.4.1 spam Example
To illustrate the use of logistic regression as a classifier, we will use the spam
dataset from the kernlab package.
# install.packages("kernlab")
library(kernlab)
data("spam")
tibble::as.tibble(spam)
## Warning: às.tibble()` was deprecated in tibble 2.0.0.

## i Please use às_tibble()` instead.
## i The signature and semantics have changed, see `?as_tibble`.
## # A tibble: 4,601 x 58
## make address all num3d our over remove internet order mail receive
## <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 0 0.64 0.64 0 0.32 0 0 0 0 0 0
## 2 0.21 0.28 0.5 0 0.14 0.28 0.21 0.07 0 0.94 0.21
## 3 0.06 0 0.71 0 1.23 0.19 0.19 0.12 0.64 0.25 0.38
## 4 0 0 0 0 0.63 0 0.31 0.63 0.31 0.63 0.31
## 5 0 0 0 0 0.63 0 0.31 0.63 0.31 0.63 0.31

## 6 0 0 0 0 1.85 0 0 1.85 0 0 0
## 7 0 0 0 0 1.92 0 0 0 0 0.64 0.96
## 8 0 0 0 0 1.88 0 0 1.88 0 0 0
## 9 0.15 0 0.46 0 0.61 0 0.3 0 0.92 0.76 0.76
## 10 0.06 0.12 0.77 0 0.19 0.32 0.38 0 0.06 0 0
## # ... with 4,591 more rows, and 47 more variables: will <dbl>, people <dbl>,
## # report <dbl>, addresses <dbl>, free <dbl>, business <dbl>, email <dbl>,
## # you <dbl>, credit <dbl>, your <dbl>, font <dbl>, num000 <dbl>, money <dbl>,
## # hp <dbl>, hpl <dbl>, george <dbl>, num650 <dbl>, lab <dbl>, labs <dbl>,
## # telnet <dbl>, num857 <dbl>, data <dbl>, num415 <dbl>, num85 <dbl>,
## # technology <dbl>, num1999 <dbl>, parts <dbl>, pm <dbl>, direct <dbl>,
## # cs <dbl>, meeting <dbl>, original <dbl>, project <dbl>, re <dbl>, ...
This dataset, created in the late 1990s at Hewlett-Packard Labs, contains 4601
emails, of which 1813 are considered spam. The remaining are not spam. (Which
for simplicity, we might call, ham.) Additional details can be obtained by using
?spam or by visiting the UCI Machine Learning Repository.
The response variable, type, is a factor with levels that label each email as
spam or nonspam. When fitting models, nonspam will be the reference level,
𝑌 = 0, as it comes first alphabetically.
is.factor(spam$type)
## [1] TRUE
levels(spam$type)
## [1] "nonspam" "spam"
Many of the predictors (often called features in machine learning) are engineered
based on the emails. For example, charDollar is the number of times an email
contains the $ character. Some variables are highly specific to this dataset, for
example george and num650. (The name and area code for one of the researchers
whose emails were used.) We should keep in mind that this dataset was created
based on emails sent to academic type researchers in the 1990s. Any results we
derive probably won’t generalize to modern emails for the general public.
To get started, we’ll first test-train split the data.
set.seed(42)
# spam_idx = sample(nrow(spam), round(nrow(spam) / 2))
spam_idx = sample(nrow(spam), 1000)
spam_trn = spam[spam_idx, ]
spam_tst = spam[-spam_idx, ]
We’ve used a somewhat small train set relative to the total size of the dataset.
In practice it should likely be larger, but this is simply to keep training time
low for illustration and rendering of this document.
fit_caps = glm(type ~ capitalTotal,

data = spam_trn, family = binomial)
fit_selected = glm(type ~ edu + money + capitalTotal + charDollar,
fit_additive = glm(type ~ .,
fit_over = glm(type ~ capitalTotal * (.),
data = spam_trn, family = binomial, maxit = 75)
We’ll fit four logistic regressions, each more complex than the previous. Note
that we’re suppressing two warnings. The first we briefly mentioned previously.
## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
Note that, when we receive this warning, we should be highly suspicious of the
parameter estimates.
coef(fit_selected)
## (Intercept) edu money capitalTotal charDollar

