Statistical Regression

and Classification
From Linear Models to
Machine Learning
CHAPMAN & HALL/CRC
Texts in Statistical Science Series
Series Editors
Joseph K. Blitzstein, Harvard University, USA
Julian J. Faraway, University of Bath, UK
Martin Tanner, Northwestern University, USA
Jim Zidek, University of British Columbia, Canada
Statistical Theory: A Concise Introduction Problem Solving: A Statistician’s Guide,
F. Abramovich and Y. Ritov Second Edition
Practical Multivariate Analysis, Fifth Edition C. Chatfield
A. Afifi, S. May, and V.A. Clark Statistics for Technology: A Course in Applied
Practical Statistics for Medical Research Statistics, Third Edition
D.G. Altman C. Chatfield

Interpreting Data: A First Course Analysis of Variance, Design, and Regression :

in Statistics Linear Modeling for Unbalanced Data,
A.J.B. Anderson Second Edition
R. Christensen
Introduction to Probability with R
K. Baclawski Bayesian Ideas and Data Analysis: An
Introduction for Scientists and Statisticians
Linear Algebra and Matrix Analysis for
Statistics R. Christensen, W. Johnson, A. Branscum,
S. Banerjee and A. Roy and T.E. Hanson

Modern Data Science with R Modelling Binary Data, Second Edition

B. S. Baumer, D. T. Kaplan, and N. J. Horton D. Collett

Mathematical Statistics: Basic Ideas and Modelling Survival Data in Medical Research,
Selected Topics, Volume I, Third Edition
Second Edition D. Collett
P. J. Bickel and K. A. Doksum Introduction to Statistical Methods for
Mathematical Statistics: Basic Ideas and Clinical Trials
Selected Topics, Volume II T.D. Cook and D.L. DeMets
P. J. Bickel and K. A. Doksum Applied Statistics: Principles and Examples
Analysis of Categorical Data with R D.R. Cox and E.J. Snell
C. R. Bilder and T. M. Loughin Multivariate Survival Analysis and Competing
Statistical Methods for SPC and TQM Risks
D. Bissell M. Crowder
Introduction to Probability Statistical Analysis of Reliability Data
J. K. Blitzstein and J. Hwang M.J. Crowder, A.C. Kimber,
T.J. Sweeting, and R.L. Smith
Bayesian Methods for Data Analysis,
Third Edition An Introduction to Generalized
B.P. Carlin and T.A. Louis Linear Models, Third Edition
A.J. Dobson and A.G. Barnett
Second Edition
R. Caulcutt Nonlinear Time Series: Theory, Methods, and
Applications with R Examples
The Analysis of Time Series: An Introduction, R. Douc, E. Moulines, and D.S. Stoffer
Sixth Edition
C. Chatfield Introduction to Optimization Methods and
Their Applications in Statistics
Introduction to Multivariate Analysis B.S. Everitt
C. Chatfield and A.J. Collins
Extending the Linear Model with R: Graphics for Statistics and Data Analysis with R
Generalized Linear, Mixed Effects and K.J. Keen
Nonparametric Regression Models, Second Mathematical Statistics
Edition K. Knight
J.J. Faraway
Introduction to Functional Data Analysis
Linear Models with R, Second Edition P. Kokoszka and M. Reimherr
J.J. Faraway
Introduction to Multivariate Analysis:
A Course in Large Sample Theory Linear and Nonlinear Modeling
T.S. Ferguson S. Konishi
Multivariate Statistics: A Practical Nonparametric Methods in Statistics with SAS
Approach Applications
B. Flury and H. Riedwyl O. Korosteleva
Readings in Decision Analysis Modeling and Analysis of Stochastic Systems,
S. French Third Edition
Discrete Data Analysis with R: Visualization V.G. Kulkarni
and Modeling Techniques for Categorical and Exercises and Solutions in Biostatistical Theory
Count Data L.L. Kupper, B.H. Neelon, and S.M. O’Brien
M. Friendly and D. Meyer
Exercises and Solutions in Statistical Theory
Markov Chain Monte Carlo: L.L. Kupper, B.H. Neelon, and S.M. O’Brien
Stochastic Simulation for Bayesian Inference,
Second Edition Design and Analysis of Experiments with R
D. Gamerman and H.F. Lopes J. Lawson

Bayesian Data Analysis, Third Edition Design and Analysis of Experiments with SAS
A. Gelman, J.B. Carlin, H.S. Stern, D.B. Dunson, J. Lawson
A. Vehtari, and D.B. Rubin A Course in Categorical Data Analysis
Multivariate Analysis of Variance and T. Leonard
Repeated Measures: A Practical Approach for Statistics for Accountants
Behavioural Scientists S. Letchford
D.J. Hand and C.C. Taylor Introduction to the Theory of Statistical
Practical Longitudinal Data Analysis Inference
D.J. Hand and M. Crowder H. Liero and S. Zwanzig
Logistic Regression Models Statistical Theory, Fourth Edition
J.M. Hilbe B.W. Lindgren
Richly Parameterized Linear Models: Stationary Stochastic Processes: Theory and
Additive, Time Series, and Spatial Models Applications
Using Random Effects G. Lindgren
J.S. Hodges Statistics for Finance
Statistics for Epidemiology E. Lindström, H. Madsen, and J. N. Nielsen
N.P. Jewell The BUGS Book: A Practical Introduction to
Stochastic Processes: An Introduction, Bayesian Analysis
Second Edition D. Lunn, C. Jackson, N. Best, A. Thomas, and
P.W. Jones and P. Smith D. Spiegelhalter
The Theory of Linear Models Introduction to General and Generalized
B. Jørgensen Linear Models
Pragmatics of Uncertainty H. Madsen and P. Thyregod
J.B. Kadane Time Series Analysis
Principles of Uncertainty H. Madsen
J.B. Kadane Pólya Urn Models
H. Mahmoud
Randomization, Bootstrap and Monte Carlo Sampling Methodologies with Applications
Methods in Biology, Third Edition P.S.R.S. Rao
B.F.J. Manly A First Course in Linear Model Theory
Statistical Regression and Classification: From N. Ravishanker and D.K. Dey
Linear Models to Machine Learning Essential Statistics, Fourth Edition
N. Matloff D.A.G. Rees
Introduction to Randomized Controlled Stochastic Modeling and Mathematical
Clinical Trials, Second Edition Statistics: A Text for Statisticians and
J.N.S. Matthews Quantitative Scientists
Statistical Rethinking: A Bayesian Course with F.J. Samaniego
Examples in R and Stan Statistical Methods for Spatial Data Analysis
R. McElreath O. Schabenberger and C.A. Gotway
Statistical Methods in Agriculture and Bayesian Networks: With Examples in R
Experimental Biology, Second Edition M. Scutari and J.-B. Denis
R. Mead, R.N. Curnow, and A.M. Hasted
Large Sample Methods in Statistics
Statistics in Engineering: A Practical Approach P.K. Sen and J. da Motta Singer
A.V. Metcalfe
Spatio-Temporal Methods in Environmental
Statistical Inference: An Integrated Approach, Epidemiology
Second Edition G. Shaddick and J.V. Zidek
H. S. Migon, D. Gamerman, and
F. Louzada Decision Analysis: A Bayesian Approach
J.Q. Smith
Beyond ANOVA: Basics of Applied Statistics
Analysis of Failure and Survival Data
R.G. Miller, Jr.
P. J. Smith
A Primer on Linear Models
Applied Statistics: Handbook of GENSTAT
J.F. Monahan
Analyses
Stochastic Processes: From Applications to E.J. Snell and H. Simpson
Theory
P.D Moral and S. Penev Applied Nonparametric Statistical Methods,
Fourth Edition
Applied Stochastic Modelling, Second Edition P. Sprent and N.C. Smeeton
B.J.T. Morgan
Data Driven Statistical Methods
Elements of Simulation P. Sprent
B.J.T. Morgan
Generalized Linear Mixed Models:
Probability: Methods and Measurement Modern Concepts, Methods and Applications
A. O’Hagan W. W. Stroup
Introduction to Statistical Limit Theory Survival Analysis Using S: Analysis of
A.M. Polansky Time-to-Event Data
Applied Bayesian Forecasting and Time Series M. Tableman and J.S. Kim
Analysis Applied Categorical and Count Data Analysis
A. Pole, M. West, and J. Harrison W. Tang, H. He, and X.M. Tu
Statistics in Research and Development, Elementary Applications of Probability Theory,
Time Series: Modeling, Computation, and Second Edition
Inference H.C. Tuckwell
R. Prado and M. West
Introduction to Statistical Inference and Its
Essentials of Probability Theory for Applications with R
Statisticians M.W. Trosset
M.A. Proschan and P.A. Shaw
Understanding Advanced Statistical Methods
Introduction to Statistical Process Control P.H. Westfall and K.S.S. Henning
P. Qiu
Statistical Process Control: Theory and Epidemiology: Study Design and
Practice, Third Edition Data Analysis, Third Edition
G.B. Wetherill and D.W. Brown M. Woodward
Generalized Additive Models: Practical Data Analysis for Designed
An Introduction with R, Second Edition Experiments
S. Wood B.S. Yandell
 Texts in Statistical Science

Statistical Regression
and Classification
From Linear Models to
Machine Learning

Norman Matloff
University of California, Davis, USA
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2017 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-4987-1091-6 (Paperback)

International Standard Book Number-13: 978-1-138-06656-5 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts
have been made to publish reliable data and information, but the author and publisher cannot assume
responsibility for the validity of all materials or the consequences of their use. The authors and publishers
have attempted to trace the copyright holders of all material reproduced in this publication and apologize
to copyright holders if permission to publish in this form has not been obtained. If any copyright material
has not been acknowledged please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information storage
or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access
www.copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc.
(CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization
that provides licenses and registration for a variety of users. For organizations that have been granted a
photocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
Contents

Preface xxix

List of Symbols xxxvii

1 Setting the Stage 1

1.1 Example: Predicting Bike-Sharing
Activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Example of the Prediction Goal: Body Fat . . . . . . . . . 2
1.3 Example: Who Clicks Web Ads? . . . . . . . . . . . . . . . 3
1.4 Approach to Prediction . . . . . . . . . . . . . . . . . . . . 4
1.5 A Note about E(), Samples and Populations . . . . . . . . . 5
1.6 Example of the Description Goal: Do
Baseball Players Gain Weight As They Age? . . . . . . . . 6
1.6.1 Prediction vs. Description . . . . . . . . . . . . . . . 7
1.6.2 A First Estimator . . . . . . . . . . . . . . . . . . . 9
1.6.3 A Possibly Better Estimator, Using a Linear Model 10
1.7 Parametric vs. Nonparametric Models . . . . . . . . . . . . 15
1.8 Example: Click-Through Rate . . . . . . . . . . . . . . . . . 15
1.9 Several Predictor Variables . . . . . . . . . . . . . . . . . . 17
1.9.1 Multipredictor Linear Models . . . . . . . . . . . . . 18

ix
x CONTENTS

1.9.1.1 Estimation of Coefficients . . . . . . . . . . 18

1.9.1.2 The Description Goal . . . . . . . . . . . . 19
1.9.2 Nonparametric Regression Estimation: k-NN . . . . 19
1.9.2.1 Looking at Nearby Points . . . . . . . . . . 20
1.9.2.2 Measures of Nearness . . . . . . . . . . . . 20
1.9.2.3 The k-NN Method, and Tuning Parameters 21
1.9.2.4 Nearest-Neighbor Analysis in the regtools
Package . . . . . . . . . . . . . . . . . . . . 21
1.9.2.5 Example: Baseball Player Data . . . . . . . 22
1.10 After Fitting a Model, How Do We Use It for Prediction? . 22
1.10.1 Parametric Settings . . . . . . . . . . . . . . . . . . 22
1.10.2 Nonparametric Settings . . . . . . . . . . . . . . . . 23
1.10.3 The Generic predict() Function . . . . . . . . . . . . 23
1.11 Overfitting, and the Variance-Bias
Tradeoff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.11.1 Intuition . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.11.2 Example: Student Evaluations of Instructors . . . . 26
1.12 Cross-Validation . . . . . . . . . . . . . . . . . . . . . . . . 26
1.12.1 Linear Model Case . . . . . . . . . . . . . . . . . . . 28
1.12.1.1 The Code . . . . . . . . . . . . . . . . . . . 28
1.12.1.2 Applying the Code . . . . . . . . . . . . . . 29
1.12.2 k-NN Case . . . . . . . . . . . . . . . . . . . . . . . 29
1.12.3 Choosing the Partition Sizes . . . . . . . . . . . . . 30
1.13 Important Note on Tuning Parameters . . . . . . . . . . . . 31
1.14 Rough Rule of Thumb . . . . . . . . . . . . . . . . . . . . . 32
1.15 Example: Bike-Sharing Data . . . . . . . . . . . . . . . . . 32
1.15.1 Linear Modeling of µ(t) . . . . . . . . . . . . . . . . 33
1.15.2 Nonparametric Analysis . . . . . . . . . . . . . . . . 38
CONTENTS xi

1.16 Interaction Terms, Including Quadratics . . . . . . . . . . . 38

1.16.1 Example: Salaries of Female Programmers
and Engineers . . . . . . . . . . . . . . . . . . . . . . 39
1.16.2 Fitting Separate Models . . . . . . . . . . . . . . . . 42
1.16.3 Saving Your Work . . . . . . . . . . . . . . . . . . . 43
1.16.4 Higher-Order Polynomial Models . . . . . . . . . . . 43
1.17 Classification Techniques . . . . . . . . . . . . . . . . . . . . 44
1.17.1 It’s a Regression Problem! . . . . . . . . . . . . . . . 44
1.17.2 Example: Bike-Sharing Data . . . . . . . . . . . . . 45
1.18 Crucial Advice: Don’t Automate, Participate! . . . . . . . . 47
1.19 Mathematical Complements . . . . . . . . . . . . . . . . . . 48
1.19.1 Indicator Random Variables . . . . . . . . . . . . . . 48
1.19.2 Mean Squared Error of an Estimator . . . . . . . . . 48
1.19.3 µ(t) Minimizes Mean Squared Prediction Error . . . 49
1.19.4 µ(t) Minimizes the Misclassification Rate . . . . . . 50
1.19.5 Some Properties of Conditional Expectation . . . . . 52
1.19.5.1 Conditional Expectation As a Random
Variable . . . . . . . . . . . . . . . . . . . . 52
1.19.5.2 The Law of Total Expectation . . . . . . . 53
1.19.5.3 Law of Total Variance . . . . . . . . . . . . 54
1.19.5.4 Tower Property . . . . . . . . . . . . . . . 54
1.19.5.5 Geometric View . . . . . . . . . . . . . . . 54
1.20 Computational Complements . . . . . . . . . . . . . . . . . 55
1.20.1 CRAN Packages . . . . . . . . . . . . . . . . . . . . 55
1.20.2 The Function tapply() and Its Cousins . . . . . . . 56
1.20.3 The Innards of the k-NN Code . . . . . . . . . . . . 58
1.20.4 Function Dispatch . . . . . . . . . . . . . . . . . . . 59
1.21 Centering and Scaling . . . . . . . . . . . . . . . . . . . . . 60
xii CONTENTS

1.22 Exercises: Data, Code and Math Problems . . . . . . . . . 61

2 Linear Regression Models 65

2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.2 The “Error Term” . . . . . . . . . . . . . . . . . . . . . . . 67
2.3 Random- vs. Fixed-X Cases . . . . . . . . . . . . . . . . . . 67
2.4 Least-Squares Estimation . . . . . . . . . . . . . . . . . . . 68
2.4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . 68
2.4.2 Matrix Formulations . . . . . . . . . . . . . . . . . . 70
2.4.3 (2.18) in Matrix Terms . . . . . . . . . . . . . . . . . 71
2.4.4 Using Matrix Operations to Minimize (2.18) . . . . . 71
2.4.5 Models without an Intercept Term . . . . . . . . . . 72
2.5 A Closer Look at lm() Output . . . . . . . . . . . . . . . . 73
2.5.1 Statistical Inference . . . . . . . . . . . . . . . . . . 74
2.6 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.6.1 Classical . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.6.2 Motivation: the Multivariate Normal
Distribution Family . . . . . . . . . . . . . . . . . . 76
2.7 Unbiasedness and Consistency . . . . . . . . . . . . . . . . . 79
2.7.1 βb Is Unbiased . . . . . . . . . . . . . . . . . . . . . . 79
2.7.2 Bias As an Issue/Nonissue . . . . . . . . . . . . . . . 80

2.7.3 βb Is Statistically Consistent . . . . . . . . . . . . . . 80

2.8 Inference under Homoscedasticity . . . . . . . . . . . . . . . 81
2.8.1 Review: Classical Inference on a Single Mean . . . . 81
2.8.2 Back to Reality . . . . . . . . . . . . . . . . . . . . . 82
2.8.3 The Concept of a Standard Error . . . . . . . . . . . 83
2.8.4 Extension to the Regression Case . . . . . . . . . . . 83
2.8.5 Example: Bike-Sharing Data . . . . . . . . . . . . . 86
CONTENTS xiii

2.9 Collective Predictive Strength of the X (j) . . . . . . . . . . 88

2.9.1 Basic Properties . . . . . . . . . . . . . . . . . . . . 88
2.9.2 Definition of R2 . . . . . . . . . . . . . . . . . . . . 90
2.9.3 Bias Issues . . . . . . . . . . . . . . . . . . . . . . . 91
2.9.4 Adjusted-R2 . . . . . . . . . . . . . . . . . . . . . . 92
2.9.5 The “Leaving-One-Out Method” . . . . . . . . . . . 94
2.9.6 Extensions of LOOM . . . . . . . . . . . . . . . . . . 95
2.9.7 LOOM for k-NN . . . . . . . . . . . . . . . . . . . . 95
2.9.8 Other Measures . . . . . . . . . . . . . . . . . . . . . 96
2.10 The Practical Value of p-Values — Small OR Large . . . . 96
2.10.1 Misleadingly Small p-Values . . . . . . . . . . . . . . 97
2.10.1.1 Example: Forest Cover Data . . . . . . . . 97
2.10.1.2 Example: Click Through Data . . . . . . . 98
2.10.2 Misleadingly LARGE p-Values . . . . . . . . . . . . 99
2.10.3 The Verdict . . . . . . . . . . . . . . . . . . . . . . . 100
2.11 Missing Values . . . . . . . . . . . . . . . . . . . . . . . . . 100
2.12 Mathematical Complements . . . . . . . . . . . . . . . . . . 101
2.12.1 Covariance Matrices . . . . . . . . . . . . . . . . . . 101
2.12.2 The Multivariate Normal Distribution Family . . . . 103
2.12.3 The Central Limit Theorem . . . . . . . . . . . . . . 104
2.12.4 Details on Models Without a Constant Term . . . . 104
2.12.5 Unbiasedness of the Least-Squares Estimator . . . . 105
2.12.6 Consistency of the Least-Squares Estimator . . . . . 106
2.12.7 Biased Nature of S . . . . . . . . . . . . . . . . . . . 108
2.12.8 The Geometry of Conditional Expectation . . . . . . 108
2.12.8.1 Random Variables As Inner Product Spaces 108
2.12.8.2 Projections . . . . . . . . . . . . . . . . . . 109
xiv CONTENTS

2.12.8.3 Conditional Expectations As Projections . 110

2.12.9 Predicted Values and Error Terms Are Uncorrelated 111
2.12.10 Classical “Exact” Inference . . . . . . . . . . . . . . 112
2.12.11 Asymptotic (p + 1)-Variate Normality of βb . . . . . 113
2.13 Computational Complements . . . . . . . . . . . . . . . . . 115
2.13.1 Details of the Computation of (2.28) . . . . . . . . . 115
2.13.2 R Functions for the Multivariate Normal
Distribution Family . . . . . . . . . . . . . . . . . . 116
2.13.2.1 Example: Simulation Computation of a
Bivariate Normal Quantity . . . . . . . . . 116
2.13.3 More Details of ’lm’ Objects . . . . . . . . . . . . . 118
2.14 Exercises: Data, Code and Math Problems . . . . . . . . . 120

3 Homoscedasticity and Other Assumptions in Practice 123

3.1 Normality Assumption . . . . . . . . . . . . . . . . . . . . . 124
3.2 Independence Assumption — Don’t
Overlook It . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.2.1 Estimation of a Single Mean . . . . . . . . . . . . . 125
3.2.2 Inference on Linear Regression Coefficients . . . . . 126
3.2.3 What Can Be Done? . . . . . . . . . . . . . . . . . . 126
3.2.4 Example: MovieLens Data . . . . . . . . . . . . . . 127
3.3 Dropping the Homoscedasticity
Assumption . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.3.1 Robustness of the Homoscedasticity Assumption . . 131
3.3.2 Weighted Least Squares . . . . . . . . . . . . . . . . 133
3.3.3 A Procedure for Valid Inference . . . . . . . . . . . . 135
3.3.4 The Methodology . . . . . . . . . . . . . . . . . . . . 135
3.3.5 Example: Female Wages . . . . . . . . . . . . . . . . 136
3.3.6 Simulation Test . . . . . . . . . . . . . . . . . . . . . 137
CONTENTS xv

3.3.7 Variance-Stabilizing Transformations . . . . . . . . . 137

3.3.8 The Verdict . . . . . . . . . . . . . . . . . . . . . . . 139
3.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 139
3.5 Computational Complements . . . . . . . . . . . . . . . . . 140
3.5.1 The R merge() Function . . . . . . . . . . . . . . . . 140
3.6 Mathematical Complements . . . . . . . . . . . . . . . . . . 141
3.6.1 The Delta Method . . . . . . . . . . . . . . . . . . . 141
3.6.2 Distortion Due to Transformation . . . . . . . . . . 142
3.7 Exercises: Data, Code and Math Problems . . . . . . . . . 143

4 Generalized Linear and Nonlinear Models 147

4.1 Example: Enzyme Kinetics Model . . . . . . . . . . . . . . 148
4.2 The Generalized Linear Model (GLM) . . . . . . . . . . . . 150
4.2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . 150
4.2.2 Poisson Regression . . . . . . . . . . . . . . . . . . . 151
4.2.3 Exponential Families . . . . . . . . . . . . . . . . . . 152
4.2.4 R’s glm() Function . . . . . . . . . . . . . . . . . . 153
4.3 GLM: the Logistic Model . . . . . . . . . . . . . . . . . . . 154
4.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . 155
4.3.2 Example: Pima Diabetes Data . . . . . . . . . . . . 158
4.3.3 Interpretation of Coefficients . . . . . . . . . . . . . 159
4.3.4 The predict() Function Again . . . . . . . . . . . . . 161
4.3.5 Overall Prediction Accuracy . . . . . . . . . . . . . . 162
4.3.6 Example: Predicting Spam E-mail . . . . . . . . . . 163
4.3.7 Linear Boundary . . . . . . . . . . . . . . . . . . . . 164
4.4 GLM: the Poisson Regression Model . . . . . . . . . . . . . 165
4.5 Least-Squares Computation . . . . . . . . . . . . . . . . . . 166
xvi CONTENTS

4.5.1 The Gauss-Newton Method . . . . . . . . . . . . . . 166

4.5.2 Eicker-White Asymptotic Standard Errors . . . . . . 168
4.5.3 Example: Bike Sharing Data . . . . . . . . . . . . . 171
4.5.4 The “Elephant in the Room”: Convergence
Issues . . . . . . . . . . . . . . . . . . . . . . . . . . 172
4.6 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 173
4.7 Computational Complements . . . . . . . . . . . . . . . . . 173
4.7.1 GLM Computation . . . . . . . . . . . . . . . . . . . 173
4.7.2 R Factors . . . . . . . . . . . . . . . . . . . . . . . . 174
4.8 Mathematical Complements . . . . . . . . . . . . . . . . . . 175
4.8.1 Maximum Likelihood Estimation . . . . . . . . . . . 175
4.9 Exercises: Data, Code and Math Problems . . . . . . . . . 176

5 Multiclass Classification Problems 179

5.1 Key Notation . . . . . . . . . . . . . . . . . . . . . . . . . . 179
5.2 Key Equations . . . . . . . . . . . . . . . . . . . . . . . . . 180
5.3 Estimating the Functions µi (t) . . . . . . . . . . . . . . . . 182
5.4 How Do We Use Models for Prediction? . . . . . . . . . . . 182
5.5 One vs. All or All vs. All? . . . . . . . . . . . . . . . . . . . 183
5.5.1 Which Is Better? . . . . . . . . . . . . . . . . . . . . 184
5.5.2 Example: Vertebrae Data . . . . . . . . . . . . . . . 184
5.5.3 Intuition . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.5.4 Example: Letter Recognition Data . . . . . . . . . . 186
5.5.5 Example: k-NN on the Letter Recognition Data . . 187
5.5.6 The Verdict . . . . . . . . . . . . . . . . . . . . . . . 188
5.6 Fisher Linear Discriminant Analysis . . . . . . . . . . . . . 188
5.6.1 Background . . . . . . . . . . . . . . . . . . . . . . . 189
5.6.2 Derivation . . . . . . . . . . . . . . . . . . . . . . . . 189
CONTENTS xvii

5.6.3 Example: Vertebrae Data . . . . . . . . . . . . . . . 190

5.6.3.1 LDA Code and Results . . . . . . . . . . . 190
5.7 Multinomial Logistic Model . . . . . . . . . . . . . . . . . . 191
5.7.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . 191
5.7.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . 192
5.7.3 Example: Vertebrae Data . . . . . . . . . . . . . . . 192
5.8 The Issue of “Unbalanced” (and Balanced) Data . . . . . . 193
5.8.1 Why the Concern Regarding Balance? . . . . . . . . 194
5.8.2 A Crucial Sampling Issue . . . . . . . . . . . . . . . 195
5.8.2.1 It All Depends on How We Sample . . . . 195
5.8.2.2 Remedies . . . . . . . . . . . . . . . . . . . 197
5.8.3 Example: Letter Recognition . . . . . . . . . . . . . 198
5.9 Going Beyond Using the 0.5 Threshold . . . . . . . . . . . . 200
5.9.1 Unequal Misclassification Costs . . . . . . . . . . . . 200
5.9.2 Revisiting the Problem of Unbalanced Data . . . . . 201
5.9.3 The Confusion Matrix and the ROC Curve . . . . . 202
5.9.3.1 Code . . . . . . . . . . . . . . . . . . . . . 203
5.9.3.2 Example: Spam Data . . . . . . . . . . . . 203
5.10 Mathematical Complements . . . . . . . . . . . . . . . . . . 203
5.10.1 Classification via Density Estimation . . . . . . . . . 203
5.10.1.1 Methods for Density Estimation . . . . . . 204
5.10.2 Time Complexity Comparison, OVA vs. AVA . . . . 205
5.10.3 Optimal Classification Rule for
Unequal Error Costs . . . . . . . . . . . . . . . . . . 206
5.11 Computational Complements . . . . . . . . . . . . . . . . . 207
5.11.1 R Code for OVA and AVA Logit Analysis . . . . . . 207
5.11.2 ROC Code . . . . . . . . . . . . . . . . . . . . . . . 210
xviii CONTENTS

5.12 Exercises: Data, Code and Math Problems . . . . . . . . . 211

6 Model Fit Assessment and Improvement 215

6.1 Aims of This Chapter . . . . . . . . . . . . . . . . . . . . . 215
6.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
6.3 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
6.4 Goals of Model Fit-Checking . . . . . . . . . . . . . . . . . 217
6.4.1 Prediction Context . . . . . . . . . . . . . . . . . . . 217
6.4.2 Description Context . . . . . . . . . . . . . . . . . . 218
6.4.3 Center vs. Fringes of the Data Set . . . . . . . . . . 218
6.5 Example: Currency Data . . . . . . . . . . . . . . . . . . . 219
6.6 Overall Measures of Model Fit . . . . . . . . . . . . . . . . 220
6.6.1 R-Squared, Revisited . . . . . . . . . . . . . . . . . . 221
6.6.2 Cross-Validation, Revisited . . . . . . . . . . . . . . 222
6.6.3 Plotting Parametric Fit Against a
Nonparametric One . . . . . . . . . . . . . . . . . . 222
6.6.4 Residuals vs. Smoothing . . . . . . . . . . . . . . . . 223
6.7 Diagnostics Related to Individual
Predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
6.7.1 Partial Residual Plots . . . . . . . . . . . . . . . . . 225
6.7.2 Plotting Nonparametric Fit Against
Each Predictor . . . . . . . . . . . . . . . . . . . . . 227
6.7.3 The freqparcoord Package . . . . . . . . . . . . . . . 229
6.7.3.1 Parallel Coordinates . . . . . . . . . . . . . 229
6.7.3.2 The freqparcoord Package . . . . . . . . . . 229
6.7.3.3 The regdiag() Function . . . . . . . . . . . 230
6.8 Effects of Unusual Observations on Model Fit . . . . . . . . 232
6.8.1 The influence() Function . . . . . . . . . . . . . . . . 232
CONTENTS xix

6.8.1.1 Example: Currency Data . . . . . . . . . . 233

6.8.2 Use of freqparcoord for Outlier Detection . . . . . . 235
6.9 Automated Outlier Resistance . . . . . . . . . . . . . . . . . 236
6.9.1 Median Regression . . . . . . . . . . . . . . . . . . . 236
6.9.2 Example: Currency Data . . . . . . . . . . . . . . . 238
6.10 Example: Vocabulary Acquisition . . . . . . . . . . . . . . . 238
6.11 Classification Settings . . . . . . . . . . . . . . . . . . . . . 241
6.11.1 Example: Pima Diabetes Study . . . . . . . . . . . . 242
6.12 Improving Fit . . . . . . . . . . . . . . . . . . . . . . . . . . 245
6.12.1 Deleting Terms from the Model . . . . . . . . . . . . 245
6.12.2 Adding Polynomial Terms . . . . . . . . . . . . . . . 247
6.12.2.1 Example: Currency Data . . . . . . . . . . 247
6.12.2.2 Example: Census Data . . . . . . . . . . . 248
6.12.3 Boosting . . . . . . . . . . . . . . . . . . . . . . . . . 251
6.12.3.1 View from the 30,000-Foot Level . . . . . . 251
6.12.3.2 Performance . . . . . . . . . . . . . . . . . 253
6.13 A Tool to Aid Model Selection . . . . . . . . . . . . . . . . 254
6.14 Special Note on the Description Goal . . . . . . . . . . . . . 255
6.15 Computational Complements . . . . . . . . . . . . . . . . . 255
6.15.1 Data Wrangling for the Currency Dataset . . . . . . 255
6.15.2 Data Wrangling for the Word Bank Dataset . . . . . 256
6.16 Mathematical Complements . . . . . . . . . . . . . . . . . . 257
6.16.1 The Hat Matrix . . . . . . . . . . . . . . . . . . . . 257
6.16.2 Matrix Inverse Update . . . . . . . . . . . . . . . . . 259
6.16.3 The Median Minimizes Mean Absolute
Deviation . . . . . . . . . . . . . . . . . . . . . . . . 260
6.16.4 The Gauss-Markov Theorem . . . . . . . . . . . . . 261
xx CONTENTS

6.16.4.1 Lagrange Multipliers . . . . . . . . . . . . . 261

6.16.4.2 Proof of Gauss-Markov . . . . . . . . . . . 262
6.17 Exercises: Data, Code and Math Problems . . . . . . . . . 264

7 Disaggregating Regressor Effects 267

7.1 A Small Analytical Example . . . . . . . . . . . . . . . . . . 268
7.2 Example: Baseball Player Data . . . . . . . . . . . . . . . . 270
7.3 Simpson’s Paradox . . . . . . . . . . . . . . . . . . . . . . . 274
7.3.1 Example: UCB Admissions Data (Logit) . . . . . . . 274
7.3.2 The Verdict . . . . . . . . . . . . . . . . . . . . . . . 278
7.4 Unobserved Predictor Variables . . . . . . . . . . . . . . . . 278
7.4.1 Instrumental Variables (IVs) . . . . . . . . . . . . . 279
7.4.1.1 The IV Method . . . . . . . . . . . . . . . 281
7.4.1.2 Two-Stage Least Squares: . . . . . . . . . . 283
7.4.1.3 Example: Years of Schooling . . . . . . . . 284
7.4.1.4 The Verdict . . . . . . . . . . . . . . . . . . 286
7.4.2 Random Effects Models . . . . . . . . . . . . . . . . 286
7.4.2.1 Example: Movie Ratings Data . . . . . . . 287
7.4.3 Multiple Random Effects . . . . . . . . . . . . . . . 288
7.4.4 Why Use Random/Mixed Effects Models? . . . . . . 288
7.5 Regression Function Averaging . . . . . . . . . . . . . . . . 289
7.5.1 Estimating the Counterfactual . . . . . . . . . . . . 290
7.5.1.1 Example: Job Training . . . . . . . . . . . 290
7.5.2 Small Area Estimation: “Borrowing from
Neighbors” . . . . . . . . . . . . . . . . . . . . . . . 291
7.5.3 The Verdict . . . . . . . . . . . . . . . . . . . . . . . 295
7.6 Multiple Inference . . . . . . . . . . . . . . . . . . . . . . . 295
7.6.1 The Frequent Occurence of Extreme Events . . . . . 295
CONTENTS xxi

7.6.2 Relation to Statistical Inference . . . . . . . . . . . . 296

7.6.3 The Bonferroni Inequality . . . . . . . . . . . . . . . 297
7.6.4 Scheffe’s Method . . . . . . . . . . . . . . . . . . . . 298
7.6.5 Example: MovieLens Data . . . . . . . . . . . . . . 300
7.6.6 The Verdict . . . . . . . . . . . . . . . . . . . . . . . 303
7.7 Computational Complements . . . . . . . . . . . . . . . . . 303
7.7.1 MovieLens Data Wrangling . . . . . . . . . . . . . . 303
7.7.2 More Data Wrangling in the MovieLens Example . . 303
7.8 Mathematical Complements . . . . . . . . . . . . . . . . . . 306
7.8.1 Iterated Projections . . . . . . . . . . . . . . . . . . 306
7.8.2 Standard Errors for RFA . . . . . . . . . . . . . . . 307
7.8.3 Asymptotic Chi-Square Distributions . . . . . . . . . 308
7.9 Exercises: Data, Code and Math Problems . . . . . . . . . 309

8 Shrinkage Estimators 311

8.1 Relevance of James-Stein to Regression Estimation . . . . . 312
8.2 Multicollinearity . . . . . . . . . . . . . . . . . . . . . . . . 313
8.2.1 What’s All the Fuss About? . . . . . . . . . . . . . . 313
8.2.2 A Simple Guiding Model . . . . . . . . . . . . . . . 313
8.2.3 Checking for Multicollinearity . . . . . . . . . . . . . 314
8.2.3.1 The Variance Inflation Factor . . . . . . . . 314
8.2.3.2 Example: Currency Data . . . . . . . . . . 315
8.2.4 What Can/Should One Do? . . . . . . . . . . . . . . 315
8.2.4.1 Do Nothing . . . . . . . . . . . . . . . . . . 315
8.2.4.2 Eliminate Some Predictors . . . . . . . . . 316
8.2.4.3 Employ a Shrinkage Method . . . . . . . . 316
8.3 Ridge Regression . . . . . . . . . . . . . . . . . . . . . . . . 316
xxii CONTENTS

8.3.1 Alternate Definitions . . . . . . . . . . . . . . . . . . 317

8.3.2 Choosing the Value of λ . . . . . . . . . . . . . . . . 318
8.3.3 Example: Currency Data . . . . . . . . . . . . . . . 319
8.4 The LASSO . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
8.4.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . 321
8.4.2 The lars Package . . . . . . . . . . . . . . . . . . . . 322
8.4.3 Example: Currency Data . . . . . . . . . . . . . . . 322
8.4.4 The Elastic Net . . . . . . . . . . . . . . . . . . . . . 324
8.5 Cases of Exact Multicollinearity,
Including p > n . . . . . . . . . . . . . . . . . . . . . . . . . 324
8.5.1 Why It May Work . . . . . . . . . . . . . . . . . . . 324
8.5.2 Example: R mtcars Data . . . . . . . . . . . . . . . 325
8.5.2.1 Additional Motivation for the Elastic Net . 326
8.6 Bias, Standard Errors and Signficance
Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
8.7 Principal Components Analysis . . . . . . . . . . . . . . . . 327
8.8 Generalized Linear Models . . . . . . . . . . . . . . . . . . . 329
8.8.1 Example: Vertebrae Data . . . . . . . . . . . . . . . 329
8.9 Other Terminology . . . . . . . . . . . . . . . . . . . . . . . 330
8.10 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 330
8.11 Mathematical Complements . . . . . . . . . . . . . . . . . . 331
8.11.1 James-Stein Theory . . . . . . . . . . . . . . . . . . 331
8.11.1.1 Definition . . . . . . . . . . . . . . . . . . . 331
8.11.1.2 Theoretical Properties . . . . . . . . . . . . 331
8.11.1.3 When Might Shrunken Estimators Be
Helpful? . . . . . . . . . . . . . . . . . . . . 332
8.11.2 Yes, It Is Smaller . . . . . . . . . . . . . . . . . . . . 332
8.11.3 Ridge Action Increases Eigenvalues . . . . . . . . . . 333
CONTENTS xxiii

8.12 Computational Complements . . . . . . . . . . . . . . . . . 334

8.12.1 Code for ridgelm() . . . . . . . . . . . . . . . . . . . 334
8.13 Exercises: Data, Code and Math Problems . . . . . . . . . 336

9 Variable Selection and Dimension Reduction 339

9.1 A Closer Look at Under/Overfitting . . . . . . . . . . . . . 341
9.1.1 A Simple Guiding Example . . . . . . . . . . . . . . 342
9.2 How Many Is Too Many? . . . . . . . . . . . . . . . . . . . 344
9.3 Fit Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
9.3.1 Some Common Measures . . . . . . . . . . . . . . . 345
9.3.2 There Is No Panacea! . . . . . . . . . . . . . . . . . 348
9.4 Variable Selection Methods . . . . . . . . . . . . . . . . . . 348
9.5 Simple Use of p-Values: Pitfalls . . . . . . . . . . . . . . . . 349
9.6 Asking “What If” Questions . . . . . . . . . . . . . . . . . . 349
9.7 Stepwise Selection . . . . . . . . . . . . . . . . . . . . . . . 351
9.7.1 Basic Notion . . . . . . . . . . . . . . . . . . . . . . 351
9.7.2 Forward vs. Backward Selection . . . . . . . . . . . 352
9.7.3 R Functions for Stepwise Regression . . . . . . . . . 352
9.7.4 Example: Bodyfat Data . . . . . . . . . . . . . . . . 352
9.7.5 Classification Settings . . . . . . . . . . . . . . . . . 357
9.7.5.1 Example: Bank Marketing Data . . . . . . 357
9.7.5.2 Example: Vertebrae Data . . . . . . . . . . 361
9.7.6 Nonparametric Settings . . . . . . . . . . . . . . . . 362
9.7.6.1 Is Dimension Reduction Important in
the Nonparametric Setting? . . . . . . . . . 362
9.7.7 The LASSO . . . . . . . . . . . . . . . . . . . . . . . 364
9.7.7.1 Why the LASSO Often Performs Subsetting 364
9.7.7.2 Example: Bodyfat Data . . . . . . . . . . . 366
xxiv CONTENTS

9.8 Post-Selection Inference . . . . . . . . . . . . . . . . . . . . 367

9.9 Direct Methods for Dimension Reduction . . . . . . . . . . 369
9.9.1 Informal Nature . . . . . . . . . . . . . . . . . . . . 369
9.9.2 Role in Regression Analysis . . . . . . . . . . . . . . 370
9.9.3 PCA . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
9.9.3.1 Issues . . . . . . . . . . . . . . . . . . . . . 371
9.9.3.2 Example: Bodyfat Data . . . . . . . . . . . 371
9.9.3.3 Example: Instructor Evaluations . . . . . . 375
9.9.4 Nonnegative Matrix Factorization (NMF) . . . . . . 377
9.9.4.1 Overview . . . . . . . . . . . . . . . . . . . 377
9.9.4.2 Interpretation . . . . . . . . . . . . . . . . 377
9.9.4.3 Sum-of-Parts Property . . . . . . . . . . . 378
9.9.4.4 Example: Spam Detection . . . . . . . . . 378
9.9.5 Use of freqparcoord for Dimension Reduction . . . . 380
9.9.5.1 Example: Student Evaluations of Instructors 380
9.9.5.2 Dimension Reduction for Dummy/R Factor
Variables . . . . . . . . . . . . . . . . . . . 381
9.10 The Verdict . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
9.11 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 383
9.12 Computational Complements . . . . . . . . . . . . . . . . . 383
9.12.1 Computation for NMF . . . . . . . . . . . . . . . . . 383
9.13 Mathematical Complements . . . . . . . . . . . . . . . . . . 386
9.13.1 MSEs for the Simple Example . . . . . . . . . . . . 386
9.14 Exercises: Data, Code and Math Problems . . . . . . . . . 387

10 Partition-Based Methods 391

10.1 CART . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
10.2 Example: Vertebral Column Data . . . . . . . . . . . . . . 394
CONTENTS xxv

10.3 Technical Details . . . . . . . . . . . . . . . . . . . . . . . . 398

10.3.1 Split Criterion . . . . . . . . . . . . . . . . . . . . . 398
10.3.2 Predictor Reuse and Statistical Consistency . . . . . 398
10.4 Tuning Parameters . . . . . . . . . . . . . . . . . . . . . . . 399
10.5 Random Forests . . . . . . . . . . . . . . . . . . . . . . . . 399
10.5.1 Bagging . . . . . . . . . . . . . . . . . . . . . . . . . 400
10.5.2 Example: Vertebrae Data . . . . . . . . . . . . . . . 400
10.5.3 Example: Letter Recognition . . . . . . . . . . . . . 401
10.6 Other Implementations of CART . . . . . . . . . . . . . . . 402
10.7 Exercises: Data, Code and Math Problems . . . . . . . . . 403

11 Semi-Linear Methods 405

11.1 k-NN with Linear Smoothing . . . . . . . . . . . . . . . . . 407
11.1.1 Extrapolation Via lm() . . . . . . . . . . . . . . . . 407
11.1.2 Multicollinearity Issues . . . . . . . . . . . . . . . . 409
11.1.3 Example: Bodyfat Data . . . . . . . . . . . . . . . . 409
11.1.4 Tuning Parameter . . . . . . . . . . . . . . . . . . . 409
11.2 Linear Approximation of Class
Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
11.2.1 SVMs . . . . . . . . . . . . . . . . . . . . . . . . . . 410
11.2.1.1 Geometric Motivation . . . . . . . . . . . . 411
11.2.1.2 Reduced convex hulls . . . . . . . . . . . . 413
11.2.1.3 Tuning Parameter . . . . . . . . . . . . . . 415
11.2.1.4 Nonlinear Boundaries . . . . . . . . . . . . 416
11.2.1.5 Statistical Consistency . . . . . . . . . . . 417
11.2.1.6 Example: Letter Recognition Data . . . . . 417
11.2.2 Neural Networks . . . . . . . . . . . . . . . . . . . . 418
11.2.2.1 Example: Vertebrae Data . . . . . . . . . . 418
xxvi CONTENTS

11.2.2.2 Tuning Parameters and Other Technical De-

tails . . . . . . . . . . . . . . . . . . . . . . 420
11.2.2.3 Dimension Reduction . . . . . . . . . . . . 421
11.2.2.4 Why Does It Work (If It Does)? . . . . . . 421
11.3 The Verdict . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
11.4 Mathematical Complements . . . . . . . . . . . . . . . . . . 424
11.4.1 Edge Bias in Nonparametric Regression . . . . . . . 424
11.4.2 Dual Formulation for SVM . . . . . . . . . . . . . . 425
11.4.3 The Kernel Trick . . . . . . . . . . . . . . . . . . . . 428
11.5 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 429
11.6 Exercises: Data, Code and Math Problems . . . . . . . . . 429

12 Regression and Classification in Big Data 431

12.1 Solving the Big-n Problem . . . . . . . . . . . . . . . . . . . 432
12.1.1 Software Alchemy . . . . . . . . . . . . . . . . . . . 432
12.1.2 Example: Flight Delay Data . . . . . . . . . . . . . 433
12.1.3 More on the Insufficient Memory Issue . . . . . . . . 436
12.1.4 Deceivingly “Big” n . . . . . . . . . . . . . . . . . . 437
12.1.5 The Independence Assumption in Big-n Data . . . . 437
12.2 Addressing Big-p . . . . . . . . . . . . . . . . . . . . . . . . 438
12.2.1 How Many Is Too Many? . . . . . . . . . . . . . . . 438
12.2.1.1 Toy Model . . . . . . . . . . . . . . . . . . 439
12.2.1.2 Results from the Research Literature . . . 440
12.2.1.3 A Much Simpler and More Direct Approach 441
12.2.1.4 Nonparametric Case . . . . . . . . . . . . . 441
12.2.1.5 The Curse of Dimensionality . . . . . . . . 443
12.2.2 Example: Currency Data . . . . . . . . . . . . . . . 443
12.2.3 Example: Quiz Documents . . . . . . . . . . . . . . 444
CONTENTS xxvii

12.2.4 The Verdict . . . . . . . . . . . . . . . . . . . . . . . 446

12.3 Mathematical Complements . . . . . . . . . . . . . . . . . . 447
12.3.1 Speedup from Software Alchemy . . . . . . . . . . . 447
12.4 Computational Complements . . . . . . . . . . . . . . . . . 448
12.4.1 The partools Package . . . . . . . . . . . . . . . . . 448
12.4.2 Use of the tm Package . . . . . . . . . . . . . . . . . 449
12.5 Exercises: Data, Code and Math Problems . . . . . . . . . 450

A Matrix Algebra 451

A.1 Terminology and Notation . . . . . . . . . . . . . . . . . . . 451
A.2 Matrix Addition and Multiplication . . . . . . . . . . . . . 452
A.3 Matrix Transpose . . . . . . . . . . . . . . . . . . . . . . . . 453
A.4 Linear Independence . . . . . . . . . . . . . . . . . . . . . . 454
A.5 Matrix Inverse . . . . . . . . . . . . . . . . . . . . . . . . . 454
A.6 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . 455
A.7 Rank of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . 456
A.8 Matrices of the Form B’B . . . . . . . . . . . . . . . . . . . 456
A.9 Partitioned Matrices . . . . . . . . . . . . . . . . . . . . . . 457
A.10 Matrix Derivatives . . . . . . . . . . . . . . . . . . . . . . . 458
A.11 Matrix Algebra in R . . . . . . . . . . . . . . . . . . . . . . 459
A.12 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . 462

Index 475
Preface

Why write yet another regression book? There is a plethora of books out
there already, written by authors whom I greatly admire, and whose work I
myself have found useful. I might cite the books by Harrell [61] and Fox [50],
among many, many excellent examples. Note that I am indeed referring to
general books on regression analysis, as opposed to more specialized work
such as [64] and [76], which belong to a different genre. My book here is
intended for a traditional (though definitely modernized) regression course,
rather than one on statistical learning.
Yet, I felt there is an urgent need for a different kind of book. So, why
is this regression book different from all other regression books? First,
it modernizes the standard treatment of regression methods. In
particular:

• The book supplements classical regression models with introductory

material on machine learning methods.
• Recognizing that these days, classification is the focus of many appli-
cations, the book covers this topic in detail, especially the multiclass
case.
• In view of the voluminous nature of many modern datasets, there is
a chapter on Big Data.
• There is much more hands-on involvement of computer usage.

Other major senses in which this book differs from others are:

• Though presenting the material in a mathematically precise man-

ner, the book aims to provide much needed applied insight for the
practicing analyst, remedying the “too many equations, too few ex-
planations” problem.

xxix
xxx PREFACE

For instance, the book not only shows how the math works for trans-
formations of variables, but also raises points on why one might refrain
from applying transformations.
• The book features a recurring interplay between parametric and non-
parametric methods. For instance, in an example involving currency
data, the book finds that the fitted linear model predicts substantially
more poorly than a k-nearest neighbor fit, suggesting deficiencies in
the linear model. Nonparametric analysis is then used to further in-
vestigate, providing parametric model assessment in a manner that is
arguably more insightful than classical residual plots.
• For those interested in computing issues, many of the book’s chap-
ters include optional sections titled Computational Complements, on
topics such as data wrangling, development of package source code,
parallel computing and so on.
Also, many of the exercises are code-oriented. In particular, in some
such exercises the reader is asked to write “mini-CRAN” functions,1
short but useful library functions that can be applied to practical
regression analysis. Here is an example exercise of this kind:

Write a function stepAR2() that works similarly to stepAIC(),

except that this new function uses adjusted R2 as its crite-
rion for adding or deleting a predictor. The call form will
be
stepAR2 ( lmobj , direction = ’ fwd ’ ,
nsteps = ncol ( lmobj $ model ) -1)
where the arguments are...

• For those who wish to go into more depth on mathematical topics,

there are Mathematical Complements sections at the end of most
chapters, and math-oriented exercises. The material ranges from
straightforward computation of mean squared error to esoteric top-
ics such as a proof of the Tower Property, E [E(V |U1 , U2 ) | U1 ] =
E(V | U1 ), a result that is used in the text.

As mentioned, this is still a book on traditional regression analysis.

In contrast to [64], this book is aimed at a traditional regression course.
Except for Chapters 10 and 11, the primary methodology used is linear
and generalized linear parametric models, covering both the Description
and Prediction goals of regression methods. We are just as interested in
1 CRAN is the online repository of user-contributed R code.
PREFACE xxxi

Description applications of regression, such as measuring the gender wage

gap in Silicon Valley, as we are in forecasting tomorrow’s demand for bike
rentals. An entire chapter is devoted to measuring such effects, including
discussion of Simpson’s Paradox, multiple inference, and causation issues.
The book’s examples are split approximately equally in terms of Description
and Prediction goals. Issues of model fit play a major role.
The book includes more than 75 full examples, using real data. But con-
cerning the above comment regarding “too many equations, too few expla-
nations,”, merely including examples with real data is not enough to truly
tell the story in a way that will be useful in practice. Rather few books
go much beyond presenting the formulas and techniques, thus leaving the
hapless practitioner to his own devices. Too little is said in terms of what
the equations really mean in a practical sense, what can be done with re-
gard to the inevitable imperfections of our models, which techniques are
too much the subject of “hype,” and so on.
As a nonstatistician, baseball great Yogi Berra, put it in his inimitable style,
“In theory there is no difference between theory and practice. In practice
there is.” This book aims to remedy this gaping deficit. It develops the
material in a manner that is mathematically precise yet always maintains
as its top priority — borrowing from a book title of the late Leo Breiman
— a view toward applications.
In other words:

The philosophy of this book is to not only prepare the analyst

to know how to do something, but also to understand what she
is doing. For successful application of data science techniques,
the latter is just as important as the former.

Some further examples of how this book differs from the other regression
books:
Intended audience and chapter coverage:
This book is aimed at both practicing professionals and use in the class-
room. It aims to be both accessible and valuable to this diversity of read-
ership.
In terms of classroom use, with proper choice of chapters and appendices,
the book could be used as a text tailored to various discipline-specific au-
diences and various levels, undergraduate or graduate. I would recommend
that the core of any course consist of most sections of Chapters 1-4 (exclud-
ing the Math and Computational Complements sections), with coverage of
xxxii PREFACE

at least introductory sections of Chapters 5, 6, 7, 8 and 9 for all audiences.

Beyond that, different types of disciplines might warrant different choices
of further material. For example:

• Statistics students: Depending on level, at least some of the Mathe-

matical Complements and math-oriented exercises should be involved.
There might be more emphasis on Chapters 6, 7 and 9.

• Computer science students: Here one would cover more of classi-

fication, machine learning and Big Data material, Chapters 5, 8, 10,
11 and 12. Also, one should cover the Computational Complements
sections and associated “mini-CRAN” code exercises.

• Economics/social science students: Here there would be heavy

emphasis on the Description side, Chapters 6 and 7, with special
emphasis on topics such as Instrumental Variables and Propensity
Matching in Chapter 7. Material on generalized linear models and
logistic regression, in Chapter 4 and parts of Chapter 5, might also
be given emphasis.

• Student class level: The core of the book could easily be used in
an undergraduate regression course, but aimed at students with back-
ground in calculus and matrix algebra, such as majors in statistics,
math or computer science. A graduate course would cover more of
the chapters on advanced topics, and would likely cover more of the
Mathematical Complements sections.

• Level of mathematical sophistication: In the main body of the

text, i.e., excluding the Mathematical Complements sections, basic
matrix algebra is used throughout, but use of calculus is minimal. As
noted, for those instructors who want the mathematical content, it is
there in the Mathematical Complements sections, but the main body
of the text requires only the matrix algebra and a little calculus.

The reader must of course be familiar with terms like confidence interval,
significance test and normal distribution. Many readers will have had at
least some prior exposure to regression analysis, but this is not assumed,
and the subject is developed from the beginning.
The reader is assumed to have some prior experience with R, but at a
minimal level: familiarity with function arguments, loops, if-else and vec-
tor/matrix operations and so on. For those without such background, there
are many gentle tutorials on the Web, as well as a leisurely introduction in
a statistical context in [21]. Those with programming experience can also
PREFACE xxxiii

read the quick introduction in the appendix of [99]. My book [100] gives
a detailed treatment of R as a programming language, but that level of
sophistication is certainly not needed for the present book.
A comment on the field of machine learning:
Mention should be made of the fact that this book’s title includes both
the word regression and the phrase machine learning. The latter phrase is
included to reflect that the book includes some introductory material on
machine learning, in a regression context.
Much has been written on a perceived gap between the statistics and ma-
chine learning communities [23]. This gap is indeed real, but work has been
done to reconcile them [16], and in any event, the gap is actually not as
wide as people think.
My own view is that machine learning (ML) consists of the development
of regression models with the Prediction goal. Typically nonparametric
(or what I call semi-parameteric) methods are used. Classification models
are more common than those for predicting continuous variables, and it
is common that more than two classes are involved, sometimes a great
many classes. All in all, though, it’s still regression analysis, involving the
conditional mean of Y given X (reducing to P (Y = 1|X) in the classification
context).
One often-claimed distinction between statistics and ML is that the former
is based on the notion of a sample from a population whereas the latter
is concerned only with the content of the data itself. But this difference
is more perceived than real. The idea of cross-validation is central to ML
methods, and since that approach is intended to measure how well one’s
model generalizes beyond our own data, it is clear that ML people do think
in terms of samples after all. Similar comments apply to ML’s citing the
variance-vs.-bias tradeoff, overfitting and so on
So, at the end of the day, we all are doing regression analysis, and this book
takes this viewpoint.
Code and software:
The book also makes use of some of my research results and associated
software. The latter is in my package regtools, available from CRAN [97].
A number of other packages from CRAN are used. Note that typically
we use only the default values for the myriad arguments available in many
functions; otherwise we could fill an entire book devoted to each package!
Cross-validation is suggested for selection of tuning parameters, but with a
warning that it too can be problematic.
xxxiv PREFACE

In some cases, the regtools source code is also displayed within the text,
so as to make clear exactly what the algorithms are doing. Similarly, data
wrangling/data cleaning code is shown, not only for the purpose of “hands-
on” learning, but also to highlight the importance of those topics.
Thanks:
Conversations with a number of people have enhanced the quality of this
book, some via direct comments on the presentation and others in discus-
sions not directly related to the book. Among them are Charles Abromaitis,
Stuart Ambler, Doug Bates, Oleksiy Budilovsky, Yongtao Cao, Tony Corke,
Tal Galili, Frank Harrell, Harlan Harris, Benjamin Hofner, Jiming Jiang,
Hyunseung Kang, Martin Mächler, Erin McGinnis, John Mount, Richard
Olshen, Pooja Rajkumar, Ariel Shin, Chuck Stone, Jessica Tsoi, Yu Wu,
Yihui Xie, Yingkang Xie, Achim Zeileis and Jiaping Zhang.
A seminar presentation by Art Owen introduced me to the application of
random effects models in recommender systems, a provocative blend of old
and new. This led to the MovieLens examples and other similar examples
in the book, as well as a vigorous new research interest for me. Art also
led me to two Stanford statistics PhD students, Alex Chin and Jing Miao,
who each read two of the chapters in great detail. Special thanks also go
to Nello Cristianini, Hui Lin, Ira Sharenow and my old friend Gail Gong
for their detailed feedback.
Thanks go to my editor, John Kimmel, for his encouragement and much-
appreciated patience, and to the internal reviewers, David Giles, Robert
Gramacy and Christopher Schmidt. Of course, I cannot put into words
how much I owe to my wonderful wife Gamis and our daughter Laura, both
of whom inspire all that I do, including this book project.
Website:
Code, errata, extra examples and so on are available at
http://heather.cs.ucdavis.edu/regclass.html.
A final comment:
My career has evolved quite a bit over the years. I wrote my dissertation
in abstract probability theory [105], but turned my attention to applied
statistics soon afterward. I was one of the founders of the Department of
Statistics at UC Davis. Though a few years later I transferred into the new
Computer Science Department, I am still a statistician, and much of my
CS research has been statistical, e.g., [98]. Most important, my interest in
regression has remained strong throughout those decades.
PREFACE xxxv

I published my first research papers on regression methodology way back

in the 1980s, and the subject has captivated me ever since. My long-held
wish has been to write a regression book, and thus one can say this work is
30 years in the making. I hope you find its goals both worthy and attained.
Above all, I simply hope you find it an interesting read.
List of Symbols and
Abbreviations

Y : the response variable

X: vector of predictor variables
e X with a 1 prepended
X:
X (j) : the j th predictor variable
n: number of observations
p: number of predictors
Yi : value of the response variable in observation i
Xi : vector of predictors in observation i
(j)
Xi : value of the j th predictor variable in observation i
A: n × (p + 1) matrix of the predictor data in a linear model
D: length-n vector of the response data in a linear model
H: the hat matrix, A(A′ A)−1 A′
µ(t): the regression function E(Y |X = t)
σ 2 (t): V ar(Y |X = t)
b(t): estimated value of µ(t)
µ
β: vector of coefficients in a linear/generalized linear model
b estimated value of β
β:
′: matrix transpose
I: multiplicative identity matrix
k-NN: k-Nearest Neighbor method
MSE: Mean Squared (Estimation) Error
MSPE: Mean Squared Prediction Error
CART: Classification and Regression Trees
SVM: Support Vector Machine
NN: neural network
PCA: Principal Components Analysis
NMF: Nonnegative Matrix Factorization
OVA: One vs. All classification

xxxvii
xxxviii PREFACE

AVA: All vs. All classification

LDA: Linear Discriminant Analysis
Chapter 1

Setting the Stage

This chapter will set the stage for the book, previewing many of the ma-
jor concepts to be presented in later chapters. The material here will be
referenced repeatedly throughout the book.

1.1 Example: Predicting Bike-Sharing

Activity

Let’s start with a well-known dataset, Bike Sharing, from the Machine
Learning Repository at the University of California, Irvine.1 Here we have
daily/hourly data on the number of riders, weather conditions, day-of-week,
month and so on. Regression analysis, which relates the mean of one vari-
able to the values of one or more other variables, may turn out to be useful
to us in at least two ways:

• Prediction:
The managers of the bike-sharing system may wish to predict rider-
ship, say for the following question:

Tomorrow, Sunday, is expected to be sunny and cool, say

62 degrees Fahrenheit. We may wish to predict the number
of riders, so that we can get some idea as to how many bikes
will need repair. We may try to predict ridership, given the
1 Available at https://archive.ics.uci.edu/ml/datasets/Bike+Sharing+Dataset.

1
2 CHAPTER 1. SETTING THE STAGE

weather conditions, day of the week, time of year and so

on.

Some bike-sharing services actually have trucks to move large num-

bers of bikes to locations that are expected to have high demand.
Prediction would be even more useful here.

• Description:

We may be interested in determining what factors affect ridership.

How much effect, for instance, does wind speed have in influencing
whether people wish to borrow a bike?

These twin goals, Prediction and Description, will arise frequently in this
book. Choice of methodology will often depend on the goal in the given
application.

1.2 Example of the Prediction Goal: Body

Fat

Prediction is difficult, especially about the future — baseball great, Yogi

Berra
The great baseball player Yogi Berra was often given to malapropisms, one
of which supposedly was the quote above. But there is more than a grain
of truth to this, because indeed we may wish to “predict” the present or
even the past.
For example, consiser the bodyfat data set, available in the R package,
mfp, available on CRAN [5]. (See Section 1.20.1 for information on CRAN
packages, a number of which will be used in this book.) Direct measurment
of body fat is expensive and unwieldy, as it involves underwater weighing.
Thus it would be highly desirable to “predict” that quantity from easily
measurable variables such as height, age, weight, abdomen circumference
and so on.
In scientific studies of ancient times, there may be similar situations in
which we “predict” past unknown quantities from present known ones.
1.3. EXAMPLE: WHO CLICKS WEB ADS? 3

Figure 1.1: Click rate vs. college rate

1.3 Example: Who Clicks Web Ads?

One of the most common applications of machine learning methods is in

marketing. Sellers wish to learn which types of people might be interested
in a given product. The reader is probably familiar with Amazon’s recom-
mender system, in which the viewer who indicates interest in a given book,
say, is shown a list of similar books.2
We will discuss recommender systems at several points in this book, be-
ginning with Section 3.2.4. A more general issue is the click-through rate
(CTR), meaning the proportion of viewers of a Web page who click on a
particular ad on the page. A simple but very engaging example was dis-
cussed online [53]. The data consist of one observation per state of the
U.S.3 There was one predictor, the proportion of college graduates in the
state, and a response variable, the CTR.
2 As a consumer, I used to ignore these, but now with the sharp decline in the num-

ber of bricks-and-mortar bookstores which I could browse, I now often find Amazon’s
suggestions useful.
3 We use the classical statistical term observation here, meaning a single data point,

in this case data for a single state. In the machine learning community, it is common to
use the term case.
4 CHAPTER 1. SETTING THE STAGE

A plot of the data, click rate vs. college rate, is in Figure 1.1. There
definitely seems to be something happening here, with a visible downward
trend to the points. But how do we quantify that? One approach to
learning what relation, if any, educational level has to CTR would be to
use regression analysis. We will see how to do so in Section 1.8.

1.4 Approach to Prediction

Even without any knowledge of statistics, many people would find it rea-
sonable to predict via subpopulation means. In the above bike-sharing
example, say, this would work as follows.
Think of the “population” of all days, past, present and future, and their
associated values of number of riders, weather variables and so on.4 Our
data set is considered a sample from this population. Now consider the
subpopulation consisting of all days with the given conditions: Sundays,
sunny skies and 62-degree temperatures.
It is intuitive that:

A reasonable prediction for tomorrow’s ridership would be the

mean ridership among all days in the subpopulation of Sundays
with sunny skies and 62-degree temperatures.

In fact, such a strategy is optimal, in the sense that it minimizes our ex-
pected squared prediction error, as discussed in Section 1.19.3 of the Math-
ematical Complements section at the end of this chapter. But what is
important for now is to note that in the above prediction rule, we are deal-
ing with a conditional mean: Mean ridership, given day of the week is
Sunday, skies are sunny, and temperature is 62.
Note too that we can only calculate an estimated conditional mean. We
wish we had the true population value, but since our data is only a sample,
we must always keep in mind that we are just working with estimates.

4 This is a somewhat slippery notion, because there may be systemic differences from

the present and the distant past and distant future, but let’s suppose we’ve resolved that
by limiting our time range.
1.5. A NOTE ABOUT E(), SAMPLES AND POPULATIONS 5

1.5 A Note about E(), Samples and Popula-

tions

To make this more mathematically precise, note carefully that in this book,
as with many other books, the expected value functional E() refers to the
population mean. Say we are studying personal income, I, for some popu-
lation, and we choose a person at random from that population. Then E(I)
is not only the mean of that random variable, but much more importantly,
it is the mean income of all people in that population.
Similarly, we can define conditional means, i.e., means of subpopulations.
Say G is gender. Then the conditional expected value, E(I | G = male) is
the mean income of all men in the population.
To illustrate this in the bike-sharing context, let’s define some variables:

• R, the number of riders

• W , the day of the week

• S, the sky conditions, e.g., sunny

• T , the temperature

We would like our prediction Q to be the conditional mean,

Q = E(R | W = Sunday, S = sunny, T = 62) (1.1)

There is one major problem, though: We don’t know the value of the right-
hand side of (1.1). All we know is what is in our sample data, whereas the
right-side of (1.1) is a population value, and thus unknown.
The difference between sample and population is of course at the
very core of statistics. In an election opinion survey, for instance, we
wish to know p, the proportion of people in the population who plan to
vote for Candidate Jones. But typically only 1200 people are sampled, and
we calculate the proportion of Jones supporters among them, pb, using that
as our estimate of p. (Note that the “hat” notation b is the traditional
one for “estimate of.”) This is why the news reports on these polls always
include the margin of error.5
5 This is actually the radius of a 95% confidence interval for p.
6 CHAPTER 1. SETTING THE STAGE

Similarly, though we would like to know the value of E(R | W = Sunday,

S = sunny, T = 62), it is an unknown population value, and thus
must be estimated from our sample data, which we’ll do later in this
chapter.
Readers will greatly profit from constantly keeping in mind this
distinction between populations and samples.
Another point is that in statistics, the populations are often rather con-
ceptual in nature. On the one hand, in the election poll example above,
there is a concrete population involved, the population of all voters. On
the other hand, consider the bike rider data in Section 1.1. Here we can
think of our data as being a sample from the population of all bikeshare
users, past, present and future.
Before going on, a bit of terminology, again to be used throughout the book:
We will refer to the quantity to be predicted, e.g., R above, as the response
variable, and the quantities used in prediction, e.g., W , S and T above, as
the predictor variables. Other popular terms are dependent variable for the
response and independent variables or regressors for the predictors. The
machine learning community uses the term features rather than predictors.

1.6 Example of the Description Goal: Do

Baseball Players Gain Weight As They
Age?

Nothing in life is to be feared, it is only to be understood. Now is the time

to understand more, so that we may fear less — Marie Curie
Though the bike-sharing data set is the main example in this chapter, it
is rather sophisticated for introductory material. Thus we will set it aside
temporarily, and bring in a simpler data set for now. We’ll return to the
bike-sharing example in Section 1.15.
This new dataset involves 1015 major league baseball players, courtesy of
the UCLA Statistics Department. You can obtain the data as the data set
mlb in freqparcoord, a CRAN package authored by Yingkang Xie and
myself [104].6 The variables of interest to us here are player weight W ,
height H and age A, especially the first two.

6 We
use the latter version of the dataset here, in which we have removed the Desig-
nated Hitters.
1.6. DO BASEBALL PLAYERS GAIN WEIGHT? 7

Here are the first few records:

> library ( freqparcoord )
> data ( mlb )
> head ( mlb )
Name Team P o s i t i o n Height
1 Adam Donachie BAL Catcher 74
2 Paul Bako BAL Catcher 74
3 Ramon Hernandez BAL Catcher 72
4 Kevin M i l l a r BAL F i r s t Baseman 72
5 C h r i s Gomez BAL F i r s t Baseman 73
6 Brian Roberts BAL Second Baseman 69
Weight Age PosCategory
1 180 2 2 . 9 9 Catcher
2 215 3 4 . 6 9 Catcher
3 210 3 0 . 7 8 Catcher
4 210 3 5 . 4 3 Infielder
5 188 3 5 . 7 1 Infielder
6 176 2 9 . 3 9 Infielder

1.6.1 Prediction vs. Description

Recall the Prediction and Description goals of regression analysis, discussed

in Section 1.1. With the baseball player data, we may be more interested
in the Description goal, such as:

Ahtletes strive to keep physically fit. Yet even they may gain
weight over time, as do people in the general population. To
what degree does this occur with the baseball players? This
question can be answered by performing a regression analysis of
weight against height and age, which we’ll do in Section 1.9.1.2.7

On the other hand, there doesn’t seem to be much of a Prediction goal

here. It is hard to imagine much need to predict a player’s weight. One
example of this, though, is working with missing data, in which we wish to
predict any value that might be unavailable.
However, for the purposes of explaining the concepts, we will often phrase
things in a Prediction context. In the baseball player example, it will turn
7 The phrasing here, “regression analysis of ... against ...,” is commonly used in this

field. The quantity before “against” is the response variable, and the ones following are
the predictors.
8 CHAPTER 1. SETTING THE STAGE

out that by trying to predict weight, we can deduce effects of height and
age. In particular, we can answer the question posed above concerning
weight gain over time.
So, suppose we will have a continuing stream of players for whom we only
know height (we’ll bring in the age variable later), and need to predict their
weights. Again, we will use the conditional mean to do so. For a player of
height 72 inches, for example, our prediction might be

c = E(W | H = 72)
W (1.2)

Again, though, this is a population value, and all we have is sample data.
How will we estimate E(W | H = 72) from that data?
First, some important notation: Recalling that µ is the traditional Greek
letter to use for a population mean, let’s now use it to denote a function
that gives us subpopulation means:

For any height t, define

µ(t) = E(W | H = t) (1.3)

which is the mean weight of all people in the population who

are of height t.
Since we can vary t, this is indeed a function, and it is known
as the regression function of W on H.

So, µ(72.12) is the mean population weight of all players of height 72.12,
µ(73.88) is the mean population weight of all players of height 73.88, and
so on. These means are population values and thus unknown, but they do
exist.
So, to predict the weight of a 71.6-inch-tall player, we would use µ(71.6) —
if we knew that value, which we don’t, since once again this is a population
value while we only have sample data. So, we need to estimate that value
from the (height, weight) pairs in our sample data, which we will denote
by (H1 , W1 ), ...(H1015 , W1015 ). How might we do that? In the next two
sections, we will explore ways to form our estimate, µ b(t). (Keep in mind
that for now, we are simply exploring, especially in the first of the following
two sections.)
1.6. DO BASEBALL PLAYERS GAIN WEIGHT? 9

1.6.2 A First Estimator

Our height data is only measured to the nearest inch, so instead of esti-
mating values like µ(71.6), we’ll settle for µ(72) and so on. A very natural
estimate for µ(72), again using the “hat” symbol to indicate “estimate of,”
is the mean weight among all players in our sample for whom height is 72,
i.e.

b(72) = mean of all Wi such that Hi = 72

µ (1.4)

b(t) at once:
R’s tapply() can give us all the µ
> library ( freqparcoord )
> data ( mlb )
> muhats <− tapply ( mlb$Weight , mlb$Height , mean)
> muhats
67 68 69 70 71 72
172.5000 173.8571 179.9474 183.0980 190.3596 192.5600
73 74 75 76 77 78
196.7716 202.4566 208.7161 214.1386 216.7273 220.4444
79 80 81 82 83
218.0714 237.4000 245.0000 240.5000 260.0000

In case you are not familiar with tapply(), here is what just happened. We
asked R to partition the Weight variable into groups according to values
of the Height variable, and then compute the mean weight in each group.
So, the mean weight of people of height 72 in our sample was 192.5600.
In other words, we would set µ b(72) = 192.5600, µb(74) = 202.4566, and so
on. (More detail on tapply() is given in the Computational Complements
section at the end of this chapter.)
Since we are simply performing the elementary statistics operation of esti-
mating population means from samples, we can form confidence intervals
(CIs). For this, we’ll need the “n” and sample standard deviation for each
height group:
> tapply ( mlb$Weight , mlb$Height , length )
67 68 69 70 71 72 73 74 75 76 77 78
2 7 19 51 89 150 162 173 155 101 55 27
79 80 81 82 83
14 5 2 2 1
> tapply ( mlb$Weight , mlb$Height , sd )
67 68 69 70 71 72
10.60660 22.08641 15.32055 13.54143 16.43461 17.56349
10 CHAPTER 1. SETTING THE STAGE

73 74 75 76 77 78
16.41249 18.10418 18.27451 19.98151 18.48669 14.44974
79 80 81 82 83
28.17108 10.89954 21.21320 13.43503 NA

Here is how that first call to tapply() worked. Recall that this function
partitions the data by the Height variables, resulting in a weight vector for
each height value. We need to specify a function to apply to each of the
resulting vectors, which in this case we choose to be R’s length() function.
The latter then gives us the count of weights for each height value, the “n”
that we need to form a CI. By the way, the NA value is due to there being
only one player with height 83, which is makes life impossible for sd(), as
it divides from “n-1.”
An approximate 95% CI for µ(72), for example, is then

17.56349
190.3596 ± 1.96 √ (1.5)
150

or about (187.6,193.2).
The above analysis takes what is called a nonparametric approach. To see
why, let’s proceed to a parametric one, in the next section.

1.6.3 A Possibly Better Estimator, Using a Linear Model

All models are wrong, but some are useful — famed statistician George Box
[In spite of ] innumerable twists and turns, the Yellow River flows east —
Confucious

So far, we have assumed nothing about the shape that µ(t) would have, if it
were plotted on a graph. Again, it is unknown, but the function does exist,
and thus it does correspond to some curve. But we might consider making
an assumption on the shape of this unknown curve. That might seem odd,
but you’ll see below that this is a very powerful, intuitively reasonable idea.
b(t) we found above. We run
Toward this end, let’s plot those values of µ
> plot ( 6 7 : 8 3 , muhats )

producing Figure 1.2.

1.6. DO BASEBALL PLAYERS GAIN WEIGHT? 11

b(t)
Figure 1.2: Plotted µ

Interestingly, the points in this plot seem to be near a straight line. Just
like the quote of Confucious above concerning the Yellow River, visually
we see something like a linear trend, in spite of the “twists and turns” of
the data in the plot. This suggests that our unknown function µ b(t) has a
linear form, i.e., that

µ(t) = c + dt (1.6)

for some constants c and d, over the range of t appropriate to human heights.
Or, in English,

mean weight = c + d × height (1.7)

Don’t forget the word mean here! We are assuming that the mean weights
in the various height subpopulations have the form (1.6), NOT that weight
itself is this function of height, which can’t be true.
This is called a parametric model for µ(t), with parameters c and d. We
will use this below to estimate µ(t). Our earlier estimation approach, in
12 CHAPTER 1. SETTING THE STAGE

Section 1.6.2, is called nonparametric. It is also called assumption-free or

model-free, since it made no assumption at all about the shape of the µ(t)
curve.
Note the following carefully:

• Figure 1.2 suggests that our straight-line model for µ(t) may be less
accurate at very small and very large values of t. This is hard to say,
though, since we have rather few data points in those two regions, as
seen in our earlier R calculations; there is only one person of height
83, for instance.
But again, in this chapter we are simply exploring, so let’s assume for
b(t) is reasonably accurate. We
now that the straight-line model for µ
will discuss in Chapter 6 how to assess the validity of this model.

• Since µ(t) is a population function, the constants c and d are popula-

tion values, thus unknown. However, we can estimate them from our
sample data. We do so using R’s lm() (“linear model”) function:8

> lmout <− lm( mlb$Weight ∼ mlb$ Height )

> lmout
Call :
lm( formula = mlb$Weight ∼ mlb$ Height )

Coefficients :
( Intercept ) mlb$ Height
−151.133 4.783

c = −151.133 and db = 4.783. We can superimpose the fitted line

This gives b
to Figure 1.2, using R’s abline() function, which adds a line with specified
slope and intercept to the currently-displayed plot:

> abline ( coef=coef ( lmout ) )

The result is shown in Figure 1.3.

Note carefully that we do not expect the line to fit the points exactly. On
the contrary, the line is only an estimate of µ(t), the condiitonal mean of
weight given height, not weight itself.

8 Details on how the estimation is done will be given in Chapter 2.

1.6. DO BASEBALL PLAYERS GAIN WEIGHT? 13

Figure 1.3: Turkish student evaluations

We would then set, for instance (using the “check” instead of the hat, so
as to distinguish from our previous estimator)

µ̌(72) = −151.133 + 4.783 × 72 = 193.2666 (1.8)

So, using this model, we would predict a slightly heavier weight than our
earlier prediction.
By the way, we need not type the above expression into R by hand. Here
is why: Writing the expression in matrix-multiply form, it is

 
1
(−151.133, 4.783) (1.9)
72
14 CHAPTER 1. SETTING THE STAGE

Be sure to see the need for that 1 in the second factor; it is used to pick
up the -151.133. Now let’s use that matrix form to show how we can
conveniently compute that value in R:9
The key is that we can exploit the fact that R’s coef() function fetches the
coefficients c and d for us:
> coef ( lmout )
( I n t e r c e p t ) mlb$ Height
−151.133291 4.783332

Recalling that the matrix-times-matrix operation in R is specified via the

%∗% operator, we can now obtain our estimated value of µ(72) as
> coef ( lmout ) %∗% c ( 1 , 7 2 )
[ ,1]
[ 1 , ] 193.2666

We can form a confidence interval from this too, which for the 95% level
will be

µ̌(72) ± 1.96 s.e.[(µ̌(72)] (1.10)

where s.e.[] signifies standard error, the estimated standard deviation of an

estimator. Here µ̌(72), being based on our random sample data, is itself
random, i.e., it will vary from sample to sample. It thus has a standard de-
viation, which we call the standard error. We will see later that s.e.[(µ̌(72)]
is obtainable using the R vcov() function:
> tmp <− c ( 1 , 7 2 )
> sqrt ( tmp %∗% vcov ( lmout ) %∗% tmp )
[ ,1]
[ 1 , ] 0.6859655
> 193.2666 + 1.96 ∗ 0.6859655
[ 1 ] 194.6111
> 193.2666 − 1.96 ∗ 0.6859655
[ 1 ] 191.9221

(More detail on vcov() and coef() as R functions is presented in Section

1.20.4 in the Computational Complements section at the end of this chap-
ter.)
9 In
order to gain a solid understanding of the concepts, we will refrain from using R’s
predict() function for now. It will be introduced later, though, in Section 1.10.3.
1.7. PARAMETRIC VS. NONPARAMETRIC MODELS 15

So, an approximate 95% CI for µ(72) under this model would be about
(191.9,194.6).

1.7 Parametric vs. Nonparametric Models

Now here is a major point: The CI we obtained from our linear model,
(191.9,194.6), was narrower than what the nonparametric approach gave
us, (187.6,193.2); the former has width of about 2.7, while the latter’s is
5.6. In other words:

A parametric model is — if it is (approximately) valid — more

powerful than a nonparametric one, yielding estimates of a re-
gression function that tend to be more accurate than what the
nonparametric approach gives us. This should translate to more
accurate prediction as well.

Why should the linear model be more effective? Here is some intuition,
say for estimating µ(72): As will be seen in Chapter 2, the lm() function
uses all of the data to estimate the regression coefficients. In our case here,
all 1015 data points played a role in the computation of µ̌(72), whereas
only 150 of our observations were used in calculating our nonparametric
estimate µb(72). The former, being based on much more data, should tend
to be more accurate.10
On the other hand, in some settings it may be difficult to find a valid para-
metric model, in which case a nonparametric approach may be much more
effective. This interplay between parametric and nonparametric models will
be a recurring theme in this book.

1.8 Example: Click-Through Rate

Let’s try a linear regression model on the CTR data in Section 1.3. The
file can be downloaded from the link in [53].
> c t r <− read . table ( ’ S t a t e CTR Date . t x t ’ ,
h e a d e r=TRUE, s e p= ’ \ t ’ )
10 Note the phrase tend to here. As you know, in statistics one usually cannot say that

one estimator is always better than another, because anomalous samples do have some
nonzero probability of occurring.
16 CHAPTER 1. SETTING THE STAGE

Figure 1.4: CTR data and fitted line

> lmout <− lm( c t r $CTR ∼ c t r $ C o l l e g e Grad )

> lmout
...
Coefficients :
( I n t e r c e p t ) c t r $ C o l l e g e Grad
0.01412 −0.01373
...

A scatter plot of the data, with the fitted line superimposed, is shown in
Figure 1.4. It was generated by the code

> plot ( c t r $ C o l l e g e Grad , c t r $CTR)

> abline ( coef=coef ( lmout ) )
1.9. SEVERAL PREDICTOR VARIABLES 17

The relation between education and CTR is interesting, but let’s put this
in perspective, by considering the standard deviation of College Grad:
> sd ( c t r $ C o l l e g e Grad )
[ 1 ] 0.04749804

So, a “typical” difference between one state and another is something like
0.05. Multiplying by the -0.01373 figure above, this translates to a difference
in click-through rate from state to state of about 0.0005. This is certainly
not enough to have any practical meaning.
So, putting aside such issues as whether our data constitute a sample from
some “population” of potential states, the data suggest that there is really
no substantial relation between educational level and CTR. The original
b cautioned that though
blog post on this data, noting the negative value of d,
this seems to indicate that the more-educated people click less, “correlation
is not causation.” Good advice, but it’s equally important to note here that
even if the effect is causal, it is tiny.

1.9 Several Predictor Variables

Now let’s predict weight from height and age. We first need some notation.
Say we are predicting a response variable Y from variables X (1) , ..., X (k) .
The regression function is now defined to be

µ(t1 , ..., tk ) = E(Y | X (1) = t1 , ..., X (k) = tk ) (1.11)

In other words, µ(t1 , ..., tk ) is the mean Y among all units (people, cars,
whatever) in the population for which X (1) = t1 , ..., X (k) = tk .
In our baseball data, Y , X (1) and X (2) might be weight, height and age,
respectively. Then µ(72, 25) would be the population mean weight among
all players of height 72 and age 25.
We will often use a vector notation

µ(t) = E(Y | X = t) (1.12)

with t = (t1 , ..., tk )′ and X = (X (1) , ..., X (k) )′ , where ′

denotes matrix
transpose.11
11 Our vectors in this book are column vectors. However, since they occupy a lot of
18 CHAPTER 1. SETTING THE STAGE

1.9.1 Multipredictor Linear Models

Let’s consider a parametric model for the baseball data,

mean weight = c + d × height + e × age (1.14)

1.9.1.1 Estimation of Coefficients

We can again use lm() to obtain sample estimates of c, d and e:

> lm( mlb$Weight ∼ mlb$ Height + mlb$Age )

...
Coefficients :
( Intercept ) mlb$ Height mlb$Age
−187.6382 4.9236 0.9115

Note that the notation mlb$Weight ∼mlb$Height + mlb$Age simply means

“predict weight from height and age.” The variable to be predicted is spec-
ified to the left of the tilde, and the predictor variables are written to the
right of it. The + does not mean addition.
A shorter formulation is

> lm( Weight ∼ Height + Age , data=mlb )

You can see that if we have many predictors, this notation is more compact
and convenient.
And, shorter still, we could write

> lm( Weight ∼ . , data=mlb [ , 4 : 6 ] )

Here the period means “all the other variables.” Since we are restricting
the data to be columns 4 and 6 of mlb, Height and Age, the period means
those two variables.

space on a page, we will often show them as transposes of rows. For instance, we will
often write (5, 12, 13)′ instead of
 
5
 12  (1.13)
13
1.9. SEVERAL PREDICTOR VARIABLES 19

So, the output shows us the estimated coefficientsis, e.g., db = 4.9236. Our
estimated regression function is

b(t1 , t2 ) = −187.6382 + 4.9236 t1 + 0.9115 t2

µ (1.15)

where t1 and t2 are height and age, respectively.

Setting t1 = 72 and t2 = 25, we find that

b(72, 25) = 189.6485

µ (1.16)

and we would predict the weight of a 72-inch tall, age 25 player to be about
190 pounds.

1.9.1.2 The Description Goal

It was mentioned in Section 1.1 that regression analysis generally has one or
both of two goals, Prediction and Description. In light of the latter, some
brief comments on the magnitudes of the estimated coefficientsis would be
useful at this point:

• We estimate that, on average (a key qualifier), each extra inch in

height corresponds to almost 5 pounds of additional weight.

• We estimate that, on average, each extra year of age corresponds to

almost a pound in extra weight.

That second item is an example of the Description goal in regression anal-

ysis. We may be interested in whether baseball players gain weight as they
age, like “normal” people do. Athletes generally make great efforts to stay
fit, but we may ask how well they succeed in this. The data here seem to
indicate that baseball players indeed are prone to some degree of “weight
creep” over time.

1.9.2 Nonparametric Regression Estimation: k-NN

Now let’s drop the linear model assumption (1.14), and estimate our re-
gression function “from scratch.” So this will be a model-free approach,
thus termed nonparametric as explained earlier.
20 CHAPTER 1. SETTING THE STAGE

Our analysis in Section 1.6.2 was model-free. But here we will need to
broaden our approach, as follows.

1.9.2.1 Looking at Nearby Points

Again say we wish to estimate, using our data, the value of µ(72, 25). A
potential problem is that there likely will not be any data points in our
sample that exactly match those numbers, quite unlike the situation in
b(72) was based on 150 data points. Let’s check:
(1.4), where µ

> z <− mlb [ mlb$ Height == 72 & mlb$Age == 2 5 , ]

> z
[ 1 ] Name Team Position
[ 4 ] Height Weight Age
[ 7 ] PosCategory
<0 rows> ( or 0−length row . names)

(Recall that in R, we use a single ampersand when “and-ing” vector quan-

tities, but use a double one for ordinary logical expressions.)
So, indeed there were no data points matching the 72 and 25 numbers. Since
the ages are recorded to the nearest 0.01 year, this result is not surprising.
But at any rate we thus cannot set µb(72, 25) to be the mean weight among
our sample data points satisfying those conditions, as we did in Section
1.6.2. And even if we had had a few data points of that nature, that would
not have been enough to obtain an accurate estimate µ b(72, 25).
Instead, we use data points that are close to the desired prediction point.
Again taking the weight/height/age case as a first example, this means
that we would estimate µ(72, 25) by the average weight in our sample data
among those data points for which height is near 72 and age is near 25.

1.9.2.2 Measures of Nearness

Nearness is generally defined as Euclidean distance:

√
distance[(s1 , s2 , ..., sk ), (t1 , t2 , ..., tk )] = ((s1 − t1 )2 + ... + (sk − tk )2
(1.17)
For instance, the distance from a player in our sample of height 72.5 and
1.9. SEVERAL PREDICTOR VARIABLES 21

age 24.2 to the point (72,25) would be

√
(72.5 − 72)2 + (24.2 − 25)2 = 0.9434 (1.18)

Note that the Euclidean distance between s = (s1 , ..., sk ) and t = (t1 , ..., tk )
is simply the Euclidean norm of the difference s − t (Section A.1).

1.9.2.3 The k-NN Method, and Tuning Parameters

The k-Nearest Neighbor (k-NN) method for estimating regression functions

is simple: Find the k data points in our sample that are closest to the
desired prediction point, and average their values of the response variable
Y.
A question arises as to how to choose the value of k. Too large a value
means we are including “nonrepresentative” data points, but too small a
value gives us too few points to average for a good estimate. We will return
to this question later, but will note that due to this nature of k, we will
call k a tuning parameter. Various tuning parameters will come up in this
book.

1.9.2.4 Nearest-Neighbor Analysis in the regtools Package

We will use the k-NN functions in my regtools package, available on CRAN

[97]. The main computation is performed by knnest(), with preparatory
nearest-neighbor computation done by preprocessx(). The call forms are
p r e p r o c e s s x ( x , kmax , x v a l=FALSE)
k n n e s t ( y , xdata , k , n e a r f=meany )

In the first, x is our predictor variable data, one column per predictor. The
argument kmax specifies the maximum value of k we wish to use (we might
try several), and xval refers to cross-validation, a concept to be introduced
later in this chapter. The essence of preprocessx() is to find the kmax
nearest neighbors of each observation in our dataset, i.e., row of x.
The arguments of knnest() are as follows. The vector y is our response
variable data; xdata is the output of preprocessx(); k is the number
of nearest neighbors we wish to use. The argument nearf specifies the
function we wish to be applied to the Y values of the neighbors; the default
is the mean, but instead we could for instance specify the median. (This
flexibility will be useful in other ways as well.)
22 CHAPTER 1. SETTING THE STAGE

The return value from knnest() is an object of class ‘knn’.

1.9.2.5 Example: Baseball Player Data

There is also a predict function associated with knnest(), with call form
predict ( kout , p r e d p t s , n e e d t o s c a l e )

Here kout is the return value of a call to knnest(), and each row of
regestpts is a point at which we wish to estimate the regression func-
tion. Also, if the points to be predicted are not in our original data, we
need to set needtoscale to TRUE.
For example, let’s estimate µ(72, 25), based on the 20 nearest neighbors at
each point.
> data ( mlb )
> library ( r e g t o o l s )
> xd <− p r e p r o c e s s x ( mlb [ , c ( 4 , 6 ) ] , 2 0 )
> kout <− k n n e s t ( mlb [ , 5 ] , xd , 2 0 )
> predict ( kout , c ( 7 2 , 2 5 ) ,TRUE)
187.4

So we would predict the weight of a 72-inches tall, age 25 player to be

about 187 pounds, not much different — in this instance — from what we
obtained earlier with the linear model.

1.10 After Fitting a Model, How Do We Use

It for Prediction?

As noted, our goal in regression analysis could be either Prediction or De-

scription (or both). How specifically does the former case work?

1.10.1 Parametric Settings

The parametric case is the simpler one. We fit our data, write down the
result, and then use that result in the future whenever we are called upon
to do a prediction.
Recall Section 1.9.1.1. It was mentioned there that in that setting, we prob-
ably are not interested in the Prediction goal, but just as an illustration,
1.10. USING THE MODEL FOR PREDICTION 23

suppose we do wish to predict. We fit our model to our data — called

our training data — resulting in our estimated regression function, (1.15).
From now on, whenever we need to predict a player’s weight, given his
height and age, we simply plug those values into (1.15).

1.10.2 Nonparametric Settings

The nonparametric case is a little more involved, because we have no explicit

equation like (1.15). Nevertheless, we use our training data in the same way.
For instance, say we need to predict the weight of a player whose height
and age are 73.2 and 26.5, respectively. Our predicted value will then be
b(73.2, 26.5). To obtain that, we go back to our training data, find the k
µ
nearest points to (73.2,26.5), and average the weights of those k players.
We would go through this process each time we are called upon to perform
a prediction.
A variation:
A slightly different approach, which is used in regtools, is as follows. De-
note our training set data as (X1 , Y1 ), ..., (Xn , Yn ), where again the Xi are
typically vectors, e.g., (height,age). We estimate our regression function at
each of the points Xi , forming µ b(Xi ), i = 1, ..., n. Then, when faced with
a new case (X, Y ) for which Y is unknown, we find the single closest Xi
to X, and guess Y to be 1 or 0, depending on whether µ b(Xi ) > 0.5. Since
b(Xi ) already incorporates the neighborhood-averaging operation, doing so
µ
for our new point would be largely redundant. Using only the single closest
point saves both computation time and storage space.

1.10.3 The Generic predict() Function

Consider this code:

> lmout <− lm( Weight ∼ Height + Age , data=mlb )
> predict ( lmout , data . frame ( Height = 7 2 , Age = 2 5 ) )
1
189.6493

We fit the model as in Section 1.9.1.1, and then predicted the weight of a
b(72, 25) for this, which of
player who is 72 inches tall and age 25. We use µ
course we could obtain as
> coef ( lmout ) %∗% c ( 1 , 7 2 , 2 5 )
[ ,1]
24 CHAPTER 1. SETTING THE STAGE

[ 1 , ] 189.6493

But the predict() function is simpler and more explicitly reflects what we
want to accomplish.
By the way, predict is a generic function. This means that R will dispatch
a call to predict() to a function specific to the given class. In this case,
lmout above is of class ’lm’, so the function ultimately executed above is
predict.lm‘. Similarly, in Section 1.9.2.5, the call to predict() goes to
predict.knn(). More details are in Section 1.20.4.
IMPORTANT NOTE: To use predict() with lm(), the latter must be
called in the data = form shown above, and the new data to be predicted
must be a data frame with the same column names.

1.11 Overfitting, and the Variance-Bias

Tradeoff

One major concern in model development is overfitting, meaning to fit such

an elaborate model that it “captures the noise rather than the signal.”
This description is often heard these days, but it is vague and potentially
misleading. We will discuss it in detail in Chapter 9, but it is of such
importance that we introduce it here in this prologue chapter.
The point is that, after fitting our model, we are concerned that it may
fit our training data well but not predict well on new data in the future.12
Let’s look into this further:

1.11.1 Intuition

To see how overfitting may occur, consider the famous bias-variance trade-
off, illustrated in the following example. Again, keep in mind that the
treatment will at this point just be intuitive, not mathematical.
Long ago, when I was just finishing my doctoral study, I had my first
experience with statistical consulting. A chain of hospitals was interested
in comparing the levels of quality of care given to heart attack patients
at its various locations. A problem was noticed by the chain regarding
straight comparison of raw survival rates: One of the locations served a
12 Note that this assumes that nothing changes in the system under study between the
time we collect our training data and the time we do future predictions.
1.11. OVERFITTING, AND THE VARIANCE-BIAS TRADEOFF 25

largely elderly population, and since this demographic presumably has more
difficulty surviving a heart attack, this particular hospital may misleadingly
appear to be giving inferior care.
An analyst who may not realize the age issue here would thus be biasing
the results. The term “bias” here doesn’t mean deliberate distortion of the
analysis, just that the model has a systemic bias, i.e., it is “skewed,” in the
common vernacular. And it is permanent bias, in the sense that it won’t
disappear, no matter how large a sample we take.
Such a situation, in which an important variable is not included in the
analysis, is said to be underfitted. By adding more predictor variables in a
regression model, in this case age, we are reducing bias.
Or, suppose we use a regression model that is linear in our predictors, but
the true regression function is nonlinear. This is bias too, and again it
won’t go away even if we make the sample size huge. This is often called
model bias by statisticians; the economists call the model misspecified.
On the other hand, we must keep in mind that our data is a sample from
a population. In the hospital example, for instance, the patients on which
we have data can be considered a sample from the (somewhat conceptual)
population of all patients at this hospital, past, present and future. A
different sample would produce different regression coefficient estimates.
In other words, there is variability in those coefficients from one sample to
another, i.e., variance. We hope that that variance is small, which gives us
confidence that the sample we have is representative.
But the more predictor variables we have, the more collective variability
there is in the inputs to our regression calculations, and thus the larger the
variances of the estimated coefficients.13 If those variances are large enough,
the bias-reducing benefit of using a lot of predictors may be overwhelmed
by the increased variability of the results. This is called overfitting.
In other words:

In deciding how many (and which) predictors to use, we have a

tradeoff. The richer our model, the less bias, but the higher the
variance.

In Section 1.19.2 it is shown that for any statistical estimator θb (that has
finite variance),

mean squred error = squared bias + variance

13 I wish to thank Ariel Shin for this interpretation.
26 CHAPTER 1. SETTING THE STAGE

Our estimator here is µb(t). This shows the tradeoff: Adding variables, such
as age in the hospital example, reduces squared bias but increases variance.
Or, equivalently, removing variables reduces variance but exacerbates bias.
It may, for example, be beneficial to accept a little bias in exchange for
a sizable reduction in variance, which we may achieve by removing some
predictors from our model.
The trick is to somehow find a “happy medium,” easier said than done.
Chapter 9 will cover this in depth, but for now, we introduce a common
method for approaching the problem:

1.11.2 Example: Student Evaluations of Instructors

In Section 9.9.5 we will analyze a dataset consisting of student evaluations

of instructors. Let’s defer the technical details until then, but here is a
sneak preview.
The main variables here consist of 28 questions on the instructor, such as
“The quizzes, assignments, projects and exams contributed to helping the
learning.” The student gives a rating of 1 to 5 on each question.
Figure 1.5 describes this data, plotting frequency of occurrence against the
questions (the 28, plus 4 others at the beginning). Again, don’t worry about
the details now, but it basically shows there are 3 kinds of instructors: one
kind gets very righ ratings on the 28 questions, across the board; one kind
gets consistently medium-high ratings; and the third kind gets low ratings
across all the questions.
This indicates that we might reduce those 28 questions to just one, in fact
any one of the 28.

1.12 Cross-Validation

The proof of the pudding is in the eating — old English saying

Toward that end, i.e., proof via “eating,” it is common to artificially create
a set of “new” data and try things out there. Instead of using all of our
collected data as our training set, we set aside part of it to serve as simulated
“new” data. This is called the validation set or test set. The remainder
will be our actual training data. In other words, we randomly partition
our original data, taking one part as our training set and the other part to
1.12. CROSS-VALIDATION 27

Figure 1.5: Turkish student evaluations (see color insert)

play the role of new data. We fit our model, or models, to the training set,
then do prediction on the test set, pretending its response variable values
are unknown. We then compare to the real values. This will give us an
idea of how well our models will predict in the future. The method is called
cross-validation.
The above description is a little vague, and since there is nothing like code
to clarify the meaning of an algorithm, let’s develop some. Here first is
code to do the random partitioning of data, with a proportion p to go to
the training set:

x v a l p a r t <− function ( data , p ) {

n <− nrow( data )
n t r a i n <− round ( p∗n )
28 CHAPTER 1. SETTING THE STAGE

t r a i n i d x s <− sample ( 1 : n , n t r a i n , replace=FALSE)

l i s t ( t r a i n=data [ t r a i n i d x s , ] ,
v a l i d=data[− t r a i n i d x s , ] )
}

R uses - in array indices for exclusion, e.g.,

> x <− c ( 5 , 1 2 , 1 3 )
> x [ −2]
[ 1 ] 5 13

Thus, using the expression -trainidxs above gives us the validation cases.
Now to perform cross-validation, we’ll consider the parametric and non-
parametric cases separately, in the next two sections.

1.12.1 Linear Model Case

To do cross-validation for linear models, we could use this code.14

1.12.1.1 The Code

# arguments :
#
# data : f u l l data
# y c o l : column number o f r e s p . v a r .
# p r e d v a r s : column numbers o f p r e d i c t o r s
# p : prop . f o r t r a i n i n g s e t
# meanabs : s e e ’ v a l u e ’ b e l o w

# value : i f meanabs i s TRUE, t h e mean a b s o l u t e

# p r e d i c t i o n e r r o r ; o t h e r w i s e , an R l i s t
# c o n t a i n i n g pred . , r e a l Y

xvallm <− function ( data , y c o l , p r e d v a r s , p , meanabs=TRUE) {

tmp <− x v a l p a r t ( data , p )
t r a i n <− tmp$ t r a i n
v a l i d <− tmp$ v a l i d
# f i t model t o t r a i n i n g d a t a
14 There are sophisticated packages on CRAN for this, such as cvTools [4]. But to

keep things simple, and to better understand the concepts, we will write our own code.
Similarly, as mentioned, we will not use R’s predict() function for the time being.
1.12. CROSS-VALIDATION 29

t r a i n y <− t r a i n [ , y c o l ]
t r a i n p r e d s <− t r a i n [ , p r e d v a r s ]
# u s i n g m a t r i x form i n lm ( ) c a l l
t r a i n p r e d s <− as . matrix ( t r a i n p r e d s )
lmout <− lm( t r a i n y ∼ t r a i n p r e d s )
# a p p l y f i t t e d model t o v a l i d a t i o n d a t a ; n o t e
# t h a t %∗% works o n l y on m a t r i c e s , not d a t a frames
v a l i d p r e d s <− as . matrix ( v a l i d [ , p r e d v a r s ] )
predy <− cbind ( 1 , v a l i d p r e d s )%∗% coef ( lmout )
r e a l y <− v a l i d [ , y c o l ]
i f ( meanabs ) return (mean( abs ( predy − r e a l y ) ) )
l i s t ( predy = predy , r e a l y = r e a l y )
}

1.12.1.2 Applying the Code

Let’s try cross-validtion on the weight/height/age data, using mean abso-

lute prediction error as our criterion for prediction accuracy:
li b r a r y ( f r e q p a r c o o r d )
data ( mlb )
xvallm ( mlb , 5 , c ( 4 , 6 ) , 2 / 3 )

> xvallm ( mlb , 5 , c ( 4 , 6 ) , 2 / 3 )

[ 1 ] 13.38045

So, on average we would be off by about 13 pounds. We might improve

upon this by using the data’s Position variable, but we’ll leave that for
later.
Keep in mind the randomness, though. We randomly split the data, and
would get a different result if we were to run the code again. This point is
explored in Exercise 1 at the end of this chapter. Also, we will later discuss
an extension, ir-fold cross-validation, in Section 2.9.6.

1.12.2 k-NN Case

Here is the code for performing cross-validation for k-NN:

# arguments :
#
# data : f u l l data
# y c o l : column number o f r e s p . v a r .
30 CHAPTER 1. SETTING THE STAGE

# k : number o f n e a r e s t n e i g h b o r s
# p : prop . f o r t r a i n i n g s e t
# meanabs : s e e ’ v a l u e ’ b e l o w

# value : i f meanabs i s TRUE, t h e mean a b s o l u t e

# p r e d i c t i o n e r r o r ; o t h e r w i s e , an R l i s t
# c o n t a i n i n g pred . , r e a l Y

xvalknn <−
function ( data , y c o l , p r e d v a r s , k , p , meanabs=TRUE) {
# c u l l o u t j u s t Y and t h e Xs
data <− data [ , c ( p r e d v a r s , y c o l ) ]
y c o l <− length ( p r e d v a r s ) + 1
tmp <− x v a l p a r t ( data , p )
t r a i n <− tmp$ t r a i n
v a l i d <− tmp$ v a l i d
v a l i d <− as . matrix ( v a l i d )
xd <− p r e p r o c e s s x ( t r a i n [ , − y c o l ] , k )
kout <− k n n e s t ( t r a i n [ , y c o l ] , xd , k )
predy <− predict ( kout , v a l i d [ , − y c o l ] ,TRUE)
r e a l y <− v a l i d [ , y c o l ]
i f ( meanabs ) return (mean( abs ( predy − r e a l y ) ) )
l i s t ( predy = predy , r e a l y = r e a l y )
}

So, how well does k-NN predict?

> library ( r e g t o o l s )
> set . s e e d ( 9 9 9 9 )
> xvalknn ( mlb , 5 , c ( 4 , 6 ) , 2 5 , 2 / 3 )
[ 1 ] 14.32817

The two methods gave similar results. However, not only must we keep
in mind the randomness of the partitioning of the data, but we also must
recognize that this output above depended on choosing a value of 25 for k,
the number of nearest neighbors. We could have tried other values of k,
and in fact could have used cross-validation to choose the “best” value.