## -1.1199744712 -1.9837988840 0.9784675298 0.0007757011 11.5772904667
However, the model can still be used to create a classifier, and we will evaluate
that classifier on its own merits.
We also, “suppressed” the warning:
## Warning: glm.fit: algorithm did not converge
In reality, we didn’t actually suppress it, but instead changed maxit to 75, when
fitting the model fit_over. This was enough additional iterations to allow
the iteratively reweighted least squares algorithm to converge when fitting the
model.
17.4.2 Evaluating Classifiers
The metric we’ll be most interested in for evaluating the overall performance
of a classifier is the misclassification rate. (Sometimes, instead accuracy is
reported, which is instead the proportion of correction classifications, so both

metrics serve the same purpose.)
1 𝑛
Misclass(𝐶,̂ Data) = ∑ 𝐼(𝑦𝑖 ≠ 𝐶(x
̂ ))
i
𝑛 𝑖=1
̂ )
0 𝑦𝑖 = 𝐶(x
̂ )) = {
𝐼(𝑦𝑖 ≠ 𝐶(x i
i ̂ )
1 𝑦𝑖 ≠ 𝐶(x i
When using this metric on the training data, it will have the same issues as RSS
did for ordinary linear regression, that is, it will only go down.
# training misclassification rate

mean(ifelse(predict(fit_caps) > 0, "spam", "nonspam") != spam_trn$type)
## [1] 0.339
mean(ifelse(predict(fit_selected) > 0, "spam", "nonspam") != spam_trn$type)
## [1] 0.224
mean(ifelse(predict(fit_additive) > 0, "spam", "nonspam") != spam_trn$type)
## [1] 0.066
mean(ifelse(predict(fit_over) > 0, "spam", "nonspam") != spam_trn$type)
## [1] 0.118
Because of this, training data isn’t useful for evaluating, as it would suggest that
we should always use the largest possible model, when in reality, that model is
likely overfitting. Recall, a model that is too complex will overfit. A model that
is too simple will underfit. (We’re looking for something in the middle.)
To overcome this, we’ll use cross-validation as we did with ordinary linear re-
gression, but this time we’ll cross-validate the misclassification rate. To do so,
we’ll use the cv.glm() function from the boot library. It takes arguments for
the data (in this case training), a model fit via glm(), and K, the number of
folds. It returns delta, a vector of the average mean-squared error (MSE) and
the bias-corrected MSE. See ?cv.glm for details.
Previously, for cross-validating RMSE in ordinary linear regression, we used
LOOCV. We certainly could do that here. However, with logistic regression,
we no longer have the clever trick that would allow us to obtain a LOOCV
metric without needing to fit the model 𝑛 times. So instead, we’ll use 5-fold
cross-validation. (5 and 10 fold are the most common in practice.) Instead of
leaving a single observation out repeatedly, we’ll leave out a fifth of the data.
Essentially we’ll repeat the following process 5 times:
• Randomly set aside a fifth of the data (each observation will only be held-
out once)
• Train model on remaining data
• Evaluate misclassification rate on held-out data
The 5-fold cross-validated misclassification rate will be the average of these

misclassification rates. By only needing to refit the model 5 times, instead of 𝑛
times, we will save a lot of computation time.
library(boot)
set.seed(1)
cv.glm(spam_trn, fit_caps, K = 5)$delta[1]
## [1] 0.2166961
cv.glm(spam_trn, fit_selected, K = 5)$delta[1]
## [1] 0.1587043
cv.glm(spam_trn, fit_additive, K = 5)$delta[1]
## [1] 0.08684467
cv.glm(spam_trn, fit_over, K = 5)$delta[1]
## [1] 0.161
Note that we’re suppressing warnings again here. (Now there would be a lot
more, since were fitting a total of 20 models.)
Based on these results, fit_caps and fit_selected are underfitting relative
to fit_additive. Similarly, fit_over is overfitting relative to fit_additive.
Thus, based on these results, we prefer the classifier created based on the logistic
regression fit and stored in fit_additive.
Going forward, to evaluate and report on the efficacy of this classifier, we’ll use
the test dataset. We’re going to take the position that the test data set should
never be used in training, which is why we used cross-validation within the

training dataset to select a model. Even though cross-validation uses hold-out
sets to generate metrics, at some point all of the data is used for training.
To quickly summarize how well this classifier works, we’ll create a confusion
matrix.
Figure 17.1: Confusion Matrix
It further breaks down the classification errors into false positives and false
negatives.
make_conf_mat = function(predicted, actual) {

table(predicted = predicted, actual = actual)
}
Let’s explicitly store the predicted values of our classifier on the test dataset.
spam_tst_pred = ifelse(predict(fit_additive, spam_tst) > 0,

"spam",
"nonspam")
spam_tst_pred = ifelse(predict(fit_additive, spam_tst, type = "response") > 0.5,
"spam",
"nonspam")
The previous two lines of code produce the same output, that is the same
predictions, since
𝜂(x) = 0 ⟺ 𝑝(x) = 0.5

Now we’ll use these predictions to create a confusion matrix.
(conf_mat_50 = make_conf_mat(predicted = spam_tst_pred, actual = spam_tst$type))
## actual
## predicted nonspam spam
## nonspam 2057 157
## spam 127 1260
P TP + FN
Prev = =
Total Obs Total Obs
table(spam_tst$type) / nrow(spam_tst)
##
## nonspam spam
## 0.6064982 0.3935018
First, note that to be a reasonable classifier, it needs to outperform the obvious

classifier of simply classifying all observations to the majority class. In this case,
classifying everything as non-spam for a test misclassification rate of 0.3935018
Next, we can see that using the classifier created from fit_additive, only
a total of 137 + 161 = 298 from the total of 3601 emails in the test set are
misclassified. Overall, the accuracy in the test set it
mean(spam_tst_pred == spam_tst$type)
## [1] 0.921133
In other words, the test misclassification is
mean(spam_tst_pred != spam_tst$type)
## [1] 0.07886698
This seems like a decent classifier…

However, are all errors created equal? In this case, absolutely not. The 137
non-spam emails that were marked as spam (false positives) are a problem. We
can’t allow important information, say, a job offer, to miss our inbox and get
sent to the spam folder. On the other hand, the 161 spam email that would
make it to an inbox (false negatives) are easily dealt with, just delete them.
Instead of simply evaluating a classifier based on its misclassification rate (or
accuracy), we’ll define two additional metrics, sensitivity and specificity. Note
that these are simply two of many more metrics that can be considered. The
Wikipedia page for sensitivity and specificity details a large number of metrics
that can be derived from a confusion matrix.
Sensitivity is essentially the true positive rate. So when sensitivity is high, the
number of false negatives is low.
TP TP
Sens = True Positive Rate = =
P TP + FN
Here we have an R function to calculate the sensitivity based on the confusion
matrix. Note that this function is good for illustrative purposes, but is easily
broken. (Think about what happens if there are no “positives” predicted.)
get_sens = function(conf_mat) {
conf_mat[2, 2] / sum(conf_mat[, 2])
}
Specificity is essentially the true negative rate. So when specificity is high, the
number of false positives is low.
TN TN
Spec = True Negative Rate = =
N TN + FP
get_spec = function(conf_mat) {
conf_mat[1, 1] / sum(conf_mat[, 1])
}
We calculate both based on the confusion matrix we had created for our classi-
fier.
get_sens(conf_mat_50)
## [1] 0.8892025
get_spec(conf_mat_50)
## [1] 0.9418498
Recall that we had created this classifier using a probability of 0.5 as a “cutoff”
for how observations should be classified. Now we’ll modify this cutoff. We’ll
see that by modifying the cutoff, 𝑐, we can improve sensitivity or specificity at
the expense of the overall accuracy (misclassification rate).
̂ 1 p(x)
̂ >c
𝐶(x) ={
0 p(x)
̂ ≤c
Additionally, if we change the cutoff to improve sensitivity, we’ll decrease speci-

ficity, and vice versa.
First let’s see what happens when we lower the cutoff from 0.5 to 0.1 to create
a new classifier, and thus new predictions.
spam_tst_pred_10 = ifelse(predict(fit_additive, spam_tst, type = "response") > 0.1,

"spam",
"nonspam")
This is essentially decreasing the threshold for an email to be labeled as spam, so

far more emails will be labeled as spam. We see that in the following confusion
matrix.
(conf_mat_10 = make_conf_mat(predicted = spam_tst_pred_10, actual = spam_tst$type))
## actual
## nonspam 1583 29
## spam 601 1388
Unfortunately, while this does greatly reduce false negatives, false positives have
almost quadrupled. We see this reflected in the sensitivity and specificity.
## [1] 0.9795342
## [1] 0.7248168
This classifier, using 0.1 instead of 0.5 has a higher sensitivity, but a much lower
specificity. Clearly, we should have moved the cutoff in the other direction.
Let’s try 0.9.
spam_tst_pred_90 = ifelse(predict(fit_additive, spam_tst, type = "response") > 0.9,

"spam",
"nonspam")
This is essentially increasing the threshold for an email to be labeled as spam,

so far fewer emails will be labeled as spam. Again, we see that in the following
confusion matrix.
(conf_mat_90 = make_conf_mat(predicted = spam_tst_pred_90, actual = spam_tst$type))