1.12.3 Choosing the Partition Sizes

One other problem, of course, is that we did have a random partition of

our data. A different one might have given substantially different results.
1.13. IMPORTANT NOTE ON TUNING PARAMETERS 31

In addition, there is the matter of choosing the sizes of the training and
validation sets (e.g., via the argument p in xvalpart()). We have a classical
tradeoff at work here: Let k be the size of our training set. If we make k
too large, the validation set will be too small for an accurate measure of
prediction accuracy. We won’t have that problem if we set k to a smaller
size, but then we are measuring the predictive ability of only k observations,
whereas in the end we will be using all n observations for predicting new
data.
The Leaving One-Out Method and its generalizations solves this problem,
albeit at the expense of much more computation. It will be presented in
Section 2.9.5.

1.13 Important Note on Tuning Parameters

Recall how k-NN works: To predict a new case for which X = t but Y is
unknown, we look at the our existing data. We find the k closest neighbors
to t, then average their Y values. That average becomes our predicted value
for the new case.
We refer to k as a tuning parameter, to be chosen by the user. Many
methods have multiple tuning parameters, making the choice a challenge.
One can of course choose their values using cross validation, and in fact the
caret package includes methods to automate the process, simultaneously
optimizing over many tuning parameters.
But cross-validation can have its own overfitting problems (Section 9.3.2).
One should not be lulled into a false sense of security.
The late Leo Breiman was suspicious of tuning parameters, and famously
praised one regression method (boosting), as “the best off-the-shelf method”
available — meaning that the method works well without tweaking tuning
parameters. His statement may have been overinterpreted regarding the
boosting method, but the key point here is that Breiman was not a fan of
tuning parameters.
A nice description of Breiman’s view was given in an obituary by Michael
Jordan, who noted [78],

Another preferred piece of Breimanesque terminology was “off-

the-shelf,” again a rather physical metaphor. Leo tended to be
suspicious of “free parameters;” procedures should work with
little or no “tuning.”
32 CHAPTER 1. SETTING THE STAGE

Breiman’s concerns about tuning parameters extended to choosing those

parameters via cross-validation. The latter is an important tool, but should
be used with a healthy dose of skepticism.
This issue will come up often, since many commonly-used methods do have
various tuning parameters; some have multiple, complex tuning parameters.
Again, this point must be kept in mind:

Optimizing for a tuning parameter is inherently prone to

overfitting, as we are optimizing for particular data. If we
have multiple tuning parameters, the potential for overfitting is
compounded.

1.14 Rough Rule of Thumb

The issue of how many predictors to use to simultaneously avoid overfitting

and still produce a good model is nuanced, and in fact this is still not fully
resolved. Chapter 9 will be devoted to this complex matter.
Until then, though it is worth using the following:15

Rough Rule of Thumb (Tukey): √ For a data set consisting

of n observations, use fewer than (n) predictors.

1.15 Example: Bike-Sharing Data

We now return to the bike-sharing data (Section 1.1). Our little excursion to
the simpler data set, involving baseball player weights and heights, helped
introduce the concepts in a less complex setting. The bike-sharing data set
is more complicated in several ways:

• Complication (a): It has more potential predictor variables.

• Complication (b): It includes some nominal (or categorical) vari-

ables, such as Day of Week. The latter is technically numeric, 0
through 6, but those codes are just names. Hence the term nominal.
In R, by the way, the formal term for such variables is factors.
15 Unfortunately, reference unknown.
1.15. EXAMPLE: BIKE-SHARING DATA 33

The problem is that there is no reason, for instance, that Sunday,

Thursday and Friday should have an ordinal relation in terms of rid-
ership just because, say, 0 < 4 < 5.

• Complication (c): It has some potentially nonlinear relations. For

instance, people don’t like to ride bikes in freezing weather, but they
are not keen on riding on really hot days either. Thus we might
suspect that the relation of ridership to temperature rises at first,
eventually reaching a peak, but declines somewhat as the temperature
increases further.

Now that we know some of the basic issues from analyzing the baseball
data, we can treat this more complicated data set.
Let’s read in the bike-sharing data. We’ll look at one of the files in that
dataset, day.csv. We’ll restrict attention to the first year,16 and since we
will focus on the registered riders, let’s shorten the name for convenience:
> s h a r <− read . csv ( ” day . c s v ” , h e a d e r=TRUE)
> s h a r <− s h a r [ 1 : 3 6 5 , ]
> names( s h a r ) [ 1 5 ] <− ” r e g ”

1.15.1 Linear Modeling of µ(t)

In view of Complication (c) above, the inclusion of the word linear in the
title of our current section might seem contradictory. But one must look
carefully at what is linear or not, and we will see shortly that, yes, we can
use linear models to analyze nonlinear relations.
Let’s first check whether the ridership/temperature relation seems nonlin-
ear, as we have speculated:
plot ( s h a r $temp , s h a r $ r e g )

The result is shown in Figure 1.6.

There seem to be some interesting groupings among the data, likely due to
the other variables, but putting those aside for now, the plot does seem to
suggest that ridership is slightly associated with temperature in the “first
rising, then later falling” form as we had guessed.
16 Thereappears to have been some systemic change in the second year, and while this
could be modeled, we’ll keep things simple by considering only the first year.
34 CHAPTER 1. SETTING THE STAGE

Figure 1.6: Ridership vs. temperature

Thus a linear model of the form

mean ridership = c + d × temperature (1.19)

would seem inappropriate. But don’t give up so quickly! A model like

mean ridership = c + d × temperature + e × temperature2 (1.20)

i.e., with a temperature-squared term added, might work fine. A negative

value for e would give us the “first up, then later down” behavior we want
our model to have.
And there is good news — the model (1.20) is actually linear! We say that
the expression is linear in the parameters, even though it is nonlinear with
respect to the temperature variable. This means that if we multiply each
of c, d and e by, say, 8, then the values of the left and right sides of the
equation both increase eightfold.
1.15. EXAMPLE: BIKE-SHARING DATA 35

Anotber way to see this is that in calling lm(), we can simply regard
squared temperature as a new variable:

> s h a r $temp2 <− s h a r $temp ˆ2

> lm( s h a r $ r e g ∼ s h a r $temp + s h a r $temp2 )

Call :
lm( formula = s h a r $ r e g ∼ s h a r $temp + s h a r $temp2 )

Coefficients :
( Intercept ) s h a r $temp s h a r $temp2
−378.9 9841.8 −6169.8

And note that, sure enough, the coefficient of the squared term, eb =
−6169.8, did indeed turn out to be negative.
Of course, we want to predict from many variables, not just temperature,
so let’s now turn to Complication (b) cited earlier, the presence of nominal
data. This is not much of a problem either.
Such situations are generally handled by setting up what are called indicator
variables or dummy variables. The former term alludes to the fact that our
variable will indicate whether a certain condition holds or not, with 1 coding
the yes case and 0 indicating no.
We could, for instance, set up such a variable for Tuesday data:

> s h a r $ t u e s <− as . integer ( s h a r $weekday == 2 )

Indeed, we could define six variables like this, one for each of the days
Monday through Saturday. Note that Sunday would then be indicated
indirectly, via the other variables all having the value 0. A direct Sunday
variable would be redundant, and in fact would present mathematical prob-
lems, as we’ll see in Chapter 8. (Actually, R’s lm() function can deal with
factor variables directly, as shown in Section 9.7.5.1. But we take the more
basic route here, in order to make sure the underlying principles are clear.)
However, let’s opt for a simpler analysis, in which we distinguish only be-
tween weekend days and weekdays, i.e. define a dummy variable that is 1
for Monday through Friday, and 0 for the other days. Actually, those who
assembled the data set already defined such a variable, which they named
workingday.17
17 More
specifically, a value of 1 for this variable indicates that the day is in the
Monday-Friday range and it is not a holiday.
36 CHAPTER 1. SETTING THE STAGE

We incorporate this into our linear model:

mean reg = c + d × temp + e × temp2 + f workingday (1.21)

There are several other dummy variables that we could add to our model,
but for this introductory example let’s define just one more:
> s h a r $ c l e a r d a y <− as . integer ( s h a r $ w e a t h e r s i t == 1 )

So, our regression model will be

mean reg = β0 + β1 temp + β2 temp2

+ β3 workingday + β4 clearday (1.22)

As is traditional, here we have used subscripted versions of the Greek letter

β to denote our equation coefficients, rather than c, d and so on.
So, let’s run this through lm():
> lmout <− lm( r e g ∼ temp+temp2+workingday+c l e a r d a y ,
data = s h a r )

The return value of lm(), assigned here to lmout, is a very complicated

R object, of class ’lm’. We shouldn’t inspect it in detail now, but let’s at
least print the object, which in R’s interactive mode can be done simply by
typing the name, which automatically calls print() on the object:18
> lmout
...
...
Coefficients :
( Intercept ) temp temp2 workingday
−1362.6 11059.2 −7636.4 686.0
clearday
518.9

Remember, the population function µ(t) is unnown, so the βi are unknown.

The above coefficients are merely sample-based estimates. For example,
using our usual “hat” notation to mean “estimate of,” we have that

βb3 = 686.0 (1.23)

18 See more detail on this in Section 1.20.4.
1.15. EXAMPLE: BIKE-SHARING DATA 37

The estimated regression function is then

b(t1 , t2 , t3 , t4 ) = −1362.6 + 11059.2t1 − 7636.4t2 + 686.0t3 + 518.9t4 (1.24)

µ

where t2 = t21 .
So, what should we predict for the number of riders on the type of day de-
scribed at the outset of this chapter — Sunday, sunny, 62 degrees Fahren-
heit? First, note that the designers of the data set have scaled the temp
variable to [0,1], as

Celsius temperature − minimum

(1.25)
maximum − minimum

where the minimum and maximum here were -8 and 39, respectively. This
form may be easier to understand, as it is expressed in terms of where the
given temperature fits on the normal range of temperatures. A Fahrenheit
temperature of 62 degrees corresponds to a scaled value of 0.525. So, our
predicted number of riders is
> coef ( lmout ) %∗% c ( 1 , 0 . 5 2 5 , 0 . 5 2 5 ˆ 2 , 0 , 1 )
[ ,1]
[ 1 , ] 2857.677

So, our predicted number of riders for sunny, 62-degree Sundays will be
about 2858. How does that compare to the average day?
> mean( s h a r $ r e g )
[ 1 ] 2728.359

So, we would predict a somewhat above-average level of ridership.

As noted earlier, one can also form confidence intervals and perform sig-
nificance tests on the βi . We’ll go into this in Chapter 2, but some brief
comments on the magnitudes and signs of the βbi are useful at this point:

• As noted, the estimated coefficient of temp2 is negative, consistent

with our intuition. Note, though, that it is actually less negative than
when we predicted reg from only temperature and its square. This
change is typical, and will be discussed in detail in Chapter 7.
• The estimated coefficient for workingday is positive. This too matches
our intuition, as presumably many of the registered riders use the
38 CHAPTER 1. SETTING THE STAGE

bikes to commute to work. The value of the estimate here, 686.0, in-
dicates that, for fixed temperature and weather conditions, weekdays
tend to have close to 700 more registered riders than weekends.

• Similarly, the coefficient of clearday suggests that for fixed temper-

ature and day of the week, there are about 519 more riders on clear
days than on other days.

1.15.2 Nonparametric Analysis

Let’s see what k-NN gives us as our predicted value for sunny, 62-degree
Sundays, say with k = 20:

> s h a r 1 <−
s h a r [ , c ( ’ workingday ’ , ’ temp ’ , ’ r e g ’ , ’ c l e a r d a y ’ ) ]
> xd <− p r e p r o c e s s x ( s h a r 1 [ , − 3 ] , 2 0 )
> kout <− k n n e s t ( s h a r 1 $ reg , xd , 2 0 )
> predict ( kout , c ( 0 , 0 . 5 2 5 , 1 ) ,TRUE)
2881.8

This is again similar to what the linear model gave us. This probably means
that the linear model was pretty good, but we will discuss this in detail in
Chapter 6.

1.16 Interaction Terms, Including Quadrat-

ics

Let’s take another look at (1.22), specifically the term involving the variable
workingday, a dummy indicating a nonholiday Monday through Friday.
Our estimate for β3 turned out to be 686.0, meaning that, holding temper-
ature and the other variables fixed, there is a mean increase of about 686.0
riders on working days.
But look at our model, (1.22). The (estimated) values of the right-hand
side will differ by 686.0 for working vs. nonworking days, no matter what
the temperature is. In other words, the working day effect is the same on
low-temprerature days as on warmer days. For a broader model that does
not make this assumption, we could add an interaction term, consisting of
a product of workingday and temp:
1.16. INTERACTION TERMS, INCLUDING QUADRATICS 39

mean reg = β0 + β1 temp + β2 temp2

+ β3 workingday + β4 clearday (1.26)
+ β5 temp × workingday (1.27)

Note that the temp2 term is also an interaction term, the interaction of the
temp variable with itself.
How does this model work? Let’s illustrate it with a new data set.

1.16.1 Example: Salaries of Female Programmers

and Engineers

This data is from the 2000 U.S. Census, consisting of 20,090 programmers
and engineers in the Silicon Valley area. The data set is included in the
freqparcoord package on CRAN [104]. Suppose we are working toward a
Description goal, specifically the effects of gender on wage income.
As with our bike-sharing data, we’ll add a quadratic term, in this case
on the age variable, reflecting the fact that many older programmers and
engineers encounter trouble finding work [108]. Let’s restrict our analysis
to workers having at least a Bachelor’s degree, and look at the variables
age, age2, sex (coded 1 for male, 2 for female), wkswrked (number of
weeks worked), ms, phd and wageinc (wage income). Other than an age2
term, we’ll start out with no interaction terms.

> library ( freqparcoord )

> data ( prgeng )
> prgeng $ age2 <− prgeng $ age ˆ2
> edu <− prgeng $educ
> prgeng $ms <− as . integer ( edu == 1 4 )
> prgeng $phd <− as . integer ( edu == 1 6 )
> prgeng $fem <− prgeng $ s e x − 1
> tmp <− prgeng [ edu >= 1 3 , ]
> pe <− tmp [ , c ( 1 , 1 2 , 9 , 1 3 , 1 4 , 1 5 , 8 ) ]
> pe <− as . matrix ( pe )

Our model is
40 CHAPTER 1. SETTING THE STAGE

mean wageinc = β0 + β1 age + β2 age2 + β3 wkswrkd

+ β4 ms + β5 phd
+ β6 fem (1.28)

We find the following:

> lm( wageinc ∼
age+age2+wkswrkd+ms+phd+fem , data=prgeng )
...
Coefficients :
( Intercept ) age age2 wkswrkd
−81136.70 3900.35 −40.33 1196.39
ms phd fem
15431.07 23183.97 −11484.49

The model probably could use some refining, for example variables we have
omitted, such as occupation. But as a preliminary statement, the results are
striking in terms of gender: With age, education and so on held constant,
women are estimated to have incomes about $11,484 lower than comparable
men.
But this analysis implicitly assumes that the female wage deficit is, for
instance, uniform across educational levels. To see this, consider (1.28).
Being female makes a β6 difference, no matter what the values of ms and
phd are. (For that matter, this is true of age too, though we won’t model
that here for simplicity.) To generalize our model in this regard, let’s define
two interaction variables, the product of ms and fem, and the product of
phd and fem.
Our model is now

mean wageinc = β0 + β1 age + β2 age2 + β3 wkswrkd

+ β4 ms + β5 phd
+ β6 fem + β7 msfem + β8 phdfem (1.29)

So, now instead of there being a single number for the “female effect,” β6 ,
we how have three:

• Female effect for holders of a Bachelor’s degree: β6

1.16. INTERACTION TERMS, INCLUDING QUADRATICS 41

• Female effect for Master’s degree holders: β6 + β7

• Female effect for PhD degree holders β6 + β8

So, let’s rerun the regression analysis:

> prgeng $msfem <− prgeng $ms ∗ prgeng $fem
> prgeng $phdfem <− prgeng $phd ∗ prgeng $fem
> lm( wageinc ∼
age+age2+wkswrkd+ms+phd+fem+msfem+phdfem ,
data=prgeng )
...
Coefficients :
( Intercept ) age age2 wkswrkd
−81216.78 3894.32 −40.29 1195.31
ms phd fem msfem
16433.67 25325.31 −10276.80 −4157.25
phdfem
−14061.64

Let’s compute the estimated values of the female effects, first for a worker
with less than a graduate degree. This is -10276.80. For the Master’s case,
the mean female effect is estimated to be -10276.80 - 4157.25 = -14434.05.
For a PhD, the figure is -10276.80 - 14861.64 = -25138.44. In other words,
Once one factors in educational level, the gender gap is seen to be even
worse than before.
Thus we still have many questions to answer, especially since we haven’t
considered other types of interactions yet. This story is not over yet, and
will be pursued in detail in Chapter 7.
Rather than creating the interaction terms “manually” as is done here, one
can use R colon operator, e.g., ms:fem, which automates the process. This
was not done above, so as to ensure that the reader fully understands the
meaning of interaction terms. But this is how it would go:

> lm( wageinc ∼ age+age2+wkswrkd+ms+phd+fem+

ms : fem+phd : fem , data=prgeng )
...
Coefficients :
( Intercept ) age age2
−81216.78 3894.32 −40.29
wkswrkd ms phd
1195.31 16433.67 25325.31
42 CHAPTER 1. SETTING THE STAGE

fem ms : fem phd : fem

−10276.80 −4157.25 −14061.64

For information on the colon and related operators, type ?formula at the
R prompt.

1.16.2 Fitting Separate Models

Suppose we have a model that includes a dummy predictor D, and we form

interaction terms between D and other predictors. In essence, this is the
same as fitting two regression models without interaction terms, one for the
subpopulation D = 1 and the other for D = 0. To see this, consider again
the census data above.
To keep things simple, let’s just one other predictor, the age variable, and
take D to be the dummy variable for female:
> data ( prgeng )
> prgeng $fem <− prgeng $ s e x − 1
> fm <− which ( prgeng $fem == 1 )
> male <− prgeng [−fm , ] # d a t a from male subpop
> f e m a l e <− prgeng [ fm , ] # d a t a from f e m a l e subpop
> lm( wageinc ∼ age , data=male )
Coefficients :
( Intercept ) age
44313.2 486.2
> lm( wageinc ∼ age , data=f e m a l e )
Coefficients :
( Intercept ) age
30551 503
> lm( wageinc ∼ age+fem+age ∗fem , data=prgeng )
Coefficients :
( Intercept ) age fem age : fem
44313.2 486.2 −13761.7 16.8

Look at that last result. For a female worker, fem and age:fem would be
equal to 1 and age, respectively. That means the coefficent for age would
be 486.2 + 16.8 = 503, which matches the 503 value obtained from running
lm() with data = female. For a male worker, fem and age:fem would
both be 0, and the age coefficent is then 486.2, matching the lm() results
for the male data. The intercept terms match similarly.
The reader may be surprised that the estimated age coefficient is higher
1.16. INTERACTION TERMS, INCLUDING QUADRATICS 43

for the women than the men. The problem is that the intercept term
is much lower for women, and the line for men is above that for women
for all reasonable values of age. At age 50, for instance, the estimated
mean for men is 44313.2 + 486.2 × 50 = 68623.2, while for women it is
30551 + 503 × 50 = 55701.
If our goal is Description, running separate regression models like this may
be much easier to interpret. This is highly encouraged. However, things
become unwieldy if we have multiple dummies; if there are d of them, we
must fit 2d separate models.

1.16.3 Saving Your Work

Readers who are running the book’s examples on their computers may find
it convenient to use R’s save() and load() functions. Our pe data above
will be used again at various points in the book, so it is worthwhile to save
it:

> save ( pe , f i l e= ’ pe . s a v e ’ )

Later — days, weeks, whatever — you can reload it by simply typing

load ( ’ pe . s a v e ’ )

Your old pe object will now be back in memory. This is a lot easier than re-
loading the original prgeng data, adding the fem, ms and phd variables,
etc.

1.16.4 Higher-Order Polynomial Models

Theoretically, we need not stop with quadratic terms. We could add cubic
terms, quartic terms and so on. Indeed, the famous Stone-Weierstrass
Theorem [123] says that any continuous function can be approximated to
any desired accuracy by some high-order polynomial.
But this is not practical. In addition to the problem of overfitting there
are numerical issues. In other words, roundoff errors in the computation
would render it meaningless at some point, and indeed lm() will refuse to
compute if it senses a situation like this. See Exercise 1 in Chapter 8.
44 CHAPTER 1. SETTING THE STAGE

1.17 Classification Techniques

Recall the hospital example in Section 1.11.1. There the response variable
is nominal, represented by a dummy variable taking the values 1 and 0,
depending on whether the patient survives or not. This is referred to as
a classification problem, because we are trying to predict which class the
population unit belongs to — in this case, whether the patient will belong
to the survival or nonsurvival class. We could set up dummy variables
for each of the hospital branches, and use these to assess whether some
were doing a better job than others, while correcting for variations in age
distribution from one branch to another. (Thus our goal here is Description
rather than directly Prediction itself.)
The point is that we are predicting a 1-0 variable. In a marketing con-
text, we might be predicting which customers are more likely to purchase
a certain product. In a computer vision context, we may want to predict
whether an image contains a certain object. In the future, if we are for-
tunate enough to develop relevant data, we might even try our hand at
predicting earthquakes.
Classification applications are extremely common. And in many cases there
are more than two classes, such as in identifying many different printed
characters in computer vision.
In a number of applications, it is desirable to actually convert a problem
with a numeric response variable into a classification problem. For instance,
there may be some legal or contractual aspect that comes into play when our
variable V is above a certain level c, and we are only interested in whether
the requirement is satisfied. We could replace V with a new variable

{
1, if V > c
Y = (1.30)
0, if V ≤ c

Classification methods will play a major role in this book.

1.17.1 It’s a Regression Problem!

Recall that the regression function is the conditional mean:

µ(t) = E(Y | X = t) (1.31)

1.17. CLASSIFICATION TECHNIQUES 45

(As usual, X and t may be vector-valued.) In the classification case, Y is an

indicator variable, which implies that we know its mean is the probability
that Y = 1 (Section 1.19.1). In other words,

µ(t) = P (Y = 1 | X = t) (1.32)

The great implication of this is that the extensive knowledge about regression
analysis developed over the years can be applied to the classification problem.
One intuitive strategy would be to guess that Y = 1 if the conditional
probability of 1 is greater than 0.5, and guess 0 otherwise. In other words,

{
1, if µ(X) > 0.5
guess for Y = (1.33)
0, if µ(X) ≤ 0.5

It turns out that this strategy is optimal, in that it minimizes the overall
misclassification error rate (see Section 1.19.4 in the Mathematical Com-
plements portion of this chapter). However, it should be noted that this
is not the only possible criterion that might be used. We’ll return to this
issue in Chapter 5.
As before, note that (1.32) is a population quantity. We’ll need to estimate
it from our sample data.

1.17.2 Example: Bike-Sharing Data

Let’s take as our example the situation in which ridership is above 3500
bikes, which we will call HighUsage:
> s h a r $ h i g h u s e <− as . integer ( s h a r $ r e g > 3 5 0 0 )

We’ll try to predict that variable. Let’s again use our earlier example, of
a Sunday, clear weather, 62 degrees. Should we guess that this will be a
High Usage day?
We can use our k-NN approach just as before. Indeed, we don’t need to
re-run preprocessx().
> kout <− k n n e s t ( as . integer ( s h a r 1 $ r e g > 3 5 0 0 ) , xd , 2 0 )
> predict ( kout , c ( 0 , 0 . 5 2 5 , 1 ) ,TRUE)
0.1
46 CHAPTER 1. SETTING THE STAGE

We estimate that there is a 10% chance of that day having HighUsage.

The parametric case is a little more involved. A model like

probability of HighUsage = β0 + β1 temp + β2 temp2

+ β3 workingday + β4 clearday (1.34)

could be used, but would not be very satisfying. The left-hand side of
(1.34), as a probability, should be in [0,1], but the right-hand side could in
principle fall far outside that range.
Instead, the most common model for conditional probability is logistic re-
gression:

probability of HighUsage = ℓ(β0 + β1 temp + β2 temp2

+ β3 workingday + β4 clearday) (1.35)

where ℓ(s) is the logistic function,

1
ℓ(s) = (1.36)
1 + e−s

Our model, then, is

1
µ(t1 , t2 , t3 , t4 ) = (1.37)
1+ e−(β0 +β1 t1 +β2 t2 +β3 t3 +β4 t4 )

where t1 is temperature, t2 is the square of temperature, and so on. We

wish to estimate µ(62, 622 , 0, 1).
Note the form of the curve, shown in Figure 1.7 The appeal of this model
is clear at a glance: First, the logistic function produces a value in [0,1], as
appropriate for modeling a probability. Second, it is a monotone increasing
function in each of the variables in (1.35), just as was the case in (1.22)
for predicting our numeric variable, reg. Other motivations for using the
logistic model will be discussed in Chapter 4.
R provides the glm() (“generalized linear model”) function for several non-
linear model families, including the logistic,19 which is designated via fam-
ily = binomial:
19 Often called “logit,” by the way.
1.18. CRUCIAL ADVICE: DON’T AUTOMATE, PARTICIPATE! 47

Figure 1.7: Logistic function

> s h a r $ h i g h u s e <− as . integer ( s h a r $ r e g > 3 5 0 0 )

> glmout <− glm( h i g h u s e ∼
temp+temp2+workingday+c l e a r d a y ,
data=shar , family=binomial )
> tmp <− coef ( glmout ) %∗% c ( 1 , 0 . 5 2 5 , 0 . 5 2 5 ˆ 2 , 0 , 1 )
> 1/(1+exp(−tmp ) )
[ ,1]
[ 1 , ] 0.1010449

So, our parametric model gives an almost identical result here to the one
arising from k-NN, about a 10% probability of HighUsage.

1.18 Crucial Advice: Don’t Automate, Par-

ticipate!

Data science should not be a “spectator sport”; the methodology is effec-

tive only if the users participate. Avoid ceding the decision making to the
computer output. For example:
48 CHAPTER 1. SETTING THE STAGE

• Statistical significance does not imply practical importance, and con-

versely.

• A model is just that — just an approximation to reality, hopefully

useful but never exact.

• Don’t rely solely on variable selection algorithms to choose your model

(Chapter 9).

• “Read directions before use” — make sure you understand what a

method really does before employing it.

1.19 Mathematical Complements

1.19.1 Indicator Random Variables

A random variable W is an indicator variable, if it is equal to 1 or 0, de-

pending on whether a certain event Q occurs or not. Two simple properties
are very useful:

• EW = P (Q)
This follows from

EW = 1 · P (Q) + 0 · P (not Q) = P (Q) (1.38)

• V ar(W ) = P (Q) · [1 − P (Q)]

True because

V ar(W ) = E(W 2 ) − (EW )2 = E(W ) − E(W 2 ) = EW (1 − EW )

(1.39)
where the second equality stems from W 2 = W (remember, W is
either 1 or 0). Then use the first bullet above!

1.19.2 Mean Squared Error of an Estimator

Say we are estimating some unknown population value θ, using an estimator

θb based on our sample data. Then a natural measure of the accuracy of
1.19. MATHEMATICAL COMPLEMENTS 49

our estimator is the Mean Squared Error (MSE),

E[(θb − θ)2 ] (1.40)

This is the squared distance from our estimator to the true value, averaged
over all possible samples.
Let’s rewrite the quantity on which we are taking the expected value:

( )2 ( )2
θb − θ = θb − E θb + E θb − θ = (θ−E
b b 2 +(E θ−θ)
θ) b 2 b
+2(θ−E b θ−θ)
θ)(E b
(1.41)
Look at the three terms on the far right of (1.41). The expected value of
b by definition of variance.
the first is V ar(θ),

As to the second term, E θb − θ is the bias of θ,

b the tendency of θb to over-
or underestimate θ over all possible samples.
What about the third term? Note first that E θb − θ is a constant, thus
factoring out of the expectation. But for what remains,

E(θb − E θ)
b =0 (1.42)

Taking the expected value of both sides of (1.41), and taking the above
remarks into account, we have

b =
MSE(θ) b + (E θb − θ)2
V ar(θ) (1.43)
= variance + bias2 (1.44)

In other words:

The MSE of θb is equal to the variance of θb plus squared bias of

b
θ.

1.19.3 µ(t) Minimizes Mean Squared Prediction Error

Claim: Consider all the functions f() with which we might predict Y from
X, i.e., Yb = f (X). The one that minimizes mean squared prediction error,
E[(Y − f (X))2 ], is the regression function, µ(t) = E(Y | X = t).
50 CHAPTER 1. SETTING THE STAGE

(Note that the above involves population quantities, not samples. Consider
the quantity E[(Y −f (X))2 ], for instance. It is the mean squared prediction
error (MSPE) over all (X, Y ) pairs in the population.)
To derive this, first ask, for any (finite-variance) random variable W , what
number c minimizes the quantity E[(W − c)2 ]? The answer is c = EW . To
see this, write

E[(W − c)2 ] = E(W 2 − 2cW + c2 ] = E(W 2 ) − 2cEW + c2 (1.45)

Setting to 0 the derivative of the right-hand side with respect to c, we find

that indeed, c = EW .
Now use the Law of Total Expectation (Section 1.19.5):

[ ]
MSPE = E[(Y − f (X))2 ] = E E((Y − f (X))2 |X) (1.46)

In the inner expectation, X is a constant, and from the statement following

(1.45) we know that the minimizing value of f (X) is “EW,” in this case
E(Y |X), i.e. µ(X). Since that minimizes the inner expectation for any X,
the overall expectation is minimized too.

1.19.4 µ(t) Minimizes the Misclassification Rate

We are concerned here with the classification context. It shows that if we

know the population distribution — we don’t, but are going through this
exercise to guide our intuition — the conditional mean provides the optimal
action in the classification context.
Remember, in this context, µ(t) = P (Y | X = t), i.e. the conditional mean
reduces to the conditional probability. Now plug in X for t, and we have
the following.
Claim: Consider all rules based on X that produce a guess Yb , taking on
values 0 and 1. The one that minimizes the overall misclassification rate
P (Yb ̸= Y ) is

{
1, if µ(X) > 0.5
Yb = (1.47)
0, if µ(X) ≤ 0.5

The claim is completely intuitive, almost trivial: After observing X, how

1.19. MATHEMATICAL COMPLEMENTS 51

should we guess Y ? If conditionally Y has a greater than 50% chance of

being 1, then guess it to be 1!
(Note: In some settings, a “false positive” may be worse than a “false
negative,” or vice versa. The reader should ponder how to modify the
material here for such a situation. We’ll return to this issue in Chapter 5.)
Think of this simple situation: There is a biased coin, with known prob-
ability of heads p. The coin will be tossed once, and we are supposed to
guess the outcome.
Let’s name your guess g (a nonrandom constant), and let C denote the
as-yet-unknown outcome of the toss (1 for heads, 0 for tails). Then the
reader should check that, no matter whether we choose 0 or 1 for g, the
probability that we guess correctly is

P (C = g) = P (C = 1)g + P (C = 0)(1 − g) (1.48)

= pg + (1 − p)(1 − g) (1.49)
= [2p − 1]g + 1 − p (1.50)

Now remember, p is known. How should we choose g, 0 or 1, in order

to maximize (1.50), the probability that our guess is correct? Inspecting
(1.50) shows that maximizing that expression will depend on whether 2p−1
is positive or negative, i.e., whether p > 0.5 or not. In the former case we
should choose g = 1, while in the latter case g should be chosen to be 0.
The above reasoning gives us the very intuitive — actually trivial, when
expressed in English — result:

If the coin is biased toward heads, we should guess heads. If the

coin is biased toward tails, we should guess tails.

Now to show the original claim, we use The Law of Total Expectation. This
will be discussed in detail in Section 1.19.5, but for now, it says this:

E(V ) = E[E(V |U )] (1.51)

i.e. the expected value of a conditional random variable is the unconditional

expectation. In the case where V is an indicator random variable, the above
reduces to

P (A) = E[P (A | U )] (1.52)

52 CHAPTER 1. SETTING THE STAGE

Returning to our original claim, write

[ ]
P (Yb = Y ) = E P (Yb = Y | X) (1.53)

In that inner probability, “p” is

P (Y = 1 | X) = µ(X) (1.54)

which completes the proof.

1.19.5 Some Properties of Conditional Expectation

Since the regression function is defined as a conditional expected value,

as in (1.3), for mathematical analysis we’ll need some properties. First, a
definition.

1.19.5.1 Conditional Expectation As a Random

Variable

For any random variables U and V with defined expectation, either of which
could be vector-valued, define a new random variable W , as follows. First
note that the conditional expectation of V given U = t is a function of t,

µ(t) = E(V | U = t) (1.55)

√
This is an ordinary function, just like, say, t. But we can turn that
ordinary function √
into a random variable by plugging in a random variable,
say Q, for t: R = Q is a random variable. Thinking along these lines, we
define the random variable version of conditional expectation accordingly.
In the function µ(t) in (1.55), we plug in U for t:

W = E(V |U ) = µ(U ) (1.56)

This W is a random variable. As a simple example, say we choose a number

U at random from the numbers 1 through 5. We then randomly choose a
second number V , from the numbers 1 through U . Then

1+t
µ(t) = E(V | U = t) = (1.57)
2
1.19. MATHEMATICAL COMPLEMENTS 53

We now form a new random variable W = (1 + U )/2.

And, since W is a random variable, we can talk of its expected value, which
turns out to be an elegant result:

1.19.5.2 The Law of Total Expectation

A property of conditional expected value, proven in many undergraduate

probability texts, is

E(V ) = EW = E[E(V | U )] (1.58)

The foreboding appearance of this equation belies the fact that it is actually
quite intuitive, as follows. Say you want to compute the mean height of all
people in the U.S., and you already have available the mean heights in each
of the 50 states. You cannot simply take the straight average of those state
mean heights, because you need to give more weight to the more populous
states. In other words, the national mean height is a weighted average of
the state means, with the weight for each state being its proportion of the
national population.
In (1.58), this corresponds to having V as height and U as state. State
coding is an integer-valued random variable, ranging from 1 to 50, so we
have

EV = E[E(V | U )] (1.59)
= EW (1.60)
∑
50
= P (U = i) E(V | U = i) (1.61)
i=1

The left-hand side, EV , is the overall mean height in the nation; E(V | U =
i) is the mean height in state i; and the weights in the weighted average
are the proportions of the national population in each state, P (U = i).
Not only can we look at the mean of W , but also its variance. By using the
various familiar properties of mean and variance, one can derive a similar
relation for variance:
54 CHAPTER 1. SETTING THE STAGE

1.19.5.3 Law of Total Variance

For scalar V ,

V ar(V ) = E[V ar(V |U )] + V ar[E(V |U )] (1.62)

One might initially guess that we only need the first term. To obtain the
national variance in height, we would take the weighted average of the state
variances. But this would not take into account that the mean heights vary
from state to state, thus also contributing to the national variance in height,
hence the second term.
This is proven in Section 2.12.8.3.

1.19.5.4 Tower Property

Now consider conditioning on two variables, say U1 and U2 . One can show
that

E [E(V |U1 , U2 ) | U1 ] = E(V | U1 ) (1.63)

Here is an intuitive interpretation of that in the height example above.

Take V , U1 and U2 to be height, state and gender, respectively, so that
E(V |U1 , U2 ) is the mean height of all people in a certain state and of a
certain gender. If we then take the mean of all these values for a certain
state — i.e. take the average of the two gender-specific means in the state
— we get the mean height in the state without regard to gender.
Again, note that we take the straight average of the two gender-specific
means, because the two genders have equal proportions. If, say, U2 were
race instead of gender, we would need to compute a weighted average of the
race-specific means, with the weights being the proportions of the various
races in the given state.
This is proven in Section 7.8.1.

1.19.5.5 Geometric View

There is an elegant way to view all of this in terms of abstract vector spaces
— (1.58) becomes the Pythagorean Theorem! — which we will address later
in Mathematical Complements Sections 2.12.8 and 7.8.1.
1.20. COMPUTATIONAL COMPLEMENTS 55

1.20 Computational Complements

1.20.1 CRAN Packages

There are thousands of useful contributed R packages available on CRAN,

the Comprehensive R Archive Network, https://cran.r-project.org. The eas-
iest way to install them is from R’s interactive mode, e.g.
> i n s t a l l . packages ( ’ f r e q p a r c o o r d ’ , ’∼/R ’ )

Here I have instructed R to download the freqparcoord package, installing

it in ∼/R, the directory where I like to store my packages.
(If you are using RStudio or some other indirect interface to R, all this can
be done from a menu, rather than using installing.packages.)
Official R parlance is package, not library, even though ironically one loads
a package using the library() function! For instance,
> library ( freqparcoord )

One can learn about the package in various ways. After loading it, for
instance, you can list its objects, such as
> l s ( ’ package : f r e q p a r c o o r d ’ )
[ 1 ] ” f r e q p a r c o o r d ” ” knndens ” ” knnreg ”
”posjitter” ” regdiag ”
[ 6 ] ” regdiagbas ” ”rmixmvnorm” ” smoothz ”
” smoothzpred ”

where we see objects (functions here) knndens() and so on. There is the
help() function, e.g.
> help ( package=f r e q p a r c o o r d )

I n f o r m a t i o n on package f r e q p a r c o o r d

Description :

Package : freqparcoord
Version : 1.1.0
Author : Norm M a t l o f f <normmatloff@gmail . com>
and Yingkang Xie
<yingkang . xie@gmail . com>
Maintainer : Norm M a t l o f f <normmatloff@gmail . com>
56 CHAPTER 1. SETTING THE STAGE

...

Some packages have vignettes, extended tutorials. Type

> vignette ()

to see what’s available.

1.20.2 The Function tapply() and Its Cousins

In Section 1.6.2 we had occasion to use R’s tapply(), a highly useful feature
of the language. To explain it, let’s start with useful function, split().
Consider this tiny data frame:
> x
gender height
1 m 66
2 f 67
3 m 72
4 f 63

Now let’s split by gender:

> xs <− s p l i t ( x , x$ g e n d e r )
> xs
$f
gender height
2 f 67
4 f 63
5 f 63

$m
gender height
1 m 66
3 m 72

Note the types of the objects:

• xs is an R list
• xs$f and xs$m are data frames, the male and female subsets of x

We could then find the mean heights for each gender this way:
1.20. COMPUTATIONAL COMPLEMENTS 57

> mean( xs $ f $ h e i g h t )
[ 1 ] 64.33333
> mean( xs $m$ h e i g h t )
[ 1 ] 69

But with tapply(), we can combine the two operations:

> tapply ( x$ h e i g h t , x$ gender , mean)
f m
64.33333 69.00000

The first argument of tapply() must be a vector, but the function that is
applied can be vector-valued. Say we want to find not only the mean but
also the standard deviation. We can do this:
> tapply ( x$ h e i g h t , x$ gender , function (w) c (mean(w) , sd (w) ) )
$f
[ 1 ] 64.333333 2.309401

$m
[ 1 ] 69.000000 4.242641

Here our function, which we defined “on the spot,” within our call to tap-
ply(), produces a vector of two components. We asked tapply() to call
that function on our vector of heights, doing so separately for each gender.
As noted in the title of this section, tapply() has “cousins.” Here is a brief
overview of some of them:
# form a m a t r i x by b i n d i n g t h e rows ( 1 , 2 ) and ( 3 , 4 )
> m <− rbind ( 1 : 2 , 3 : 4 )
> m
[ ,1] [ ,2]
[1 ,] 1 2
[2 ,] 3 4
# a p p l y t h e sum ( ) f u n c t i o n t o each row
> apply (m, 1 ,sum)
[1] 3 7
# a p p l y t h e sum ( ) f u n c t i o n t o each column
> apply (m, 2 ,sum)
[1] 4 6

> l <− l i s t ( a = c ( 3 , 8 ) , b = 1 2 )
> l
$a
58 CHAPTER 1. SETTING THE STAGE

[1] 3 8
$b
[ 1 ] 12
# a p p l y sum ( ) t o each e l e m e n t o f t h e l i s t ,
# f o r m i n g a new l i s t
> lapply ( l ,sum)
$a
[ 1 ] 11
$b
[ 1 ] 12
# do t h e same , b u t t r y t o r e d u c e t h e r e s u l t
# to a vector
> sapply ( l ,sum)
a b
11 12

1.20.3 The Innards of the k-NN Code

Here are simplified versions of the code:

p r e p r o c e s s x <− function ( x , kmax , x v a l=FALSE) {
r e s u l t $x <− x
tmp <− FNN : : get . knnx ( data=x , query=x , k=kmax+x v a l )
n n i <− tmp$nn . index
r e s u l t $ i d x s <− n n i [ , ( 1 + x v a l ) : ncol ( n n i ) ]
r e s u l t $ x v a l <− x v a l
r e s u l t $kmax <− kmax
c l a s s ( r e s u l t ) <− ’ preknn ’
result
}

The code is essentially just a wrapper for calls to the FNN package on
CRAN, which does nearest-neighbor computation.
k n n e s t <− function ( y , xdata , k , n e a r f=meany )
{
i d x s <− xdata $ i d x s
i d x <− i d x s [ , 1 : k ]
# s e t i d x r o w s [ [ i ] ] t o row i o f i d x , t h e i n d i c e s o f
# t h e n e i g h b o r s o f t h e i −t h o b s e r v a t i o n
i d x r o w s <− m a t r i x t o l i s t ( 1 , i d x )
# now do t h e kNN smoothing
1.20. COMPUTATIONAL COMPLEMENTS 59

# f i r s t , form t h e n e i g h b o r h o o d s
x <− xdata $x
xy <− cbind ( x , y )
n y c o l <− ncol ( y ) # how many c o l s i n xy a r e y ?
# f t n t o form one n e i g h b o r h o o d ( x and y v a l s )
form1nbhd <− function ( idxrow ) xy [ idxrow , ]
# now form a l l t h e n e i g h b o r h o o d s
nearxy <−
lapply ( idxrows , function ( idxrow ) xy [ idxrow , ] )
# now nearxy [ [ i ] ] i s t h e rows o f x c o r r e s p o n d i n g t o
# neighbors of x [ i , ] , together with the associated
# Y values

# now f i n d t h e e s t i m a t e d r e g r e s s i o n f u n c t i o n v a l u e s
# a t each p o i n t i n t h e t r a i n i n g s e t
r e g e s t <− sapply ( 1 : nrow( x ) ,
function ( i ) n e a r f ( x [ i , ] , nearxy [ [ i ] ] ) )
r e g e s t <−
i f ( n y c o l > 1 ) t ( r e g e s t ) e l s e as . matrix ( r e g e s t )
xdata $ r e g e s t <− r e g e s t
xdata $ n y c o l <− n y c o l
xdata $y <− y
xdata $k <− k
c l a s s ( xdata ) <− ’ knn ’
xdata
}

1.20.4 Function Dispatch

The return value from a call to lm() is an object of R’s S3 class structure;
the class, not surprisingly, is named ‘lm’. It turns out that the functions
coef() and vcov() mentioned in this chapter are actually related to this
class, as follows.
Recall our usage, on the baseball player data:
> lmout <− lm( mlb$Weight ∼ mlb$ Height )
> coef ( lmout ) %∗% c ( 1 , 7 2 )
[ ,1]
[ 1 , ] 193.2666

The call to coef extracted the vector of estimated regression coefficents

60 CHAPTER 1. SETTING THE STAGE

(which we also could have obtained as lmout$coefficents). But here is

what happened behind the scenes:
The R function coef() is a generic function, which means it’s just a place-
holder, not a “real” function. When we call it, the R interpreter says,

This is a generic function, so I need to relay this call to the one

associated with this class, ‘lm’. That means I need to check
whether we have a function coef.lm(). Oh, yes we do, so let’s
call that.

That relaying action is referred to in R terminology as the original call

being dispatched to coef.lm().
This is a nice convenience. Consider another generic R function, plot().
No matter what object we are working with, the odds are that some kind
of plotting function has been written for it. We can just call plot() on the
given object, and leave it to R to find the proper call. (This includes the
‘lm’ class; try it on our lmout above!)
Similarly, there are a number of R classes on which coef() is defined, and
the same is true for vcov().
One generic function we will use quite often, and indeed have already used
in this chapter, is summary(). As its name implies, it summarizes (what
the function’s author believes) are the most important characteristics of the
object. So, when this generic function is called on an ‘lm’ object, the call
is dispatched to summary.lm(), yielding estimated coefficients, standard
errors and so on.
Another generic function to be used often here is predict(), from Sec-
tion 1.10.3. In the example there, lmout was of class ‘lm’, so the call to
predict() was dispatched to predict.lm().

1.21 Centering and Scaling

It is common in many statistical methods to center and scale the data. Here
we subtract from each variable the sample mean of that variable. This
process is called centering. Typically one also scales each predictor, i.e.
divides each predictor by its sample standard deviation. Now all variables
will have mean 0 and standard deviation 1.
It is clear that this is very useful for k-NN regression. Consider the ex-
1.22. EXERCISES: DATA, CODE AND MATH PROBLEMS 61

ample later in this chapter involving Census data. Without at least scal-
ing, variables that are very large, such as income, would dominate the
nearest-neighbor computations, and small but important variables such as
age would essentially be ignored. The knnest() function that we will be
using does do centering and scaling as preprocessing for the predictor vari-
ables.
In a parametric setting such as linear models, centering and scaling has the
goal of reducing numerical roundoff error.
In R, the centering/scaling operation is done with the scale() function.
In order to be able to reverse the process later, the means and standard
deviations are recorded as R attributes:
> m <− rbind ( 1 : 2 , 3 : 4 )
> m
[ ,1] [ ,2]
[1 ,] 1 2
[2 ,] 3 4
> m1 <− s c a l e (m)
> m1
[ ,1] [ ,2]
[ 1 , ] −0.7071068 −0.7071068
[ 2 , ] 0.7071068 0.7071068
attr ( , ” s c a l e d : c e n t e r ” )
[1] 2 3
attr ( , ” s c a l e d : s c a l e ” )
[ 1 ] 1.414214 1.414214
> attr (m1, ’ s c a l e d : c e n t e r ’ )
[1] 2 3

1.22 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. In Section 1.12.1.2, the reader was reminded that the results of a cross-
validation are random, due to the random partitioning into training and
test sets. Try doing several runs of the linear and k-NN code in that section,
comparing results.
2. Extend (1.28) to include interaction terms for age and gender, and age2
62 CHAPTER 1. SETTING THE STAGE

and gender. Run the new model, and find the estimated effect of being
female, for a 32-year-old person with a Master’s degree.
3. Consider the bodyfat data mentioned in Section 1.2. Use lm() to form
a prediction equation for density from the other variables (skipping the
first three), and comment on whether use of indirect methods in this way
seems feasible.
4. In Section 1.19.5.2, we gave this intuitive explanation:

In other words, the national mean height is a weighted average

of the state means, with the weight for each state being its
proportion of the national population. Replace state by gender
in the following.

(a) Write English prose that relates the overall mean height of people and
the gender-specific mean heights.
(b) Write English prose that relates the overall proportion of people taller
than 70 inches to the gender-specific proportions.

Mini-CRAN and other computational problems:

5. In Section 1.12, we used R’s negative-index capability to form the train-
ing/test set partitioning. Show how we could use the R function setdiff()
to do this as an alternate approach.
6. We saw in this chapter, e.g., in Figure 1.3, how R’s abline() function
can be used to add a straight line to a plot. What about adding a quadratic
function?

(a) Write an R function with call form

a b c c u r v e ( coef , x i n t )

where coef is a vector of the coeficients a, b and c in the polynomial

a + bt + ct2 (1.64)

and xint is a 2-element vector that gives the range of the horizontal
axis for t. The function superimposes the quadratic curve onto the
existing graph. Hint: Use R’s curve() function.
(b) Fit a quadratic model to the click-through data, and use your abc-
curve() function on the scatter plot for that data.
1.22. DATA, CODE AND MATH PROBLEMS 63

Math problems:
7. Suppose the joint density of (X, Y ) is 3s2 e−st , 1 < s < 2, 0 < t < ∞.
Find the regression function µ(s) = E(Y |X = s).
8. For (X, Y ) in the notation of Section 1.19.3, show that the predicted
value µ(X) and the predicton error Y − µ(X) are uncorrelated.
9. Suppose X is a scalar random variable with density g. We are interested
in the nearest neighbors to a point t, based on a random sample X1 , ..., Xn
from g. Find Lk , the cumulative distribution function of the distance of
the k th -nearest neighbor to t.
Chapter 2

Linear Regression Models

In this chapter we go into the details of linear models. Let’s first set some
notation, to be used here and in the succeeding chapters.

2.1 Notation

Some notation here will be used throughout the book, so it is important to

get a firm understanding. (Note the List of Symbols and Abbreviations at
the front of this book, for easy reference.)
Let Y be our response variable, and let X = (X (1) , X (2) , ..., X (p)) )′ de-
note the vector of our p predictor variables. Using our weight/height/age
baseball player example from Chapter 1 as our running example here, we
would have p = 2, and Y , X (1) and X (2) would be weight, height and age,
respectively.
The quantities Y, X (1) , ..., X p denote the values of these random variables
in the population.1 For instance, in the baseball example, Y is weight, so it
represents the weight of a player chosen randomly from the population. But
there is also notation for the values of these variables in our data, thought
of as a sample from the population:
Our sample consists of n data points, X1 , X2 , ..., Xn , each a p-element pre-
dictor vector, and Y1 , Y2 , ..., Yn , associated scalars. In the baseball example,
n was 1015. Also, the third player had height 72, was of age 30.78, and
1 Sometimes the population is rather conceptual, as discussed in Section 1.5.

65
66 CHAPTER 2. LINEAR REGRESSION MODELS

weighed 210. So,

( )
72
X3 = (2.1)
30.78

and

Y3 = 210 (2.2)

Write the Xi in terms of their components:

Xi = (Xi , ..., Xi )′
(1) (p)
(2.3)

So, again using the baseball player example, the height, age and weight of
(1) (2)
the third player would be X3 , X3 and Y3 , respectively.
And just one more piece of notation: We sometimes will need to augment
a vector with a 1 element at the top, such as we did in (1.9). Our notation
for this will consist of a tilde above the symbol, For instance, (2.1) becomes

 
1
f3 = 
X 72  (2.4)
30.78

So, our linear model is, for a p-element vector t = (t1 , ..., tp )′ ,

µ(t) = β0 + β1 t1 + .... + βp tp = e
t′β (2.5)

In the baseball example, with both height and weight as predictors:

µ(height,age) = β0 + β1 height + β2 age (2.6)

 
β0
= (1, height, age)′  β1  (2.7)
β2
2.2. THE “ERROR TERM” 67

2.2 The “Error Term”

Define

ϵ = Y − µ(X) (2.8)

This is the error we would make if we were to predict Y from X, if we some-

how knew the population regression function µ(t). It is also interpretable
as the collective effect of all predictors of Y that are not in our
model (and for which we usually do not have data).
We could then write (2.5) as

Y = β0 + β1 t1 + .... + βp tp + ϵ (2.9)

This is the more common way to define the linear regression model, and ϵ,
a term in that model, is called the error term. We will sometimes use this
formulation, but the primary one is (2.5).