## actual
## nonspam 2136 537
## spam 48 880
This is the result we’re looking for. We have far fewer false positives. While
sensitivity is greatly reduced, specificity has gone up.
## [1] 0.6210303
## [1] 0.978022
While this is far fewer false positives, is it acceptable though? Still probably
not. Also, don’t forget, this would actually be a terrible spam detector today
since this is based on data from a very different era of the internet, for a very
specific set of people. Spam has changed a lot since the 90s! (Ironically, machine
learning is probably partially to blame.)
This chapter has provided a rather quick introduction to classification, and thus,
machine learning. For a more complete coverage of machine learning, An Intro-
duction to Statistical Learning is a highly recommended resource. Additionally,
R for Statistical Learning has been written as a supplement which provides ad-
ditional detail on how to perform these methods using R. The classification and
logistic regression chapters might be useful.
We should note that the code to perform classification using logistic regression
is presented in a way that illustrates the concepts to the reader. In practice, you
may prefer to use a more general machine learning pipeline such as caret in R.
This will streamline processes for creating predictions and generating evaluation
metrics.
17.5 R Markdown
• logistic.Rmd

Chapter 18
Beyond
“End? No, the journey doesn’t end here.”

— J.R.R. Tolkien
• Understand the roadmap to continued education about models and the R

programming language.
18.1 What’s Next

So you’ve completed STAT 420, where do you go from here? Now that you un-
derstand the basics of linear modeling, there is a wide world of applied statistics
waiting to be explored. We’ll briefly detail some resources and discuss how they
relate to what you have learned in STAT 420.
18.2 RStudio
RStudio has recently released version 1.0! This is exciting for a number of rea-
sons, especially the release of R Notebooks. R Notebooks combine the RMark-
down you have already learned with the ability to work interactively.
18.3 Tidy Data

In this textbook, much of the data we have seen has been nice and tidy. It was
rectangular where each row was an observation and each column was a variable.
453
454 CHAPTER 18. BEYOND
This is not always the case! Many packages have been developed to deal with
data, and force it into a nice format, which is called tidy data, that we can then
use for modeling. Often during analysis, this is where a large portion of your
time will be spent.
The R community has started to call this collection of packages the Tidyverse.
It was once called the Hadleyverse, as Hadley Wickham has authored so many
of the packages. Hadley is writing a book called R for Data Science which
describes the use of many of these packages. (And also how to use some to
make the modeling process better!) This book is a great starting point for
diving deeper into the R community. The two main packages are dplyr and
tidyr both of which are used internally in RStudio.
18.4 Visualization
In this course, we have mostly used the base plotting methods in R. When work-
ing with tidy data, many users prefer to use the ggplot2 package, also devel-
oped by Hadley Wickham. RStudio provides a rather detailed “cheat sheet” for
working with ggplot2. The community maintains a graph gallery of examples.
Use of the manipulate package with RStudio gives the ability to quickly change
a static graphic to become interactive.
18.5 Web Applications
RStudio has made it incredibly easy to create data products through the use
of Shiny, which allows for the creation of web applications with R. RStudio
maintains an ever-growing tutorial and gallery of examples.
18.6 Experimental Design
In the ANOVA chapter, we briefly discussed experimental design. This topic

could easily be its own class, and is currently an area of revitalized interest with
the rise of A/B testing. Two more classic statistical references include Statistics
for Experimenters by Box, Hunter, and Hunter as well as Design and Analysis of
Experiments by Douglas Montgomery. There are several R packages for design
of experiments, listed in the CRAN Task View.
18.7. MACHINE LEARNING 455
18.7 Machine Learning

Using models for prediction is the key focus of machine learning. There are
many methods, each with its own package, however R has a wonderful package
called caret, Classification And REgression Training, which provides a unified
interface to training these models. It also contains various utilities for data
processing and visualization that are useful for predictive modeling.
Applied Predictive Modeling by Max Kuhn, the author of the caret package,
is a good general resource for predictive modeling, which obviously utilizes R.
An Introduction to Statistical Learning by James, Witten, Hastie, and Tibshi-
rani is a gentle introduction to machine learning from a statistical perspective
which uses R and picks up right where this course stops. This is based on the
often referenced The Elements of Statistical Learning by Hastie, Tibshirani, and
Friedman. Both are freely available online.
18.7.1 Deep Learning
While it probably isn’t the best tool for the job, R now has the ability to train
deep neural networks via TensorFlow.
18.8 Time Series

In this class we have only considered independent data. What if data is depen-
dent? Time Series is the area of statistics which deals with this issue, and could
easily span multiple courses.
The primary textbook for STAT 429: Time Series Analysis at the University of
Illinois that is free is:
• Time Series Analysis and Its Applications: With R Examples by Shumway

and Stoffer
Some tutorials:
• Little Book of R for Time Series

• Quick R: Time Series and Forecasting
• TSA: Start to Finish Examples
When performing time series analysis in R you should be aware of the many
packages that are useful for analysis. It should be hard to avoid the forecast
and zoo packages. Often the most difficult part will be dealing with time and
date data. Make sure you are utilizing one of the many packages that help with
this.
456 CHAPTER 18. BEYOND
18.9 Bayesianism
In this class, we have worked within the frequentist view of statistics. There is
an entire alternative universe of Bayesian statistics.
Doing Bayesian Data Analysis: A Tutorial with R, JAGS, and Stan by John
Kruschke is a great introduction to the topic. It introduces the world of prob-
abilistic programming, in particular Stan, which can be used in both R and
Python.
18.10 High Performance Computing

Often R will be called a “slow” language, for two reasons. One, because many
do not understand R. Two, because sometimes it really is. Luckily, it is easy
to extend R via the Rcpp package to allow for faster code. Many modern R
packages utilize Rcpp to achieve better performance.
18.11 Further R Resources

Also, don’t forget that previously in this book we have outlined a large number
of R resources. Now that you’ve gotten started with R many of these will be
much more useful.
If any of these topics interest you, and you would like more information, please
don’t hesitate to start a discussion on the forums!
:)
Chapter 19
Appendix
457

Learn R For Applied Statistics

Uploaded by

Copyright:

Available Formats

Learn R For Applied Statistics

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Learn R For Applied Statistics

Uploaded by

Copyright:

Available Formats

Applied Statistics with R

3 Data and Programming 21

3.3.1 Control Flow . . . . . . . . . . . . . . . . . . . . . . . . . 51

5 Probability and Statistics in R 67

7 Simple Linear Regression 89

8 Inference for Simple Linear Regression 123

9 Multiple Linear Regression 153

10 Model Building 185

11 Categorical Predictors and Interactions 195

12 Analysis of Variance 231

12.1 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231

12.2 Two-Sample t-Test . . . . . . . . . . . . . . . . . . . . . . . . . . 232

12.3 One-Way ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . 235

12.3.1 Factor Variables . . . . . . . . . . . . . . . . . . . . . . . 242

12.3.2 Some Simulation . . . . . . . . . . . . . . . . . . . . . . . 243

12.3.3 Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

12.4 Post Hoc Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

12.5 Two-Way ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . 249

12.6 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

13 Model Diagnostics 261

13.1 Model Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . 261

13.2 Checking Assumptions . . . . . . . . . . . . . . . . . . . . . . . . 263

13.2.1 Fitted versus Residuals Plot . . . . . . . . . . . . . . . . . 264

13.2.2 Breusch-Pagan Test . . . . . . . . . . . . . . . . . . . . . 270

13.2.3 Histograms . . . . . . . . . . . . . . . . . . . . . . . . . . 272

13.2.4 Q-Q Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . 273

13.2.5 Shapiro-Wilk Test . . . . . . . . . . . . . . . . . . . . . . 280

13.3 Unusual Observations . . . . . . . . . . . . . . . . . . . . . . . . 282

13.3.1 Leverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284

13.3.2 Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290

13.3.3 Influence . . . . . . . . . . . . . . . . . . . . . . . . . . . 292

13.4 Data Analysis Examples . . . . . . . . . . . . . . . . . . . . . . . 294

13.4.1 Good Diagnostics . . . . . . . . . . . . . . . . . . . . . . . 294

13.4.2 Suspect Diagnostics . . . . . . . . . . . . . . . . . . . . . 298

13.5 R Markdown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301

16 Variable Selection and Model Building 383

16.3 Higher Order Terms . . . . . . . . . . . . . . . . . . . . . . . . . 408

17 Logistic Regression 417

18.6 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . 454

Welcome to Applied Statistics with R!

1.1 About This Book

We use the quantity 𝑝 to refer to the number of 𝛽 parameters in a linear model,

Figure 1.1: This work is licensed under a Creative Commons Attribution-

2.1 Getting Started

• R, the actual programming language.

• On Windows: Alt + Shift + K

• Hadley Wickham Style Guide from Advanced R

2.2 Basic Calculations

Addition, Subtraction, Multiplication, and Division

2.3 Getting Help

• Describe what you expect the code to do.

2.4 Installing Packages

Data and Programming

3.1 Data Types

3.2 Data Structures

Dimension Homogeneous Heterogeneous

Many operations in R make heavy use of vectors. Vectors in R are indexed

As an aside, there is a long history of the assignment operator in R, partially

stay consistent. Also, if working on a larger collaborative project, you should