2.3 Random- vs. Fixed-X Cases

We will usually consider the Xi and Yi to be random samples from some

population. If we have data on people and are predicting weight from height
and age, for instance, that means weight, height and age are random vari-
ables, since we are sampling people at random. Thus (X1 , Y1 ), ..., (Xn , Yn )
are independent and identically distributed (i.i.d.), with their distribution
being that of the population. If, for instance, 23% of people in our popula-
tion are taller than 72 inches, then

(1)
P (Xi > 75) = 0.23 (2.10)

according to the population. This is a random-X setting, meaning that

both the Xi and Yi are random.
But there are some situations in which the X values are fixed by design,
known as a fixed-X setting. This might be the case in chemistry research for
instance, in which we decide in advance to perform experiments at specific
levels of concentration of some chemicals.
68 CHAPTER 2. LINEAR REGRESSION MODELS

Recall that the regression function is

µ(t) = E(Y | X = t) (2.11)

i.e., we are dealing with the conditional distribution of Y given X. So, in

many cases, it doesn’t matter whether our data arose in an random-X vs.
fixed-X setting; in the random-X case, once we condition on X, we are in
the same situation as in the fixed-X case. It does matter in some situations,
though, as will be seen in Chapter 5.
By the way, in fixed-X settings, the X values are often chosen to form an
orthogonal design. In this context, it is assumed that each predictor variable
has mean 0 (by subtracting their means, if necessary), and that the data
vectors for different predictor variables are orthogonal in the linear algebra
sense: The inner product of the vector of values of predictor j and the one
for predictor k is 0:

∑
n
(j) (k)
Xi Xi = 0, for all j ̸= k (2.12)
i=1

This simplified computation in the pre-computer days, and may result in

better interpretability of the βi .
The mean-0 property also implies (Section 2.12.4) that the regression model
does not have a constant term, i.e., for t = (t1 , ..., tp )′ ,

µ(t) = β1 t1 + .... + βp tp (2.13)

We will take this to be part of our definition of the term orthogonal design.

2.4 Least-Squares Estimation

Linear regression analysis is sometimes called least-squares estimation. Let’s

first look at how this evolved.

2.4.1 Motivation

This discussion will begin at the population level. As noted in Section

1.19, setting f (X) = µ(X) minimizes the mean squared prediction error,
2.4. LEAST-SQUARES ESTIMATION 69

i.e., minimizes

E[(Y − f (X))2 ] (2.14)

over all functions f . And since our assumption is that µ(t) = e

t ′ β, we can
set

f (t) = e
t′ b (2.15)

above, and thus also say that b = β minimizes

e ′ b))2 ]
E[(Y − X (2.16)

over all vectors b.

Now let us consider sample analogs of the above. If W1 , ..., Wn is a sample
from
∑n a population having mean EW, the sample analog of EW is W =
( i=1 Wi )/n; one is the average value of W in the population, and the other
is the average value of W in the sample. Let’s write this correspondence as

EW ←→ W (2.17)

It will be crucial to always keep in mind the distinction between population

values and their sample estimates, especially when we discuss overfitting in
detail.
Similarly, for any fixed b, (2.16) is a population quantity, the average
squared error using b for prediction in the population (recall Section 1.5).
The population/sample correspondence here is

1∑
n
′
e ′ b))2 ] ←→
E[(Y − X fi b)2
(Yi − X (2.18)
n i=1

where the right-hand side is the average squared error using b for prediction
in the sample.
So, since β is the value of b minimizing (2.16), it is intuitive to take our
b to be the value of b that minimizes (2.18). Hence the term
estimate, β,
least squares.
70 CHAPTER 2. LINEAR REGRESSION MODELS

To find the minimizing b, we could apply calculus, taking the partial deriva-
tives of (2.18) with respect to bi , i = 0, 1, ..., p, set them to 0 and solve.
Fortunately, R’s lm() does all that for us, but it’s good to know what is
happening inside. Also, this will give the reader more practice with matrix
expressions, which will be important in some parts of the book.

2.4.2 Matrix Formulations

Use of matrix notation in linear regression analysis greatly compactifies

and clarifies the presentation. You may find that this requires a period of
adjustment at first, but it will be well worth the effort.
As usual, use p for the number of predictor variables, and let A denote the
n × (p + 1) matrix of X values in our sample,

 ′ 
f1
X
 f2
′ 
 X 
A=  (2.19)
 ... 
′
fn
X

and let D be the n × 1 vector of Y values,2

 
Y1
 Y2 
D= 
 ...  (2.20)
Yn

In the baseball example, row 3 of A is

(1, 30.78, 72) (2.21)

and the third element of D is 210.

2 The matrix is typically called X in regression literature, but there are so many

symbols here using “X” that it is clearer to call the matrix something else. The same
comment applies to the vector D.
2.4. LEAST-SQUARES ESTIMATION 71

2.4.3 (2.18) in Matrix Terms

Our first order of business will be to recast the right-hand side of (2.18) as
′
a matrix expression. To start, look at the quantities Xfi b, i = 1, ..., n there
in (2.18). Stringing them together in matrix form, we get

 ′ 
f1
X
 f2
′ 
 X 
  b = Ab (2.22)
 ... 
′
f
Xn

Now consider the n summands in (2.18), before squaring. Stringing them

into a vector as in Section A.9, we get

D − Ab (2.23)

We need just one more step: Recall (see (A.15)) that for a vector a =
(a1 , ..., ak )′ ,

∑
k
a2k = a′ a (2.24)
i=1

In other words, (2.18) (except for the 1/n factor) is actually

(D − Ab)′ (D − Ab) (2.25)

Now that we have this in matrix form, we can go about finding the optimal
b.

2.4.4 Using Matrix Operations to Minimize (2.18)

Remember, we will set βb to whatever value of b minimizes (2.25). Thus

we need to take the derivative of that expression with respect to b, and set
the result to 0. Here we can draw upon the matrix derivative material in
Section A.10.
72 CHAPTER 2. LINEAR REGRESSION MODELS

Specifically, in the context of (A.40), set u = D − Ab, M = −A, w = D

and v = b. This tells us that the derivative of (2.25) with respect to b is

−2A′ (D − Ab) (2.26)

Setting this to 0, we have

A′ D = A′ Ab (2.27)

Solving for b we have our answer:

βb = (A′ A)−1 A′ D (2.28)

This is what lm() calculates! In essence, it does

solve ( t ( a ) %∗% a , t ( a ) %∗% d )

(See further comment on the calculation in Section 2.13.3.)

2.4.5 Models without an Intercept Term

In some cases, it is appropriate to omit the β0 term from (2.5). The deriva-
tion in this setting is the same as before, except that (2.19) becomes

 
X1 ′
 X2 ′ 
A=
 ... 
 (2.29)
Xn ′

which differs from (2.19) only in that this new matrix does not have a
column of 1s.
This may arise when modeling some physical or chemical process, for in-
stance, in which theoretical considerations imply that µ(0) = 0. A much
more common use of such a model occurs as follows.
In computation for linear models, the data are typically first centered and
scaled (Section 1.21). It turns out that centering forces the intercept term,
βb0 , to 0. This is shown in Section 2.12.4. but our point now is that that
action does not change the other βbi , i > 0. So, centering does no harm —
2.5. A CLOSER LOOK AT LM() OUTPUT 73

if our goal is Description, we usually are not interested in the uncentered

value of β0 — and it may be computationally beneficial as noted above.
In R, we can request that lm() omit the intercept term in the model by
including the symbol -1 in the predictor list. For example,
> x <− rnorm ( 1 0 0 0 , 2 5 , 8 )
> y <− x + rnorm( 1 0 0 0 )
> lm( y ∼ x−1)

Call :
lm( formula = y ∼ x − 1 )

Coefficients :
( Intercept ) x
0.04575 0.99927

> lm( y ∼ x−1)

Call :
lm( formula = y ∼ x − 1 )

Coefficients :
x
1.001

2.5 A Closer Look at lm() Output

Since the last section was rather abstract, let’s get our bearings by taking
a closer look at the output in the baseball example:3
> lmout <− lm( mlb$Weight ∼ mlb$ Height + mlb$Age )
> summary( lmout )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) −187.6382 1 7 . 9 4 4 7 −10.46 < 2 e −16
mlb$ Height 4.9236 0.2344 2 1 . 0 0 < 2 e −16
mlb$Age 0.9115 0.1257 7 . 2 5 8 . 2 5 e −13

3 Note the use of the ellipsis . . ., indicating that portions of the output have been
omitted, for clarity.
74 CHAPTER 2. LINEAR REGRESSION MODELS

( I n t e r c e p t ) ∗∗∗
mlb$ Height ∗∗∗
mlb$Age ∗∗∗
−−−
S i g n i f . codes :
0 ∗∗∗ 0 . 0 0 1 ∗∗ 0 . 0 1
∗ 0.05 . 0.1 1
...
M u l t i p l e R−s q u a r e d : 0 . 3 1 8 ,
Adjusted R−s q u a r e d : 0 . 3 1 6 6
...

There is a lot here! Let’s get an overview, so that the material in the coming
sections will be better motivated.

2.5.1 Statistical Inference

The lm() output is heavily focused on statistical inference — forming con-

fidence intervals and performing significance tests — and the first thing you
may notice is all those asterisks: The estimates of the intercept, the height
coefficient and the age coefficient all are marked with three stars, indicating
a p-value of less than 0.001.
Those p-values correspond to tests of the hypothesis

H0 : βi = 0 (2.30)

under assumptions to be discussed shortly. But first, what are the practical
implications of these p-values?
Look at the coefficient for height, for example. The test of the hypothesis
that β1 = 0, i.e., no height effect on weight, has a p-value of 2 × 10−10 ,
extremely small. Thus the hypothesis would be resoundingly rejected, and
one could say, “Height has a significant effect on weight.” Not surprising at
all, though the finding for age might be more interesting, in that we expect
athletes to keep fit, even as they age.
We could form a confidence interval for β2 , for instance, by adding and
subtracting 1.96 times the associated standard error,4 which is 0.1257 in
this case. Our resulting CI would be about (0.66,1.16), indicating that
4 The standard error of an estimator was defined in Section 1.6.3.
2.6. ASSUMPTIONS 75

the average player gains between 0.66 and 1.16 pounds per year. So even
baseball players gain weight over time!
We will return to this vital topic of misuse of p-values in Section 2.10.

2.6 Assumptions

But where did this come from? Surely there must be some assumptions
underlying these statistical inference procedures. What are they?

2.6.1 Classical

The classical assumptions, to which the reader may have some prior expo-
sure, are:

• Linearity: There is some vector β for which Equation (2.5) holds for
all t.

• Normality: The assumption is that, conditional on the vector of

predictor variables X, the response variable Y has a normal distribu-
tion.
In the weight/height/age example, this would mean, for instance, that
within the subpopulation of all baseball players of height 72 and age
25, weight is normally distributed.

• Homoscedasticity: Just as we define the regression function in

terms of the conditional mean,

µ(t) = E(Y | X = t) (2.31)

we can define the conditional variance function

σ 2 (t) = V ar(Y | X = t) (2.32)

The homoscedasticity assumption is that σ 2 (t) does not depend on t.

In the weight/height/age example, this would say that the variance
in weight among, say, 70-inches-tall 22-year-olds is the same as that
among the subpopulation of those of height 75 inches and age 32.

• Independence: The Yi are (conditionally) independent.

76 CHAPTER 2. LINEAR REGRESSION MODELS

By the way, note that the above assumptions all concern the structure of
the population. In addition, we assume that in the sample drawn from that
population, the observations are independent.

2.6.2 Motivation: the Multivariate Normal

Distribution Family

We’ll discuss the real-world propriety of the above three assumptions

• Linearity
• Normality (conditional, of Y given X)
• Homoscedasticity

in later chapters, but it is worthwhile to take a look now at the motivation

for those assumptions. This will help us understand where the assumptions
came from originally, and help us to visualize how departures from the
assumptions can arise.
For simplicity, we will focus here on the case p = 1, i.e., just one predictor
variable.
All of the assumptions are intuitively reasonable as choices for a simple
model (which, by the way, is extremely widely used). But they also all
follow if the vector (Y, X)′ has a bivariate normal distribution. What is
this?
Consider the example in which Y and X are human weight and height,
respectively. To say Y is normally distributed means its density follows the
familiar “bell-shaped curve,” with equation5

1
e− 2 ((t−c)/d)
1 2
f (t) = √ (2.33)
2πd

where c and d are the population mean and standard deviation. The same
would be true for a normal distribution for X. But what does it mean for
Y and X to be jointly normal, i.e., have a bivariate normal distribution?
The bivariate normal density takes the shape of a three-dimensional bell,
as in Figure 2.1. (The figure is adapted from Romaine Francois’ old R
5 There
are other families of bell-shaped curves, such as the Cauchy, so the normal
density form is not “the” bell-shaped one.
2.6. ASSUMPTIONS 77

Figure 2.1: Bivariate normal density, ρ = 0.2 (see color insert)

Graphics Gallery.) A density function is a population-level entity, but if

we were to draw a sample from such a population and form a two-variable
histogram from it, the result would look like this 3-D bell shape.
The mathematical density form is pretty complex,

 
1 (s−µ1 )2 (t−µ2 )2 2ρ(s−µ1 )(t−µ2)
1 − + −

2 2
σ1 2
σ2 σ1 σ2
f (s, t) = e 2(1−ρ ) , (2.34)
2πσ1 σ2 1 − ρ2

for −∞ < s, t < ∞, where the µi and σi are the means and standard
deviations of X and Y , and ρ is the correlation between X and Y (Section
2.12.1).
Now, let’s see how this relates to our linear regression assumptions above,
Linearity, Normality and Homoscedasticity. Since the regression function
by definition is the conditional mean of Y given X, we need the conditional
density. That means holding X constant, so we treat s in (2.34) as a con-
78 CHAPTER 2. LINEAR REGRESSION MODELS

stant. (This will give us the conditional density, except for a multiplicative
constant.) We’ll omit the messy algebraic details, but the point is that in
the end, (2.34) reduces to a function that

• has the form (2.33), thus giving us the (conditional) Normality prop-
erty;

• has mean as a linear function of s, yielding the Linearity property;

and

• has variance independent of s, thus giving us Homoscedasticity.

The specific linear function in the second bullet above can be shown to be

σ2
E(Y | X = s) = ρ (s − µ1 ) + µ2 (2.35)
σ1

and the constant-variance property in the third bullet is specifically

V ar(Y | X = s) = (1 − ρ)2 σ22 (2.36)

which is indeed independent of s.

In other words, if X and Y have a bivariate normal distribution, all of our
linear regression assumptions follow. This is the original motivation for the
assumptions, though as mentioned they are plausible in many applications.
One can gain further insight by examining the effect of ρ. Toward this end,
take another look at Figure 2.1, in particular the level sets of the graphed
function, meaning the points of constant height. In other words, what
happens if we slice that mound horizontally, parallel to the s-t plane?
That’s equivalent to setting the exponent in (2.34) to some constant. Some
readers may recognize this as the equation of an ellipse. Then the major,
i.e., longer, axis of that ellipse turns out to be the regression line!
Moreover, consider what will happen as we increase ρ toward 1.0 (in the
picture, ρ = 0.2). The minor axis will shrink, and the 3-D bell shape will
become flatter and flatter, closing in on the regression line. See the picture
for ρ = 0.8, Figure 2.2. This is reflected in the conditional variance in (2.36)
going to 0 as ρ goes to 1.
These issues, and properties of the general multivariate normal distribution,
are discussed in Section 2.12.2.
2.7. UNBIASEDNESS AND CONSISTENCY 79

Figure 2.2: Bivariate normal density, ρ = 0.8 (see color insert)

2.7 Unbiasedness and Consistency


Here we discuss two statistical properties of the least-squares estimator β.

2.7.1 β Is Unbiased

One of the central concepts in the early development of statistics was un-
biasedness. As you’ll see, to some degree it is only historical baggage, but
on the other hand it does become quite relevant in some contexts here.
To explain the concept, say we are estimating some population value θ,
using an estimator θ based on our sample. Remember, θ is a random
variable — if we take a new sample, we get a new value of θ.  So, some
samples will yield a θ that overestimates θ, while in other samples θ will

come out too low.
The pioneers of statistics believed that a nice property for θ to have would
be that on average, i.e., averaged over all possible samples, θ comes out
80 CHAPTER 2. LINEAR REGRESSION MODELS

“just right”:

E θb = θ (2.37)

This seems like a nice property for an estimator to have (though far from
mandatory, as we’ll see below), and sure enough, our least-squares estimator
has that property:

E βb = β (2.38)

Note that since this is a vector equation, the unbiasedness is meant for the
individual components. In other words, (2.38) is a compact way of saying

E βbj = βj , j = 0, 1, ..., p (2.39)

This is derived in the Mathematical Complements portion of this chapter,

Section 2.12.5. Note that we do not need the Normality or Homoscedasticity
assumption for this result.

2.7.2 Bias As an Issue/Nonissue

Arguably the pioneers of statistics shouldn’t have placed so much emphasis

on unbiasednedness. Most statistical estimators have some degree of bias
(though the amount of bias is usually small and goes to 0 as the sample size
n grows). Even the much-heralded unbiased nature of the classical defini-
tion of sample variance, Equation (2.41) below, is somewhat misleading, as
the sample standard deviation, arguably just as important a quantity, is bi-
ased (Section 2.12.7). And other than least-squares, none of the regression
function estimators in common use, such as k-NN, is unbiased.
Nevertheless, bias can be an issue in some contexts, as will be seen later in
this chapter.

2.7.3 βb Is Statistically Consistent

In contrast to unbiasedness, which as argued above may not be a generally

necessary goodness criterion for an estimator, there is a more basic property
that we would insist that almost any estimator have, consistency: As the
sample size n goes to infinity, then the sample estimate θb goes to θ. This is
2.8. INFERENCE UNDER HOMOSCEDASTICITY 81

not a very strong property, but it is a minimal one. It is shown in Section

2.12.6 that the least-squares estimator βb is indeed a consistent estimator of
β. Again, we do not need the Normality or Homoscedasticity assumption
for this result.

2.8 Inference under Homoscedasticity

Let’s see what the homoscedasticity assumption gives us.

2.8.1 Review: Classical Inference on a Single Mean

You may have noticed the familiar Student-t distribution mentioned in the
output of lm() above. Before proceeding, it will be helpful to review this
situation from elementary statistics.

We have a random sample W1 , ..., Wn from a population hav-

ing mean ν = EW and variance η 2 . Suppose W is normally
distributed in the population. Form

1∑
n
W = Wi (2.40)
n i=1

and

1 ∑
n
S2 = (Wi − W )2 (2.41)
n − 1 i=1

Then

W −ν
T = √ (2.42)
S/ n
has a Student-t distribution with n − 1 degrees of freedom (df).
This is then used for statistical inference on ν. We can form
√ a
95% confidence interval by adding and subtracting c × S/ n to
W , where c is the point of the upper-0.025 area for the Student-t
distribution with n − 1 df.
Under the normality assumption, such inference is exact; a 95%
confidence interval, say, has exactly 0.95 probability of contain-
ing ν.
82 CHAPTER 2. LINEAR REGRESSION MODELS

2.8.2 Back to Reality

The normal distribution model is just that, a model, not expected to be

exact. It rarely happens, if ever at all, that a population distribution is
exactly normal. Human weight, for instance, cannot be negative and cannot
be a million pounds; it is bounded, unlike normal distributions, whose
support is (−∞, ∞). So “exact” inference using the Student-t distribution
as above is not exact after all.
If n is large, the assumption of a normal population becomes irrelevant:
The Central Limit Theorem (CLT, Section 2.12.3) tells us that

W −ν
√ (2.43)
η/ n

has an approximate N(0,1) distribution even though the distribution of W

is not normal. We then must show that if we replace η by S in (2.43),
the result will still be approximately normal. This follows from something
called Slutsky’s Theorem and the fact that S goes to η as n → ∞.6 Thus
we can perform (approximate) statistical inference on ν using (2.42) and
N(0,1), again without assuming that W has a normal distribution.
For instance, since the upper 2.5% tail of the N(0,1) distribution starts at
1.96, an approximate 95% confidence interval for ν would be

S
W ± 1.96 √ (2.44)
n

What if n is small? We could use the Student-t distribution anyway, but

we would have no idea how accurate it would be. We could not even use
the data to assess the normality assumption on which the t-distribution is
based, as we would have too little data to do so.
The normality assumption for the Wi , then, is of rather little value, and as
explained in the next section, is of even less value in the regression context.
One possible virtue, though, of using Student-t would be that it gives a
wider interval than does N(0,1). For example, for n = 28, our confidence
interval would be

s
W ± 2.04 √ (2.45)
n
6 In its simpler form, the theorem says that if U converges to a normal distribution
n
and Vn → v as n → ∞, then Un /Vn also is asymptotically normal.
2.8. INFERENCE UNDER HOMOSCEDASTICITY 83

instead of (2.44). The importance of this is that using S instead of η adds

further variability to (2.42), which goes away as n → ∞ but makes (2.43)
overly narrow. Using a Student-t value might compensate for that, though
it may also overcompensate.

2.8.3 The Concept of a Standard Error

The following concept will be used repeatedly throughout the book.

If θb is an approximately normally distributed estimator of a

population value θ, then an approximate 95% confidence inter-
val for θ is

θb ± 1.96 s.e.(θ)
b (2.46)

whereas in Section 1.6.3, the notation s.e.() denotes “standard

error of.”
√
In (2.44), the standard error of W is S/ n.

2.8.4 Extension to the Regression Case

The discussion in the last section concerned inference for a mean. What
about inference for regression functions (which are conditional means)?
The first point to note is this:

• Under the classical assumption that the conditional distribution of Y

given X is normal, then βb has an exact multivariate normal distribu-
tion. (This follows from the properties in Section 2.12.2.)

• The distribution of the least-squares estimator βb is approximately

(p + 1)-variate normal, without assuming normality.7

That second bullet again follows from the CLT. (Since we are looking at
fixed-X regression here, we need a non-identically distributed version of the
7 The statement is true even without assuming homoscedasticity, but we won’t drop
that assumption until the next chapter.
84 CHAPTER 2. LINEAR REGRESSION MODELS

CLT.) Consider for instance a typical component of A′ D in (2.28),

∑
n
(j)
Xi Yi (2.47)
i=1

This is a sum of independent terms, thus approximately normal. In (2.28),

we are working with a random vector, D, and in such a context, the cor-
responding CLT result is multivariate normal. So, βb has an asymptotic
(p + 1)-variate normal distribution. (A more formal derivation is presented
in Section 2.12.11.)
To perform statistical inference, we need the approximate covariance matrix
b from which we can obtain standard errors of the βbj .
(Section 2.12.1) of β,
The standard way to do this is by assuming homoscedasticity.
So, lm() assumes that in (2.32), the function σ 2 (t) is constant in t. For
brevity, then, we will simply refer to it as σ 2 . Note that this plus our
independence assumption implies the following about the (conditional) co-
variance matrix of D:

Cov(D|A) = σ 2 I (2.48)

where I is the identity matrix.

To avoid (much) clutter, define

B = (A′ A)−1 A′ (2.49)

Then by the properties of covariance matrices (Equation (2.79)),

Cov(βb |A) = Cov(BD) (2.50)

= B Cov(D|A) B ′ (2.51)
= σ 2 BB ′ (2.52)

Fortunately, the various properties of matrix transpose (Section A.3) can

be used to show that

BB ′ = (A′ A)−1 (2.53)

2.8. INFERENCE UNDER HOMOSCEDASTICITY 85

Thus

b = σ 2 (A′ A)−1
Cov(β) (2.54)

That’s a nice (surprisingly) compact expression, but the quantity σ 2 is an

unknown population value. It thus must be estimated, as we estimated η 2
by S 2 in Section 2.8.1. And again, an unbiased estimator is available. So,
we take as our estimator of σ 2

1 ∑ n
s2 = (Yi − X b2
ei′ β) (2.55)
n − p − 1 i=1

which can be shown to be unbiased.

If the normality assumption were to hold, then quantities like

βbi − βi
√ (2.56)
s aii

would have an exact Student-t distribution with n−p−1 degrees of freedom,

where aii is the (i, i) element of (A′ A)−1 .8
But as noted, this is usually an unrealistic assumption, and we instead rely
on the CLT. Putting the above together, we have:

b
The conditional distribution of the least-squares estimator β,
given A, is approximately multivariate normal (Section 2.6.2)
with mean β and approximate covariance matrix

s2 (A′ A)−1 (2.57)

Thus the standard error of βbj is the square root of element (j, j)
of this matrix (counting the top-left element as being in row 0,
column 0).
Similarly, suppose we are interested in some linear combination
λ′ β of the elements of β, estimating it by λ′ βb Section (A.4). By
(2.80), the standard error is then the square root of
8 A common interpretation of the number of degrees of freedom here is, “We have n

data points, but must subtract one degree of freedom for each of the p + 1 estimated
parameters.”
86 CHAPTER 2. LINEAR REGRESSION MODELS

s2 λ′ (A′ A)−1 λ (2.58)

And as before, we might as well calculate s2 with a denominator of n, as

opposed to the n − p − 1 expression above.
Recall from Chapter 1, by the way, that R’s vcov() function gives us the
matrix (2.57), both for lm() and also for some other regression modeling
functions that we will encounter later.
Before going to some examples, note that the conditional nature of the
statements above is not an issue, even in random-X settings. Say for in-
stance we form a 95% confidence interval for some quantity, conditional on
A. Let V be an indicator variable for the event that the interval contains
the quantity of interest. Then

P (V = 1) = E[P (V = 1 | A)] = E(0.95) = 0.95 (2.59)

Thus the unconditional coverage probability is still 0.95.

2.8.5 Example: Bike-Sharing Data

Let’s form some confidence intervals from the bike-sharing data.

> lmout <− lm( r e g ∼ temp+temp2+workingday+c l e a r d a y ,

data=s h a r )
> summary( lmout )
....
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) −1362.56 2 3 2 . 8 2 −5.852 1 . 0 9 e −08
temp 11059.20 9 8 8 . 0 8 1 1 . 1 9 3 < 2 e −16
temp2 −7636.40 1 0 1 3 . 9 0 −7.532 4 . 0 8 e −13
workingday 685.99 71.00 9 . 6 6 1 < 2 e −16
clearday 518.95 69.52 7 . 4 6 5 6 . 3 4 e −13
...
M u l t i p l e R−s q u a r e d : 0 . 6 5 4 8 , Adjusted R−s q u a r e d : 0 . 6 5 1

We estimate that a working day adds about 686 riders to the day’s ridership.
An approximate 95% confidence interval for the population value for this
2.8. INFERENCE UNDER HOMOSCEDASTICITY 87

effect is

685.99 ± 1.96 × 71.00 = (546.83, 825.15) (2.60)

This is a disappointingly wide interval, but it shouldn’t surprise us. After

all, it is based on only 365 data points.
Given the nonlinear effect of temperature in our model, finding a relevant
confidence interval here is a little more involved. Let’s compare the mean
ridership for our example in the last chapter — 62 degree weather, a Sunday
and sunny — with the same setting but with 75 degrees.
The difference in (population!) mean ridership levels between these two
settings is9

(β0 + β1 0.679+ β2 0.6792 + β3 0 + β1 1) − (β0 + β1 0.525 + β2 0.5252 + β3 0 + β1 1)

= β1 0.154 + β2 0.186

Our sample estimate for that difference in mean ridership between the two
types of days is then obtained as follows:
> lamb <− c ( 0 , 0 . 1 5 4 , 0 . 1 8 6 , 0 , 0 )
> t ( lamb ) %∗% coef ( lmout )
[ ,1]
[ 1 , ] 282.7453

or about 283 more riders on the warmer day. For a confidence interval, we
need a standard error. So, in (2.58), take λ = (0, 0.154, 0.186, 0, 0)′ . Our
standard error is then obtained via
> sqrt ( t ( lamb ) %∗% vcov ( lmout ) %∗% lamb )
[ ,1]
[ 1 , ] 47.16063

Our confidence interval for the difference between 75-degree and 62-degree
days is

282.75 ± 1.96 · 47.16 = (190.32, 375.18) (2.61)

9 Recall that the dataset here uses a scaled version of temperature; see page 37.
88 CHAPTER 2. LINEAR REGRESSION MODELS

Again, a very wide interval, but it does appear that a lot more riders show
up on the warmer days.
The value of s is itself probably not of major interest, as its use is usually in-
direct, in (2.57). However, we can determine it if need be, as lmout$residuals
contains the residuals, i.e., the sample prediction errors

′
Yi − X b i = 1, 2, ..., n
fi β, (2.62)

Using (2.55), we can find s:

> s <− sqrt (sum( lmout$ r e s i d u a l s ˆ 2 ) / (365 −4 −1))

> s
> s
[ 1 ] 626.303

2.9 Collective Predictive Strength of the X (j)

The R2 quantity in the output of lm() is a measure of how well our model
b a sample quantity, estimates the population
predicts Y . Yet, just as β,
quantity β, one would reason that the R2 value printed out by lm() must
estimate a population quantity too. In this section, we’ll make that concept
precise, and deal with a troubling bias problem.
We will also introduce an alternative form of the cross-validation notion
discussed in Section 1.12.

2.9.1 Basic Properties

Note carefully that we are working with population quantities here, gen-
erally unknown, but existent nonetheless. Note too that, for now, we are
NOT assuming normality or homoscedasticity. In fact, even the assump-
tion of having a linear regression function will be dropped for the moment.
The context, by the way, is random-X regression (Section 2.3).
Suppose we somehow knew the exact population regression function µ(t).
Whenever we would encounter a person/item/day/etc. with a known X
but unknown Y , we would predict the latter by µ(X). Define ϵ to be the
2.9. COLLECTIVE PREDICTIVE STRENGTH OF THE X (J) 89

prediction error

ϵ = Y − µ(X) (2.63)

It can be shown (Section 2.12.9) that µ(X) and ϵ are uncorrelated, i.e.,
have zero covariance. We can thus write

V ar(Y ) = V ar[µ(X)] + V ar(ϵ) (2.64)

With this partitioning, it makes sense to say:

The quantity

V ar[µ(X)]
ω= (2.65)
V ar(Y )

is the proportion of variation of Y explainable by X.

Section 2.12.9 goes further:

Define

√
ρ= ω (2.66)

Then ρ is the correlation between our predicted value µ(X) and

the actual Y .

Again, the normality and homoscedasticity assumptions are NOT needed

for these results. In fact, they hold for any regression function, not just
one satisfying the linear model. This includes the “fancy” nonparametric
techniques such as CART (Chapter 10) and SVM and neural networks
(Chapter 11).
90 CHAPTER 2. LINEAR REGRESSION MODELS

2.9.2 Definition of R2

The quantity R2 output by lm() is the sample analog of ρ2 :

R2 is the squared sample correlation between the actual re-

b R2 is a consis-
e ′ β.
sponse values Yi and the predicted values Xi
2
tent estimator of ρ .

Exactly how is R2 defined? From (2.64) and (2.65), we see that

V ar[ϵ]
ρ2 = 1 − (2.67)
V ar(Y )

Since Eϵ = 0, we have

V ar(ϵ) = E(ϵ2 ) (2.68)

The latter is the average squared prediction error in the population, whose
sample analog is the average squared error in our sample. In other words,
using our “correspondence” notation from before,

1∑
n
E(ϵ2 ) ←→ (Yi − X b2
ei′ β) (2.69)
n i=1

Now considering the denominator in (2.67), the sample analog is

1∑
n
V ar(Y ) ←→ (Yi − Y )2 (2.70)
n i=1

∑n
where of course Y = ( i=1 Yi )/n.
And that is R2 :

∑n b2
e ′ β)
1
(Yi − X i
R =1−
2 n
∑i=1
n (2.71)
1
n i=1 (Yi − Y )2

(Yes, the 1/n factors do cancel, but it will be useful to leave them there.)
As a sample estimate of the population ρ2 , the quantity R2 would appear
to be a very useful measure of the collective predictive ability of the X (j) .
2.9. COLLECTIVE PREDICTIVE STRENGTH OF THE X (J) 91

However, the story is not so simple, and curiously, the problem is actually
bias.

2.9.3 Bias Issues

R2 can be shown to be biased upward, not surprising in light of the fact

b In
that we are predicting on the same data that we had used to calculate β.
the extreme, we could fit an n − 1 degree polynomial in a single predictor,
with the curve passing through each data point, producing R2 = 1, even
though our ability to predict future data would be very weak.
The bias can be severe if p is a substantial portion of n. (In the above
polynomial example, we would have p = n − 1, even though we started
with p = 1.) This is the overfitting problem mentioned in the last chapter,
to be treated in depth in later chapters. But for now, let’s see how bad the
bias can be, using the following simulation code:
si mr2 <− function ( n , p , n r e p s ) {
r 2 s <− vector ( length=n r e p s )
for ( i i n 1 : n r e p s ) {
x <− matrix (rnorm( n∗p ) , ncol=p )
y <− x %∗% rep ( 1 , p ) + rnorm( n , sd=sqrt ( p ) )
r 2 s [ i ] <− g e t r 2 ( x , y )
}
hist ( r2s )
}

g e t r 2 <− function ( x , y ) {
smm <− summary(lm( y ∼ x ) )
smm$ r . s q u a r e d
}

Here we are simulating a population in which

Y = X (1) + ... + X (p) + ϵ (2.72)

so that β consists of a 0 followed by p 1s. We set the X (j) to have variance

1, and ϵ has variance p. This gives ρ2 = 0.50. Hopefully R2 will usually
be near this value. To assess this, I ran simr2(25,8,1000), i.e., n = 25
and p = 8, with 1000 repetitions of the experiment. The result is shown in
Figure 2.3.
92 CHAPTER 2. LINEAR REGRESSION MODELS

Figure 2.3: Plotted R2 values, n = 25

These results are not encouraging at all! The R2 values are typically around
0.7, rather than 0.5 as they should be. In other words, R2 is typically giving
us much too rosy a picture as to the predictive strength of our X (j) .
Of course, it should be kept in mind that I deliberately chose a setting that
produced substantial overfitting — 8 predictors for only 25 data points,
which is probably too many predictors.
Running the simulation with n = 250 should show much better behavior.
The results are shown in Figure 2.4. This is indeed much better. Note,
though, that the upward bias is still evident, with values more typically
above 0.5 than below it.
Note too that R2 seems to have large variance, even in the case of n = 250.
Thus in samples in which p/n is large, we should not take our sample’s
value of R2 overly seriously.

2.9.4 Adjusted-R2

The adjusted-R2 statistic is aimed at serving as a less biased version of

ordinary R2 . Its derivation is actually quite simple, though note that we
2.9. COLLECTIVE PREDICTIVE STRENGTH OF THE X (J) 93

Figure 2.4: Plotted R2 values, n = 250

do need to assume homoscedasticity.

Under the latter assumption, V ar() = σ 2 in (2.67). Then the numerator
in (2.71) is biased, which we know from (2.55) can be fixed by using the
factor 1/(n−p−1) instead of 1/n. Similarly, we know that the denominator
will be unbiased if we divide by 1/(n − 1) instead of 1/n. Those changes do
NOT make (2.71) unbiased; the ratio of two unbiased estimators is generally
biased. However, the hope is that this new version of R2 , called adjusted
R2 , will have less bias than the original.
The formula is

1
n 2
  β)
n−p−1 i=1 (Yi −Xi
2
Radj =1− 1
n (2.73)
n−1 i=1 (Yi − Y )2

We can explore this using the same simulation code as above. We simply
change the line
smm$ r . s q u a r e d
94 CHAPTER 2. LINEAR REGRESSION MODELS

Figure 2.5: Plotted adjusted R2 values, n = 25

to
smm$ a d j . r . s q u a r e d

Rerunning simr2(25,8,1000), we obtain the result shown in Figure 2.5.

This is a good sign! The values are more or less centered around 0.5, as
they should be (though there is still a considerable amount of variation).
See the Computational Complements section at the end of this chapter for
a note on how R2 and the adjusted version are computed from the output
of lm().

2.9.5 The “Leaving-One-Out Method”

Our theme here in Section 2.9 has been assessing the predictive ability of
our model, with the approach described so far being the R2 measure. But
recall that we have another measure: Section 1.12 introduced the concept
of cross-validation for assessing predictive ability. We will now look at a
variant of that method.
2.9. COLLECTIVE PREDICTIVE STRENGTH OF THE X (J) 95

First, a quick review of cross-validation: Say we have n observations in

our data set. With cross-validation, we randomly partition the data into
a training set and a validation set, of k and n − k observations, respec-
tively. We fit our model to the training set, and use the result to predict
the response variable in the validation set, and then see how well those
predictions turned out.
Clearly there is an issue of the choice of k. If k is large, our validation
set will be too small to obtain an accurate estimate of predictive ability.
That is not a problem if k is small, but then we have a subtler problem:
We are getting an estimate of strength of our model when constructed on
k observations, but in the end we wish to use all n observations.
One solution is the Leaving One Out Method (LOOM). Here we set k =
n − 1, but apply the training/validation process to all possible (n − 1, 1)
partitions. The name alludes to the fact that LOOM repeatedly omits one
observation, predicting it from fitting the model to the remaining observa-
tion. It is hoped that this gives us “the best of both worlds”: We have n
validation points, the best possible, and the training sets are of size n − 1,
i.e., nearly full-sized.

2.9.6 Extensions of LOOM

Instead of the Leaving One Out Method, we might leave out k observations
instead of just 1, known as k-fold cross-validation. In other words, for each
possible subsct of k observations, we predict those k by the remaining n−k.
This gives us many more test sets, at a cost of more computation.
There is theoretical evidence [128] that as the sample size n goes to infin-
ity, cross-validation will only provide statistical consistency if k-fold cross-
validation is used with k/n → 1.

2.9.7 LOOM for k-NN

LOOM is directly invocable in the regtools package. If in calling the

preprocesx() function, one specifies xval = TRUE, the computation of
nearest neighbors to a data point won’t include the point itself.
96 CHAPTER 2. LINEAR REGRESSION MODELS

2.9.8 Other Measures

A number of other measures of predictive ability are in common use, notably

Mallows’ Cp and the Akaike Information Criterion. These will be treated
in Chapter 9.

2.10 The Practical Value of p-Values — Small

OR Large

“Sir Ronald [Fisher] has befuddled us, mesmerized us, and led us down the
primrose path” — Paul Meehl, professor of psychology and the philosophy
of science

When the concept of significance testing, especially the 5% value for α, was
developed in the 1920s by Sir Ronald Fisher, many prominent statisticians
opposed the idea — for good reason, as we’ll see below. But Fisher was so
influential that he prevailed, and thus significance testing became the core
operation of statistics.
So, today significance testing is deeply entrenched in the field, and even
though it is widely recognized as faulty. many continue to engage in the
practice.10 Most modern statisticians understand this, It was eloquently
stated in a guide to statistics prepared for the U.S. Supreme Court by two
prominent scholars, one a statistician and the other a law professor [80]:

Statistical significance depends on the p-value, and p-values de-

pend on sample size. Therefore, a ‘significant’ effect could be
small. Conversely, an effect that is ‘not significant’ could be
large. By inquiring into the magnitude of an effect, courts can
avoid being misled by p-values. To focus attention where it be-
longs — on the actual size of an effect and the reliability of the
statistical analysis — interval estimates may be valuable. See-
ing a plausible range of values for the quantity of interest helps
describe the statistical uncertainty in the estimate.

The basic problem is that a significance test is answering the wrong ques-
tion. Say in a regression analysis we are interested in the relation between
10 Many are forced to do so, e.g., to comply with government standards in pharmaceu-

tical testing. My own approach in such situations is to quote the test results but then
point out the problems, and present confidence intervals as well.
2.10. THE PRACTICAL VALUE OF P-VALUES 97

X (1) and Y . Our test might have as null hypothesis

H0 : β1 = 0 (2.74)

But we probably know a priori that there is at least some relation between
the two variables; β1 cannot be 0.000000000... to infinitely many decimal
places. So we already know that H0 is false.11 The better approach is to
form a confidence interval for β1 , so that we can gauge the size of β1 , i.e.,
the strength of the relation.
Note carefully that this does not mean avoiding making a deci-
sion. The point is to make an informed decision, rather than letting the
machine make a decision for you that may not be useful.

2.10.1 Misleadingly Small p-Values

Many researchers are ecstatic when they find a tiny p-value. But actually,
that p-value may be rather meaningless.

2.10.1.1 Example: Forest Cover Data

For instance, consider another UCI data set, Forest Cover, which involves
a remote sensing project. The goal was to predict which one of seven types
of ground cover exists in a certain inaccessible location, using variables that
can be measured by satellite. One of the variables is Hillside Shade at Noon
(HS12).
For this example, I restricted the data to Cover Types 1 and 2, and took
a random subset of 1000 observations to keep the example manageable. I
named the resulting data frame f2512. The logistic model here is

1
P (Cover Type 2) = (2.75)
1 + e−(β0 +β1 HS12)

Here is the glm() output, with column 8 being HS12 and column 56 being
a dummy variable indicating Cover Type 2:
> glmout <−
glm( f 2 5 1 2 [ , 5 6 ] ∼ f 2 5 1 2 [ , 8 ] , family=binomial )
11 A similar point holds for the F-test in lm() output, which tests that all the βi are
0, i.e., H0 : β1 = β2 = . . . βp = 0.
98 CHAPTER 2. LINEAR REGRESSION MODELS

> summary( glmout )

...
Coefficients :
Estimate Std . E r r o r z v a l u e Pr ( >| z | )
( I n t e r c e p t ) −2.147856 0 . 6 3 4 0 7 7 −3.387 0 . 0 0 0 7 0 6 ∗∗∗
f2512 [ , 8] 0.014102 0.002817 5 . 0 0 7 5 . 5 3 e −07 ∗∗∗
...

The triple-star result for β1 would indicate that HS12 is a “very highly
significant” predictor of cover type. Yet we see that β c1 , 0.014102, is tiny.
HS12 is in the 200+ range, with sample means 227.1 and 223.4 for the two
cover types, differing only by 3.7. Multiplying the latter by 0.014102 gives a
value of about 0.052, which is swamped in (2.75) by the β c0 term, -2.147856.
In plain English: HS12 has almost no predictive power for Cover Type, yet
the test declares it “very highly significant.”
The confidence interval for β1 here is

0.014102 ± 1.96 · 0.002817 = (0.00858, 0.01962) (2.76)

The fact that the interval excludes 0 is irrelevant. The real value of the
interval here is that it shows that β1 is quite small; even the right-hand end
point is tiny.

2.10.1.2 Example: Click Through Data

In the informal analysis of the Click-Through Rate data in Section 1.8, we

stated that it appeared that educational level and CTR have no substantial
relation. Now that we have formal tools available, let’s revisit that.
> summary(lm( c t r $CTR ∼ c t r $ C o l l e g e Grad ) )
...
Coefficients :
Estimate Std . E r r o r t v a l u e
( Intercept ) 0.0141233 0.0005969 23.659
c t r $ C o l l e g e Grad −0.0137300 0 . 0 0 2 4 3 3 4 −5.642
Pr ( >| t | )
( Intercept ) < 2 e −16 ∗∗∗
c t r $ C o l l e g e Grad 8 . 2 8 e −07 ∗∗∗
...
Mult . R−s q u a r e d : 0 . 3 9 3 8 , Adj . R−s q u a r e d : 0 . 3 8 1 5
...
2.10. THE PRACTICAL VALUE OF P-VALUES 99

Again putting aside the issue of whether this data on U.S. states can be
considered a random sample from some population, the computer output
here declares the education variable to be “very highly significant.”
And indeed, the R-squared values are rather good. In fact, recalling that
R-squared is the squared sample correlation between Y and X (in the one-
predictor case), we see that that correlation is about -0.63. That value was
cited in the online article from which I obtained this data [53], which stated
“This data presents a stunning -0.63 rate between [educational level and
CTR].”
Yet all of those numbers are leading us astray. As we saw in Section 1.8,
the effect of education on CTR is quite negligible. So, again, one should
treat significance tests with great skepticism.
Note too our “sample size” n here seems small, only 51. On the one hand,
that would suggest a substantial upward bias in R-squared, as discussed
previously.
On the other hand, each CTR value is based on hundreds of thousands of
clicks, more than 31 million in all. In this sense we have a similar large-data
problem as in the forest cover data in the last section.

2.10.2 Misleadingly LARGE p-Values

Just as it is wrong to treat very small p-values as automatically showing

“significance,” it is equally dangerous to treat large p-values as evidence
of no important effect in the given quantity of interest, say a coeficient βi .
We might simply not have enough data to say anything at all about that
quantity.
This can occur even in very large samples. A notable case is dummy
variables. Say X (5) is a dummy variable and we are interested in estimating
β5 , to gauge the impact this variable has on Y . But we might have only a
few observations in which X (5) = 1. Intuitively this will prevent us from
obtaining a good estimate of β5 , even if n is very large. This makes sense
analytically, as follows.
The sample variance of X (5) will be q(1 − q), where q = P (X (6) = 1). The
latter will be small, resulting in a small corresponding element in A′ A in
Section 2.8, if we center the data. That produces a large value in (A′ A)−1 ,
thus a large standard error for βb5 .
So, a large p-value doesn’t necessarily indicate that the effect of interest is
100 CHAPTER 2. LINEAR REGRESSION MODELS

small.

2.10.3 The Verdict

So, a small p-value does not necessarily imply an important effect, and a
large p-value likewise should not be treated as showing lack of an important
effect. Very misleading results can occur by relying on p-values.
This book recommends using confidence intervals instead of p-values. The
two key points are:

• The width of the interval tells us the accuracy of our estimator.

• The location of the interval gives us an estimate of the size of the

effect.

Say our interval for βi excludes 0 but the location is very near 0. Then
the effect is probably small, and we should probably not call the effect
“signifcant.” On the other hand, if the interval is near 0 or even contains
it, but the interval is very wide, the latter property tells us that we just
don’t have much to say about the effect.
Some would object to our claim above that we almost always know a priori
that H0 is false. They might point to tests of the form

H0 : θ ≤ c, H1 : θ > c (2.77)

One major problem with that argument is that there is almost always mea-
surement error, due say to either finite-precision machine measurements
or sampling bias (sampling a narrower or skewed population than we had
intended). In addition, there is still the problem in which, say, θ > c but
with θ − c being so tiny that the difference is negligible.

2.11 Missing Values

The “dirty little secret” about data analysis is that most data is dirty. Some
data is erroneous, and/or it is often missing altogether.
On the one hand, R is very good about missing values, which are coded
as NA. It checks data for NAs (which comes at a cost of somewhat slower
2.12. MATHEMATICAL COMPLEMENTS 101

execution), and tries to accommodate them, typically by omitting obser-

vations that have at least one NA value. R actually has a function, com-
plete.cases(), to flag such observations.
Though seemingly reasonable, there can be real problems with the complete-
cases approach. The problem is that it can create biases. If we are studying
income and the richest people tend not to fill out the income blank in a
survey, the bias effect is obvious. But it can be more subtle than that.
For instance, one of the earliest methods for dealing with missing data is to
simply replace an NA by the mean of that variable. But this attentuates
the relations between variables. And again, in the case of the richer people
tending to omit the income portion of a survey, clearly replacing NA by the
mean income produces biases as well.
This problem has been the subject of intense study by top researchers [91],
yet there is no really good solution. The problem is that one must take
into account the mechanism that led to the NA values. Various levels of
randomness have been proposed, with the easiest one to solve being Missing
Completely at Random (MCAR). As its name implies, the fact that the
value Qij of variable j in observation i is NA is assumed to be statistically
independent of both Qij and the other Qik — a very stringent assumption.
Less stringent assumptions have been formulated as well, but are difficult
to verify.
The regtools package includes several functions that assume MCAR, for
the linear model, principal components and the log-linear model. Several
packages that attack the problem in a more sophisticated way are available
on CRAN, such as Amelia II. Further discussion is beyond the scope of this
book.

2.12 Mathematical Complements

2.12.1 Covariance Matrices

The covariance between random variables X and Y is defined as

Cov(X, Y ) = E[(X − EX)(Y − EY )] (2.78)

Suppose that typically when X is larger than its mean, Y is also larger than
its mean, and vice versa for below-mean values. Then (2.78) will likely be
positive. In other words, if X and Y are positively correlated (a term we
102 CHAPTER 2. LINEAR REGRESSION MODELS

will define formally below but keep intuitive for now), then their covariance
is positive. Similarly, if X is often smaller than its mean whenever Y is
larger than its mean and vice versa, the covariance and correlation between
them will be negative. All of this is roughly speaking, of course, since it
depends on how much and how often X is larger or smaller than its mean,
etc.
For a random vector U = (U1 , ..., Uk )′ , its covariance matrix Cov(U ) is the
k×k matrix whose (i, j) element is Cov(Ui , Uj ). Also, for a constant matrix
A with k columns, AU is a new random vector, and one can show that

Cov(AU ) = A Cov(U ) A′ (2.79)

In the special case in which A is a row vector a′ , this reduces to

V ar(a′ U ) = a′ Cov(U ) a (2.80)

Covariance does measure how much or little X and Y vary together, but
it is hard to decide whether a given value of covariance is “large” or not.
For instance, if we are measuring lengths in feet and change to inches, then
(2.78) shows that the covariance will increase by 12 if the unit change is
just in X, and by 122 = 144 if the change is in Y as well. Thus it makes
sense to scale covariance according to the variables’ standard deviations.
Accordingly, the correlation between two random variables X and Y is
defined by

Cov(X, Y )
ρ(X, Y ) = √ √ (2.81)
V ar(X) V ar(Y )

So, correlation is unitless, i.e. does not involve units like feet, pounds, etc.
And it can be shown (Section 2.12.8.1) that‘

−1 ≤ ρ(X, Y ) ≤ 1 (2.82)

These are all population values. The sample analog of (2.78) is

1∑
n
(Xi − X)(Yi − Y ) (2.83)
n i=1
2.12. MATHEMATICAL COMPLEMENTS 103

This is calculated by R’s cov() function, though as explained in Section

2.7.2, the function uses n − 1 instead of n as the divisor, following tradition.
An example of cov() is given in Section 2.13.2.1.

2.12.2 The Multivariate Normal Distribution Family

In Section 2.6.2, we discussed the bivariate normal distribution family, and

its role in motivating the classical assumptions of linear regression analysis.
Consider the case of general p. If (Y, X ′ )′ has a p + 1-variate normal distri-
bution, it can be shown that the Linearity, Normality and Homoscedasticity
assumptions hold for the regression of Y on any subset of the predictors in
X. In fact, by the symmetry of the situation, the same is true for regressing
any X (i) on any of the other X (j) and/or Y . This fact will be useful at
some points in coming chapters.
Matrix notation allows a compact representation of the multivariate normal
density:

′ −1
c e−0.5(t−µ) Σ (t−µ)
(2.84)

where µ and Σ are the mean vector and covariance matrix of the given
random vector, and c is a constant needed to make the density integrate to
1.0.
The multivariate normal distribution family has many interesting proper-
ties:

• Property A:
Suppose the random vector V has a multivariate normal distribution
with mean ν and covariance matrix Γ. Let A be a constant matrix
with the same number of columns as the length of V . Then the
random vector W = AV is also multivariate normal distributed, with
mean Aν and covariance matrix

Cov(W ) = AΓA′ (2.85)

Note carefully that the remarkable part of that last statement is that
W , the new random vector, also has a multivariate normal distribu-
tion, “inheriting” it from V . The statement about W ’s mean and
covariance matrix are true even if V does not have a multivariate
normal distribution, as we saw in Section 2.12.1.
104 CHAPTER 2. LINEAR REGRESSION MODELS

• Property B:
In general, if two random variables T and U have 0 correlation, this
does not imply they are independent. However, if they have a bivari-
ate normal distribution, the independence does hold.

• Property C:
Suppose B is a k × k idempotent matrix, i.e. B 2 = B, and suppose
U is a k-variate normally distributed random vector with B EU = 0
and with covariance matrix σ 2 I. Then the quantity

U ′ BU/σ 2 (2.86)

has a chi-squared distribution with degrees of freedom equal to rank(B).

R functions for multivariate normal distributions are discussed in Section

2.13.2.

2.12.3 The Central Limit Theorem

Roughly speaking, the Central Limit Theorem states that sums of random
variables have an approximately normal distribution. More formally, if
Ui , i = 1, 2, 3, ... are i.i.d., each with mean µ and variance σ 2 , then the
cumulative distribution function of

U1 + ... + Un − nµ
√ (2.87)
σ n

goes to that of N(0,1) as n → ∞.

The multivariate version is true as well. A sum of i.i.d. random vectors will
have an approximately multivariate normal distribution.
There are versions of the CLT for independent, non-identically distributed
random variables as well, under assumptions such as the Lindeberg Condi-
tion.

2.12.4 Details on Models Without a Constant Term

Here we will fill in the details of some claims made in Section 2.4.5. First,
let’s see why βb0 is forced to 0 if we center the data. Expand (2.18) (without
2.12. MATHEMATICAL COMPLEMENTS 105

the 1/n factor) to

∑
n
(1) (p)
(Yi − b0 − b1 Xi − ... − bp Xi )2 (2.88)
i=1

Set the partial derivative with respect to b0 to 0:

∑
n
(1) (p)
0= (Yi − b0 − b1 Xi − ... − bp Xi )(−1) (2.89)
i=1

But any centered variable will sum to 0, in this case Y and each X (j) , so
(2.89) becomes

0 = nb0 (2.90)

Since we take βb to be the value of b that minimizes (2.88), we see that

βb0 = 0 (2.91)

demonstrating the claimed result.

2.12.5 Unbiasedness of the Least-Squares Estimator

We will show that βb is conditionally unbiased,

E(βb | X1 , ..., Xn ) = β (2.92)

under the linearity assumption

e ′β
E(Y | X) = µ(X) = X (2.93)

This approach has the advantage of including the fixed-X case, and it also
implies the unconditional case for random-X, since

E βb = E[E(βb | X1 , ..., Xn )] = Eβ = β (2.94)

by the Law of Total Expectation, (1.58).

106 CHAPTER 2. LINEAR REGRESSION MODELS

So let’s derive (2.92). First note that Equation (2.93) tells us that

E(D | A) = Aβ (2.95)

where A and D are as in Section 2.4.2.

Now using (2.28) we have

E(βb | X1 , ..., Xn ) = E[βb | A] (2.96)

′ −1 ′
= E[(A A) A D|A] (2.97)
= (A′ A)−1 A′ E(D |A) (2.98)
= (A′ A)−1 A′ Aβ) (2.99)
= β (2.100)

thus showing that βb is unbiased.

2.12.6 Consistency of the Least-Squares Estimator

We’ll make use of a famous theorem:

Strong Law of Large Numbers (SLLN): Say W1 , W2 , ... are i.i.d.

(scalar or vector) with common mean EW. Then

1∑
n
lim Wi = EW, with probability 1 (2.101)
n→∞ n
i=1

Though independent, non-identically distributed versions of the SLLN do

exist, they have rather technical conditions, so we will assume the random-
X setting here.
Below, the role of Wi will sometimes be played by the vectors

(Yi , Xi , ..., Xi )′
(1) (p)
(2.102)

and sometimes by individual scalars. We will assume that the various

expectations exist, e.g., E(XX ′ ) below.
2.12. MATHEMATICAL COMPLEMENTS 107

Armed with that fundamental theorem in probability theory, rewrite (2.28)

as

( )−1
1 ′ 1
βb = AA ( A′ D) (2.103)
n n

e notation here for augmenting with

To avoid clutter, we will not use the X
a 1 element at the top of a vector. Assume instead that the 1 is X (1) .
1 ′
By the SLLN, the (i,j) element of nA A converges as n → ∞:

1 ∑ (i) (j)
n
1 ′
(A A)ij = Xk Xk → E[X (i) X (j) ] = [E(XX ′ )]ij (2.104)
n n
k=1

i.e.,

1 ′
A A → E(XX ′ ) (2.105)
n
The vector A′ D is a linear combination of the columns of A′ (Section A.4),
with the coefficients of that linear combination being the elements of the
vector D. Since the columns of A′ are Xk , k = 1, ..., n, we then have

∑
n
A′ D = Yk Xk (2.106)
k=1

and thus

1 ′
A D → E(Y X) (2.107)
n
The latter quantity is

E [E(Y X | X)] = E [XE(Y | X)] (2.108)

= E [X(X ′ β)] (2.109)
= E [X(X ′ Iβ)] (2.110)
= E [(XX ′ )Iβ)] (2.111)
= E(XX ′ ) β (2.112)
(2.113)
108 CHAPTER 2. LINEAR REGRESSION MODELS

So, we see that βb converges to

[E(XX ′ )]−1 E(XY ) = [E(XX ′ )]−1 E(XX ′ β) = [E(XX ′ )]−1 E(XX ′ )β = β

(2.114)

2.12.7 Biased Nature of S

It was stated in Section 2.7.2 that S, even with the n − 1 divisor, is a biased
estimator of η, the population standard deviation. We’ll derive that here.

0 < V ar(S) (2.115)

= E(S 2 ) − (ES)2 (2.116)
= η 2 − (ES)2 (2.117)

since S 2 is an unbiased estimator of η 2 . So,

ES < η (2.118)

2.12.8 The Geometry of Conditional Expectation

Readers with a good grounding in vector spaces may find the material in
this section helpful to their insight. It is recommended that the reader
review Section 1.19.5 before continuing.12

2.12.8.1 Random Variables As Inner Product Spaces

Consider the set of all scalar random variables U defined in some probability
space that have finite second moment, i.e. E(U 2 ) < ∞. This forms a linear
space: The sum of two such random variables is another random variable
with finite second moment, as is a scalar times such a random variable.
12 It should be noted that the treatment here will not be fully mathematically rigorous.

For instance, we bring in projections below, without addressing the question of the
conditions for their existence.
2.12. MATHEMATICAL COMPLEMENTS 109

We can define an inner product on this space. For random variables S and
T in this space, define

(S, T ) = E(ST ) (2.119)

This defines the norm

√
||S|| = (S, S)1/2 = E(S 2 ) (2.120)

So, if ES = 0, then

√
||S|| = V ar(S) (2.121)

Many properties for regression analysis can be derived quickly from this
vector space formulation. Let’s start with (2.82).
The famous Cauchy-Schwartz Inequality for inner product spaces states that
for any vectors x and y, we have

|(x, y)| ≤ ||x|| ||y|| (2.122)

It is left as an exercise to the reader to show that this implies (2.82).

2.12.8.2 Projections

Inner product spaces also have the notion of a projection. Suppose we have
an inner product space V, and subspace W. Then for any vector x, the
projection z of x onto W is defined to be the closest vector to x in W. An
important property is that we have a “right triangle,” i.e.

(z, x − z) = 0 (2.123)

We say that z and x − z are orthogonal. And the Pythagorean Theorem

holds:

||x||2 = ||z||2 + ||x − z||2 (2.124)

110 CHAPTER 2. LINEAR REGRESSION MODELS

2.12.8.3 Conditional Expectations As Projections

In regression terms, the discussion in Section 1.19.3 shows that the regres-
sion function, E(Y | X) = µ(X) has the property that

µ(X) = argming E[(Y − g(X))2 ] = argming ||Y − g(X)||2 (2.125)

as g ranges over all functions of X. Therefore, by definition, µ(X) is the

projection of Y onto the subspace consisting of all random variables with
finite variance that are functions of X. This view can be very useful.
We can also use (2.124) to derive the Law of Total Variance, (1.62). For
convenience in present notation, rewrite that equation as

V ar(Y ) = E[V ar(Y |X)] + V ar[E(Y |X)] (2.126)

The derivation will be less cluttered if we restrict attention to the case

EY = 0. (For the general case, define a new random variable W = Y −EY ,
and apply the mean-0 result, left as an exercise for the reader.) Note that
by the Law of Total Expectation (Section 1.19.5.2), this implies the µ(X)
also has mean 0.
Then (2.124) and (2.121) say that

[ ]
V ar(Y ) = E[µ(X)2 ] + E (Y − µ(X))2 (2.127)

Recalling that Eµ(X) = 0, the first term in (2.127) is

V ar[µ(X)] = V ar[E(Y |X)] (2.128)

which is exactly the second term in (1.62).

Now rewrite the second term in (2.127) using (1.58):

[ ] [ ]
E (Y − µ(X))2 = E{E (Y − µ(X))2 | X } (2.129)
= E[V ar(Y |X)] (2.130)

And, that last expression is exactly the first term in (1.62)! So, we are done
with the derivation.
2.12. MATHEMATICAL COMPLEMENTS 111

2.12.9 Predicted Values and Error Terms Are Uncor-

related

Assume a random-X context, and take x in (2.123) to be Y , so in that

equation

z = µ(X) (2.131)

and thus

E[µ(X)(Y − µ(X))] = 0 (2.132)

In other words, our prediction µ(X) is uncorrelated with our prediction

error, Y − µ(X).
The above concerns the population level, but a similar argument can be
made at the sample level for linear models. Here we will assume a fixed-X
model (conditioning on X in the random-X case), and once again use the
notation of Section 2.4.2.
Define

b ei βb
ϵi = Yi − X (2.133)

Also define b
ϵ to be the vector of the b
ϵi .
ϵi and βbj is 0 for any i and
The claim is then that the correlation between b
j. Again, a vector space argument can be made. In this case, take the full
vector space to be Rn , the space in which D roams, and the subspace will
be that spanned by the columns of A.
The vector Aβb is in that subspace, and because b = βb minimizes (2.25), Aβb
is then the projection of D onto that subspace. Again, that makes D − Aβb
and Aβb orthogonal, i.e.

ϵ′ Aβb = (D − Aβ)
b b ′ Aβb = 0 (2.134)

Since this must hold for all A, we see that each b

ϵi is uncorrelated with any
b
component of β.
112 CHAPTER 2. LINEAR REGRESSION MODELS

2.12.10 Classical “Exact” Inference

Again, assume a fixed-X setting. Classically we assume not only the linear
model, independence of the observations and homoscedasticity, but also
normality: Conditional on Xi , the response Yi has a normal distribution.
Since A is constant, (2.28) shows that βb has an exact multivariate normal
distribution.
All this gives rise to the classical hypothesis testing structure, such as the
use of the Student-t test for H0 : βi = 0. How does this work?
Our test statistic for H0 : βi = 0 is (2.56), i.e.

βbi − βi
√ (2.135)
s aii

Does this actually have a Student-t

√ distribution under H0 ? That distribu-
tion is defined as the ratio Z/ Q/r, where Z is a N(0,1) random variable,
Q is a chi-squared random variable with r degrees of freedom, and Z and
Q are independent.
Here, the roles of Z and Q will be played by

βbi − βi
Z= √ (2.136)
σ aii

and

Q = (n − p + 1)s2 /σ 2 (2.137)

from (2.55). Z is clearly N(0,1)-distributed, so it remains to show that Z

and Q are independent, and that Q has the claimed chi-squared distribu-
tion.
Are Z and Q independent? Yes: First note that the numerator in (2.55) is
ϵ′ b
b ϵ and βb are uncorre-
ϵ. Then recall that we found in Section 2.12.9 that b
lated. Since we are now assuming normality, that uncorrelatedness implies
that b b
ϵ is independent of β.
We then must show that Q from (2.137) is chi-squared distributed. This
can be seen as follows.
2.12. MATHEMATICAL COMPLEMENTS 113

The numerator in (2.55) is

b ′ (D − Aβ)
(D − Aβ) b (2.138)

But

D − Aβb = (I − H)D (2.139)

where H is the famous hat matrix,

H = A(A′ A)−1 A′ (2.140)

(Exercise 8). Furthermore,

b = Aβ − Aβ = 0
(I − H) ED = E(D − Aβ) (2.141)

b
by the linear model and the unbiasedness of β.
Finally, since A has full rank p + 1, the same will hold for H. Meanwhile,
I has rank n. Thus I − H will have rank n − p + 1 (Exercise (9)).
We then apply Property C in Section (2.12.2), establishing that Q has a
chi-squared distribution.

2.12.11 Asymptotic (p + 1)-Variate Normality of βb

Here we show that asymptotically βb has a (p + 1)-vartiate normal distri-

bution, and importantly, derive the corresponding asymptotic covariance
matrix, even without the normality assumption for Y given X, and for that
matter, without the homoscedasticity assumption. We assume the random-
X setting,13 and as in Section 2.12.6, avoid clutter by incorporating the 1
element of Xe into X.

First, define the actual prediction errors we would have if we knew the true
population value of β and were to predict the Yi from the Xi ,

ϵi = Yi − Xi′ β (2.142)
13 The derivation could be done for the fixed-X case, but we would need to use a CLT
for non-identically distributed random variables, and it would get messy.
114 CHAPTER 2. LINEAR REGRESSION MODELS

Let G denote the vector of the ϵi :

G = (ϵ1 , ..., ϵn )′ (2.143)

Then

D = Aβ + G (2.144)

√ b
We first show that the distribution of n(β −β) converges to (p+1)-variate
normal with mean 0.
Multiplying both sides of (2.144) by (A′ A)−1 A′ , we have

βb = β + (A′ A)−1 A′ G (2.145)

Thus

√ √
n(βb − β) = (A′ A)−1 n A′ G (2.146)

Using Slutsky’s Theorem and (2.105), the right-hand side has the same
asymptotic distribution as

√ 1
[E(XX ′ )]−1 n ( A′ G) (2.147)
n

We also have that

∑
n
A′ G = ϵi Xi (2.148)
i=1

This is a sum of i.i.d. terms with mean 0, the latter fact coming from

E(ϵX) = E[E(ϵX|X)] = 0 (2.149)

√
since E(ϵ|X) = 0. So the CLT says that n · (A′ G/n) is asymptotically
normal with mean 0 and covariance matrix equal to that of ϵX.
Putting this information together with (2.146), we have:
2.13. COMPUTATIONAL COMPLEMENTS 115

βb is asymptotically (p + 1)-variate normal with mean β and

covariance matrix

1
[E(XX ′ )]−1 Cov(ϵX)[E(XX ′ )]−1 (2.150)
n

But we can go still further: By the properties of covariance matrices (Ex-

ercise 12),

Cov(ϵX) = E(ϵ2 XX ′ ) − E(ϵX) E(ϵX)′ = E(ϵ2 XX ′ ) (2.151)

So, the general, heteroscedastic asymptotic covariance matrix of βb is

1
[E(XX ′ )]−1 E[ϵ2 XX ′ ][E(XX ′ )]−1 (2.152)
n

This version will turn out to be useful in Chapter 3.

2.13 Computational Complements

2.13.1 Details of the Computation of (2.28)

For the purpose of reducing roundoff error, linear regression software typ-
ically uses the QR decomposition in place of the actual matrix inversion
seen in (2.28). See Section A.5.
There is also the issue of whether the matrix inverse in (2.28) exists. Con-
sider again the example of female wages in Section 1.16.1. Suppose we
construct dummy variables for both male and female, and say, also use age
as a predictor:
> data ( prgeng )
> prgeng $ f e m a l e <− as . integer ( prgeng $ s e x == 1 )
> prgeng $male <− as . integer ( prgeng $ s e x == 2 )
> lm( wageinc ∼ age + male + f ema le , data=prgeng )
...
Coefficients :
( Intercept ) age male female
44178.1 489.6 −13098.2 NA
116 CHAPTER 2. LINEAR REGRESSION MODELS

How did that NA value arise? Think of the matrix A, denoting its column
j as cj . We have

• c1 consists of all 1s

• c2 contains the age values

• c3 has 1s for the men, 0s for the women

• c4 has 0s for the men, 1s for the women

Here is the key point:

c1 = c3 + c4 (2.153)

In other words,

1 · c1 + (−1) · c3 + (−1) · c4 = 0 (2.154)

Thus (the reader may wish to review Section A.7) A is of less than full
rank, as are A′ and A′ A. Therefore A′ A is not invertible.
There is the notion of generalized matrix inverse to deal with this issue,
useful in Analysis of Variance Models, but for our purposes, proper choice
of dummy variables solves the problem.

2.13.2 R Functions for the Multivariate Normal

Distribution Family

In R the density, cdf and quantiles of the multivariate normal distribution

are given by the functions dmvnorm(), pmvnorm() and qmvnorm() in
the library mvtnorm. You can simulate a multivariate normal distribution
by using mvrnorm() in the library MASS.

2.13.2.1 Example: Simulation Computation of a

Bivariate Normal Quantity

Consider a vector X = (X1 , X2 )′ having a bivariate normal distribution

with mean vecor (1, 1)′ , with standard deviations 1 for each component,
and correlation 0.5 between the components. Say we are interested in the
2.13. COMPUTATIONAL COMPLEMENTS 117

quantity E(|X1 − X2 |). This cannot be determined analytically, but it is

easy via simulation:

> l i b r a r y (MASS)
> mu <− c ( 1 , 1 )
> s i g <− rbind ( c ( 1 , 0 . 5 ) , c ( 0 . 5 , 1 ) )
> sig
[ ,1] [ ,2]
[1 ,] 1.0 0.5
[2 ,] 0.5 1.0
> x <− mvrnorm ( n=100 , mu=mu, Sigma=s i g )
> head ( x )
[ ,1] [ ,2]
[ 1 , ] 2.5384881 2.8800789
[ 2 , ] 1.4566831 2.1817883
[ 3 , ] −0.3286932 −0.2016951
[ 4 , ] 1.6158710 1.2448996
[ 5 , ] 2.0325496 0.1370805
[ 6 , ] 0.9100862 0.9779601
> mean( abs ( x [ , 1 ] − x [ , 2 ] ) )
[ 1 ] 0.8767933

So E(|X1 − X2 |) is about 0.88. Note the word about, though, as this is

only our sample estimate, not the population. To see this more concretely,
let’s get the estimate of the covariance matrix:

> cov ( x )
[ ,1] [ ,2]
[ 1 , ] 1.134815 0.3996990
[ 2 , ] 0.399699 0.9384828

This is substantially off the correct values:

( )
1 0.5
(2.155)
0.5 1

Since this is a simulation and thus we have the luxury of knowing the exact
covariance matrix and setting n, we see that we need to set a much larger
value of n.
118 CHAPTER 2. LINEAR REGRESSION MODELS

2.13.3 More Details of ’lm’ Objects

Since a call to summary() on an ”lm” object yields, among other things,

R2 and adjusted R2 , one might think that these quantities are components
of the object. There actually are no such components, and these quantities
are computed by summary.lm().
This will give us an opportunity to learn more about R objects, especially
those of ”lm” class. Let’s take a look.
> lmout <− lm( Weight ∼ Height + Age , data=mlb )
> names( lmout )
[1] ”coefficients” ” residuals ”
[3] ”effects” ” rank ”
[ 5 ] ” f i t t e d . values ” ” assign ”
[ 7 ] ” qr ” ” df . r e s i d u a l ”
[9] ” xlevels ” ”call”
[ 1 1 ] ” terms ” ” model ”

Some of these are themselves names of S3 objects, yes, objects-within-

objects!
> lmout$model
Weight Height Age
1 180 74 2 2 . 9 9
2 215 74 3 4 . 6 9
3 210 72 3 0 . 7 8
...
> s t r ( lmo$model )
’ data . frame ’ : 3 obs . o f 2 v a r i a b l e s :
$ y : num 1 2 4
$ x : int 1 2 3
− attr ( ∗ , ” terms ”)= C l a s s e s ’ terms ’ ,
’ f o r m u l a ’ length 3 y ∼ x
. . .. − attr ( ∗ , ” v a r i a b l e s ”)= l a n g u a g e l i s t ( y , x )
...

Another way to inspect an object is str()

> s t r ( lmout )
L i s t o f 12
$ c o e f f i c i e n t s : Named num [ 1 : 3 ] −187.638 4 . 9 2 4 0 . 9 1 2
.. − attr ( ∗ , ”names”)= c h r [ 1 : 3 ] ” ( I n t e r c e p t ) ”
” Height ” ”Age”
$ residuals : Named num [ 1 : 1 0 1 5 ] −17.66 6 . 6 7 1 5 . 0 8
2.13. COMPUTATIONAL COMPLEMENTS 119

1 0 . 8 4 −16.34 . . .
.. − attr ( ∗ , ”names”)= c h r [ 1 : 1 0 1 5 ] ” 1 ” ” 2 ” ” 3 ” ” 4 ” . . .
$ effects : Named num [ 1 : 1 0 1 5 ] −6414.8 −352.5
−124.8 1 0 . 8 −16.5 . . .
.. − attr ( ∗ , ”names”)= c h r [ 1 : 1 0 1 5 ] ” ( I n t e r c e p t ) ”
” Height ” ”Age” ”” . . .
$ rank : int 3
$ f i t t e d . v a l u e s : Named num [ 1 : 1 0 1 5 ] 198 208 195
199 204 . . .
.. − attr ( ∗ , ”names”)= c h r [ 1 : 1 0 1 5 ] ” 1 ” ” 2 ” ” 3 ”
”4” . . .
$ assign : int [ 1 : 3 ] 0 1 2
$ qr : List of 5
. . $ qr : num [ 1 : 1 0 1 5 , 1 : 3 ] −31.8591 0 . 0 3 1 4 0 . 0 3 1 4
0.0314 0.0314 . . .
. . .. − attr ( ∗ , ”dimnames”)= L i s t o f 2
...

The fitted.values component will be used occasionally in this book. The

ith one is

b(Xi )
µ (2.156)

which can be viewed in two ways:

• It is the value of the estimated regression function at the ith observa-

tion in our data set.
• It is the value that we would predict for Yi if we did not know Yi .
(We do know it, of course.)

In the context of the second bullet, our prediction error for observation i is

Yi − µ
b(Xi ) (2.157)

Recall that this is known as the ith residual. These values are available to
us in lmout$residuals.
With these various pieces of information in lmout, we can easily calculate
R2 in (2.71). The numerator there, for instance, involves the sum of the
squared residuals. The reader can browse through these and other compu-
tations by typing
120 CHAPTER 2. LINEAR REGRESSION MODELS

> edit (summary . lm)

2.14 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. Consider the census data in Section 1.16.1.

(a) Form an approximate 95% confidence interval for β6 in the model

(1.28).

(b) Form an approximate 95% confidence interval for the gender effect
for Master’s degree holders, β6 + β7 , in the model (1.28).

2. The full bikeshare dataset spans 3 years’ time. Our analyses here have
only used the first year. Extend the analysis in Section 2.8.5 to the full
data set, adding dummy variables indicating the second and third year.
Form an approximate 95% confidence interval for the difference between
the coefficients of these two dummies.
3. Suppose we are studying growth patterns in children, at k particular
ages. Denote the height of the ith child in our sample data at age j by
Hij , with Hi = (Hi1 , ..., Hik )′ denoting the data for child i. Suppose the
population distribution of each Hi is k-variate normal with mean vector µ
and covariance matrix Σ. Say we are interested in successive differences in
heights, Dij = Hi,j+1 − Hij , j = 1, 2, ..., k − 1. Define Di = (Di1 , ..., Dik )′ .
Explain why each Di is (k−1)-variate normal, and derive matrix expressions
for the mean vector and covariance matrices.
4. In the simulation in Section 2.9.3, it is claimed that ρ2 = 0.50. Confirm
this through derivation.
5. In the census example in Section 1.16.2, find an appropriate 95% con-
fidence interval for the difference in mean incomes of 50-year-old men and
50-year-old women. Note that the data in the two subgroups will be inde-
pendent.
Mini-CRAN and other problems:
6. Write a function with call form
2.14. DATA, CODE AND MATH PROBLEMS 121

mape ( lmout )

where lmout is an ’lm’ object returned from a call to lm(). The function
will compute the mean absolute prediction error,

1∑
n
|Yi − µ
b(Xi )| (2.158)
n i=1

You’ll need the material in Section 2.13.3.

Math problems:
7. Show that (2.122) implies (2.82).
8. Derive (2.138).
9. Suppose B is a k × k symmetric, idempotent matrix. Show that

rank(I − B) = k − rank(B) (2.159)

(Suggestion: First show that the eigenvalues of an idempotent matrix must

be either 0 or 1.)
10. In the derivation of (1.62) in Section (2.12.8.3), we assumed for con-
venience that EY = 0. Extend this to the general case, by defining
W = Y − EY and then applying (1.62) to W and X. Make sure to justify
your steps.
11. Suppose we have random variables U and V , with equal expected values
and each with variance 1. Let ρ denote the correlation between them. Show
that

lim P (|U − V | < ϵ) = 1 (2.160)

ρ→1

for any ϵ > 0. (Hint: Use Markov’s Inequality, P (T > c) ≤ ET /c for any
nonnegative random variable T and any positive constant c.)
12. For a scalar random variable U , a famous formula is

V ar(U ) = E(U 2 ) − (EU )2 (2.161)

122 CHAPTER 2. LINEAR REGRESSION MODELS

Derive the analog for the covariance matrix of a random vector U ,

Cov(U ) = E(U U ′ ) − (EU )(EU )′ (2.162)

Chapter 3

Homoscedasticity and
Other Assumptions in
Practice

This chapter will take a practical look at the classical assumptions of linear
regression models. While most of the assumptions are not very important
for the Prediction goal, assumption (d) below both matters for Description
and has a remedy. Again, this is crucial for the Description goal, be-
cause otherwise our statistical inference may be quite inaccurate.
To review, the assumptions are:

(a) Linearity: The conditional mean is linear in parameters β.

e =e
E(Y | X t) = e
t′ β (3.1)

(b) Normality: The conditional distribution of Y given X is normal.

(c) Independence: The data (Xi , Yi ) are independent across i.

(d) Homoscedasticity:

V ar(Y | X = t) (3.2)

is constant in t.

123
124 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

Verifying assumption (a), and dealing with substantial departures from it,
form the focus of an entire chapter, Chapter 6. So, this chapter will focus
on assumptions (b)-(d).
The bulk of the material will concern (d).

3.1 Normality Assumption

We already discussed (b) in Section 2.8.4, but the topic deserves further
comment. First, let’s review what was found before. (We continue to use
the notation from that chapter, as in Section 2.4.2.)

Neither normality nor homoscedasticity is needed for βb to be unbiased, con-

sistent and asymptotically normal. Standard statistical inference proce-
dures do assume homoscedasticity, but we’ll return to the latter issue in
Section 3.3. For now, let’s concentrate on the normality assumption. Re-
taining the homoscedasticity assumption for the moment, we found in the
last chapter that:

b
The conditional distribution of the least-squares estimator β,
given A, is approximately multivariate normal distributed with
mean β and approximate covariance matrix

s2 (A′ A)−1 (3.3)

Thus the standard error of βbj is the square root of element j

of this matrix (counting the top-left element as being in row 0,
column 0).
Similarly, suppose we are interested in some linear combination
b
λ′ β of the elements of β (Section A.4), estimating it by λ′ β.
The standard error is the square root of

s2 λ′ (A′ A)−1 λ (3.4)

The reader should not overlook the word asymptotic in the above. With-
out assumption (b) above, our inference procedures (confidence intervals,
significance tests) are indeed valid, but only approximately. On the other
hand, the reader should be cautioned (as in Section 2.8.1) that so-called
“exact” inference methods assuming normal population distributions, such
as the Student-t distribution and the F distribution, are themselves only
approximate, since true normal distributions rarely if ever exist in real life.
3.2. INDEPENDENCE ASSUMPTION — DON’T OVERLOOK IT 125

In other words:

We must live with approximations one way or the other, and

the end result is that the normality assumption is not very im-
portant.

3.2 Independence Assumption — Don’t

Overlook It

Statistics books tend to blithely say things like “Assume the data are inde-
pendent and identically distributed (i.i.d.),” without giving any comment
to (i) how they might be nonindependent and (ii) what the consequences
are of using standard statstical methods on nonindependent data. Let’s
take a closer look at this.

3.2.1 Estimation of a Single Mean

√
Note the denominator S/ n in (2.42). This is the standard error of W ,
i.e., the estimated standard deviation of that quantity. That in turn comes
from a derivation you may recall from statistics courses,

1 ∑n
V ar(W ) = V ar( Wi ) (3.5)
n2 i=1

1 ∑
n
= V ar(Wi )) (3.6)
n2 i=1
1 2
= σ (3.7)
n

and so on. In going from the first equation to the second, we are making
use of the usual assumption that the Wi are independent.
But suppose the Wi are correlated. Then the correct equation is

∑n ∑
n ∑
V ar( Wi ) = V ar(Wi ) + 2 Cov(Wi , Wj ) (3.8)
i=1 i=1 1≤i<j≤n
126 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

It is often the case that our data are positively correlated. Many data sets,
for instance, consist of multiple measurements on the same person, say 10
blood pressure readings for each of 100 people. In such cases, the covariance
terms in (3.8) will be positive, and (3.7) will yield too low a value. Thus
the denominator in (2.42) will be smaller than it should be. That means
that our confidence interval (2.44) will be too small (as will be p-values), a
serious problem in terms of our ability to do valid inference.
Here is the intuition behind this: Although we have 1000 blood pressure
readings, the positive intra-person correlation means that there is some
degree of repetition in our data. Thus we don’t have “1000 observations
worth” of data, i.e., our effective n is less than 1000. Hence our confidence
interval, computed using n = 1000, is overly optimistic.
Note that W will still be an unbiased and consistent estimate of ν.1 In
other words, W is still useful, even if inference procedures computed from
it may be suspect.

3.2.2 Inference on Linear Regression Coefficients

All of this applies to inference on regression coefficients as well. If our data

is correlated, i.e., rows within (A, D) are not independent, then (2.54) will
be incorrect, because the off-diagonal elements in Cov(Y ) won’t be 0s. And
if they are positive, (2.54) will be “too small,” and the same will be true for
(3.3). Again, the result will be that our confidence intervals and p-values
will be too small, i.e., overly optimistic. In such a situation, then our βb will
still be useful, but our inference procedures will be suspect.

3.2.3 What Can Be Done?

This is a difficult problem. Some possibilities are:

• Simply note the dependency problem, e.g., in our report to a client,

and state that though our estimates are valid (in the sense of statis-
tical consistency), we don’t have reliable standard errors.
• Somehow model the dependency, i.e., model the off-diagonal elements
of Cov(Y ). For instance, in the blood pressure case, we might try
mixed effects models [77] [52]. Or if the dependency is due to a time
variable, one might use time series models.
1 Mathematically, this claim about consistency would have to be posed more rigor-
ously, with conditions on the quantities Cov(Wi , Wj ), say as ]i − j| → ∞.
3.2. INDEPENDENCE ASSUMPTION — DON’T OVERLOOK IT 127

• Collapse the data in some way to achieve independence.

An example of this last point is presented in the next section.

3.2.4 Example: MovieLens Data

The MovieLens data (http://grouplens.org/ ) consists of ratings of various

movies by various users. The main file of the 100K version, which we’ll an-
alyze here, consists of columns User ID, Movie ID, Rating and Timestamp.
There is one row per rating. If a user has rated, say eight movies, then
he/she will have eight rows scattered about in the data matrix. Of course,
most users have not rated most movies. There is also a file for user data,
with age, gender and so on,2 and another for movie data, showing title,
genres etc.
Let Yij denote the ratings of user i, j = 1, 2, ..., Ni , where Ni is the number
of movies rated by this user. In the example here, we are not taking into
account which movies the user rates, just analyzing general user behavior.
We are treating the users as a random sample from a conceptual population
of all potential users, and similarly for the movies.
As with the blood pressure example above, for fixed i, the Yij are not
independent, since they come from the same user. Some users tend to give
harsher ratings, others tend to give favorable ones. But we can form

1 ∑
Ni
Ti = Yij (3.9)
Ni j=1

the average rating given by user i, and now we have independent random
variables. And, if we treat the Ni as random too, and i.i.d., then the Ti are
i.i.d., enabling standard statistical analyses.
For instance, we can run the model, say,

mean rating = β0 + β1 age + β2 gender (3.10)

and then pose questions such as “Do older people tend to give lower rat-
ings?” Let’s see what this gives us.
2 Unfortunately, in recent editions of the data, this is no longer included.
128 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

The data are in two separate files: u.data contains the ratings, and u.user
contains the demographic information. Let’s read in the first file:
> ud <− read . table ( ’ u . data ’ , h e a d e r=FALSE, s e p= ’ \ t ’ )
> head ( ud )
V1 V2 V3 V4
1 196 242 3 881250949
2 186 302 3 891717742
3 22 377 1 878887116
4 244 51 2 880606923
5 166 346 1 886397596
6 298 474 4 884182806

The first record in the file has user 196 rating movie 242, giving it a 3, and
so on. (The fourth column is a timestamp.) There was no header on this
file (nor on the next one we’ll look at below), and the field separator was a
TAB.
Now, let’s find the Ti . For each unique user ID, we’ll find the average rating
given by that user, making use of the tapply() function (Section 1.20.2):
> z <− tapply ( ud$V3 , ud$V1 , mean)
> head ( z )
1 2 3 4 5 6
3.610294 3.709677 2.796296 4.333333 2.874286 3.635071

We are telling R, “Group the ratings by user, and find the mean of each
group.” So user 1 (not to be confused with the user in the first line of
u.data, user 196) gave ratings that averaged 3.610294 and so on.
Now we’ll read in the demographics file:
> uu <− read . table ( ’ u . u s e r ’ , h e a d e r=F , s e p= ’ | ’ )
# no names i n t h e o r i g data , so add some
> names( uu ) <− c ( ’ u s e r i d ’ , ’ age ’ , ’ g e n d e r ’ ,
’ occup ’ , ’ z i p ’ )
> head ( uu )
u s e r i d age g e n d e r occup zip
1 1 24 M t e c h n i c i a n 85711
2 2 53 F o t h e r 94043
3 3 23 M w r i t e r 32067
4 4 24 M t e c h n i c i a n 43537
5 5 33 F o t h e r 15213
6 6 42 M e x e c u t i v e 98101
3.2. INDEPENDENCE ASSUMPTION — DON’T OVERLOOK IT 129

Now append our Ti to do this latter data frame, and run the regression:
> uu$ ge n d e r <− as . integer ( uu$ g e n d e r == ’M’ )
> uu$avg r a t <− z
> head ( uu )
u s e r i d age g e n d e r occup zip avg r a t
1 1 24 0 t e c h n i c i a n 85711 3 . 6 1 0 2 9 4
2 2 53 0 o t h e r 94043 3 . 7 0 9 6 7 7
3 3 23 0 w r i t e r 32067 2 . 7 9 6 2 9 6
4 4 24 0 t e c h n i c i a n 43537 4 . 3 3 3 3 3 3
5 5 33 0 o t h e r 15213 2 . 8 7 4 2 8 6
6 6 42 0 e x e c u t i v e 98101 3 . 6 3 5 0 7 1
> q <− lm( avg r a t ∼ age + gender , data=uu )
> summary( q )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) 3 . 4 7 2 5 8 2 1 0 . 0 4 8 2 6 5 5 7 1 . 9 4 7 < 2 e −16 ∗∗∗
age 0.0033891 0.0011860 2 . 8 5 8 0 . 0 0 4 3 6 ∗∗
gender 0.0002862 0.0318670 0.009 0.99284
...
M u l t i p l e R−s q u a r e d : 0 . 0 0 8 6 1 5 ,
Adjusted R−s q u a r e d : 0 . 0 0 6 5 0 5

This is again an example of how misleading signficance tests can be. The
age factor here is “double star,” so the standard response would be “Age is
a highly signficant predictor of movie rating.” But that is not true at all. A
10-year difference in age only has an impact of about 0.03 on ratings, which
are on the scale 1 to 5. And the R-squared values are tiny. So, while the
older users tend to give somewhat higher ratings, the effect is negligible.
On the other hand, let’s look at what factors may affect which kinds of
users post more. Consider the model

mean number of ratings = β0 + β1 age + β2 gender (3.11)

Run the analysis:

# g e t r a t i n g s count f o r each u s e r
> uu$ n i <− tapply ( ud$V3 , ud$V1 , length )
> summary(lm( n i ∼ age + gender , data=uu ) )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
130 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

( Intercept ) 120.6782 10.9094 11.062 < 2 e −16 ∗∗∗

age −0.7805 0.2681 −2.912 0 . 0 0 3 6 8 ∗∗
gender 16.8124 7.2029 2.334 0.01980 ∗
...

There is a somewhat more substantial age effect here, with older people
posting fewer ratings. A 10-year age increase brings about 8 fewer postings.
Is that a lot?
> mean( uu$ n i )
[ 1 ] 106.0445

So there is a modest decline in posting activity.

But look at gender. Women rate an average of 16.8 fewer movies than men,
rather substantial.

3.3 Dropping the Homoscedasticity

Assumption

For an example of problems with the homoscedasticity assumption, again

consider weight and height. It is intuitive that tall people have more vari-
ation in weight than do short people, for instance. We can confirm that in
our baseball player data. Let’s find the sample standard deviations for each
height group (restricting to the groups with over 50 observations), seen in
Section 1.6.2):
> library ( freqparcoord )
> data ( mlb )
> m70 <− mlb [ mlb$ Height >= 70 & mlb$ Height <= 7 7 , ]
> s d s <− tapply (m70$Weight , m70$Height , sd )
> plot ( 7 0 : 7 7 , s d s )

The result is shown in Figure 3.1. The upward trend is clearly visible, and
thus the homoscedasticity assumption is not reasonable.
The vector βb in (2.28) is called the ordinary least-squares (OLS) estimator
of β, in contrast to weighted least-sqaures (WLS), to be discussed shortly.
Statistical inference on β using OLS is based on (2.54),

b = σ 2 (A′ A)−1
Cov(β) (3.12)
3.3. DROPPING THE HOMOSCEDASTICITY ASSUMPTION 131

Figure 3.1: Standard deviations of weight, by height group

which is in turn based on the homoscedasticity assumption — that V ar(Y | X =

t) is constant in t. Yet that assumption is rarely if ever valid.
Given the inevitable nonconstancy of (2.32), there are questions that must
be raised:

• Do departures from constancy of (2.32) in t substantially impact the

validity of statistical inference procedures that are based on (2.54)?

• Can we somehow estimate the function σ 2 (t), and then use that in-
formation to perform a WLS analysis?

• Can we somehow modify (2.54) for the heteroscedastic case?

These points will be addressed in this section.

3.3.1 Robustness of the Homoscedasticity Assumption

In statistics parlance, we ask,

132 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

Is classical inference on the vector of regression coeficients β ro-

bust to the homoscedasticity assumption, meaning that there is
not much effect on the validity of our inference procedures (con-
fidence intervals, significance tests) unless the setting is quite
profoundly heteroscedastic?

We can explore this idea via simulation in known settings. For instance,
let’s investigate settings in which

σ(t) = |µ(t)|q (3.13)

where q is a parameter to vary in the investigation. We’ve discussed previ-

ously that typically in practice, σ 2 (t) should increase with µ(t) if the latter
is positive, so for q > 0 Equation (3.13) should serve as a good test bed in
which to explore the effects of heteroscedasticity
This includes several important cases in which the value of q is implied by
famous models:

• q = 0: Homoscedasticity.

• q = 0.5: Conditional distribution of Y given X is Poisson.

• q = 1: Conditional distribution of Y given X is exponential.

Below is the code. It simulates a setting in which

E(Y | X = t) = β1 t1 + ... + βp tp (3.14)

and with conditional variance as in (3.13). We construct a nominal 95%

confidence interval for β1 in each simulated sample of size n, and then
calculate the proportion of the nreps intervals that contain 1.0, the true
value of β1 .
s i mh e t <− function ( n , p , nreps , sdpow ) {
bh1s <− vector ( length=n r e p s )
s e s <− vector ( length=n r e p s )
for ( i i n 1 : n r e p s ) {
x <− matrix (rnorm( n∗p ) , ncol=p )
meany <− x %∗% rep ( 1 , p )
s d s <− abs ( meany ) ˆ sdpow
y <− meany + rnorm( n , sd=s d s )
3.3. DROPPING THE HOMOSCEDASTICITY ASSUMPTION 133

n p q conf. lvl.
100 5 0.0 0.94683
100 5 0.5 0.92359
100 5 1.0 0.90203
100 5 1.5 0.87889
100 5 2.0 0.86129

Table 3.1: Heteroscedasticity effect simulation

lmout <− lm( y ∼ x )

bh1s [ i ] <− coef ( lmout ) [ 2 ]
s e s [ i ] <− sqrt ( vcov ( lmout ) [ 2 , 2 ] )
}
mean( abs ( bh1s − 1 . 0 ) < 1 . 9 6 ∗ s e s )
}

The simulation finds the true confidence level (providing nreps is set to a
large value) corresponding to a nominal 95% confidence interval. Table 3.1
shows the results of a few runs, all with nreps set to 100,000. We see that
there is indeed an effect on the true confidence level.

3.3.2 Weighted Least Squares

If one knows the function σ 2 (t) (at least up to a constant multiple), one
can perform a weighted least-squares (WLS) analysis. Here, instead of
minimizing (2.18), one minimizes

1∑ 1
n
′
fi b)2
(Yi − X (3.15)
n i=1 wi

(without the 1/n factor, of course), where

wi = σ 2 (Xi ) (3.16)

Just as one can show that in the homoscedastic case, OLS gives the opti-
mal (minimum-variance unbiased) estimator, the same is true for WLS in
134 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

heteroscedastic settings, provided we know the function σ 2 (t).

R’s lm() function has an optional weights argument for specifying the wi .
But needless to say, this situation is not common. To illustrate this point,
consider the classical inference procedure for a single mean, reviewed in
Section 2.8.1. If we don’t know the population mean ν, we are even less
likely to know the population variance η 2 . The same holds in the regression
context, concerning conditional means and conditional variances.
One option would be to estimate the function σ(t) using nonparametric
regression techniques. This was first proposed in [122] and later studied
extensively in work by Raymond Carroll, including [32].
For instance, we can use our k-NN function knnest() in the regtools
package [97], with nearf = vary. The latter specifies to calculate the
sample variances of Y in the neighborhoods of our X values. This gives us
estimates of σ 2 (Xi ), exactly what we need.
With the baseball data, though, we can just estimate the variances as we
did for Figure 3.1. Let’s run the analysis first with, and then without,
weights, and check how much difference weighting makes:

> mlb <− mlb [ , c ( 4 , 6 , 5 ) ]

> m70 <− mlb [ mlb$ Height >= 70 & mlb$ Height <= 7 7 , ]
> v a r s <− tapply (m70$Weight , m70$Height , var )
> wts <− 1 / v a r s [ m70$Height −69]
> summary(lm(m70$Weight ∼ m70$Height , weights=wts ) )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) −160.5523 2 1 . 7 8 7 0 −7.369 3 . 7 8 e −13
m70$ Height 4.9119 0 . 2 9 7 2 1 6 . 5 2 8 < 2 e −16

( I n t e r c e p t ) ∗∗∗
m70$ Height ∗∗∗
...
> summary(lm(m70$Weight ∼ m70$ Height ) )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) −162.8544 2 2 . 7 1 8 2 −7.168 1 . 5 4 e −12
m70$ Height 4.9438 0 . 3 0 8 7 1 6 . 0 1 3 < 2 e −16

( I n t e r c e p t ) ∗∗∗
m70$ Height ∗∗∗
3.3. DROPPING THE HOMOSCEDASTICITY ASSUMPTION 135

...

The weighted analysis, the “true” one (albeit with the weights being only
approximate), did give slightly different results than those of OLS, including
the standard errors.
It should be noted that the estimated conditional variances seem to flatten
somewhat toward the right end, and are based on smaller sample sizes at
both ends:
> vars
70 71 72 73 74
183.3702 270.0965 308.4762 269.3699 327.7612
75 76 77
333.9579 399.2606 341.7576
> tapply (m70$Weight , m70$Height , length )
70 71 72 73 74 75 76 77
51 89 150 162 173 155 101 55

3.3.3 A Procedure for Valid Inference

Fortunately, there exists a rather simple procedure, originally developed by

Eicker and later refined by White [43] [137]. This section will present the
methodology, and test it on data.

3.3.4 The Methodology

It was found in Section 2.12.11 that the general, heteroscedastic asymptotic

covariance matrix of βb is

1
[E(XX ′ )]−1 E[ϵ2 XX ′ ][E(XX ′ )]−1 (3.17)
n

Each of the factors is easily estimated by sample analogs:

b ′ 1
E(XX ) = A′ A (3.18)
n

∑ n
b 2 XX ′ ) = 1
E(ϵ ϵ2 Xi Xi′
b (3.19)
n i=1 i
136 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

where as before

b ei βb
ϵi = Yi − X (3.20)

The expression in (3.19) is rather unwieldy, but it is easy to program, and

most important, we’re in business! We can now conduct valid statistical
inference even in the heteroscedastic case.
Code implementing (3.19) (or refinements of it) is available in R’s car [49]
and sandwich packages [95], as the functions hccm() and vcovHC(),
respectively. (These functions also offer various refinements of the method.)
They are drop-in replacements for the standard vcov(), so that for instance
we could use vcovHC() in place of vcov() on page 87.

3.3.5 Example: Female Wages

Let’s try this on the census data, Section 1.16.1. We’ll find the standard
error of βb6 , the coefficient of the fem variable.
Keep in mind that in β and β, b the first element is number 0, so βb6 is the
seventh. From Section 2.8.4, then, the estimated variance of βb6 is given in
b a matrix obtain-
the (7,7) element of the estimated covariance matrix of β,
able through R’s vcov() function under the homoscedasticity assumption.
The standard error of βb is then the square root of that (7,7) element. As
mentioned, vcovHC() is a drop-in replacement for vcov() in using the
Eicker method, without assuming homoscedasticity.
Continuing with the data frame prgeng in that example, we have

> lmout <− lm( wageinc ∼ age+age2+wkswrkd+ms+phd+fem ,

data=prgeng )
> sqrt ( vcov ( lmout ) [ 7 , 7 ] )
[ 1 ] 705.2994
> l i b r a r y ( sandwich )
> sqrt ( vcovHC ( lmout ) [ 7 , 7 ] )
[ 1 ] 593.85

That is quite a difference. Apparently, using OLS was conservative in this

case, i.e., was causing us to have wider confidence intervals than necessary
and was inflating p-values.
3.3. DROPPING THE HOMOSCEDASTICITY ASSUMPTION 137

n p q conf. lvl.
100 5 0.0 0.95176
100 5 0.5 0.94928
100 5 1.0 0.94910
100 5 1.5 0.95001
100 5 2.0 0.95283

Table 3.2: Heteroscedasticity simulation

3.3.6 Simulation Test

Let’s see if it works, at least in the small simulation experiment in Section

3.3.1. We use the same code as before, simply replacing the call to vcov()
by one to vcovHC(). The results, shown in Table 3.2, are excellent.

3.3.7 Variance-Stabilizing Transformations

Most classical treatments of regression analysis devote a substantial amount

of space to transformations of the data. For instance, one might replace
Y by ln Y , and possibly apply the log to the predictors as well. There are
several reasons why this might be done:

(a) The distribution of Y given X may be skewed, and applying the log
may make it more symmetric, thus more normal-like.

(b) Log models may have some meaning relevant to the area of applica-
tion, such as elasticity models in economics.

(c) Applying the log may convert a heteroscedastic setting to one that is
close to homoscedastic.

One of the themes of this chapter has been that the normality assumption
is not of much practical importance, which would indicate that Reason
(a) above may not be so useful. Reason (b) is domain-specific, and thus
outside the scope of this book. But Reason (c) relates directly to our
current discussion on heteroscedasticity. Here is how transformations come
into play.
138 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

The Delta Method (Section 3.6.1) says, roughly, that if the random variable
W is approximately normal with a small coefficient of variation (ratio of
standard deviation to mean), and g is a smooth function, then the new
random variable g(W ) is also approximately normal, with mean g(EW )
and variance

[g ′ (EW )]2 V ar(W ) (3.21)

Let’s consider that in the context of (3.13). Assuming that the regression
function is always positive, (3.13) reduces to

σ(t) = µq (t) (3.22)

Now, suppose (3.22) holds with q = 1. Take g(t) = ln(t). Then since

d 1
ln t = (3.23)
dt t

we see that (3.21) becomes

1
· µ2 (t) = 1 (3.24)
µ2 (t)

In other words V ar(ln Y | X = t) is approximately 1 for all t, and we are√

back to the homoscedastic case. Similarly, if q = 0.5, then setting g(t) = t
would give us approximate homoscedasticity.
However, this method has real drawbacks: Distortion of the model, diffi-
culty interpreting the coefficients and so on.
Let’s look at a very simple model that illustrates the distortion issue. (It
is further explored in Section 3.6.2.) Suppose X takes on the values 1 and
2. Given X = 1, Y is either 2 or 1/2, with probability 1/2 each. If X = 2,
then Y is either 4 or 1/4, with probability 1/2 each. Let U = log2 Y .
Let µY and µU denote the regression functions of Y and U on X. Then

µU (1) = 0.5 · 1 + 0.5 · (−1) = 0 (3.25)

and similarly µU (2) = 0 as well. So, there is no relation between U and

X at all! Yet the relation between Y and X is quite substantial. The
transformation has destroyed the latter relation.
3.4. FURTHER READING 139

Of course, this example is highly contrived, and one can construct examples
with the opposite effect. Nevertheless, it shows that a log transformation
can indeed bring about considerable distortion. This is to be expected in a
sense, since the log function flattens out as we move to the right. Indeed,
the U.S. Food and Drug Administration once recommended against using
transformations.3

3.3.8 The Verdict

While the examples here do not constitute a research study (the reader is
encouraged to try the code in other settings, simulated and real), an overall
theme is suggested.
In principle, WLS provides more efficient estimates and correct statistical
inference. What are the implications?
If our goal is Prediction, then forming correct standard errors is typically
of secondary interest, if at all. And unless there is really strong variation in
the proper weights, having efficient estimates is not so important. In other
words, for Prediction, OLS may be fine.
The picture changes if the goal is Description, in which case correct stan-
dard errors may be important. Variance-stabilizing transformations may
cause problems, and while one might estimate WLS weights via nonpara-
metric regression methods as mentioned above, these may be too sensitive
to sampling variation. But the method of Section 3.3.3 is now commonly
available in statistical software packages, and is likely to be the best way
to cope with heteroscedasticity.

3.4 Further Reading

The MovieLens dataset (Section 3.2.4) is a very popular example for work
in the field of recommender systems, in which user ratings of items are
predicted. A comprehensive treatment is available in [1].

3 Quoted in The Log Transformation Is Special, Statistics in Medicine, Oliver Keene,

1995, 811-819. That author takes the opposite point of view.
140 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

3.5 Computational Complements

3.5.1 The R merge() Function

The MovieLens data set in Section 3.2.4 is interesting in that it consists of

three main files, not one. We’ve already discussed the first two files; the
third, u.item, contains data on the movies (name, release date, genre).
In Section 3.2.4, we had not collapsed the u.data file by taking averages
of movie ratings across each user. Had we wanted to retain the ability to
distinguish between movies, paying attention to genre and so on, we would
have needed to combine the u.data and u.item files. How can that be
done?
The key point is that the two files have a column in common, the user ID.
R’s merge() function can then be used to exploit this.
The merge() function does as its name implies, merging two data frames.
In the simple case, there must be a column having the same name in each
of the two input data frames. Let’s call that column k. Then whenever
that column has the same value in row i of one of the input data frames
and row j of the other, we form an output row that concatenates those
two rows. (Of course, the common column is not duplicated.) In computer
science, this is called a join operation.
The code for the movie join is seen here:
> ud <− read . table ( ’ u . data ’ , h e a d e r=F , s e p= ’ \ t ’ )
> uu <− read . table ( ’ u . u s e r ’ , h e a d e r=F , s e p= ’ | ’ )
> udu <− merge( ud , uu , by . x=1,by . y=1)
> head ( udu )
V1 V2 . x V3 . x V4 . x V2 . y V3 . y V4 . y V5
1 1 1 5 874965758 24 M t e c h n i c i a n 85711
2 1 23 4 875072895 24 M t e c h n i c i a n 85711
3 1 223 5 876892918 24 M t e c h n i c i a n 85711
4 1 171 5 889751711 24 M t e c h n i c i a n 85711
5 1 16 5 878543541 24 M t e c h n i c i a n 85711
6 1 73 3 876892774 24 M t e c h n i c i a n 85711

Here by.x and by.y specify the position of the common column within the
two input data frames.
The first four columns of the output data frame are those of ud, while the
last four come from uu. The latter actually has five columns, but the first
of them is the user ID, the column of commonality between uu and ud.
3.6. MATHEMATICAL COMPLEMENTS 141

Using this merged data frame, we could now do various analyses involving
user-movie interactions, with demographic variables as covariates. Note,
though, that the people who assembled the MovieLens dataset were wise
to make three separate files, because one unified file would have lots of
redundant information, and thus would take up much more room. This
may not be an issue with this 100K version of the data, but MovieLens also
has a version of size 20 million.

3.6 Mathematical Complements

3.6.1 The Delta Method

Say we have an estimator θb of a scalar population quantity θ, such that

the estimator is asymptotically normally distributed. Denote the asymp-
totic variance of θb by AV ar(θ),
b meaning that the cumulative distribution
√ b b as n → ∞.
function of n(θ − θ) converges to the cdf of N (0, AV ar(θ))
For some suitable function f , we form a new estimator,

b
W = f (θ) (3.26)

for the quantity f (θ). We will assume at first for convenience that θ is
scalar-valued. Then roughly speaking, the Delta Method gives us a way to
show that W is also asymptotically normal, and most importantly, provides
us with the asymptotic variance of W :

b
AV ar(W ) = [g(θ)]2 AV ar(θ) (3.27)

where g is the derivative of f .

A standard error for W is then computed by substituting θb for θ in (3.27),
and then taking the square root.
The method extends to the case of multivariate θ. Here (3.27) becomes

b
G′ ACov(θ)G (3.28)

where G is the gradient (column vector of partial derivatives) of f , evaluated

at θ.
142 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

An intuitive derivation of (3.27) is simple. We expand f into a Taylor series,

taking the first terms:

f (t) ≈ f (a) + g(a)(t − a) (3.29)

for t near a. Setting t = θb and a = θ, we have

W ≈ f (θ) + g(θ)(θb − θ) (3.30)

Taking variances of both sides, and recalling that θ is a constant, we obtain

(3.27). Equation (3.28) then follows the same reasoning, accompanied by
(2.85). We will stop with the intuition here and not go through the formal
details.

3.6.2 Distortion Due to Transformation

Consider this famous inequality:

Jensen’s Inequality: Suppose h is a convex function,4 and V is

a random variable for which the expected values in (3.31) exist.
Then

E[h(V )] ≥ h(EV ) (3.31)

In our context, h is our transformation in Section 3.3.7, and the E() are
conditional means, i.e., regression functions. In the case of the log transform
(and the square-root transform), h is concave-down, so the sense of the
inequality is reversed:

E[ln Y |X = t] ≤ ln(E(Y |X = t) (3.32)

Since equality will hold only in trivial cases, we see that the regression
function of ln Y will be smaller than the log of the regression function of
Y.
4 This is “concave up,” in the calculus sense.
3.7. EXERCISES: DATA, CODE AND MATH PROBLEMS 143

Say we assume that

E(Y |X = t) = eβ0 +β1 t (3.33)

and reason that this implies that a linear model would be reasonable for
ln Y :

E(ln Y |X = t) = β0 + β1 t (3.34)

Jensen’s Inequality tells us that such reasoning may be risky. In fact, if we

are in a substantially heteroscedastic setting (for Y , not ln Y ), the discrep-
ancy between the two sides of (3.32) could vary a lot with t, potentially
producing quite a bit of distortion to the shape of the regression curve. This
follows from a result of Robert Becker [12]. who expresses the difference
between the left- and right-hand sides of (3.31) in terms of V ar(V ).

3.7 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. This problem concerns the InstEval dataset in the lme4 package, which
is on CRAN [9]. Here users are students, and “items” are instructors, who
are rated by students. Perform an analysis similar to that in Section 3.2.4.
Note that you may need to do some data wrangling first, e.g., change R
factors to numeric type.
Mini-CRAN and other computational problems:
2. In Section 3.3.2, the standard errors for the OLS and WLS estimates
differed by about 8%. Using R’s pnorm() function, explore how much
impact this would have on true confidence level in a nominal 95% confidence
interval.
3. This problem concerns the bike-sharing data (Section 1.1).

(a) Check heteroscedasticity here, using the nonparametric regression ap-

proach of Section 3.3.2: Run knnest() twice, once to estimate µ(t)
and then to estimate σ 2 (t), evaluated for t = X1 , ..., Xn . Plot the
latter values against former ones.
144 CHAPTER 3. HOMOSCEDASTICITY ETC. IN PRACTICE

(b) Write general code for this, in the form of a function with call form
p l o t m u s i g 2 ( y , xdata , k )

where the arguments are as in knnest().

Math problems:
4. Consider a fixed-X regression setting (Section 2.3) with replication,
meaning that more than one Y is observed for each X value. (Here we
have just one predictor, p = 1.) So our data are

(Xi , Yi1 , ..., Yini ), i = 1, ..., r (3.35)

Assume that X > 0 and

V ar(Y | X = t) = X q σ 2 (3.36)

for known q.
Find an unbiased estimator for σ 2 .
5. (Presumes background in Maximum Likelihood estimation.) Assume
the setting of Problem 4, but with q unknown and to be estimated from
the data. The conditional distribution of Y given X is assumed normal
with mean β0 + β1 X. Also, for simplicity, assume n1 = ..., nr = m. Write
an R function with call form
lmq ( y , x )

with y being the r × m matrix of Y values and x being the vector of X

values. The code will estimate β, σ 2 and q by the method of Maximum
Likelihood, by calling the R function mle().
6. Suppose we have one predictor variable, but our response variable Y is
actually a vector, of length 2,

Y = (Y (1) , Y (2) )′ (3.37)

This gives us two regression functions,

µi (t) = E(Y (i) | X = t), i = 1, 2 (3.38)

3.7. DATA, CODE AND MATH PROBLEMS 145

If we use a linear model, our β coefficients now are doubly subscripted,

µi (t) = β0i + β1i t, i = 1, 2 (3.39)

Now suppose we are interested in the ratios µ2 (t)/µ1 (t), which we estimate
by

βb02 + βb12 t
(3.40)
βb01 + βb11 t

Use the Delta Method to derive the asympotic variance of this estimator.
Chapter 4

Generalized Linear and

Nonlinear Models

Consider our bike-sharing data (e.g., Section 1.1), which spans a time period
of several years. On the assumption that ridership trends are seasonal, and
that there is no other time trend (e.g., no long-term growth in the program),
then there would be a periodic relation between ridership R and G, the day
in our data; here G would take the values 1, 2, 3, ..., with the top value
being, say, 3 × 365 = 1095 for three consecutive years of data.1 Assuming
that we have no other predictors, we might try fitting the model with a
sine term:

mean R = β0 + β1 sin(2π · G/365) (4.1)

Just as adding a quadratic term didn’t change the linearity of our model in
Section 1.16.1 with respect to β, the model (4.1) is linear in β too. In the
notation of Section 2.4.2, as long as we can write our model as

mean D = A β (4.2)

then by definition the model is linear: Multiplying β by a constant changes

1 We’ll ignore the issue of leap years here, to keep things simple.

147
148 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

mean D by that constant. In the bike data model above, A would be

 
1 sin(2π · 1/365)
 1 sin(2π · 2/365) 
A=
 ...

 (4.3)
1 sin(2π · 1095/365)

But in this example, we have a known period, 365. In some other periodic
setting, the period might be unknown, and would need to be estimated
from our data. Our model might be, say,

mean Y = β0 + β1 sin(2π · X/β2 ) (4.4)

where β2 is the unknown period. This does not correspond to (4.2). The
model is still parametric, but is nonlinear.
Nonlinear parametric modeling, then, is the topic of this chapter. We’ll de-
velop procedures for computing least squares estimates, and forming con-
fidence intervals and p-values, again without assuming homoscedasticity.
The bulk of the chapter will be devoted to the Generalized Linear Model
(GLM), which is a widely-used broad class of nonlinear regression mod-
els. Two important special cases of the GLM will be the logistic model
introduced briefly in Section 1.1, and Poisson regression.

4.1 Example: Enzyme Kinetics Model

Data for the famous Michaelis-Menten enzyme kinetics model is available

in the nlstools package on CRAN [11]. For the data set vmkm, we predict
the reaction rate V from substrate concentration S. The model used was
suggested by theoretical considerations to be

β1 t
E(V | S = t) = (4.5)
β2 + t

In the second data set, vmkmki,2 an addiitonal predictor I, inhibitor

concentration, was added:
> data ( vmkmki )
2 There were 72 observation in this data, but the last 12 appear to be anomalous
(gradient 0 in all elements), and thus were excluded.
4.1. EXAMPLE: ENZYME KINETICS MODEL 149

> head ( vmkmki )

S I v
1 200 0 . 0 0 1 8 . 1
2 200 0 . 0 0 1 8 . 8
3 200 6 . 2 5 1 7 . 7
4 200 6 . 2 5 1 8 . 1
5 200 1 2 . 5 0 1 6 . 4
6 200 1 2 . 5 0 1 7 . 6

Their new model was

β1 t
E(V | S = t, I = u) = (4.6)
t + β2 (1 + u/β3 )

We’ll fit the model using R’s nls() function:

> library ( n l s t o o l s )
> data ( vmkmki )
> r e g f t n <− function ( t , u , b1 , b2 , b3 )
b1 ∗ t / ( t + b2 ∗ ( 1 + u/b3 ) )

All nonlinear least-squares algorithms are iterative: We make an initial

guess at the least-squares estimate, and from that, use the data to update
the guess. Then we update the update, and so on, iterating until the guesses
converge. In nls(), we specify the initial guess for the parameters, using the
start argument, an R list.3 Let’s set that up, and then run the analysis:
> b s t a r t <− l i s t ( b1=1,b2=1, b3=1)

The values 1 here were arbitrary, not informed guesses at all. Domain
expertise can be helpful.
> z <− n l s ( v ∼ r e g f t n ( S , I , b1 , b2 , b3 ) , data=vmkmki ,
s t a r t=l i s t ( b1=1,b2=1, b3 =1))
> z
N o n l i n e a r r e g r e s s i o n model
model : v ∼ r e g f t n ( S , I , b1 , b2 , b3 )
data : vmkmki
b1 b2 b3
18.06 15.21 22.28
r e s i d u a l sum−of −s q u a r e s : 1 7 7 . 3

3 This
also gives the code a chance to learn the names of the parameters, needed for
computation of derivatives.
150 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

Number o f i t e r a t i o n s t o c o n v e r g e n c e : 11
Achieved c o n v e r g e n c e t o l e r a n c e : 4 . 9 5 1 e −06

c1 = 18.06 etc.
So, β
We can apply summary(), coef() and vcov() to the output of nls(),
just as we did earlier with lm(). For example, here is the approximate
covariance matrix of the coefficient vector:
> vcov ( z )
b1 b2 b3
b1 0 . 4 7 8 6 7 7 6 1 . 3 7 4 9 6 1 0 . 8 9 3 0 4 3 1
b2 1 . 3 7 4 9 6 1 2 7 . 5 6 8 8 3 7 1 1 . 1 3 3 2 8 2 1
b3 0 . 8 9 3 0 4 3 1 1 1 . 1 3 3 2 8 2 2 9 . 1 3 6 3 3 6 6

This assumes homoscedasticity. Under that assumption, an approximate

95% confidence interval for β1 would be

√
18.06 ± 1.96 0.4786776 (4.7)

One can use the approach in Section 3.3.3 to adapt nls() to the het-
eroscedastic case, and we will do so in Section 4.5.2.

4.2 The Generalized Linear Model (GLM)

GLMs are generalized versions of linear models, in the sense that, although
the regression function µ(t) is of the form t′ β, it is some function of a linear
function of t′ β.

4.2.1 Definition

To motivate GLMs, first recall again the logistic model, introduced in

(1.36). We are dealing with a classification problem, so the Y takes on
the values 0 and 1. Let X = (X (1) , X (2) , ..., X (p) )′ denote the vector of our
predictor variables.
Our model is

µ(t) = E(Y | X (1) = t1 , ..., X (p)) = tp ) = q(β0 + β1 t1 + ... + βp tp ) (4.8)

4.2. THE GENERALIZED LINEAR MODEL (GLM) 151

where

1
q(s) = (4.9)
1 + e−s

and t = (t1 , ..., tp )′ .4

The key point is that even though the right-hand side of (4.8) is not linear
in t, it is a function of a linear expression in β, hence the term generalized
linear model (GLM).
So, GLMs actually form a broad class of models. We can use many different
functions q() instead of the one in (4.9); for each such function, we have
a different GLM. The particular GLM we wish to use is specified by the
inverse of the function q(t), called the link function.
In addition, GLM assumes a specified parametric class for the conditional
distribution of Y given X, which we will denote FY |X . In the logistic
case, the assumption is trivially that the distribution is Bernoulli, i.e.,
binomial with number of trials equal to 1. Having such assumptions enables
Maximum Likelihood estimation (Section 4.8.1).

4.2.2 Poisson Regression

As another example, the Poisson regression model assumes that FY |X is a

Poisson distribution and sets q() = exp(), i.e.,

µ(t) = E(Y | X (1) = t1 , ..., X (p)) = tp ) = eβ0 +β1 t1 +...+βp tp (4.10)

Why the use of exp()? The model for the most part makes sense without
the exponentiation, i.e.,

µ(t) = E(Y | X (1) = t1 , ..., X (p)) = tp ) = β0 + β1 t1 + ... + βp tp (4.11)

But many analysts hesitate to use (4.11) as the model, as it may generate
negative values. They view that as a problem, since Y ≥ 0 (and P (Y >
0) > 0), so we have µ(t) > 0. The use of exp() in (4.10) meets that
objection.
4 Recall from Section 1.17.1 that the classification problem is a special case of regres-
sion.
152 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

Many who work with Poisson distributions find the Poisson relation

σ 2 (t) = µ(t) (4.12)

to be overly restrictive. They would like to use a model that is “Poisson-

like” but with > instead of = in (4.12),

σ 2 (t) > µ(t) (4.13)

a condition called overdispersion. This arises if we have a mixture of Poisson

distributions, i.e., the variable in question has a Poisson distribution in
various subpopulations of our population (Exercise 4).
One such model is that of the negative binomial. Though this distribution
is typically presented as the number of Bernoulli trials to attain r successes,
in this context we simply use it as a general distribution model. It can be
shown that a gamma-distributed mixture of Poisson random variables has
a negative binomail distribution.
The function glm.nb in the MASS package performs a negative binomial
regression analysis.

4.2.3 Exponential Families

In general, the core of GLM assumes that FY |X belongs to an exponen-

tial family [136]. This is formally defined as a parametric family whose
probability density/mass function has the form

exp[η(θ)T (x) − A(θ) + B(x)] (4.14)

where θ is the parameter vector and x is a value taken on by the random

variable. Though this may seem imposing, it suffices to say that the above
formulation includes many familiar distribution families such as Bernoulli,
binomial, Poisson, exponential and normal. In the Poisson case, for in-
stance, setting η(θ) = log λ, T (x) = x, A(θ) = −λ and B(x) = − log(x!)
yields the expression

e−λ λx
(4.15)
x!
the famous form of the Poisson probability mass function.
4.2. THE GENERALIZED LINEAR MODEL (GLM) 153

As previously noted, GLM terminology centers around the link function,

which is the functional inverse of our function q() above. For Poisson re-
gression, the link function is the inverse of exp(), i.e., log(). For logit, set
u = q(s) = (1 + exp(−s))−1 , and solve for s, giving us the link function,

u
link(u) = ln (4.16)
1−u

4.2.4 R’s glm() Function

Of course, the glm() function does all this for us. For ordinary usage, the
call is the same as for lm(), except for one extra argument, family. In the
Poisson regression case, for example, the call looks like
glm( y ∼ x , family = poisson )

The family argument actually has its own online help entry:
> ? family
family package : s t a t s
R Documentation

Family O b j e c t s for Models

Description :
...

Usage :

family ( o b j e c t , . . . )

binomial ( l i n k = ” l o g i t ” )
gaussian ( l i n k = ” i d e n t i t y ” )
Gamma( l i n k = ” i n v e r s e ” )
inverse . gaussian ( l i n k = ” 1/muˆ2 ” )
poisson ( l i n k = ” l o g ” )
quasi ( l i n k = ” i d e n t i t y ” , v a r i a n c e = ” c o n s t a n t ” )
quasibinomial ( link = ” l o g i t ” )
q u a s i p o i s s o n ( link = ” log ” )

...
154 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

Ah, so the family argument is a function! There are built-in ones we can
use, such as the poisson one we used above, or a user could define her own
custom function.
Well, then, what are the arguments to such a function? A key argument is
link, which is obviously the link function q −1 () discussed above, which we
found to be log() in the Poisson case.
For a logistic model, as noted earlier, FY |X is binomial with number of
trials m equal to 1. Recall that the variance of a binomial random variable
with m trials is mr(1−r), where r is the “success” probability on each trial,
Recall too that the mean of a 0-1-valued random variable is the probability
of a 1. Putting all this together, we have

σ 2 (t) = µ(t)[1 − µ(t)] (4.17)

Sure enough, this appears in the code of the built-in function binomial():

> binomial
function ( l i n k = ” l o g i t ” )
{
...
v a r i a n c e <− function (mu) mu ∗ ( 1 − mu)

Let’s now turn to details of two of the most widely-used models, the logistic
and Poisson.

4.3 GLM: the Logistic Model

The logistic regression model, introduced in Section 1.1, is by far the most
popular nonlinear regression method. Here we are predicting a response
variable Y that takes on the values 1 and 0, indicating which of two classes
our unit belongs to. As we saw in Section 1.17.1, this indeed is a regression
situation, as E(Y | X = t) reduces to P (Y = 1 | X = t).
The model, again, is

1
P (Y = 1 | X = (t1 , ..., tp )) = (4.18)
1 + e−(β0 +β1 t1 +....+βp tp )
4.3. GLM: THE LOGISTIC MODEL 155

4.3.1 Motivation

We noted in Section 1.1 that the logistic model is appealing for two reasons:
(a) It takes values in [0,1], as a model for probabilities should, and (b) it
is monotone in the predictor variables, as in the case of a linear model, a
common situation in practice.
But there’s even more reason to choose the logistic model. It turns out
that the logistic model is implied by many familiar distribution families. In
other words, there is often good theoretical justification for using the logit.
To illustrate that, consider a very simple example of text classification,
involving Twitter tweets. Suppose we wish to automatically classify tweets
into those involving financial issues and all others. We’ll do that by having
our code check whether a tweet contains words from a list of financial terms
we’ve set up for this purpose, say bank, rate and so on.
Here Y is 1 or 0, for the financial and nonfinancial classes, and X is the
number of occurrences of terms from the list. Suppose that from past
data we know that among financial tweets, the number of occurrences of
words from this list has a Poisson distribution with mean 1.8, while for
nonfinancial tweets the mean is 0.2. Mathematically, that says that FX|Y =1
is Poisson with mean 1.8, and FX|Y =0 is Poisson with mean 0.2. (Be sure to
distinguish the situation here, in which FX|Y is a Poisson distribution, from
Poisson regression, in which it is assumed that FY |X is Poisson.) Finally,
suppose 5% of all tweets are financial.
Recall once again (Section 1.17.1) that in the classification case, our regres-
sion function takes the form

µ(t) = P (Y = 1 | X = t) (4.19)

Let’s calculate this function:

P (Y = 1 and X = t)
P (Y = 1 | X = t) = (4.20)
P X = t)
P (Y = 1 and X = t)
=
P (Y = 1 and X = t or Y = 0 and X = t)
π P (X = t | Y = 1)
=
π P (X = t | Y = 1) + (1 − π) P (X = t | Y = 0)

where π = P (Y = 1) is the population proportion of individuals in class 1.

156 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

The numerator in (4.20) is

e−1.8 1.8t
0.05 · (4.21)
t!

and similarly the denominator is

e−1.8 1.8t e−0.2 0.2t

0.05 · + 0.95 · (4.22)
t! t!

Putting this into (4.20) and simplifying, we get

1
P (Y = 1 | X = t) = 1 t (4.23)
1 + 19e1.6( 9 )
1
= (4.24)
1 + exp(log 19 + 1.6 − t log 9)

That last expression is of the form

1
(4.25)
1 + exp[−(β0 + β1 t)]

with

β0 = − log 19 − 1.6 (4.26)

and

β1 = log 9 (4.27)

In other words the setting in which FX|Y is Poisson implies the logistic
model!
This is true too if FX|Y is an exponential distribution. Since this is a
continuous distribution family rather than a discrete one, the quantities
P (X = t|Y = i) in (4.23) must be replaced by density values:

P (Y = 1 | X = t) =
4.3. GLM: THE LOGISTIC MODEL 157

π f1 (X = t | Y = 1)
(4.28)
π f1 (X = t | Y = 1) + (1 − π) f0 (X = t | Y = 0)

where the within-class densities of X are

fi (w) = λi e−λi w , i = 0, 1 (4.29)

After simplifying, we again obtain a logistic form.

Most important, consider the multivariate normal case (Section 2.6.2): Say
for groups i = 0, 1, FX|Y =i is a multivariate normal distribution with mean
vector µi and covariance matrix Σ, where the latter does not have a sub-
script i. This is a generalization of the classical two-sample t-test setting,
in which two (scalar) populations are assumed to have possibly different
means but the same variance.5 Again using (4.28), and going through a lot
of algebra, we find that again P (Y = 1 | X = t) turns out to have a logistic
form,

1
P (Y = 1 | X = t) = ′ (4.30)
1+ e−(β0 +β t)

with

1
β0 = log(1 − π) − log π + (µ′1 µ1 − µ′0 µ0 ) (4.31)
2

and

β = (µ0 − µ1 )′ Σ−1 (4.32)

where t is the vector of predictor variables, the β vector is broken down into
(β0 , β), and π is P (Y = 1). The messy form of the coefficients here is not
important; instead, the point is that we find that the multivariate normal
model implies the logistic model, giving the latter even more justification.
In summary:

Not only is the logistic model intuitively appealing because it is

a monotonic function with values in (0,1), but also because it
5 Itis also the setting for Fisher’s Linear Discriminant Analysis, to be discussed in
Section 5.6.
158 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

is implied by various familiar parametric models for the within-

class distribution of X.

No wonder the logit model is so popular!

4.3.2 Example: Pima Diabetes Data

Another famous UCI data set is from a study of the Pima tribe of Native
Americans, involving factors associated with diabetes. There is data on 768
women.6 Let’s predict diabetes from the other variables:
> pima <− read . csv ( ’ pima−i n d i a n s −d i a b e t e s . data ’ )
> head ( pima )
NPreg Gluc BP Thick I n s u l BMI Genet Age Diab
1 6 148 72 35 0 3 3 . 6 0 . 6 2 7 50 1
2 1 85 66 29 0 2 6 . 6 0 . 3 5 1 31 0
3 8 183 64 0 0 2 3 . 3 0 . 6 7 2 32 1
4 1 89 66 23 94 2 8 . 1 0 . 1 6 7 21 0
5 0 137 40 35 168 4 3 . 1 2 . 2 8 8 33 1
6 5 116 74 0 0 2 5 . 6 0 . 2 0 1 30 0
# Diab = 1 means has d i a b e t e s
> l o g i t o u t <− glm( Diab ∼ . , data=pima , family=binomial )
> summary( l o g i t o u t )
...
Coefficients :
Estimate Std . E r r o r z v a l u e
( I n t e r c e p t ) −8.4046964 0 . 7 1 6 6 3 5 9 −11.728
NPreg 0.1231823 0.0320776 3.840
Gluc 0.0351637 0.0037087 9.481
BP −0.0132955 0 . 0 0 5 2 3 3 6 −2.540
Thick 0.0006190 0.0068994 0.090
Insul −0.0011917 0 . 0 0 0 9 0 1 2 −1.322
BMI 0.0897010 0.0150876 5.945
Genet 0.9451797 0.2991475 3.160
Age 0.0148690 0.0093348 1.593
Pr ( >| z | )
( I n t e r c e p t ) < 2 e −16 ∗∗∗
NPreg 0 . 0 0 0 1 2 3 ∗∗∗
Gluc < 2 e −16 ∗∗∗
BP 0.011072 ∗
6 The data set is at https://archive.ics.uci.edu/ml/datasets/Pima+Indians+Diabetes.
I have added a header record to the file.
4.3. GLM: THE LOGISTIC MODEL 159

Thick 0.928515
Insul 0.186065
BMI 2 . 7 6 e −09 ∗∗∗
Genet 0 . 0 0 1 5 8 0 ∗∗
Age 0.111192
...

4.3.3 Interpretation of Coefficients

In nonlinear regression models, the parameters βi do not have the simple

marginal interpretation they enjoy in the linear case. Statements like we
made in Section 1.9.1.2, “We estimate that, on average, each extra year of
age corresponds to almost a pound in extra weight,” are not possible here.
However, in the nonlinear case, the regression function is still defined as
the conditional mean, which in the logit case reduces to the conditional
probability of a 1. Practical interpretation is definitely still possible, if
slightly less convenient.
Consider for example the estimated Glucose coefficient in our diabetes data
above, 0.035. Let’s apply that to the people similar to the first person in
the data set:
> pima [ 1 , ]
NPreg Gluc BP Thick I n s u l BMI Genet Age Diab
1 6 148 72 35 0 3 3 . 6 0 . 6 2 7 50 1

Ignore the fact that this woman has diabetes. Let’s consider the subpopu-
lation of all women with the same characteristics as this one, i.e., all who
have had 6 pregnancies, a glucose level of 148 and so on, through an age of
50. The estimated proportion of women with diabetes in this subpopulation
is

1
(4.33)
1+ e−(8.4047+0.1232·6+...+0.0149·50)

We don’t have to plug these numbers in by hand, of course:

> l <− function ( t ) 1/(1+exp(−t ) )
> pima1 <− u n l i s t ( pima [ 1 , − 9 ] )
> l ( coef ( l o g i t o u t ) %∗% c ( 1 , pima1 ) )
[ ,1]
[ 1 , ] 0.7217266
160 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

Note that pima[1,-9] is actually a data frame (having been derived from
a data frame), so in order to multiply it, we needed to make a vector out
of it, using unlist().
So, we estimate that about 72% of women in this subpopulation have dia-
betes. But what about the subpopulation of the same characteristics, but
of age 40 instead of 50?

> w <− pima1

> w[ ’ Age ’ ] <− 40
> l ( coef ( l o g i t o u t ) %∗% c ( 1 ,w) )
[ ,1]
[ 1 , ] 0.6909047

Only about 69% of the younger women have diabetes.

So, there is an effect of age on developing diabetes, but only a mild one; a
10-year increase in age only increased the chance of diabetes by about 3.1%.
However, note carefully that this was for women having a given set of the
other factors, e.g., 6 pregnancies. Let’s look at a different subpopulation,
those with 2 pregnancies and a glucose level of 120, comparing 40- and
50-year-olds:

> u <− pima1

> u [ 1 ] <− 2
> u [ 2 ] <− 100
> v <− u
> v [ 8 ] <− 40
> l ( coef ( l o g i t o u t ) %∗% c ( 1 , u ) )
[ ,1]
[ 1 , ] 0.2266113
> l ( coef ( l o g i t o u t ) %∗% c ( 1 , v ) )
[ ,1]
[ 1 , ] 0.2016143

So here, the 10-year age effect was somewhat less, about 2.5%. A more
careful analysis would involve calculating standard errors for these numbers,
but the chief point here is that the effect of a factor in nonlinear situations
depends on the values of the other factors.

P (Y = 0 | X (1) = t1 , ..., X (p)) = tp ) = 1 − ℓ(β0 + β1 t1 + ... + βp tp ) (4.34)

4.3. GLM: THE LOGISTIC MODEL 161

Some analysts like to look at the log-odds ratio,

P (Y = 1 | X (1) = t1 , ..., X (p)) = tp )

log (4.35)
P (Y = 0 | X (1) = t1 , ..., X (p)) = tp )

in this case the logaritihm of the ratio of the probability of having and not
having the disease. By Equation (4.8), this simplifies to

β0 + β1 t1 + ... + βp tp (4.36)

a linear function. Thus, in interpreting the coefficients output from a lo-

gistic analysis, it is convenient to look at this log-odds ratio, as it gives us
a single marginal-effect number for each factor. This may be sufficient for
the application at hand, but a more thorough analysis should consider the
effects of the factors on the probabilities themselves.

4.3.4 The predict() Function Again

In the previous section, we evaluated the estimated regression function (and

thus predicted values as well) the straightforward but messy way, e.g.,
> l ( coef ( l o g i t o u t ) %∗% c ( 1 , v ) )

The easy way is to use R’s predict() function:

> predict ( o b j e c t=l o g i t o u t , newdata=pima [ 1 , − 9 ] ,
type= ’ r e s p o n s e ’ )
1
0.7217266

We saw that in Section 1.10.3 for objects of ’lm’ class. But in our case
here, we invoked it on logitout. What is the class of that object?
> class ( l o g i t o u t )
[ 1 ] ”glm” ”lm”

So, it is an object of class ’glm’, and, we see, the latter is a subclass of the
class ’lm’. For that subclass, the predict() function, i.e., predict.glm(),
there is an extra argument (actually several), type. The value of that
argument that we want here is type = ’response’, alluding to the fact
that we want a prediction on the scale of the response variable, Y .
162 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

4.3.5 Overall Prediction Accuracy

How well can we predict in the Pima example above? For the best measure,
we should use cross validation or something similar, but we can obtain a
quick measure as follows.
The value returned by glm() has class ’glm’, which is actually a subclass of
’lm’. The latter, and thus the former, includes a component fitted.values,
the ith of which is

b(Xi )
µ (4.37)

i.e., the estimated value of the regression function at observation i. If we

did not know Yi , we would predict it to be 1 or 0, depending on whether
b(Xi ) is greater than or less than 0.5. In R terms, that predicted value is
µ
simply

round ( l o g i t o u t $ f i t t e d . v a l u e s [ i ] )

Using the fact that the proportion of 1s in a vector of 1s and 0s is simply the
mean value in that vector, we have that the overall probability of correct
classification is

> mean( pima$Diab == round ( l o g i t o u t $ f i t t e d . v a l u e s ) )

0.7825521

That seems pretty good (though again, it is biased upward and cross val-
idation would give us a better estimate), but we must compare it against
how well we would do without the covariates. We reason as follows. First,

> mean( pima$Diab )

[ 1 ] 0.3489583

Most of the women do not have diabetes, so our strategy, lacking covariate
information, would be to always guess that Y = 0. We will be correct a
proportion

> 1 − 0.3489583
[ 1 ] 0.6510417

of the time. Thus our 78% accuracy using covariates does seem to be an
improvement.
4.3. GLM: THE LOGISTIC MODEL 163

4.3.6 Example: Predicting Spam E-mail

One application of these methods is text classification. In our example here,

the goal is machine prediction of whether an incoming e-mail message is
span, i.e., unwanted mail, typically ads.
We’ll use the spam dataset from the UCI Machine Learning Data Repos-
itory. It is also available from the CRAN package ElemStatLearn [64],
which we will use here, but note that the UCI version includes a word list.
It has data on 4601 e-mail messages, with 57 predictors. The first 48 of
those predictors consist of frequencies of 48 words. Thus the first column,
for instance, consists of the proportions of Word 1 in each of the 4601 mes-
sages, with the total number of words in a message as the base in each
case. The remaining predictors involve measures such as the numbers of
consecutive capital letters in words.
The last column is an R factor with levels spam and e-mail. This R type
is explained in Section 4.7.2, and though glm() can handle such variables,
for pedagogical reasons, let’s use dummies for a while. (We will begin using
factors directly in Section 5.6.3.)
Let’s fit a logistic model.

> l i b r a r y ( ElemStatLearn )
> data ( spam )
> spam$spam <− as . integer ( spam$spam == ’ spam ’ )
> glmout <− glm( spam ∼ . , data=spam ,
family=binomial )
> summary( glmout )
...

Coefficients :
Estimate Std . E r r o r z v a l u e
( I n t e r c e p t ) −1.569 e+00 1 . 4 2 0 e −01 −11.044
A. 1 −3.895 e −01 2 . 3 1 5 e −01 −1.683
A. 2 −1.458 e −01 6 . 9 2 8 e −02 −2.104
A. 3 1 . 1 4 1 e −01 1 . 1 0 3 e −01 1.035
A. 4 2 . 2 5 2 e+00 1 . 5 0 7 e+00 1.494
A. 5 5 . 6 2 4 e −01 1 . 0 1 8 e −01 5.524
A. 6 8 . 8 3 0 e −01 2 . 4 9 8 e −01 3.534
A. 7 2 . 2 7 9 e+00 3 . 3 2 8 e −01 6.846
A. 8 5 . 6 9 6 e −01 1 . 6 8 2 e −01 3.387
...
Pr ( >| z | )
164 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

( Intercept ) < 2 e −16 ∗∗∗

A. 1 0.092388 .
A. 2 0.035362 ∗
A. 3 0.300759
A. 4 0.135168
A. 5 3 . 3 1 e −08 ∗∗∗
A. 6 0.000409 ∗∗∗
A. 7 7 . 5 7 e −12 ∗∗∗
A. 8 0.000707 ∗∗∗
...

Let’s see how accurately we can predict with this model:

> mean( spam$spam == round ( glmout$ f i t t e d . v a l u e s ) )

[ 1 ] 0.9313193

Not bad at all. But much as we are annoyed by spam, we hope that a
genuine message would not be likely to be culled out by our spam filter.
Let’s check:

> spamnot <− which ( spam$spam == 0 )

> mean( round ( glmout$ f i t t e d . v a l u e s [ spamnot ] ) == 0 )
[ 1 ] 0.956241

So if a message is real, it will have a 95% chance of getting past the spam
filter.

4.3.7 Linear Boundary

In (4.18), which values of t = (t1 , ..., tp )′ will cause us to gues Y = 1 and

which will result in a guess of Y = 0? The boundary occurs when (4.18)
has the value 0.5. In other words, the boundary consists of all t such that

β0 + β1 t1 + .... + βp tp = 0 (4.38)

So, the boundary has linear form, a hyperplane in p-dimensional space. This
may seem somewhat abstract now, but it will have value later on.
4.4. GLM: THE POISSON REGRESSION MODEL 165

4.4 GLM: the Poisson Regression Model

Since in the Pima data (Section 4.3.2) the number of pregnancies is a count,
we might consider predicting it using Poisson regression.7 Here’s how we
can do this with glm():

> p o i s o u t <− glm( NPreg ∼ . , data=pima , family=poisson )

> summary( p o i s o u t )
...
Coefficients :
Estimate Std . E r r o r z v a l u e
( Intercept ) 0.2963661 0.1207149 2.455
Gluc −0.0015080 0 . 0 0 0 6 7 0 4 −2.249
BP 0.0011986 0.0010512 1.140
Thick 0.0000732 0.0013281 0.055
Insul −0.0003745 0 . 0 0 0 1 8 9 4 −1.977
BMI −0.0002781 0 . 0 0 2 7 3 3 5 −0.102
Genet −0.1664164 0 . 0 6 0 6 3 6 4 −2.744
Age 0.0319994 0.0014650 21.843
Diab 0.2931233 0.0429765 6.821
Pr ( >| z | )
( Intercept ) 0.01408 ∗
Gluc 0.02450 ∗
BP 0.25419
Thick 0.95604
Insul 0.04801 ∗
BMI 0.91896
Genet 0 . 0 0 6 0 6 ∗∗
Age < 2 e −16 ∗∗∗
Diab 9 . 0 7 e −12 ∗∗∗
...

On the other hand, even if we believe that our count data follow a Poisson
distribution, there is no law dictating that we use Poisson regression, i.e.,
the model (4.10). As mentioned following that equation, the main motiva-
tion for using exp() in that model is to ensure that our regression function
is nonnegative, conforming to the nonnegative nature of Poisson random
variables. This is not unreasonable, but as noted in a somewhat different
context in Section 3.3.7, transformations — in this case, the use of exp()
— can produce distortions. Let’s try the “unorthodox” model, (4.11):
7 It
may seem unnatural to predict this, but as noted before, predicting any variable
may be useful if data on that variable may be missing.
166 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

> q u a s i o u t <− glm( NPreg ∼ . , data=pima ,

family=quasi ( v a r i a n c e=”muˆ2 ” ) , s t a r t=rep ( 1 , 9 ) )

This “quasi” family is a catch-all option, specifying a linear model but here
allowing us to specify a Poisson variance function:

V ar(Y | X = t) = [µ(t)]2 (4.39)

with µ(t) = t′ β. This is (4.11), not the standard Poisson regression model,
but worth trying anyway.
Well, then, which model performed better? As a rough, quick look, ignoring
issues of overfitting and the like, let’s consider R2 . This quantity is not
calculated by glm(), but recall from Section 2.9.2 that R2 is the squared
correlation between the predicted and actual Y values. This quantity makes
sense for any regression situation, so let’s calculate it here:
> cor ( p o i s o u t $ f i t t e d . v a l u e s , p o i s o u t $y ) ˆ 2
[ 1 ] 0.2314203
> cor ( q u a s i o u t $ f i t t e d . v a l u e s , q u a s i o u t $y ) ˆ 2
[ 1 ] 0.3008466

The “unorthodox” model performed better than the “official” one! We

cannot generalize from this, but it does show again that one must use
transformations carefully.

4.5 Least-Squares Computation

A point made in Section 1.4 was that the regression function, i.e., the con-
ditional mean, is the optimal predictor function, minimizing mean squared
prediction error. This still holds in the nonlinear (and even nonparametric)
case. The problem is that in the nonlinear setting, the least-squares estima-
tor does not have a nice, closed-form solution like (2.28) for the linear case.
Let’s see how we can compute the solution through iterative approximation.

4.5.1 The Gauss-Newton Method

Denote the nonlinear model by

E(Y | X = t) = g(t, β) (4.40)

4.5. LEAST-SQUARES COMPUTATION 167

where both t and β are possibly vector-valued. In (4.5), for instance, t is a

scalar but β is a vector. The least-squares estimate βb is the value of b that
minimizes

∑
n
[Yi − g(Xi , b)]2 (4.41)
i=1

Many methods exist to minimize (4.41), most of which involve derivatives

with respect to b. (The reason for the plural derivatives is that there is a
partial derivative for each of the elements of b.)
The best intuitive explanation of derivative-based methods, which will also
prove useful in a somewhat different context later in this chapter, is to set
up a Taylor series approximation for g(Xi , b) (as in Section 3.6.1):

b + h(Xi , β)
g(Xi , b) ≈ g(Xi , β) b ′ (b − β)
b (4.42)

where h(Xi , b) is the derivative vector of g(Xi , b) with respect to b, and the
prime symbol, as usual, means matrix transpose (not a derivative). The
b is of course yet unknown, but let’s put that matter aside for
value of β,
now. Then (4.41) is approximately

∑
n
b + h(Xi , β)
[Yi − g(Xi , β) b ′ βb − h(Xi , β)
b ′ b]2 (4.43)
i=1

At iteration k we take our previous iteration bk−1 to be an approximation

b and make that substitution in (4.43), yielding
to β,

∑
n
[Yi − g(Xi , bk−1 ) + h(Xi , bk−1 )′ bk−1 − h(Xi , bk−1 )′ b]2 (4.44)
i=1

Our bk is then the value that minimizes (4.44) over all possible values of b.
But why is that minimization any easier than minimizing (4.41)? To see
why, write (4.44) as

∑
n
[Yi − g(Xi , bk−1 ) + h(Xi , bk−1 )′ bk−1 −h(Xi , bk−1 )′ b]2 (4.45)
| {z }
i=1
168 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

(4.45) (2.18)
Yi − g(Xi , bk−1 ) + h(Xi , bk−1 )′ bk−1 Yi
h(Xi , bk−1 ) X̃i

Table 4.1: Nonlinear/linear correspondences

This should look familiar. It has exactly the same form as (2.18), with the
correspondences shown in Table 4.1. In other words, what we have in (4.45)
is a linear regression problem!
In other words, we can find the minimizing b in (4.45) using lm(). There
is one small adjustment to be made, though. Recall that in (2.18), the
quantity Xei includes a 1 term (Section 2.1), i.e., the first column of A in
(2.19) consists of all 1s. That is not the case in Table 4.1 (second row, first
column), which we need to indicate in our lm() call. We can do this via
specifying “-1” in the formula part of the call (Section 2.4.5).
Another issue is the computation of h(). Instead of burdening the user with
this, it is typical to compute h() using numerical approximation, e.g., using
R’s numericDeriv() function or the numDeriv package [57].

4.5.2 Eicker-White Asymptotic Standard Errors

As noted, nls() assumes homoscedasticity, which generally is a poor as-

sumption (Section 2.6). It would be nice, then, to somehow apply the
Eicker-White method (Section 3.3.3), which is for linear models, to the
nonlinear case. Actually, it is remarkably easy to do that adaptation.
The key is to note the linear approximation (4.5.1). One way to look at
this is that it has already set things up for us to use the Delta Method,
which uses a linear approximation. Thus we can apply Eicker-White to the
lm() output, say using vcovHC(), as in Section 3.3.4.
Below is code along these lines. It requires the user to run nlsLM(), an
alternate version of nls() in the CRAN package minpack.lm [45].8
8 This version is needed here because it provides the intermediate quantities we need

from the computation. However, we will see in Section 4.5.4 that this version has other
important advantages as well.
4.5. LEAST-SQUARES COMPUTATION 169

li b r a r y ( minpack . lm)
li b r a r y ( sandwich )

# u s e s o u t p u t o f nlsLM ( ) o f t h e minpack . lm p a c k a g e
# t o g e t an a s y m p t o t i c c o v a r i a n c e m a t r i x w i t h o u t
# assuming h o m o s c e d a s t i c i t y

# arguments :
#
# n l s l m o u t : r e t u r n v a l u e from nlsLM ( )
#
# value : approximate covariance matrix f o r the
# e s t i m a t e d parameter v e c t o r

n l s v c o v h c <− function ( n l s l m o u t ) {
# n o t a t i o n : g ( t , b ) i s t h e r e g r e s s i o n model ,
# where x i s t h e v e c t o r o f v a r i a b l e s f o r a
# g i v e n o b s e r v a t i o n ; b i s t h e e s t i m a t e d parameter
# vector ; x i s the matrix of p r e d i c t o r v a l u e s
b <− coef ( n l s l m o u t )
m <− n l s l m o u t $m
# y − g:
resid <− m$ resid ( )
# row i o f hmat w i l l be d e r i v o f g ( x [ i , ] , b )
# with respect to b
hmat <− m$ g r a d i e n t ( )
# c a l c u l a t e t h e a r t i f i c i a l ” x ” and ” y ” o f
# the algorithm
f a k e x <− hmat
f a k e y <− resid + hmat %∗% b
# −1 means no c o n s t a n t term i n t h e model
lmout <− lm( f a k e y ∼ f a k e x − 1 )
vcovHC ( lmout )
}

In addition to nice convergence behavior, the advantage for us here of nl-

sLM() over nls() is that the former gives us access to the quantities we
need in (4.45), especially the matrix of h() values. We then apply lm() one
more time, to get an object of type ”lm”, needed by vcovHC().
Applying this to the enzyme data, we have
> nlsvcovhc ( z )
fakex1 fakex2 fakex3
170 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

fakex1 0.4708209 1.706591 2.410712

fakex2 1.7065910 10.394496 20.314688
fakex3 2.4107117 20.314688 53.086958

This is rather startling. Except for the estimated variance of βb1 , the esti-
mated variances and covariances from Eicker-White are much larger than
what nls() found under the assumption of homoscedasticity.
Of course, with only 60 observations, both of the estimated covariance
matrices must be “taken with a grain of salt.” So, let’s compare the two
approaches by performing a simulation. Here

1
E(Y | X = t) = (4.46)
t′ β

where t = (t1 , t2 )′ and β = (β1 , β2 )′ . We’ll take the components of X

to be independent and exponentially distributed with mean 1.0, with the
heteroscedasticity modeled as being such that the standard deviation of Y
given X is proportional to the regression function value at X. We’ll use as
a check the fact that, 90% of the time, a N(0,1) variable is less than 1.28
Here is the simulation code:
sim <− function ( n , n r e p s ) {
b <− 1 : 2
r e s <− r e p l i c a t e ( nreps , {
x <− matrix ( rexp ( 2 ∗n ) , ncol =2)
meany <− 1 / ( x %∗% b )
y <− meany + ( runif ( n ) − 0 . 5 ) ∗ meany
xy <− data . frame ( x , y )
n l o u t <− n l s (X3 ∼ 1 / ( b1∗X1+b2∗X2 ) ,
data=xy , s t a r t=l i s t ( b1 = 1 , b2 =1))
b <− coef ( n l o u t )
vc <− vcov ( n l o u t )
vchc <− n l s v c o v h c ( n l o u t )
z1 <− ( b [ 1 ] − 1 ) / sqrt ( vc [ 1 , 1 ] )
z2 <− ( b [ 1 ] − 1 ) / sqrt ( vchc [ 1 , 1 ] )
c ( z1 , z2 )
})
print (mean( r e s [ 1 , ] < 1 . 2 8 ) )
print (mean( r e s [ 2 , ] < 1 . 2 8 ) )
}

And here is a run of the code:

4.5. LEAST-SQUARES COMPUTATION 171

> sim ( 2 5 0 , 2 5 0 0 )
[ 1 ] 0.6188
[ 1 ] 0.9096

That’s quite a difference! Eicker-White worked well, whereas assuming

homoscedasticity fared quite poorly. (Similar results were obtained for
n = 100.)

4.5.3 Example: Bike Sharing Data

In our bike-sharing data (Section 1.1), there are two kinds of riders, reg-
istered and casual. We may be interested in factors determining the mix,
i.e.,

registered
(4.47)
registered + casual

Since the mix proportion is between 0 and 1, we might try the logistic
model, introduced in (1.36) in the context of classification. Note, though,
that the example here does not involve a classification problem. so we
should not reflexively use glm() as before. Indeed, that function not only
differs from our current situation in that here Y takes on values in [0,1]
rather than in {0,1}, but also glm() assumes

V ar(Y | X =) = µ(t)(1 − µ(t)) (4.48)

(as implied by Y being in {0,1}), which we have no basis for assuming

here. Thus use of glm(), at least in the form we have seen so far, would
be inappropriate. Here are the results:
> s h a r <− read . csv ( ” day . c s v ” , h e a d e r=T)
> s h a r $temp2 <− s h a r $temp ˆ2
> s h a r $summer <− as . integer ( s h a r $ s e a s o n == 3 )
> s h a r $ p r o p r e g <− s h a r $ r e g / ( s h a r $ r e g+s h a r $ c n t )
> names( s h a r ) [ 1 5 ] <− ” r e g ”
> l i b r a r y ( minpack . lm)
> l o g i t <− function ( t1 , t2 , t3 , t4 , b0 , b1 , b2 , b3 , b4 )
1 / ( 1 + exp(−b0 − b1∗ t 1 −b2∗ t 2 −b3∗ t 3 −b4∗ t 4 ) )
> z <− nlsLM ( p r o p r e g ∼
l o g i t ( temp , temp2 , workingday , summer , b0 , b1 , b2 , b3 , b4 ) ,
data=shar , s t a r t=l i s t ( b0=1,b1=1,b2=1,b3=1,b4 =1))
172 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

> summary( z )
...
Parameters :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
b0 −0.083417 0 . 0 2 0 8 1 4 −4.008 6 . 7 6 e −05 ∗∗∗
b1 −0.876605 0 . 0 9 3 7 7 3 −9.348 < 2 e −16 ∗∗∗
b2 0 . 5 6 3 7 5 9 0.100890 5 . 5 8 8 3 . 2 5 e −08 ∗∗∗
b3 0 . 2 2 7 0 1 1 0 . 0 0 6 1 0 6 3 7 . 1 8 0 < 2 e −16 ∗∗∗
b4 0 . 0 1 2 6 4 1 0.009892 1.278 0.202
...

As expected, on working days, the proportion of registered riders is higher,

as we are dealing with the commute crowd on those days. On the other
hand, the proportion doesn’t seem to be much different during the sum-
mer, even though the vacationers would presumably add to the casual-rider
count.
But are those standard errors trustworthy? Let’s look at the Eicker-White
versions:
> sqrt ( diag ( n l s v c o v h c ( z ) ) )
fakex1 fakex2 fakex3 fakex4
0.021936045 0.090544374 0.092647403 0.007766202
fakex5
0.007798938

Again, we see some substantial differences.

4.5.4 The “Elephant in the Room”: Convergence

Issues

So far we have sidestepped the fact that any iterative method runs the risk
of nonconvergence. Or it might converge to some point at which there is
only a local minimum, not the global one — worse than nonconvergence,
in the sense that the user might be unaware of the situation.
For this reason, it is best to try multiple, diverse sets of starting values.
In addition, there are refinements of the Gauss-Newton method that have
better convergence behavior, such as the Levenberg-Marquardt method.
Gauss-Newton sometimes has a tendency to “overshoot,” producing too
large an increment in b from one iteration to the next. Levenberg-Marquardt
generates smaller increments. Interestingly it is a forerunner of ridge re-
4.6. FURTHER READING 173

gression that we’ll discuss in Chapter 8. It is implemented in the CRAN

package minpack.lm, which we used earlier in this chapter.

4.6 Further Reading

For more on the Generalized Linear Model, see for instance [47] [3] [41].
Exponential random graph models use logistic regression and similar tech-
niques to analyze relations between nodes in a network, say connections be-
tween friends in a group of people [60] [75]. The book by Luke [93] presents
various R tools for random graphs, and serves as a short introduction to
field.

4.7 Computational Complements

4.7.1 GLM Computation

Though estimation in GLM uses Maximum Likelihood, it can be shown

that the actual computation can be done by extending the ideas underlying
least-squares models. We use a weight function, which works as follows.
Let’s review Section 3.3.2, which discussed weighted least squares in the
case of a linear model. Using our usual notation µ(t) = E(Y | X = t) and
σ 2 (t) = V ar(Y | X = t), the optimal estimator of β is the value of b that
minimizes

∑
n
1 ′
fi b)2
(Yi − X (4.49)
fi )
σ 2 (X
i=1

Of course, generally σ 2 (t) is unknown, but it will be estimated.

Now consider the case of Poisson regression. One of the famous properties
of the Poisson distribution family is that the variance equals the mean.
Thus (4.43) becomes

∑
n
1
[Yi − g(Xi , b)]2 (4.50)
i=1
g(Xi , b)
174 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

Then (4.45) bceoms

∑
n
1
[Yi − g(Xi , bk−1 ) + h(Xi , bk−1 )′ bk−1 −h(Xi , bk−1 )′ b]2
i=1
g(Xi , bk−1 ) | {z }
(4.51)
and we again solve for the new iterate bk by calling lm(), this time making
use of the latter’s weights argument (Section 3.3.2).
We iterate as before, but now the weights are updated at each iteration
too. For that reason, the process is known as iteratively reweighted least
squares.

4.7.2 R Factors

To explain this R feature, let’s look at the famous iris dataset included in
the R package:
> head ( i r i s )
S e p a l . Length S e p a l . Width P e t a l . Length P e t a l . Width S p e c i e s
1 5.1 3.5 1.4 0.2 setosa
2 4.9 3.0 1.4 0.2 setosa
3 4.7 3.2 1.3 0.2 setosa
4 4.6 3.1 1.5 0.2 setosa
5 5.0 3.6 1.4 0.2 setosa
6 5.4 3.9 1.7 0.4 setosa
> i s <− i r i s $ S p e c i e s
> class ( is )
[1] ” factor ”
> str ( is )
F a c t o r w/ 3 l e v e l s ” s e t o s a ” , ” v e r s i c o l o r ” , . . :
1 1 1 1 1 1 1 1 1 1 ...
> table ( i s )
is
setosa versicolor virginica
50 50 50
> mode( i s )
[ 1 ] ” numeric ”
> levels ( is )
[ 1 ] ” setosa ” ” versicolor ” ” virginica ”
4.8. MATHEMATICAL COMPLEMENTS 175

We see that is is basically a numeric vector, with its first few values being
1, 1, 1. But these codes have names, known as levels, such as ’setosa’ for
the code 1.
In some cases, all that machinery actually gets in the way. If so, we can
convert to an ordinary vector, e.g.,
> s <− as . numeric ( i s )

4.8 Mathematical Complements

4.8.1 Maximum Likelihood Estimation

This is a method of statistical estimation. It is fairly generally applicable

to estimation in parametric models, not just in a GLM or regression con-
text, but let’s look at the latter for concreteness, making things even more
concrete by looking at a very simple example, n = 2.
Our model will be Poisson regression, (4.10), and say our data are (Y1 , X1 ) =
(3, 20) and (Y2 , X2 ) = (8, 16). Our estimation will be conditional on the
Xi , so they will be treated as constants.
The philosophy of MLE is to ask, “What values of the βi in (4.10) would
maximize the probability of our data occurring?” Using (4.10) and (4.15),
and the independence of our data, that probability is

e−λ1 λ31 e−λ2 λ82

· (4.52)
3! 8!
where

λ1 = eβ0 +β1 ·20 , λ2 = eβ0 +β1 ·16 (4.53)

So, to answer the question of maximizing the probability of our data, we

would maximize (4.52) with respect to β0 and β1 , keeping in mind (4.53).
The resulting values are called the Maximum Likelihood Estimators of the
βi .
MLEs form a key part of the theory of mathematical statistics. Under
certain conditions, they are optimal [136]. However, this is well beyond the
scope of this book.
176 CHAPTER 4. NONLINEAR/GENERALIZED LINEAR MODELS

4.9 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. Conduct a negative binomial regression analysis on the Pima data.
Compare to the results in Section 4.3.2.
2. Consider the spam example in Section 4.3.6. Find approximate 95%
confidence intervals for the effect of the presence of word A.1, if none of the
other words is present, and the nonword predictor variables A.49, A.50 and
so on are all 0. Then do the same for word A.2. Finally, find a confidence
interval for the difference of the two effects.
Mini-CRAN and other computational problems:
3. Though the logit model is plausible for the case Y = 0, 1, as noted
in Section 4.3.1, we could try modeling µ(t) as linear. We would then call
lm() instead of glm(), and simply predict Y to be whichever value in {0,1}
b(t) is closest to.
that µ

(a) Write R functions with call forms

b i n l i n ( i n d a t a , yname )
predict . b i n l i n ( b i n l i n o b j , newpts )

Here indata is a data frame, and yname is the name of the variable
there that will be taken as the response variable, Y , a vector of 1s
and 0s; the predictors will be the remaining columns. The function
binlin() calls lm(), and changes its class to ’binlin’, a subclass of
’lm’.
The function predict.binlin() then acts on binlinobj, an object
of class ’binlin’, predicting on the rows of the data frame newpts
(which must have the same column names as indata). The return
value will be a vector of 1s and 0s, computed as in the approach
proposed above.
(b) Try this approach on the spam prediction example of Section 4.3.6.
Using cross-validation, fit both a logit model and a linear one to the
training data, and see which one has better prediction accuracy on
the test set.

Math problems:
4. Suppose Ui , i = 1, 2 are independent random variables with means λi .
4.9. DATA, CODE AND MATH PROBLEMS 177

Let J be a random variable, independent of the Ui , which takes on the

values 1 and 0 with probability q and 1 − q. Define

W = UJ (4.54)

Show that W satisfies the overdispersion condition (4.13). (Hint: Use

(1.62).)
Chapter 5

Multiclass Classification
Problems

We introduced the classification problem in Section 1.17, and then covered

the logistic model for classification in some detail in Section 4.3. But in the
classification problems we’ve discussed so far, we have assumed just two
classes. The patient either has the disease in question, or not; the customer
chooses to buy a certain item, or not; and so on.
In many applications, we have multiple classes. We may, for instance,
be considering several different diseases that a patient might have.1 In
computer vision applications, the number of classes can be quite large, say
face recognition with data on a large number of people. Let m denote the
number of classes, and label them 0, 1, ..., m - 1.

5.1 Key Notation

The notation is a bit more complex than before, but still quite simple. The
reader is urged to read this section carefully in order to acquire a solid
grounding for the remaining material.
Say for instance we wish to do machine recognition of handwritten digits,
so we have 10 classes, with our variables being various patterns in the
1 For a classification problem, the classes must be mutually exclusive. In this case,

there would be the assumption that the patient does not have more than one of the
diseases.

179
180 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

pixels, e.g., the number of (approximately) straight line segments. Instead

of having a single response variable Y as before, we would now have 10
of them, setting Y (i) to be 1 or 0, according to whether the given digit is
i, for i = 0, 1, ..., 9. We could run 10 logistic regression models, and then
use each one to estimate the probability that our new image represents a
certain digit.
In general, as above, let Y (i) , i = 0, ..., m − 1 be the indicator variables for
the classes, and define the class probabilities

πi = P (Y (i) = 1), i = 0, 1..., m − 1 (5.1)

Of course, we must have

∑
m−1
πi = 1 (5.2)
i=0

We will still refer to Y , now meaning the value of i for which Y (i) = 1.
Note that in this chapter, we will be concerned primarily with the Predic-
tion goal, rather than Description.

5.2 Key Equations

Equations (4.20) and (4.28), and their generalizations, will play a key role
here. Let’s relate our new multiclass notation to what we had in the two-
class case before. If m = 2, then:

• What we called Y (1) above was just called Y in our previous discussion
of the two-class case.

• The class probability π1 here was called simply π previously.

Now, let’s review from the earlier material. (Keep in mind that typically X
will be vector-valued, i.e., we have more than one predictor variable.) For
m = 2:

• The quantity of interest is µ(t) = P (Y = 1 | X = t).

5.2. KEY EQUATIONS 181

• If X has a discrete distribution, then

µ(t) = P (Y = 1 | X = t)
π P (X = t | Y = 1)
=
π P (X = t | Y = 1) + (1 − π) P (X = t | Y = 0)
(5.3)

• If X has a continuous distribution, then

π f1 (t)
µ(t) = P (Y = 1 | X = t) = (5.4)
π f1 (t) + (1 − π) f0 (t)

where the within-class densities of X are f1 and f0 .2

• Sometimes it is more useful to use the following equivalence to (5.4):

1
µ(t) = P (Y = 1 | X = t) = (5.5)
1−π f0 (t)
1+ π f1 (t)

Note that, in keeping with the notion that classification amounts to a re-
gression problem (Section 1.17.1), we have used our regression function
notation µ(t) above.
Things generalize easily to the multiclass case. We are now interested in
the quantities

P (Y = i) = µi (t) = P (Y (i) = 1 | X = t), i = 0, 1, ..., m − 1 (5.6)

For continuous X, (5.4) becomes

πi fi (t)
P (Y = i) = µi (t) = P (Y (i) = 1 | X = t) = ∑m−1 (5.7)
j=0 πj fj (t)

2 Another term for the class probabilities π is prior probabilities. Readers familar
i
with the debate over Bayesian versus frequentist approaches to statistics may wonder
if we are dealing with (subjective) Bayesian analyses here. Actually, that is not the
case; we are not working with “gut feeling” probabilities as in (nonempirical) Bayesian
methods. There is some connection, in the sense that (5.3) and (5.4) make use of Bayes’
Rule, but the latter is standard for all statisticians, frequentist and Bayesian alike. Note
by the way that probabilities like (5.4) are often termed posterior probabilities, again
sounding Bayesian but again actually Bayesian/frequentist-neutral.
182 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

5.3 Estimating the Functions µi (t)

We have already estimated µ(t) in, say, (5.5) using logit models. We can
do the same for (5.6), running a logit analysis for each i, and indeed will
do so later.
Another possibility would be to take a nonparametric approach. For in-
stance, in the two-class case, one could estimate µ(t) = P (Y = 1|X = t) to
be the proportion of neighbors of t in our training data that have Y = 1. A
less direct, but sometimes useful approach is to estimate the fi (t) in (5.5)
and (5.6), and then plug our estimates into (5.5). For instance, one can do
this using a k-nearest neighbor method, outlined in Section 5.10.1 of the
Mathematical Complements section at the end of this chapter.

5.4 How Do We Use Models for Prediction?

In Section 1.10, we discussed the specifics of predicting new cases, in which

we know X but not Y , after fitting a model to training data, in which both
X and Y are known (our training data). The parametric and nonparametric
cases were slightly different.
Here in the multiclass setting, as noted, we now have multiple functions
µi (t), i = 0, 1, ..., m − 1 that need to be estimated from our training data,
as opposed to just µ(t) as before. Given a new case for which X = t, we
guess the class Y to be the value of i for which µ bi (t) is largest, i.e.,

Yb = arg max µ
bi (t) (5.8)
i

In the two-class case, this reduces to:

Given X. set Yb = 1 if and only if µ(X) > 0.5 (5.9)

It should be noted, though, that some nonparametric methods do not ex-

plicitly estimate µi (t), and instead only estimate “boundaries” involving
those functions. These methods will be discussed in Chapter 11.
5.5. ONE VS. ALL OR ALL VS. ALL? 183

5.5 One vs. All or All vs. All?

Let’s consider the Vertebral Column data from the UC Irvine Machine
Learning Repository.3 Here there are m = 3 classes: Normal, Disk Hernia
and Spondylolisthesis. The predictors are, as described on the UCI site,
“six biomechanical attributes derived from the shape and orientation of the
pelvis.” Consider two approaches we might take to predicting the status of
the vertebral column, based on logistic regression:

• One vs. All (OVA): Here we predict each class against all the other
classes. In the vertebrae data, we would fit 3 logit models to our train-
ing data, predicting each of the 3 classes, one at a time. So, first we
would fit a logit model to predict Normal vs. Other, the latter mean-
ing the Disk Hernia and Spondylolisthesis classes combined. Next
we would predict Disk Hernia vs. Other, with the latter now being
Normal and Spondylolisthesis combined, and the third model would
be Spondylolisthesis vs. Other.
The ith model would regress Y (i) against the 6 predictor variables,
yielding µbi (t), i = 0, 1, 2. To predict Y for X = tc , we would guess Y
to be whatever i has the largest value of µ bi (tc ), i.e., the most likely
class, given the predictor values.

• All vs. All (AVA): Here we would fit 3 logit models again, but with
one model for each possible pair of classes. Our first model would
pit class 0 against class 1, meaning that we would restrict our data
to only those cases in which the class is 0 or 1, then predict class 0
versus 1 in that restricted data set. Our second logit model would
restrict to the classes 0 and 2, and predict 0, while the last model
would be for classes 1 and 2, predicting 1. (We would still use our 6
predictor variables in each model.) In each case, we tally which class
“wins”; in the case in which we pit class 0 against class 1, our model
might predict that the given new data point is of class 1, thus tally
a win for that class. Then, whichever class “gets the most votes” in
this process is our final predicted class. (If there is a tie, we could
employ various tiebreaking procedures.)

Note that it was just coincidence that we have the same number of models
in the OVA and AVA approaches here (3 each). In general, with m classes,
we will run m logistic models (or k-NN or whatever type of regression
3 https://archive.ics.uci.edu/ml/datasets/Vertebral+Column
184 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

modeling we like) under OVA, but C(m, 2) = m(m − 1)/2 models under
AVA.4
Code for OVA and AVA is given in Section 5.11.1.

5.5.1 Which Is Better?

Clearly, AVA involves a lot of computation, in a more complex form if not

increased time. (See the Mathematical Complements section at the end of
this chapter for details.) At least at first glance, AVA would not seem to
have much to offer to make up for that. Indeed, it may have other problems
as well.
For instance, since each of its models uses much less than our full data,
the resulting estimated coefficients will likely be less accurate than what
we calculate under OVA. And if m is large, we will have so many pairs
that at least some will likely be especially inaccurate. (This notion will be
discussed in Section 7.6.) And yet some researchers report they find AVA
to work better in some settings.
To better understand the situation, let’s consider an example and draw
upon some intuition.

5.5.2 Example: Vertebrae Data

Here we analyze the vertebrae data first introduced in Section 5.5, apply-
ing the OVA and AVA methods to a training set of 225 randomly chosen
records, then predicting the remaining records.5 We’ll use the OVA and
AVA logistic code from the regtools package.6
> library ( r e g t o o l s )
> v e r t <− read . table ( ’ V e r t e b r a e /column 3C. dat ’ ,
he a d e r=FALSE)
> v e r t $V7 <− as . numeric ( v e r t $V7) − 1a
# for reproducible results
> set . s e e d ( 9 9 9 9 )
4 Here the notation C(r, s) means the number of combinations one can form from r

objects, taking them s at a time.

5 To avoid clutter, some messages, “glm.fit: fitted probabilities numerically 0 or 1

occurred,” have been removed, here and below. The warnings should not present a
problem.
6 Code for them is also shown in the Computational Complements section at the end

of this chapter.
5.5. ONE VS. ALL OR ALL VS. ALL? 185

> t r n i d x s <− sample ( 1 : 3 1 0 , 2 2 5 )

# a r g s a r e m and (X,Y) data , Y l a s t
> ovout <− o v a l o g t r n ( 3 , v e r t [ t r n i d x s , ] )
> predy <− o v a l o g p r e d ( ovout , v e r t [− t r n i d x s , 1 : 6 ] )
> mean( predy == v e r t [− t r n i d x s , 7 ] )
[ 1 ] 0.8941176
> avout <− a v a l o g t r n ( 3 , v e r t [ t r n i d x s , ] )
> predy <− a v a l o g p r e d ( 3 , avout , v e r t [− t r n i d x s , 1 : 6 ] )
> mean( predy == v e r t [− t r n i d x s , 7 ] )
[ 1 ] 0.8823529

Note that ovalogpred() requires that Y be coded 0, 1, ..., m − 1, hence the

call to as.numeric().
The two correct-classification rates here are, of course, subject to sampling
error, but in any case AVA did not seem superior.

5.5.3 Intuition

To put this in context, consider the artificial example in Figure 5.1, adapted
from Friedman [51]. Here we have m = 3 classes, with p = 2 predictors.
For each class, the bulk of the distribution of the predictor vectors mass for
that group is assumed to lie within one of the circles.
Now suppose a logistic model were used here. It implies that the prediction
boundary between our two classes is linear (Section 4.3.7). The figure
shows that a logit model would fare well under AVA, because for any pair
of classes, there is a straight line (pictured) that separates that pair of
classes well. But under OVA, we’d have a problem; though a straight
line separates the top circle from the bottom two, there is no straight line
that separates the bottom-left circle well from the other two very well; the
boundary between that bottom-left circle and the other two would be a
curve.
Keep that word curve in mind, as it will arise below. The problem here, of
course is that the logit, at least in the form implied above, is not a good
model in such a situation, so that OVA vs. AVA is not the real issue. In
other words, if AVA does do better than OVA on some dataset, it may be
due to AVA’s helping us overcome model bias. We will explore this in the
next section.
186 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

Figure 5.1: Three artificial regression lines

5.5.4 Example: Letter Recognition Data

Following up on the notion at the end of the last section that AVA may
work to reduce model bias, i.e., that AVA’s value occurs in settings in which
our model is not very good, let’s look at an example in which we know the
model is imperfect.
The UCI Letters Recognition data set7 uses various summaries of pixel
patterns to classify images of capital English letters. A naively applied
logistic model may sacrifice some accuracy here, due to the fact that the
predictors do not necessarily have monotonic relations with the response
variable, the class identity.
Actually, the naive approach doesn’t do too poorly:
> l i b r a r y ( mlbench )
> data ( L e t t e r R e c o g n i t i o n )
> l r <− L e t t e r R e c o g n i t i o n
> l r [ , 1 ] <− as . numeric ( l r [ , 1 ] ) − 1
7 https://archive.ics.uci.edu/ml/datasets/Letter+Recognition; also available in the R
package mlbench [87].
5.5. ONE VS. ALL OR ALL VS. ALL? 187

> # t r a i n i n g and t e s t s e t s
> l r t r n <− l r [ 1 : 1 4 0 0 0 , ]
> l r t e s t <− l r [ 1 4 0 0 1 : 2 0 0 0 0 , ]
> o l o g o u t <− o v a l o g t r n ( 2 6 , l r t r n [ , c ( 2 : 1 7 , 1 ) ] )
> ypred <− o v a l o g p r e d ( o l o g o u t , l r t e s t [ , − 1 ] )
> mean( ypred == l r t e s t [ , 1 ] )
[ 1 ] 0.7193333

We will see shortly that one can do considerably better. But for now, we
have a live candidate for a “poor model example,” on which we can try
AVA:
> a l o g o u t <− a v a l o g t r n ( 2 6 , l r t r n [ , c ( 2 : 1 7 , 1 ) ] )
> ypred <− a v a l o g p r e d ( 2 6 , a l o g o u t , l r t e s t [ , − 1 ] )
> mean( ypred == l r t e s t [ , 1 ] )
[ 1 ] 0.8355

That is quite a difference! So, apparently AVA fixed a poor model. But of
course, its better to make a good model in the first place. Based on our
previous observation that the boundaries may be better approximated by
curves than lines, let’s try a quadratic model.
A full quad model will have all squares and interactions among the 16
predictors. But there are 16·15/2+16 = 136 of them! That risks overfitting,
so let’s settle for just adding in the squares of the predictors:
> for ( i i n 2 : 1 7 ) l r <− cbind ( l r , l r [ , i ] ˆ 2 )
> l r t r n <− l r [ 1 : 1 4 0 0 0 , ]
> l r t e s t <− l r [ 1 4 0 0 1 : 2 0 0 0 0 , ]
> o l o g o u t <− o v a l o g t r n ( 2 6 , l r t r n [ , c ( 2 : 3 3 , 1 ) ] )
> ypred <− o v a l o g p r e d ( o l o g o u t , l r t e s t [ , − 1 ] )
> mean( ypred == l r t e s t [ , 1 ] )
[ 1 ] 0.8086667

Ah, much better. Not quite as good as AVA, but the difference is proba-
bly commensurate with sampling error, and we didn’t even try interaction
terms.

5.5.5 Example: k-NN on the Letter Recognition Data

As a basis of comparison to the above analyses of the letter recognition

data, let’s try k-NN. Note that with a nonparametric method such as k-
NN, there is no geometric issue as with Figure 5.1. Also, for convenience
we will not do cross-validation here.
188 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

Recall that with the knnest() function in regtools, we need a numeric Y .

Our data here has Y as an R factor. So, we’ll use the dummies package
to conveniently generate the m = 26 dummy variables for Y .
> xd <− p r e p r o c e s s x ( l r [ , − 1 ] , 5 0 )
> l i b r a r y ( dummies )
> y <− dummy( l r [ , 1 ] )
> xd <− p r e p r o c e s s x ( l r [ , − 1 ] , 5 0 )
> kout <− k n n e s t ( y , xd , 5 0 )

Now kout$regest has estimated class probabilities. To change those to

predictions using (5.8), we’ll use R’s apply() (Section 1.20.2) and which.max()
functions, and then convert back to characters via R’s built-in LETTERS
vector:
> tmp <− apply ( kout$ r e g e s t , 1 , which .max)
> knnpred <− LETTERS[ tmp ]

How well did we do?

> mean( knnpred == l r $ l e t t r )
[ 1 ] 0.9121

So it appears that neither OVA nor AVA solves the problems of a logit
model here.

5.5.6 The Verdict

With proper choice of model, OVA may do as well as AVA, if not better.
And a paper supporting OVA, [119] contends that some of the pro-AVA
experiments in the research literature were not done properly.
Clearly, though, our letters recognition example shows that AVA is worth
considering. We will return to this issue later.

5.6 Fisher Linear Discriminant Analysis

Sir Ronald Fisher (1890–1962) was one of the pioneers of statistics. He

called his solution to the multiclass problem linear discriminant analysis
(LDA), now considered a classic.
It is assumed that within class i, the vector of predictor variables X has a
multivariate normal distribution with mean vector µi and covariance matrix
5.6. FISHER LINEAR DISCRIMINANT ANALYSIS 189

Σ (Section 2.6.2). Note that the latter does not have a subscript i, i.e., in
LDA the covariance matrix for X is assumed the same within each class.

5.6.1 Background

To explain this method, let’s review some material from Section 4.3.1.
Let’s first temporarily go back to the two-class case, and use our past
notation:

Y = Y (1) , π = π1 (5.10)

For convenience, let’s reproduce (5.4) here:

π f1 (t)
P (Y = 1 | X = t) = (5.11)
π f1 (t) + (1 − π) f0 (t)

5.6.2 Derivation

As noted in Section 4.3.1, after substituting the multivariate normal density

for the fi in (5.11), we find that

1
P (Y = 1 | X = t) = ′ (5.12)
1+ e−(β0 +β t)

with

1
β0 = log(1 − π) − log π + (µ′1 µ1 − µ′0 µ0 ) (5.13)
2

and

β = (µ0 − µ1 )′ Σ−1 (5.14)

Intuitively, if we observe X = t, we should predict Y to be 1 if

P (Y = 1 | X = t) > 0.5 (5.15)

190 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

and this was shown in Section 1.17.1 to be the optimal strategy.8 Combining
this with (5.12), we predict Y to be 1 if

1
′ > 0.5 (5.16)
1+ e−(β0 +β t)

which simplifies to

′
β t > −β0 (5.17)

So it turns out that our decision rule is linear in t, hence the term linear in
linear discriminant analysis.9
Without the assumption of equal covariance matrices, (5.17) turns out to
be quadratic in t, and is called quadratic discriminant analysis.

5.6.3 Example: Vertebrae Data

Let’s apply this to the vertebrae data, which we analyzed in Section 5.5.2,
now using the lda() function. The latter is in the MASS library that is
built-in to R.

5.6.3.1 LDA Code and Results

The lda() function assumes that the class variable is an R factor, so we

won’t convert to numeric codes this time. Here is the code:
> l i b r a r y (MASS)
> v e r t <− read . table ( ’ V e r t e b r a e /column 3C. dat ’ ,
h e a d e r=FALSE)
> l d a o u t <− l d a (V7 ∼ . , data=v e r t ,CV=TRUE)
> mean( l d a o u t $ c l a s s == v e r t $V7)
[ 1 ] 0.8096774

That CV argument tells lda() to predict the classes after fitting the model,
using (5.4) and the multivariate normal means and covariance matrix that
is estimated from the data. Here we find a correct-classification rate of
about 81%. This is biased upward, since we didn’t bother here to set up
8 Again assuming equal costs of the two types of misclassification.
9 The word discriminant alludes to our trying to distinguish between Y = 1 and
Y = 0.
5.7. MULTINOMIAL LOGISTIC MODEL 191

separate training and test sets, but even then we did not do as well as our
earlier logit analysis. Note that in the latter, we didn’t assume a common
covariance matrix within each class, and that may have made the difference.
Of course, we could also try quadratic versions of LDA.

5.7 Multinomial Logistic Model

Within the logit realm, one might also consider multinomial logistic regres-
sion. This is similar to fitting m separate logit models, as we did in OVA
above, with a somewhat different point of view, motivated by the log-odds
ratio introduced in Section 4.3.3.

5.7.1 Model

The model now is to assume that the log-odds ratio for class i relative to
class 0 has a linear form,

P (Y = i | X = t) γi
log = log = β0i + β1i t1 + ... + βpi , i = 1, 2, ..., m − 1
P (Y = 0 | X = t) γ0
(5.18)
(Here i begins at 1 rather than 0, as each of the classes 1 through m − 1 is
being compared to class 0.)
Note that this is not the same model as we used before, though rather
similar in appearance.
The βji can be estimated via Maximum Likelihood, yielding

bi
γ
log = βb0i + βb1i t1 + ... + βbpi (5.19)
b0
γ

We then apply exp(), yielding the ratios γ bi /b

γ0 , after which the individual
bi can be solved algebraically, using the constraint
probabilities γ

∑
m−1
bi = 1
γ (5.20)
i=0
192 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

5.7.2 Software

There are several CRAN packages that implement this method. We’ll use
the nnet [121] package here. Though it is primarily for neural networks
analysis, it does implement multinomial logit, and has the advantage that
it is compatible with R’s stepAIC() function, used in Chapter 9.

5.7.3 Example: Vertebrae Data

Let’s try it out on the vertebrae data. Note that multinom() assumes
that the response Y is an R factor.
> v e r t <− read . table ( ’ column 3C. dat ’ , h e a d e r=FALSE)
# v e r t $V7 l e f t as a f a c t o r , not numeric
> l i b r a r y ( nnet )
> mnout <− multinom (V7 ∼ . , data=v e r t )
...
> c f <− coef ( mnout )
> cf
( Intercept ) V1 V2 V3
NO −20.23244 −4.584673 4.485069 0.03527065
SL −21.71458 1 6 . 5 9 7 9 4 6 −16.609481 0 . 0 2 1 8 4 5 1 3
V4 V5 V6
NO 4 . 7 3 7 0 2 4 0 . 1 3 0 4 9 2 2 7 −0.005418782
SL −16.390098 0 . 0 7 7 9 8 6 4 3 0 . 3 0 9 5 2 1 2 5 0

Since m = 3, we expect 3-1 = 2 sets of estimated coefficients, which indeed

we have above. The class of the return value, mnout in our example here,
is ’multinom’.
To illustrate how prediction of new cases would then work, let’s predict an
old case, vert[1,]. Remember, in R, many classes have methods for generic
functions such as print() and plot(), and especially in the regression con-
text, predict(). There is indeed a ”multinom” method for the R generic
predict() function:
> vt1 <− v e r t [ 1 , − 7 ]
> predict ( mnout , vt1 )
[ 1 ] DH
L e v e l s : DH NO SL

But it is more informative to determine the estimated conditional class

probabilities:
5.8. THE ISSUE OF “UNBALANCED” (AND BALANCED) DATA 193

> vt1
V1 V2 V3 V4 V5 V6
1 6 3 . 0 3 2 2 . 5 5 3 9 . 6 1 4 0 . 4 8 9 8 . 6 7 −0.25
> u <− exp ( c f %∗% c ( 1 , as . numeric ( vt1 ) ) )
> u
[ ,1]
1 0.130388474
2 0.006231212
> c ( 1 , u ) / sum( c ( 1 , u ) )
[ 1 ] 0.879801760 0.114716009 0.005482231

What happened here? Consider first the computation

c f %∗% c ( 1 , as . numeric ( vt1 ) )

(The call to as.numeric() is needed because vt1 is an R data frame rather

than a vector.)
According to our model (5.18), the above code computes the log-odds ratios.
Applying exp() gives us the raw odds ratios. Then we use (5.20) to solve
for the individual probabilities P (Y = i|X = t), which are 0.879801760 and
so on.
In this case, since we are just “predicting” our first observation, we need
not go through all that trouble above:
> mnout$ f i t t e d . v a l u e s [ 1 , ]
0 1 2
0.879801760 0.114716009 0.005482231

But for truly new data, the above sequence of operations will give us the
estimated class probabilities, which as mentioned are more informative than
merely a predicted class.

5.8 The Issue of “Unbalanced” (and Balanced)

Data

Here we will discuss a topic that is usually glossed over in treatments of

the classification problem, and indeed is often misunderstood, with much
questionable handwringing over “the problem of unbalanced data.” This
will be explained, and it will be seen below that often the real problem is
that the data are balanced.
194 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

For concreteness and simplicity, consider the two-class problem of predict-

ing whether a customer will purchase a certain item (Y = 1) or not (Y = 0),
based on a single predictor variable, X, the customer’s age. Suppose also
that most of the customers are older.

5.8.1 Why the Concern Regarding Balance?

Though one typically is interested in the overall rate of incorrect classifi-

cation, we may also wish to estimate rates of “false positives” and “false
negatives,” and to also gauge how well we can predict in certain subpop-
ulations. In our customer purchasing example, for instance, we wish to
ask, What percentage of the time do we predict that the customer does not
purchase the item, among cases in which the purchase actually is made?
And how well do we predict among the older customers? One problem is
that, although our overall misclassification rate is low, we may do poorly
on conditional error rates of this sort. This may occur, for example, if we
have unbalanced data, as follows.
Suppose only 1.5% of the customers in the population opt to buy the prod-
uct.10 The concern among some machine learning researchers and practi-
tioners is that, with random sampling (note the qualifier), the vast majority
of the data in our sample will be from the class Y = 0, thus giving us “un-
balanced” data. Then our statistical decision rule will likely just predict
almost everything to be Class 0, and thus may not predict the other class
well. Let’s take a closer look at this.
For the time being, assume that we have a random sample from the overall
population, rather than separate random samples from the two subclasses,
as discussed below.
Say we are using a logit model for µ(t). If the model is accurate throughout
the range of X, unbalanced data is not really a problem. The fact that one
class, say Y = 1, occurs rarely will likely increase standard errors of the
estimated regression coefficients, but it will not be a fundamental issue.
We will still have statistically consistent estimators (Section 2.7.3) of the
coefficients of β.
On the other hand, say we do classification using nonparametric density
estimation (Section 5.10.1). Since even among older customers, rather few
buy the product, we won’t have much data from Class 1, so our estimate
10 As mentioned earlier in the book, in some cases it may be difficult to define a

target population, even conceptually. There is not much that can be done about this,
unfortunately.
5.8. THE ISSUE OF “UNBALANCED” (AND BALANCED) DATA 195

of fb1 probably won’t be very accurate. Thus Equation (5.5) then suggests
we have a problem. We still have statistically consistent estimation, but
for finite samples that may not be enough. Nevertheless, short of using a
parametric model, there really is no solution to this.
Ironically, a more pressing issue is that we may have data that is too bal-
anced. Then we will not even have statistically consistent estimation. This
is the subject of our next section.

5.8.2 A Crucial Sampling Issue

In this chapter, we have often dealt with expressions such as P (Y = 1) and

P (Y = 1 |X = t). These seem straightforward, but actually they may be
undefined, due to our sampling design, as we’ll see here.
In our customer behavior context, P (Y = 1) is the unconditional proba-
bility that a customer will buy the given item. If it is equal to 0.12, for
example, that means that 12% of all customers purchase this item. By
contrast, P (Y = 1 | X = 38) is a conditional probability, and if it is equal
to 0.18, this would mean that among all people of age 38, 18% of them buy
the item.
The quantities π = P (Y = 1) and 1 − π = P (Y = 0) play a crucial role,
as can be seen immediately in (5.3) and (5.4). Let’s take a closer look
at this. Continuing our customer-age example, X (age) has a continuous
distribution, so (5.4) applies. Actually, it will be more useful to look at the
equivalent equation, (5.5).

5.8.2.1 It All Depends on How We Sample

Say our training data set consists of records on 1000 customers. Let N1
and N0 denote the number of people in our data set who did and did not
purchase the item, with N1 + N0 = 1000. If our data set can be regarded
as a statistical random sample from the population of all customers, then
we can estimate π from the data. If for instance 141 of the customers in
our sample purchased the item, then we would set

N1
b=
π = 0.141 (5.21)
1000

The trouble is, though, that the expression P (Y = 1) may not even make
sense with some data. Consider two sampling plans that may have been
196 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

followed by whoever assembled our data set.

(a) He sampled 1000 customers from our full customer database.11

(b) He opted to sample 500 customers from those who purchased the
item, and 500 from those who did not buy it.

Say we are using the density estimation approach to estimate P (Y | X = t),

in (5.5). In sampling scheme (a), N1 and N0 are random variables, and as
noted we can estimate π by the quantity N1 /1000. But in sampling scheme
(b), we have no way to estimate π from our data.
Or, suppose we opt to use a logit model here. It turns out that we will run
into similar trouble in sampling scheme (b), as follows. From (4.18) and
(5.5), write the population relation

β0 + β1 t1 + ... + βp tp = ln(π/(1 − π)) + ln[f1 (t)/f0 (t)] (5.22)

where t = (t1 , ..., tp )′ . Since this must hold for all t, we see that β0 =
ln(π/(1 − π)). So if we follow design (b) above, our estimator, not knowing
better, will assume (a), and estimate π to be 0.5. However, under design
(b), βi , i > 0 will not change, because the fi are within-class densities, and
their ratio will still be estimated properly. only β0 changes. In other words,
our logit-estimation software will produce the wrong constant term, but be
all right on the other coefficients.
In summary:

Under sampling scheme (b), we are obtaining the wrong βb0 ,

though the other βbi are correct.

If our goal is merely Description rather than Prediction, this may not be a
concern, since we are usually interested only in the values of βi , i > 0. But
if Prediction is our goal, as we are assuming in this chapter, we do have a
serious problem, since we will need all of the estimated coefficients in order
to estimate P (Y |X = t) in (4.18).
A similar problem arises if we use the k-Nearest Neighbor method. Sup-
pose for instance that the true value of π is low, say 0.06, i.e., only 6%
of customers buy the product. Consider estimation of P (Y | X = 38).
11 Or,this was our entire customer database, which we are treating as a random sample
from the population of all customers.
5.8. THE ISSUE OF “UNBALANCED” (AND BALANCED) DATA 197

Under the k-NN approach, we would find the k closest observations in our
sample data to 38, and estimate P (Y | X = 38) to be the proportion of
those neighboring observations in which the customer bought the product.
The problem is that under sampling scenario (b), there will be many more
among those neighbors who bought the product than there “should” be.
Our analysis won’t be valid.
So, all the focus on unbalanced data in the literature is arguably misplaced.
As we saw in Section 5.8.1, it is not so much of an issue in the parametric
case, and in any event there really isn’t much we can do about it. At least,
things do work out as the sample size grows. By contrast, with sampling
scheme (b), we have a permanent bias, even as the sample size grows.
Scenario (b) is not uncommon. In the UCI Letters Recognition data set
mentioned earlier for instance, there are between 700 and 800 cases for each
English capital letter, which does not reflect that wide variation in letter
frequencies. The letter ’E’, for example, is more than 100 times as frequent
as the letter ’Z’, according to published data (see below).
Fortunately, there are remedies, as we will now see.

5.8.2.2 Remedies

As noted, use of “unnaturally balanced” data can seriously bias our classi-
fication process. In this section, we turn to remedies.
It is assumed here that we have an external data source for the class prob-
abilities πi . For instance, in the English letters example above, there is
much published data, such as at the Web page Practical Cryptography.12
It turns out that πA = 0.0855, πB = 0.0160, πC = 0.0316 and so on.
So, if we do have external data on the πi (or possibly want to make some
“what if” speculations), how do we adjust our code output to correct the
error?
For LDA, R’s lda() function does the adjustment for us, using its priors
argument. That code is based on the relation (4.31), which we now see is
a special case of (5.22).
The latter equation shows how to deal with the logit case as well: We
simply adjust the βb0 that glm() gives us as follows.

12 http://practicalcryptography.com/cryptanalysis/letter-frequencies-various-

languages/english-letter-frequencies/.
198 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

(a) Subtract ln(N1 /N0 ).

(b) Add ln[π)/(1 − π))], where π is the true class probability.

Note that for an OVA m-class setting, we estimate m logistic regression

functions, adjusting βb0 in each case. The function ovalogtrn() includes an
option for this.
What about nonparametric settings? Equation (5.5) shows us how to make
the necessary adjustment, as follows:

(a) Our software has given us an estimate of the left-hand side of that
equation for any t.

(b) We know the value that our software has used for its estimate of
(1 − π)/π, which is N0 /N1 .

(c) Using (a) and (b), we can solve for the estimate of f1 (t)/f0 (t).

(d) Now plug the correct estimate of (1 − π)/π, and the result of (c),
back into (5.5) to get the proper estimate of the desired conditional
probability.

Code for this is straightforward:

c l a s s a d j u s t <− function ( econdprobs , w r o n g r a t i o ,
trueratio ) {
f r a t i o s <− ( 1 / e c o n d p r o b s − 1 ) ∗ ( 1 / w r o n g r a t i o )
1 / (1 + t r u e r a t i o ∗ f r a t i o s )
}

Note that if we are taking the approach described in the paragraph labeled
“A variation” in Section 1.10.2, we do this adjustment only at the stage
in which we fit the training data. No further adjustment at the prediction
stage is needed.

5.8.3 Example: Letter Recognition

Let’s try the k-NN analysis on the letter data. First, some data prep:
> l i b r a r y ( mlbench )
> data ( L e t t e r R e c o g n i t i o n )
> l r <− L e t t e r R e c o g n i t i o n
5.8. THE ISSUE OF “UNBALANCED” (AND BALANCED) DATA 199

> # code Y v a l u e s
> l r [ , 1 ] <− as . numeric ( l r [ , 1 ] ) − 1
> # t r a i n i n g and t e s t s e t s
> l r t r n <− l r [ 1 : 1 4 0 0 0 , ]
> l r t e s t <− l r [ 1 4 0 0 1 : 2 0 0 0 0 , ]

As discussed earlier, this data set has approximately equal frequencies for
all the letters, which is unrealistic. The regtools package contains the
correct frequencies [97], obtained from the Practical Cryptography Web
site cited before. Let’s load those in:
We continue our analysis from Section 5.5.4.
> l i b r a r y ( mlbench )
> data ( L e t t e r R e c o g n i t i o n )
> library ( r e g t o o l s )
> tmp <− table ( L e t t e r R e c o g n i t i o n [ , 1 ] )
> w r o n g p r i o r s <− tmp / sum( tmp )
> data ( l t r f r e q s )
> l t r f r e q s <− l t r f r e q s [ order ( l t r f r e q s [ , 1 ] ) , ]
> t r u e p r i o r s <− l t r f r e q s [ , 2 ] / 100

(Recall from Footnote 2 that the term priors refers to class probabili-
ties, and that word is used both by frequentists and Bayesians. It is not
“Bayesian” in the sense of subjective probability.)
So, here is the straightforward analysis, taking the letter frequencies as they
are, with 50 neighbors:
> xdata <− p r e p r o c e s s x ( l r t r n [ , − 1 ] , 5 0 )
> t r n o u t <− knntrn ( l r t r n [ , 1 ] , xdata , 2 6 , 5 0 )
> tmp <− predict ( t r n o u t , l r t e s t [ , − 1 ] )
> ypred <− apply ( as . matrix ( tmp ) , 1 , which .max) −
> mean( ypred == l r t e s t [ , 1 ] )
[ 1 ] 0.8641667

In light of the fact that we have 26 classes, 86% accuracy is pretty good.
But it’s misleading: We did take the trouble of separating into training and
test sets, but as mentioned, the letter frequencies are unrealistic. How well
would our classifier do in the “real world”? To simulate that, let’s create a
second test set with correct letter frequencies:
> newidxs <−
sample ( 0 : 2 5 , 6 0 0 0 , replace=T, prob=t r u e p r i o r s )
> l r t e s t 1 <− l r t e s t [ newidxs , ]
200 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

Now we can try our classifier on this more realistic data:

> ypred <− knnpred ( t r n o u t , l r t e s t 1 [ , − 1 ] )
> mean( ypred == l r t e s t 1 [ , 1 ] )
[ 1 ] 0.7543415

Only about 75%. But in order to prepare for the real world, we can make
use of the truepriors argument in knntrn()
> t r n o u t 1 <− knntrn ( l r t r n [ , 1 ] , xdata , 2 6 , 5 0 , t r u e p r i o r s )
> ypred <− predict ( t r n o u t 1 , l r t e s t 1 [ , − 1 ] )
> mean( ypred == l r t e s t 1 [ , 1 ] )
[ 1 ] 0.8787988

Ah, very nice!

5.9 Going Beyond Using the 0.5 Threshold

As we have seen, the optimal rule (knowing population functions) in the

two-class case is given by (5.9). (That formulation assumes that population
quantities are known; in practice, of course, we must use estimates, i.e.,
b(t) rather than µ(t).) But here “optimal” referred to minimizing overall
µ
misclassification rate, and more detailed analysis may be appropriate.

5.9.1 Unequal Misclassification Costs

Say we are trying to determine whether a patient has a particular disease,

based on a vector X of various test results, demographic variables and so
on for this patient. Denote the value of X by tc , and suppose our estimate
of P (Y (1) = 1 | X = tc ) is 0.02. We estimate that this patient has only a
2% chance of having the disease. This isn’t very high, so we might simply
stop there.
But in the case of a catastrophic disease, the misclassification costs may
not be equal. Failing to detect the disease when it’s present may be a much
more serious error than ordering further medical tests that turn out to be
negative. What can be done to address this?
Informal approach:
The physician may have a hunch, based on information not in X and thus
not in our sample data, that leads her to suspect that the patient does have
5.9. GOING BEYOND USING THE 0.5 THRESHOLD 201

the disease. The physician may thus order further tests and so on, in spite
of the low estimated probability.
Remember, our estimated P (Y (i) = 1 | X = tc ) can be used as just one of
several components that may enter into our final decision.
Automatic classification:
In many applications today, our classification process will be automated,
done entirely by machine. Consider the example in Section 4.3.1 of clas-
sifying subject matter of Twitter tweets, say into financial tweets and all
others, a two-class setting. Here again there may be unequal misclassifica-
tion costs, depending on our goals. If so, the prescription (5.9), i.e.,

{
1, if µ(X) > 0.5
guess for Y = (5.23)
0, if µ(X) ≤ 0.5

is not what we want, as it implicitly assumes equal costs. If we wish to

automate, we’ll probably need to set up a formal cost structure. This will
result in the 0.5 threshold above changing to something else. We will use
this idea the next section.

5.9.2 Revisiting the Problem of Unbalanced Data

In Section 5.8, it was argued that if our goal is to minimize the overall
misclassification rate, the problem of “unbalanced” data is not really a
problem in the first place (or if it is a problem, it’s insoluble). But as
pointed out in Section 5.9.1, we may be more interested in correct prediction
for some classes than others, so the overall misclassification rate is not our
primary interest.
In the Mathematical Complements section at the end of this chapter, a
simple argument shows that we should guess Y = 1 if

ℓ0 1
µ(X) ≥ = (5.24)
ℓ0 + ℓ1 1 + l1 /l0

where the li are our misclassification costs (“losses”). All that matters is
their ratio. For instance, say we consider guessing Y = 1 when in fact
Y = 0 (cost l0 ) to be 3 times worse an error than guessing Y = 0 when
202 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

actually Y = 1 (cost l1 ). Then our new threshold is

1
1 = 0.75 (5.25)
1+ 3

So, we set our threshold at 0.75 rather than 0.5. This more stringent
criterion for guessing Y = 1 means we take such action less often, thus
addressing our concern that a false positive is a very serious error.

5.9.3 The Confusion Matrix and the ROC Curve

In an m-class setting, the confusion matrix is defined as follows: For 0 ≤

i, j ≤ m − 1, the (i + 1, j + 1) element is the number of cases in which the
actual class was i and we predicted it to be j. This tells us at a glance on
which classes we are predicting well or poorly.
Now consider what would happen, say in the case m = 2 for simplicity, if
we were to try various values for the threshold in (5.9) instead of just 0.5.
Let’s call the threshold h, meaning that we guess Y = 1 if µ(X) > h.
As we vary h, the confusion matrix would change. As we reduce h starting
from near 1.0, we would have more true positive guesses (Y = Yb = 1) but
the number of false positives (Y = 0, Yb = 1) would increae.
The Receiver Operating Characteristic (ROC) curve plots the rate of true
positives (TPR) against the proportion of false positives (FPR), as h varies.
Since we may not have specific misclassification costs in mind, the ROC
curve allows us to explore the effects of using different values of h.
Specifically, we plot estimates of

P (guess Y = 1 when Y = 1) = P (µ(X) > h | Y = 1) (5.26)

versus

P (guess Y = 1 when Y = 0) = P (µ(X) > h | Y = 0) (5.27)

Look at the disease diagnosis case, for example, where having the disease is
coded Y = 1. Then TPR is the proportion of time we would guess that the
patient has the disease, among those patients who actually have it, versus
incorrectly guessing presence of the disease among those who don’t have it.
5.10. MATHEMATICAL COMPLEMENTS 203

5.9.3.1 Code

The code for plotting ROC is presented in the Computational Complements

section at the end of this chapter. It is written from the point of view of
clarity of the ROC process, rather than efficiency.
Its call form is simple:
r o c ( x , y , r e g e s t , nh )

Here x is the matrix/data frame of X values, y is the vector of Y values,

b(Xi ) values. The optional argument nh is the
and regest is the vector of µ
number of h values to plot, evenly spaced in (0,1). Note that regest could
come from a logistic model, k-NN analysis and so on.

5.9.3.2 Example: Spam Data

Let’s continue the computation we made on the spam data in Section 4.3.6.
The object glmout was the output from fitting a logit model. The Y
variable was in column 58 of the data frame, and all the other columns
were used as predictors. So, our call is
> r o c ( spam [ , − 5 8 ] , spam [ , 5 8 ] , glmout$ f i t t e d . v a l u e s )

The plot is shown in Figure 5.2. Note that the quantity h does not appear
explicitly. Instead, it is used to generate the (FPR,TPR) pairs, one pair
per value of h.
The curve in this case rises steeply for small values of FPR. In general, the
steeper the better, because it indicates that we can obtain a good TPR rate
at a small “price,” i.e., by tolerating just a small amount of false positives.
In this case, this does occur, not surprising in light of the high rate of
correct classification we found in our earlier analysis of this data.

5.10 Mathematical Complements

5.10.1 Classification via Density Estimation

Since classification amounts to a regression problem, we can use nonpara-

metric regression methods such as k-Nearest Neighbor if we desire a non-
parametric approach, as seen above. However, Equations (5.4) and (5.7)
204 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

Figure 5.2: ROC curve for spam data

suggest that one approach to the classification problem would be to esti-

mate the within-class densities fi . Actually, this approach is not commonly
used, as it is difficult to get good estimates, especially if the number of pre-
dictors is large. However, we will examine it in this section anyway, as it
will yield some useful insights.

5.10.1.1 Methods for Density Estimation

Say for simplicity that X is one-dimensional. You are already familiar

with one famous nonparametric method for density estimation — the his-
togram!13 What is more commonly done, though is a variant of that.
Say we are estimating f (), the density of X, with corresponding cdf F (t).
From introductory calculus we have

F (t + h) − F (t − h)
f (t) ≈ (5.28)
2h
13 However, since a density integrates to 1.0, we should scale our histogram accordingly.
In R’s hist() function, we specify this by setting the argument freq to FALSE.
5.10. MATHEMATICAL COMPLEMENTS 205

for small h > 0. Since we can estimate the cdf directly from the data, our
estimate is

#(t − h, t + h)/n
fb(t) = (5.29)
2h
#(t − h, t + h)
= (5.30)
2hn

where # stands for the number of Xi in the given interval and n is the
sample size.
For k-NN, with k neighbors, do the following. In the denominator of (5.29),
set h equal to the distance from t to the furthest neighbor, and set the
numerator to k.
The above is for the case p = 1. As noted, things become difficult in higher
dimensions, and none of the R packages go past p = 3, due to high variances
in the estimated values [40].
For an in-depth treatment of density estimation, see [127].

5.10.2 Time Complexity Comparison, OVA vs. AVA

The running time of an algorithm is typically described in terms of “big-O”

notation. Consider finding the inverse of a k × k matrix, for instance. It
turns out that this takes about ck 3 time steps, where c is some constant. We
usually don’t care about the specific value of c, because we are concerned
mainly with growth rates. How does the running time grow as k increases?
If for exanple the number of rows and columns k is doubled, we see that
the time to invert the matrix will increase by a factor of 8, regardless of the
value of c.
Let’s apply that to the OVA/AVA situation. For fixed number of predictor
variables p, here is a rough time estimate. For a logit model, the compu-
tation will be proportional to the number of cases n (due to computing
various sums over all cases). Say our training data is approximately bal-
anced in terms of sizes of the classes, so that the data corresponding to
class i has about n/m cases in it.
Then the computation for one pair of classes in AVA will take O(n/m)
time, but there will be O(m2 ) pairs, so the total amount of computation
will be O(m2 × n/m) = O(mn). This is potentially much more than the
corresponding O(n) time complexity for OVA.
206 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

However, so far we haven’t factored p into our analysis. If we do this, then

the results become very dependent on the type of classifier we are using
— logit, k-NN, SVM (Chapter 11 and so on. Such analysis becomes quite
complex, but it can certainly be the case that for some types of estimators
the speed advantage of OVA over AVA is not as great.

5.10.3 Optimal Classification Rule for

Unequal Error Costs

Let ℓ0 denote our cost for guessing Y to be 1 when it’s actually 0, and
define ℓ1 for the opposite kind of error. Now reason as follows as to what
we should guess for Y , knowing that X = tc . For convenience, write

p = P (Y = 1 | X = tc ) (5.31)

Suppose we guess Y to be 1. Then our expected cost is

(1 − p)ℓ0 (5.32)

If on the other hand we guess Y to be 0, our expected cost is

pℓ1 (5.33)

So, our strategy could be to choose our guess to be the one that gives us
the smaller of (5.32) and (5.33):

{
1, if (1 − p)ℓ0 ≤ pℓ1
guess for Y = (5.34)
0, if (1 − p)ℓ0 > pℓ1

Solving for p = µ(tc ), we have

{
1, if (1 − p)ℓ0 ≤ pℓ1
guess for Y = (5.35)
0, if (1 − p)ℓ0 > pℓ1

In other words, given X, we guess Y to be 1 if

ℓ0
µ(X) ≥ (5.36)
ℓ0 + ℓ1
5.11. COMPUTATIONAL COMPLEMENTS 207

5.11 Computational Complements

5.11.1 R Code for OVA and AVA Logit Analysis

To make this concrete, here is code for the two approaches:

# One−v s .− A l l (OVA) and A l l −v s . A l l (AVA) ,
# l o g i t models

# arguments :

# m: number o f c l a s s e s
# t r n x y : X, Y t r a i n i n g s e t ; Y i n l a s t column ;
# Y coded 0 , 1 , . . . ,m−1 f o r t h e m c l a s s e s
# p r e d x : X v a l u e s from which t o p r e d i c t Y v a l u e s
# t s t x y : X, Y t e s t s e t , same f o rm a t

#####################################################
# ovalogtrn : generate estimated regression functions
#####################################################

# arguments :

# m: as above
# t r n x y : as abo ve

# value :

# m a t r i x o f t h e b e t a h a t v e c t o r s , one p e r column

o v a l o g t r n <− function (m, t r n x y ) {

p <− ncol ( t r n x y )
x <− as . matrix ( t r n x y [ , 1 : ( p − 1 ) ] )
y <− t r n x y [ , p ]
outmat <− NULL
for ( i i n 0 : (m−1)) {
ym <− as . integer ( y == i )
b e t a h a t <− coef (glm(ym ∼ x , family=binomial ) )
outmat <− cbind ( outmat , b e t a h a t )
}
outmat
}
208 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

#####################################################
# o v a l o g p r e d : p r e d i c t Ys from new Xs
#####################################################

# arguments :
#
# co efmat : c o e f . matrix , o u t p u t from o v a l o g t r n ( )
# p r e d x : as abov e
#
# value :
#
# v e c t o r o f p r e d i c t e d Y v a l u e s , i n { 0 , 1 , . . . ,m−1} ,
# one e l e m e n t f o r each row o f p r e d x

o v a l o g p r e d <− function ( coefmat , predx ) {

# g e t e s t r e g f t n v a l u e s f o r each row o f p r e d x
# and each c o l o f c o e f m a t ; v a l s from
# coefmat [ , ] i n tmp [ , i ]
tmp <− as . matrix ( cbind ( 1 , predx ) ) %∗% coefmat
tmp <− l o g i t ( tmp )
apply ( tmp , 1 , which .max) − 1
}

#####################################################
# avalogtrn : generate estimated regression functions
#####################################################

# arguments :

# m: as above
# t r n x y : as abo ve

# value :

# m a t r i x o f t h e b e t a h a t v e c t o r s , one p e r column ,
# i n t h e o r d e r o f combin ( )

a v a l o g t r n <− function (m, t r n x y ) {

p <− ncol ( t r n x y )
n <− nrow( t r n x y )
x <− as . matrix ( t r n x y [ , 1 : ( p − 1 ) ] )
y <− t r n x y [ , p ]
outmat <− NULL
5.11. COMPUTATIONAL COMPLEMENTS 209

i j s <− combn (m, 2 )

d o r e g <− function ( i j ) {
i <− i j [ 1 ] − 1
j <− i j [ 2 ] − 1
tmp <− rep ( −1 ,n )
tmp [ y == i ] <− 1
tmp [ y == j ] <− 0
y i j <− tmp [ tmp != −1]
x i j <− x [ tmp != −1 ,]
coef (glm( y i j ∼ x i j , family=binomial ) )
}
coefmat <− NULL
for ( k i n 1 : ncol ( i j s ) ) {
coefmat <− cbind ( coefmat , d o r e g ( i j s [ , k ] ) )
}
coefmat
}

#####################################################
# a v a l o g p r e d : p r e d i c t Ys from new Xs
#####################################################

# arguments :
#
# m: as above
# co efmat : c o e f . matrix , o u t p u t from a v a l o g t r n ( )
# p r e d x : as abov e
#
# value :
#
# v e c t o r o f p r e d i c t e d Y v a l u e s , i n { 0 , 1 , . . . ,m−1} ,
# one e l e m e n t f o r each row o f p r e d x

a v a l o g p r e d <− function (m, coefmat , predx ) {

i j s <− combn (m, 2 ) # as i n a v a l o g t r n ( )
n <− nrow( predx )
ypred <− vector ( length = n )
for ( r i n 1 : n ) {
# p r e d i c t t h e r t h new o b s e r v a t i o n
xrow <− c ( 1 , u n l i s t ( predx [ r , ] ) )
# wins [ i ] t e l l s how many t i m e s c l a s s i −1
# has won
wins <− rep ( 0 ,m)
210 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

for ( k i n 1 : ncol ( i j s ) ) {
i <− i j s [ 1 , k ] # c l a s s i −1
j <− i j s [ 2 , k ] # c l a s s j −1
bhat <− coefmat [ , k ]
mhat <− l o g i t ( bhat %∗% xrow )
i f ( mhat >= 0 . 5 ) wins [ i ] <− wins [ i ] + 1 e l s e
wins [ j ] <− wins [ j ] + 1
}
ypred [ r ] <− which .max( wins ) − 1
}
ypred
}

l o g i t <− function ( t ) 1 / (1+exp(−t ) )

For instance, under OVA, we call ovalogtrn() on our training data, yielding
a logit coefficient matrix having m columns; the ith column will consist of
the estimated coefficients from fitting a logit model predicting Y (i) . We
then use this matrix as input for predicting Y in all future cases that come
our way, by calling ovalogpred() whenever we need to do a prediction.
Under AVA, we do the same thing, calling avalogtrn() and avalogpred().

5.11.2 ROC Code

Here is the code for ROC plotting:

# s i m p l e i m p l e m e n t a t i o n o f ROC, meant t o show t h e
# p r i n c i p l e s r a t h e r than be e f f i c i e n t

# arguments :

# x : m a t r i x / d a t a frame o f X v a l u e s
# y : v e c t o r o f Y v a l u e s (0 or 1)
# regest : vector of estimated regression function
# v a l u e s ; t h e f i t t e d . v a l u e s component
# from glm ( ) and l o g i t
# nh : number o f v a l u e s o f t h r e s h o l d t o p l o t

r o c <− function ( x , y , r e g e s t , nh=100) {

# find the indices of the
# Y = 0 and Y = 1 c a s e s
y 0 i d x s <− which ( y == 0 )
5.12. EXERCISES: DATA, CODE AND MATH PROBLEMS 211

y 1 i d x s <− which ( y == 1 )
# and t h e e s t i m a t e d v a l u e s o f P(Y = 1 | X)
# for those cases
r e g e s t 0 <− r e g e s t [ y 0 i d x s ]
r e g e s t 1 <− r e g e s t [ y 1 i d x s ]
# try various threshold values h
increm <− 1/nh
h <− ( 1 : ( nh −1)) / nh
# s e t v e c t o r s f o r t h e FPR, TPR v a l u e s
f p r v a l s <− vector ( length = nh−1)
t p r v a l s <− vector ( length = nh−1)
# f o r each p o s s i b l e t h r e s h o l d , f i n d FPR, TPR
for ( i i n 1 : ( nh −1)) {
f p r v a l s [ i ] <− mean( r e g e s t 0 > h [ i ] )
t p r v a l s [ i ] <− mean( r e g e s t 1 > h [ i ] )
}
plot ( f p r v a l s , t p r v a l s , x l a b= ’FPR ’ , y l a b= ’TPR ’ )
}

5.12 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. Plot ROC curves for each of the three classes in the Vertebral Column
data analyzed in this chapter. Try both logistic and k-NN approaches.
(Also see Exercise 4 below.)
2. Consider the OVA vs. AVA comparison, with cross-validation, in Section
5.5.4. Re-run the analysis, recording run time (system.time()).
3. Try multinomial logit on the Letter Recognition data, comparing it to
the results in Section 5.5.4. Then re-run it after adding squared versions of
the predictors, as was also done in that section.
Mini-CRAN and other computational problems:
4. Write a function with call form
m u l t i r o c ( x , y , r e g e s t m a t , nh=100)
212 CHAPTER 5. MULTICLASS CLASSIFICATION PROBLEMS

to plot multiple ROC curves. Here x and nh are as in roc() (Section

roccode); regestmat is a matrix version of regest in roc(); and y is the
class data. The latter could either be an integer taking values in 0, 1, ..., m−
1 or an R factor.
As to the arrangement of these multiple plots, you might show them one at
a time, as with parvsnonparplot() in Chapter 6, or a grid of plots, say
by setting the mfrow argument in R’s par() function.
5. Here we will consider the “OVA and AVA” approaches to the multiclass
problem (Section 5.5), using the UCBAdmissions data set that is built in to
R. This data set comes in the form of a counts table, which can be viewed
in proportion terms via
UCBAdmissions / sum( UCBAdmissions )

For the sake of this experiment, let’s take those cell proportions to be pop-
ulation values, so that for instance 7.776% of all applicants are male, apply
to departmental program F and are rejected. The accuracy of our classifi-
cation process then is not subject to the issue of variance of estimators of
logistic regression coefficients or the like.

(a) Which would work better in this population, OVA or AVA, say in
terms of overall misclassification rate?
[Computational hint: First convert the table to an artificial data
frame:
ucbd <− as . data . frame ( UCBAdmissions )

(b) Write a general function

o v a a v a t b l <− function ( t b l , yname )

that will perform the computation in part (a) for any table tbl, with
the class variable having the name yname, returning the two mis-
classification rates. Note that the name can be retrieved via
names( attr ( t b l , ’ dimnames ’ ) )

6. Write a function to compute the confusion matrix for the output of

multinom(), Section 5.7. The call form will be
nmconfmat ( mnobj )
5.12. DATA, CODE AND MATH PROBLEMS 213

where mnobj is an object of class ’multinom’, which is the type returned

by multinom(). The function will return a matrix, defined as follows.
For 0 ≤ i, j ≤ m − 1, the (i + 1, j + 1) element of the matrix will be the
proportion of cases in which the actual class is i and the predicted class is
j.
Math problems:
7. Extend (5.29) to the case p = 2.
8. Section 5.9.2 showed how to handle the problem of unequal costs for the
two-class case. Show how to extend this to m classes.
Chapter 6

Model Fit Assessment and

Improvement

The famous Box quote from our first chapter is well worth repeating:

All models are wrong, but some are useful — famed statistician
George Box

We have quite a bit of powerful machinery to fit parametric models. But

are they any good on a given data set? We’ll discuss this subject here in
this chapter.

6.1 Aims of This Chapter

Most regression books have a chapter on diagnostics, methods for assessing

model fit and checking assumptions needed for statistical inference (confi-
dence intervals and significance tests).
In this chapter we are concerned only with the model itself. For instance,
in a model that is linear in the predictor variables, how accurate is that
linearity assumption? Are there extreme or erroneous observations that
mar the fit of our model?
We are not concerned here with assumptions that only affect inference, as

215
216 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

those were treated in Chapter 2.1

6.2 Methods

There is a plethora of diagnostic methods! Entire books could and have

been written on this topic. Here I will treat only a few such methods here
— some classical, some of my own — with the choice of methods presented
stemming from these considerations:

• This book generally avoids statistical methods that rely on assuming

that our sample data is drawn from a normally distributed popula-
tion.2 Accordingly, the material here on unusual observations does
not make such an assumption.
• Intuitive clarity of a method is paramount. If a method can’t be ex-
plained well to, say, a reasonably numerate but nonstatistician client,
then I prefer to avoid it.

6.3 Notation

As before, say we have data (Xi , Yi ), i = 1, ..., n. Here the Xi are p-

component vectors,

Xi = (Xi , ..., Xi )′
(1) (p)
(6.1)

and the Yi are scalars (including the case Y = 0, 1, ..., m − 1 in classification

applications). We typically won’t worry too much in this chapter whether
the n observations are independent. As usual, let

µ(t) = E(Y | X = t) (6.2)

b(t) denote its estimate from

be our population regression function, and let µ
our sample data.
1 That chapter showed that the assumption of normal distributions is not very impor-

tant, and Chapter 3 presented methods for dealing with nonhomogeneous variance. The
third assumption, statistical independence, was not covered there, and indeed will not
be covered elsewhere in the book, in the sense of methods for assessing independence;
there are not many such methods, and typically they depend on their own assumptions,
thus “back to Square One.”
2 “Rely on” here means that the method is not robust to the normality assumption.
6.4. GOALS OF MODEL FIT-CHECKING 217

6.4 Goals of Model Fit-Checking

What do we really mean when we ask whether a model fits a data set well?
Our answer ought to be as follows:

Possible Fit-Checking Goal:

b(t) is near µ(t) for all t.

Our model fits well if µ

That criterion is of course only conceptual; we don’t know the values of

µ(t), so it’s an impossible criterion to truly verify. Nevertheless, it may
serve well as a goal, and our various model-checking methods will be aimed
at that goal.
Note how the issue of overfitting comes in here. The above Goal could have
b(t) is near µ(t) for t = Xi , i = 1, 2, ..., n.” As noted before,
said, “...if µ
with p = 1, for example, we could fit a polynomial model of degree n − 1
and have a “perfect” fit, obviously misleading.
Part of the answer to our goals question goes back to the twin regression
goals of Prediction and Description. We’ll explore this in the following
sections.

6.4.1 Prediction Context

If our regression goal is Prediction and we are doing classification, our above
Fit-Checking Goal may be much too stringent. Say for example m = 2, just
′
two classes, 0 and 1. Let (Ynew , Xnew )′ denote our new observation, with
Xnew known but Y unknown and to be predicted. We will guess Y = 1 if
b(Xnew ) > 0.5.
µ
If µ(Xnew ) is 0.9 but µb(Xnew ) = 0.62, we will still make the correct guess,
Y = 1, even though our regression function estimate is well off the mark.
Similarly, if µ(Xnew ) is near 0 (or less than 0.5, actually), we will make the
proper guess for Y as long as our estimated value µ b(Xnew ) is under 0.5.
Still, other than the classification case, the above Fit-Checking Goal is
appropriate. Errors in our estimate of the population regression function
will impact our ability to predict.
218 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

6.4.2 Description Context

Good model fit is especially important when our regression goal is Descrip-
tion. We really want assurance that the estimated regression coefficients
represent the true regression function well. since we will be using them to
describe the underlying process.

6.4.3 Center vs. Fringes of the Data Set

Consider a Prediction setting, in the classification case. In Section 6.4.1

above, we saw that we actually can afford a rather poor model fit in regions
of the predictor space in which the population regression function is near 1
or 0.
The same reasoning shows, though, that having a good fit in regions where
b(t) and µ(t) are on
µ(t) is mid-range, say in (0.25,0.75) is important. If µ
opposite sides of the number 0.5 (i.e., one below 0.5 and the other above),
we will make the wrong decision even, though we may still be lucky and
guess Y correctly. Say for instance µ(t) = 0.6. The correct decision would
b(t) = 0.42, our decision will be to guess Y = 0.
be to guess Y = 1, but if µ
However, our guess could turn out to be correct anyway.
In classification contexts, the p-variate density of X is often “mound-
shaped,” if not bell-shaped, within each class. (In fact, many clustering
algorithms are aimed at this situation.) For such data, the regions of most
sensitivity in the above sense will be the areas near the lines/curves sep-
arating the pairs of mounds. (Recall Figure 5.1.) The fringes of the data
set, far from these pairwise boundaries, will be regions in which model fit
is less important, again assuming a Prediction goal.
In regression contexts (continous Y , count data etc.), the full data set will
tend to be mound-shaped. Here good estimation will be important for
Prediction and Description throughout the entire region. However, one
must keep in mind that model fit will typically be better near the center
b(t) may be approximately linear near
of the data than at the fringes, i.e., µ
the center but depart considerably from linear on the fringes. Moreover,
the observations at the fringes typically have the heaviest impact on the
estimated coefficients. This latter consideration is of course of great import
in the Description case.
We will return to these considerations at various points in this chapter.
6.5. EXAMPLE: CURRENCY DATA 219

6.5 Example: Currency Data

Fong and Ouliaris [48] did an analysis of relations between currency rates for
Canada, Germany, France, the UK and Japan. This was in pre-European
Union days, with the currency names being the Canadian dollar, the Ger-
man mark, the French franc, the British pound and the Japanese yen. The
mark and franc are gone today, of course.
An example question of interest is do the currencies move up and down
together? We will assess this by predicting the Japanese yen from the
others.
The data can be downloaded at http://qed.econ.queensu.ca/jae/1995-v10.3/
fong-ouliaris/. The data set does require some wrangling, which we show
in Section 6.15.1 in the Computational Complements material at the end
of this chapter. In the end, we have a data frame curr. Here is the top
part of the data frame:
> head ( c u r r )
Canada Mark Franc Pound Yen
1 0.9770 2.575 4.763 0.41997 301.5
2 0.9768 2.615 4.818 0.42400 302.4
3 0.9776 2.630 4.806 0.42976 303.2
4 0.9882 2.663 4.825 0.43241 301.9
5 0.9864 2.654 4.796 0.43185 302.7
6 0.9876 2.663 4.818 0.43163 302.5

This is time series data, and the authors of the above paper do a very
sophisticated analysis along those lines. So, the data points, such as for the
pound, are not independent through time. But since we are just using the
data as an example and won’t be doing inference (confidence intervals and
significance tests), we will not worry about that here.
Let’s start with a straightforward linear model:
> f o u t <− lm( Yen ∼ . , data=c u r 1 )
> summary( f o u t )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( Intercept ) 102.855 14.663 7.015 5 . 1 2 e −12 ∗∗∗
Can −45.941 1 1 . 9 7 9 −3.835 0.000136 ∗∗∗
Mark 147.328 3.325 44.313 < 2 e −16 ∗∗∗
Franc −21.790 1 . 4 6 3 −14.893 < 2 e −16 ∗∗∗
Pound −48.771 1 4 . 5 5 3 −3.351 0.000844 ∗∗∗
220 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

...
M u l t i p l e R−s q u a r e d : 0 . 8 9 2 3 , Adjusted R−s q u a r e d : 0 . 8 9 1 8

Not surprisingly, this model works well, with an adjusted R-squared value
of about 0.89. The signs of the coefficients are interesting, with the yen
seeming to fluctuate opposite to all of the other currencies except for the
German mark. Of course, professional financial analysts (domain experts,
in the data science vernacular) should be consulted as to the reasons for
such relations, but here we will proceed without such knowledge.
It may be helpful to scale our data so as to better understand the roles
of the predictors, though, so as to make all the predictors commensurate
(Section 1.21). Each predictor will be divided by its standard deviation
(and have its mean subtracted off first), so all the predictors have standard
deviation 1.0:
> c u r r 1 <− as . matrix ( c u r r ) # t o e n a b l e scale ()
> c u r r 1 [ , − 5 ] <− s c a l e ( c u r r 1 [ , − 5 ] )
> f o u t 1 <− lm( Yen ∼ . , data=c u r r 1 )
> summary( f o u t 1 )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( Intercept ) 224.9451 0.6197 362.999 < 2 e −16 ∗∗∗
Can −5.6151 1 . 4 6 4 1 −3.835 0.000136 ∗∗∗
Mark 57.8886 1.3064 44.313 < 2 e −16 ∗∗∗
Franc −34.7027 2 . 3 3 0 1 −14.893 < 2 e −16 ∗∗∗
Pound −5.3316 1 . 5 9 0 9 −3.351 0.000844 ∗∗∗
...

So the German and French currencies appear to have the strongest rela-
tion to the yen. This, and the signs (positive for the mark, negative for
the franc), form a good example of the use of regression analysis for the
Description goal.
In the next few sections, we’ll use this example to illustrate the basic con-
cepts.

6.6 Overall Measures of Model Fit

We’ll look at two broad categories of fit assessment methods. The first
will consist of overall measures, while the second will involve relating fit to
individual predictor variables.
6.6. OVERALL MEASURES OF MODEL FIT 221

6.6.1 R-Squared, Revisited

We have already seen one overall measure of model fit, the R-squared value
(Section 2.9). As noted before, its cousin, Adjusted R-squared, is considered
more useful, as it is aimed at compensating for overfitting.
For the currency data above, the two R-squared values (ordinary and ad-
justed) were 0.8923 and 0.8918, both rather high. Note that they didn’t
differ much from each other, as there were well over 700 observations, which
should easily handle a model with only 4 predictors (a topic we’ll discuss
in Chapter 9).
Recall that R-squared, whether a population value or the sample estimate
reported by lm(), is the squared correlation between Y and its predicted
value µ(X) or µb(X), respectively. Thus it can be calculated for any method
of regression function estimation, not just the linear model. In particular,
we can apply the concept to k-Nearest Neighbor methods.
The point of doing this with k-NN is that the latter in principle does not
have model-fit issues. Whereas our linear model for the currency data
assumes a linear relationship between the yen and the other currencies, k-
NN makes no assumptions on the form of the regression function. If k-NN
were to have a substantially larger R-squared value than that of our linear
model, then we may be “leaving money on the table,” i.e., not fitting as
well as we could with a more sophisticated parametric model.3
This indeed seems to be the case:4
> library ( r e g t o o l s )
> c u r r 2 <− c u r r 1 [ − 7 6 2 , ]
> xdata <− p r e p r o c e s s x ( c u r r 2 [ , − 5 ] , 2 5 , x v a l=TRUE)
> kout <− k n n e s t ( c u r r 2 [ , 5 ] , xdata , 2 5 )
> cor ( kout$ r e g e s t , c u r r 2 [ , 5 ] ) ˆ 2
[ ,1]
[ 1 , ] 0.9817131

This would indicate that, in spite of a seemingly good fit for our linear
model, it does not adequately describe the currency fluctuation process.
So, our linear model, which seemed so nice at first, is missing something.
Maybe we can determine why via the methods in the sections below. But
3 We could of course simply use k-NN in the first place. But this would not give

us the Description usefulness of the parametric model, and also would give us a higher
estimation variance, since the parametric model is pretty good. See Section 1.7.
4 As noted in the data wrangling, the last row has some NA values, so we will omit it.
222 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

first, another measure of fit:

6.6.2 Cross-Validation, Revisited

As discussed before, this involves dividing our n data points into subsets of
r and n − r points. We do our data fitting on the first partition, then use
the results to predict the second partition. The motivation is to solve the
bias problems cited above.
One common variant is m-fold cross validation, where we divide the data
into m equal-sized subsets, and take r = n − n/m. We then perform the
above procedure m times. With m = n, we have the LOOM technique (Sec-
tion 2.9.5). This gives us more accuracy than using just one partitioning,
at the expense of needing much more computation. When we look at only
one partitioning, cross-validation is sometimes called the holdout method,
as we are “holding out” n − r data points from our fit.
In many discussions of fitting regression models, cross-validation is pre-
sented as a panacea. This is certainly not the case, however, and the reader
is advised to use it with caution. We will discuss this further in Section
9.3.2.

6.6.3 Plotting Parametric Fit Against a

Nonparametric One

b values of the linear model against those of k-NN:

Let’s plot the µ
> library ( r e g t o o l s )
> p a r v s n o n p a r p l o t ( f o u t 1 , kout )

The result is shown in Figure 6.1. It suggests that the linear model is
overestimating the regression function at times when the yen is very low or
very high, and possibly underestimating in the moderate range.
We must view this cautiously, though. First, of course, there is the issue
of sampling variation; the apparent model bias effects here may just be
sampling anomalies.
Second, k-NN itself is subject to some bias at the edges of a data set. This
will be discussed in detail in Section 11.1 (and a remedy presented for it),
but basically what happens is that k-NN tends to overestimate µ(t) when
the value is low and underestimate when µ(t) is high. The implication in
the currency case, k-NN tends to overestimate for low values of the yen,
6.6. OVERALL MEASURES OF MODEL FIT 223

Figure 6.1: Estimation of regression values, two methods

and underestimate at the high end. This can be addressed by doing locally-
linear smoothing, an option offered by knnest(), but let’s not use it for
now. And in any event, this k-NN edge bias effect would not entirely explain
the patterns we see in the figure.
The “hook shape” at the left end, and a “tail” in the middle suggest odd
nonlinear effects, possibly some local nonlinearities, which k-NN is picking
up but which the linear model misses.

6.6.4 Residuals vs. Smoothing

In any regression analysis, the quantities

(Xi )
r i = Yi − µ (6.3)
224 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

are traditionally called the residual values, or simply the residuals. They
are of course the prediction errors we obtain when fitting our model and
then predicting our Yi from our Xi . The smaller these values are in absolute
value, the better, but also we hope that they may inform us of inaccuracies
in our model, say nonlinear relations between Y and our predictor variables.
In the case of a linear model, the residuals are

ri = Yi − βb0 − βb1 Xi − ... − βbp Xi

(1) (p)
(6.4)

Many diagnostic methods for checking linear regression models are based
on residuals. In turn, their convenient computation typically involves first
computing the hat matrix, about which there is some material in the Math-
ematical Complements section at the end of this chapter.
The generic R function plot() can be applied to any object of class ”lm”
(including the subclass ”glm”). Let’s do that with fout1:
> plot ( f o u t 1 )
Hit <Return> t o see next plot :
Hit <Return> t o see next plot :
Hit <Return> t o see next plot :
Hit <Return> t o see next plot :

We obtain a series of graphs, displayed sequentially. Most of them involve

more intricate concepts than we’ll use in this book (recall Section 6.2), but
let’s look at the first plot, shown in Figure 6.2. The “hook” and “tail” are
visible here too.
Arguably the effects are clearer in Figure 6.1. This is due to the fact
that the latter figure is plotting smoothed values, not residuals. In other
words, residuals and smoothing play complementary roles to each other:
Smoothing-based plots can more easily give us “the big picture,” but resid-
uals may enable us to spot some fine details.
In any case, it’s clear that the linear model does not tell the whole story.

6.7 Diagnostics Related to Individual

Predictors

It may be that the relationship with the response variable Y is close to linear
for some predictors X (i) but not for others. How might we investigate this?
6.7. DIAGNOSTICS RELATED TO INDIVIDUAL PREDICTORS 225

Figure 6.2: Residuals against linear fitted values

6.7.1 Partial Residual Plots

We might approach this question by simply plotting a scatter diagram of Y

against each predictor variable. However, the relation of Y with one X (i)
may change in the presence of another X (j) . A more sophisticated approach
may be partial residual plots, also known as component + residual plots.
These would be easy to code on one’s own, but the crPlot() function
in the car package [49] does the job nicely for us. Continuing with the
currency data, we try

> library ( car )

> crPlots ( fout1 )

The resulting graph is shown in Figure 6.3. Before discussing these rather
bizarre results, let’s ask what these plots are depicting.
Here is how the partial-residual method works. The partial residuals for a
predictor X (j) are defined to be
226 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

Figure 6.3: Partial residuals plot, currency data

6.7. DIAGNOSTICS RELATED TO INDIVIDUAL PREDICTORS 227

= ri + βbj Xi
(j)
pi (6.5)
b b (1) b
= Yi − β0 − β1 X − ... − βj−1 X
(j−1) b
− βj+1 X
(j+1) b
− ... − βp X
(p)
i i i i

for i = 1, 2, ..., n.
In other words, we started with the residuals (6.4), but removed the linear
term contributed by predictor j, i.e., removed βbj Xi . We then plot the pi
(j)

against predictor j, to try to discern a relation. In effect we are saying,

Cancel that linear contribution of predictor j. Let’s start fresh

with this predictor, and see how adding it in a possibly nonlinear
form might extend the collective predictive ability of the other
predictors.

If the resulting graph looks nonlinear, we may profit from modifying our
model to one that reflects a nonlinear relation.
In that light, what might we glean from Figure 6.3? First, we see that the
only “clean” relations are the one for the franc and the one for the mark.
No wonder, then, that we found earlier that these two currencies seemed to
have the strongest linear relation to the yen. There does seem to be some
nonlinearity in the case of the franc, with a more negative slope for low
franc values, and this may be worth pursuing, say by adding a quadratic
term.
For the Canadian dollar and the pound, though, the relations don’t look
“clean” at all. On the contrary, the points in the graphs clump together
much more than we typically encounter in scatter plots.
But even the mark is not off the hook (pardon the pun), as the “hook” shape
noticed earlier is here for that currency, and apparently for the Canadian
dollar as well. So, whatever odd phenomenon is at work may be related to
these two currencies,

6.7.2 Plotting Nonparametric Fit Against

Each Predictor

As noted, one approach would be to draw many scatter diagrams, plotting

Y individually against each X (i) . But scatter diagrams are, well, scattered.
A better way is to plot the smoothed nonparametric fit, say using k-NN
228 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

Figure 6.4: Nonparametric fit against the mark

as is done below. against each predictor. The regtools function non-

parvsxplot() does this, plotting one graph for each predictor, presented
in succession with user prompts:

> n o n p a r v s x p l o t ( kout )
next plot
next plot
next plot
next plot

The graph for for the mark is shown in Figure 6.4. Oh my gosh! With the
partial residual plots, the mark and the franc seemed to be the only “clean”
ones. Now we see that the situation for the mark is much more complex.
The same is true for the other predictors (not shown here). This is indeed
a difficult data set.
Again, note that the use of smoothing has brought these effects into better
focus, as discussed in Section 6.6.4.
6.7. DIAGNOSTICS RELATED TO INDIVIDUAL PREDICTORS 229

6.7.3 The freqparcoord Package

Another graphical approach uses the freqparcoord package, written by

Yingkang Xie and me [104]. To explain this, we must first discuss the
notion of parallel coordinates, a method for visualizing multidimensional
data in a 2-dimensional graph.
6.7.3.1 Parallel Coordinates

The approach dates back to the 1800s, but was first developed in depth
in modern times, notably by Alfred Inselberg and Ed Wegman; see [74].
The method is motivated by the problem that scatter plots work fine for
displaying a paird of variables, but there is no direct multidimensional
analog. The use of parallel coordinates allows us to visualize many variables
at once.
The general method of parallel coordinates is quite simple. Here we draw p
vertical axes, one for each variable. For each of our n data points, we draw
a polygonal line from each axis to the next. The height of the line on axis
j is the value of variable j for this data point.
Figure 6.5 shows a very simple example, showing the polygonal lines repre-
senting two people. The first person, in the upper line, is 70 inches tall, is
25 years old, and weighs 72.73 kilograms, while the second person has val-
ues, 62, 66 and 95.45. By the way, though we have not centered and scaled
the data here (Section 1.21), some sort of scaling is typically applied, so
that the graph is more balanced in scale.
As we will see, parallel coordinates plots often enable analysts to obtain
highly valuable insights in their data, by exposing very telling patterns.
Several R functions to create parallel coordinates plots are available, such
as parcoord() in the MASS package included in base R; parallelplot()
in the lattice graphics package [125]; and ggparcoord in GGally, a
ggplot2-based graphics package [126].
As Inselberg pointed out, in mathematical terms the plot performs a trans-
formation mapping p-dimensional points to p − 1-segment lines. There is
an elegant geometric theory arising from this, but for us the practical effect
is that we can visualize how our p variables vary together.

6.7.3.2 The freqparcoord Package

One major problem will parallel coordinates is that if the number of data
points n is large, our plot will consist of a chaotic jumble of lines, maybe
230 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

Figure 6.5: Simple parallel coordinates plot

even with the “black screen problem,” meaning that so much has been
plotted that the graph is mostly black, no defining features.
One solution to that problem is taken by freqparcoord. It plots only
the most frequently-occurring lines.5 It starts with all the lines drawn
by ggparcoord(), but removes most of them, retaining only the most
frequently-occurring, i.e., the most representative ones.

6.7.3.3 The regdiag() Function

A function in the freqparcoord package, regdiag(), applies this to re-

gression diagnostics. The first variable plotted, i.e., the first vertical axis,
is what we call the divergences, meaning the differences beween the para-
5 There is an issue of what is meant by “most frequent.” If our variables take on only

integer values, this is clear, but if the variables are continuous, things are a little more
involved; it may well be the case that no two lines are exactly the same. Here we might
group together lines that are near each other. What freqparcoord does is use k-NN for
this, estimating the p-dimensional joint density function, then plotting only those points
that have the highest values of this function.
6.7. DIAGNOSTICS RELATED TO INDIVIDUAL PREDICTORS 231

Figure 6.6: Freqparcoord plot, currency data

metric and nonparametric estimates of the population regression function,

blinmod (Xi ) − µ
µ bknn (Xi ), i = 1, ..., n (6.6)

The other axes represent our predictor variables. Vertical measures are
numbers of standard deviations from the mean of the given variable.
The code
> library ( freqparcoord )
> regdiag ( fout1 )

produces the graph in Figure 6.6.

Here we plot five variables, consisting of the divergences and the four cur-
rencies. Each variable is centered and scaled.
There are three groups, thus three subgraphs, for the upper 10%, middle
80% and lower 10% of the divergence values, labeled 1, 2 and 3 on the
right margin. (The scaling of the variables, including the divergences, is
done on the basis of the data as a whole, not within the subgroups.) So for
232 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

instance the upper subgraph describes data points Xi at which the linear
model greatly underestimates the true regression function.
What we see, then, is that in regions in which the linear model underes-
timates, the Canadian dollar tends to be high and the mark low, with an
opposite relation for the region of overestimation. Note that this is not the
same as saying that the correlation between those two currencies is neg-
ative; on the contrary, running cor(curr1) shows their correlation to be
positive and tiny, about 0.01. This suggests that we might try adding a
dollar/mark interaction term to our model, though the effect here seems
mild, with peaks and valleys of only about 1 standard deviation..

6.8 Effects of Unusual Observations on Model

Fit

Suppose we are doing a study of rural consumer behavior in a small country

C in the developing world. One day, a few billionaires discover the idyllic
beauty of rural C and decide to move there. We almost certainly would
want to exclude data on these interlopers from our analysis. Second, most
data contain errors. Obviously these must be excluded too, or corrected if
possible. These two types of observations are sometimes collectively called
outliers, or simply unusual.
Though we may hear that Bill Gates has moved to C, and we can clean
the data to remove the obvious errors, such as a human height of 25 feet
or a negative weight. other extreme values or errors may not jump out at
us. Thus it would be useful to have methods that attempt to find such
observations in some organized, mechanical way.

6.8.1 The influence() Function

Base R includes a very handy function, influence(). We input an object

of type ”lm” or ”glm”, and it returns an R list. One of the components
of that list, coefficients, is just what we want: It has a column for each
βbj , and a row for each observation in our data set. Row i, column j tells
us how much βbj would change if observation i were deleted from the data
set.6 If the change is large, we should take a close look at observation i, to
6 The entire computation does not need to be done from scratch. The Sherman-

Morrison-Woodbury formula provides a shortcut. See the Mathematical Complements

section at the end of this chapter.
6.8. EFFECTS OF UNUSUAL OBSERVATIONS ON MODEL FIT 233

determine whether it is “unusual.”

6.8.1.1 Example: Currency Data

Let’s take a look, continuing our analysis in Section 6.5:

> i n f c f s <− influence ( f o u t 1 ) $coef

> head ( i n f c f s )
( Intercept ) Can Mark Franc
1 −0.01538183 0 . 0 3 1 1 4 3 6 8 −0.009961471 −0.07634899
2 −0.02018040 0 . 0 3 7 4 3 1 3 5 −0.018141461 −0.09488785
3 −0.02196501 0 . 0 3 5 8 3 0 0 0 −0.024654614 −0.08885551
4 −0.02877846 0 . 0 2 5 7 3 9 2 6 −0.050914092 −0.08862127
5 −0.02693571 0 . 0 2 5 6 4 7 9 9 −0.046012297 −0.08261002
6 −0.02827297 0 . 0 2 5 2 4 0 5 1 −0.050186868 −0.08733993
Pound
1 0.07751692
2 0.10129967
3 0.10190745
4 0.13201466
5 0.12140985
6 0.13027910

So, if we were to remove the second data point, this says that βb0 would
decline by 0.02018040, βb1 would increase by 0.03743135, and so on. Let’s
check to be sure:

> coef ( f o u t 1 )
( Intercept ) Can Mark Franc
224.945099 −5.615144 57.888556 −34.702731
Pound
−5.331583
> coef (lm( Yen ∼ . , data=c u r r 1 [ − 2 , ] ) )
( Intercept ) Can Mark Franc
224.965279 −5.652575 57.906698 −34.607843
Pound
−5.432882
> −5.652575 + 0.037431
[ 1 ] −5.615144

Ah, it checks. Now let’s find which points have large influence.
234 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

A change in an estimated coefficient should be considered “large” only rel-

ative to the standard error, so let’s scale accordingly, dividing each change
by the standard error of the corresponding coefficient:7
> se <− sqrt ( diag ( vcov ( f o u t 1 ) ) )
> i n f c f s <− i n f c f s %∗% diag ( 1 /se )

So, how big do the changes brought by deletions get in this data? And for
which observations does this occur? Let’s take a look.8
> i a <− abs ( i n f c f s )
> max( i a )
[ 1 ] 0.1928661
> f 1 5 <− function ( rw ) any ( rw > 0 . 1 5 )
> i a 1 5 <− apply ( i a , 1 , f 1 5 )
> names( i a 1 5 ) <− NULL
> which ( i a 1 5 )
[ 1 ] 744 745 747 748 749 750 751 752 753 754 755
[ 1 2 ] 756 757 758 759 760 761

Here we (somewhat arbitrarily) decided to identify which deletions of ob-

servations would result in an absolute change of some coefficient of more
than 0.15.
Now this is interesting. There are 761 observations in this data set, and
now we find that all of the final 18 (and more) are influential. Let’s look
more closely:
> t a i l ( ia , 5 )
[ ,1] [ ,2] [ ,3] [ ,4]
757 0 . 0 5 5 8 5 5 3 8 0 . 1 3 1 1 1 1 0 0.1572411 0.1305925
758 0 . 0 5 8 5 1 0 8 7 0 . 1 2 9 4 4 1 2 0.1563013 0.1259741
759 0 . 0 5 8 3 8 8 1 3 0 . 1 3 8 6 6 1 4 0.1629851 0.1358875
760 0 . 0 5 8 1 8 7 3 0 0 . 1 4 2 9 9 5 1 0.1654354 0.1385146
761 0 . 0 5 6 2 6 2 1 2 0 . 1 3 1 6 8 8 4 0.1534207 0.1211305
[ ,5]
757 0 . 0 2 1 3 0 0 1 7 7
758 0 . 0 1 5 7 0 1 0 0 5
759 0 . 0 2 0 4 3 1 3 9 1
760 0 . 0 1 9 9 6 2 5 0 2
761 0 . 0 0 6 7 9 3 6 7 3
7 R’s diag() function is quite versatile. In our first call here, we are extracting the

diagonal elements of the estimated covariance matrix of our coefficients, placing the
result in a vector. In the second call, we are creating a diagonal matrix from a vector.
8 The reader may wish to review Section 1.20.2 before continuing.
6.8. EFFECTS OF UNUSUAL OBSERVATIONS ON MODEL FIT 235

So, the influence of these final observations was on the coefficients of the
Canadian dollar, the mark and the franc — but not on the one for the
pound.
Something special was happening in those last time periods. It would be
imperative for us to track this down with currency experts.
Each of the observations has something like a 0.15 impact, and intuitively,
removing all of these observations should cause quite a change. Let’s see:
> c u r r 3 <− c u r r 1 [ − ( 7 4 4 : 7 6 1 ) , ]
> lm( Yen ∼ . , data=c u r r 3 )
...
Coefficients :
( Intercept ) Canada Mark
225.780 −10.271 52.926
Franc Pound
−27.126 −6.431

> fout1
...
Coefficients :
( Intercept ) Canada Mark
224.945 −5.615 57.889
Franc Pound
−34.703 −5.332

These are very substantial changes! The coefficient for the Canadian cur-
rency almost doubled, and even the pound’s value changed almost 30%.
That latter is a dramatic difference, in view of the fact that each individual
observation had only about a 2% influence on the pound.
A collection of advanced influence measures is provided by another R func-
tion, whose name is, not surprisingly, influence.measures().

6.8.2 Use of freqparcoord for Outlier Detection

The versatile freqparcoord package can also be used for outlier detection.
Here we find the least-frequent points, rather than the ones with highest
frequemcy as before. Continuing the currency example, we run this code:
> f r e q p a r c o o r d ( c u r r 1 ,m=−5,method= ’ maxdens ’ ,
k e e p i d x s =1)$ i d x s
[ 1 ] 547 548 549 551 550
236 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

The argument m indicates how many observations we wish to be reported,

i.e., how many polygonal lines to plot. A negative value, combined with an
argument method=’maxdens’, specifies that we want to plot the least
frequent cases, if keepidxs is not NULL.9 The idxs component of the
return value gives us the indices of the possible outliers.
Again, a set of consecutive observations turned out to be troubling, And
for instance, we might check curr1[547,] for possible errors or other unde-
sirable characteristics.

6.9 Automated Outlier Resistance

The term robust in statistics generally means the ability of methodology to

withstand problematic conditions. Linear regression models, for instance,
are said to be “robust to the normality” assumption, meaning that (at least
large-sample) inference on the coefficients will work well even though the
distribution of Y given X is not normal.
But here we are concerned with robustness of regression models to outliers.
Such methods are called robust regression. There are many such methods,
one of which is median regression, to be discussed next.

6.9.1 Median Regression

Suppose we wish to estimate the mean of some variable in some population,

but we are concerned about unusual observations. As an alternative, we
might consider estimating the median value of the variable, which will be
much less sensitive to unusual observations. Recall our hypothetical exam-
ple earlier, in which we were interested in income distributions. If one very
rich person moves into the area, the mean may be affected substantially —
but the median would likely not change at all.
We thus say that the median is robust to unusual data points. One can do
the same thing to make regression analysis robust in this sense.
Denote the conditional median of Y given X = t, i.e., median(Y | X = t),
by ν(t). It turns out (see the Mathematical Complements section at the
end of this chapter) that

ν(t) = argminm E(|Y − m| | X = t) (6.7)

9 If not NULL, this argument states the variable on which we want the results sorted.
6.9. AUTOMATED OUTLIER RESISTANCE 237

In other words, in contrast to the regression function, i.e., the conditional

mean, which minimizes mean squared prediction error, the conditional me-
dian minimizes mean absolute error.
Remember, as with regression, we are estimating an entire function here,
as t varies. A nonparametric approach to this would be to use knnest()
with nearf set to
function ( predpt , nearxy )
{
y c o l <− ncol ( nearxy )
median( nearxy [ , y c o l ] )
}

However, in this chapter we are primarily concerned with parametric mod-

els. So, we might, in analogy to the linear regression model, make the
assumption that (6.7) has the familiar linear form

ν(t) = β0 + β1 t1 + ... + βp tp (6.8)

Solving this at the sample level is a linear programming problem, which has
been implemented in the CRAN package quantreg [81]. As the package
name implies, we can estimate general conditional quantile functions, not
just the conditional median. The argument tau of the rq() function spec-
ifies what quantile we want, with the value 0.5, i.e., the median, being the
default.
It is important to understand that ν(t) is not the regression function, i.e.,
not the conditional mean. Thus rq() is not estimating the same quantity
as is lm(). Thus the term quantile regression, in this case the term me-
dian regression, is somewhat misleading here. But we can use ν(t) as an
alternative to µ(t) in one of two senses:

(a) We may believe that ν(t) is close to µ(t). They will be exactly the
same, of course, if the conditional distribution of Y given X is sym-
metric, at least if the unusual observations are excluded.

(b) We may take the point of view that the conditional median is just as
meaningful as the conditional mean (no pun intended this time), so
why not simply model ν(t) in the first place?

Sense (a) above will be particularly relevant here.

238 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

It should be noted, though, that there are reasons not to use median re-
gression. If estimating the mean of a normally distributed random variable,
the sample median’s efficiency relative to the sample mean, meaning the
ratio of asymptotic variances, can be shown to be only 2/π ≈ 0.64. In other
words, by using the median regression model rather than a linear one, there
is not only the question of whether the models are close to reality, but also
that we risk having estimators with large standard errors.

6.9.2 Example: Currency Data

Let’s apply rq() to the currency data:

> qout <− rq ( Yen ∼ . , data=c u r r 1 )
> qout
...
Coefficients :
( Intercept ) Can Mark Franc
2 2 4 . 5 1 7 8 9 9 −11.038238 5 3 . 8 5 4 0 0 5 −27.443584
Pound
−5.320035
...
> fout1
...
Coefficients :
( Intercept ) Can Mark
224.945 −5.615 57.889
Franc Pound
−34.703 −5.332

The results are strikingly similar to what we obtained in Section 6.8.1.1 by

calling lm() with the bad observations at the end of the data set removed.
In other words,

Median regression can be viewed as an automated method for

removing (the effects of) the unusual data points.

6.10 Example: Vocabulary Acquisition

The Wordbank data, http://wordbank.stanford.edu/, concerns child vocabu-

lary development, in not only English but also a number of other languages,
6.10. EXAMPLE: VOCABULARY ACQUISITION 239

such as Cantonese and Turkish. These are mainly toddlers, ages from about
a year to 2.5 years.
Let’s read in the English set:
> e n g l <− read . csv ( ’ E n g l i s h . c s v ’ )
# take a look
> head ( e n g l )
data i d age l a n g u a g e form b i r t h order e t h n i c i t y
1 1 24 E n g l i s h WS First Asian
2 2 19 E n g l i s h WS Second Black
3 3 24 E n g l i s h WS First Other
4 4 18 E n g l i s h WS First White
5 5 24 E n g l i s h WS First White
6 6 19 E n g l i s h WS First Other
sex mom ed measure vocab demo
1 Female Graduate p r o d u c t i o n 337 A l l Data
2 Female College production 384 A l l Data
3 Male Some Secondary p r o d u c t i o n 76 A l l Data
4 Male Secondary p r o d u c t i o n 19 A l l Data
5 Female Secondary p r o d u c t i o n 480 A l l Data
6 Female Some C o l l e g e p r o d u c t i o n 313 A l l Data
n demo l a b e l
1 5498 A l l Data ( n = 5 4 9 8 )
2 5498 A l l Data ( n = 5 4 9 8 )
3 5498 A l l Data ( n = 5 4 9 8 )
4 5498 A l l Data ( n = 5 4 9 8 )
5 5498 A l l Data ( n = 5 4 9 8 )
6 5498 A l l Data ( n = 5 4 9 8 )

We have a number of R factors here, such as birth order, so we need to do

some data wrangling here. The details are presented in Section 6.15.2. The
result is
> head ( e n c c )
age b i r t h order mom ed vocab male a s i a n b l a c k
1 24 1 20 337 0 1 0
2 19 2 16 384 0 0 1
3 24 1 10 76 1 0 0
4 18 1 12 19 1 0 0
5 24 1 12 480 0 0 0
6 19 1 14 313 0 0 0
l a t i n o othernonwhite
1 0 0
240 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

2 0 0
3 0 1
4 0 0
5 0 0
6 0 1

Running knnest() (not shown), there seemed to be an approximately linear

relation between vocabulary and age (for the age range studied). Let’s run
a linear regression analysis:
> summary(lm( vocab ∼ . , data=e n c c ) )
...
Coefficients :
Estimate Std . E r r o r t v a l u e
( Intercept ) −475.8761 2 3 . 5 0 5 9 −20.245
age 33.3968 0.6422 52.005
b i r t h order −20.5033 3 . 0 8 9 1 −6.637
mom ed 3.8680 0.9780 3.955
male −48.9628 5 . 4 8 4 2 −8.928
asian −16.0745 1 7 . 2 9 4 7 −0.929
black 1.1094 10.2321 0.108
latino −75.5545 1 3 . 0 1 7 2 −5.804
o t h e r n o n w h i t e −53.9844 1 5 . 1 2 9 1 −3.568
Pr ( >| t | )
( Intercept ) < 2 e −16 ∗∗∗
age < 2 e −16 ∗∗∗
b i r t h order 3 . 8 4 e −11 ∗∗∗
mom ed 7 . 8 5 e −05 ∗∗∗
male < 2 e −16 ∗∗∗
asian 0.352741
black 0.913671
latino 7 . 2 1 e −09 ∗∗∗
o t h e r n o n w h i t e 0 . 0 0 0 3 6 6 ∗∗∗
...
Mult . R−s q u a r e d : 0 . 5 1 3 2 , Adj . R−s q u a r e d : 0.5118

Age is recorded in months. So, the kids seemed to be learning about 33

words per month during the studied age range. Latino kids seem to start
slowly, possibly due to speaking Spanish at home. Consistent with the
general notion that girls develop faster than boys, the latter have a slower
start. Having a lot of older siblings also seems to be related negatively,
possibly due to the child being one of several competing for the parents’
attention. Having a mother with more education had a modest positive
6.11. CLASSIFICATION SETTINGS 241

effect, about 19 words for every extra 5 years of schooling.

A word on the standard errors: As each child was measured multiple times
as he/she aged, the observations are not independent, and the true standard
errors are larger than those given.
Let’s try median regression instead:
> library ( quantreg )
> rq ( vocab ∼ . , data=e n c c )
...
Coefficients :
( Intercept ) age b i r t h order
−581.73077 37.84615 −19.50000
mom ed male asian
4.00000 −43.42308 −20.07692
black l a t i n o othernonwhite
−13.73077 −67.65385 −49.65385

(There was also a warning, “Solution may be nonunique.”)

The robust results here are similar to what we obtained earlier, but with
some modest shifts.
It is often worthwhile to investigate other quantiles than the median. Trying
that for age only:
> plot ( c ( 1 2 , 3 0 ) , c ( 0 , 8 0 0 ) , type = ”n” , x l a b = ” age ” ,
y l a b = ” vocab ” )
> abline ( coef ( rq ( vocab ∼ age , data=encc , tau = 0 . 1 ) ) )
> abline ( coef ( rq ( vocab ∼ age , data=encc , tau = 0 . 5 ) ) )
> abline ( coef ( rq ( vocab ∼ age , data=encc , tau = 0 . 9 ) ) )

As seen in Figure 6.7, the middle-level children start out knowing many
fewer words than the most voluble ones, but narrow the gap over time.
By contrast, the kids with smaller vocabularies start out around the same
level as the middle kids, but actually lose ground over time, suggesting that
educational interventions may be helpful.

6.11 Classification Settings

Since we treat classification as a special case of regression, we can use the

same fit assessment methods, though in some cases some adapting of them
is desirable.
242 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

Figure 6.7: Vocabulary vs. age

6.11.1 Example: Pima Diabetes Study

Let’s illustrate with the Pima diabetes data set from Section 4.3.2.
> pima <− read . csv ( ’ pima−i n d i a n s −d i a b e t e s . data ’ )

It goes without saying that with any data set, we should first do proper
cleaning.10 This data is actually a very good example. Let’s first try the
freqparcoord package:
> library ( fr eqparcoord )
> f r e q p a r c o o r d ( pima [ , −9] , −10)

Here we display those 10 data points (predictors only, not response variable)
whose estimated joint density is lowest, thus qualifying as “unusual.”
The graph is shown in Figure 6.8. Again we see a jumble of lines, but look
at the big dips in the variables BP and BMI, blood pressure and Body
Mass Index. They seem unusual. Let’s look more closely at blood pressure:
10 And of course we should have done so for the other data earlier in this chapter, but
we will keep the first analyses simple.
6.11. CLASSIFICATION SETTINGS 243

> table ( pima$BP)

0 24 30 38 40 44 46 48 50 52 54 55
35 1 2 1 1 4 2 5 13 11 11 2
56 58 60 61 62 64 65 66 68 70 72 74
12 21 37 1 34 43 7 30 45 57 44 52
75 76 78 80 82 84 85 86 88 90 92 94
8 39 45 40 30 23 6 21 25 22 8 6
95 96 98 100 102 104 106 108 110 114 122
1 4 3 3 1 2 3 2 3 1 1

One cannot have a blood pressure of 0, yet 35 women in our data set are
reported as such. The value 24 is suspect too, but the 0s are wrong for
sure. What about BMI?
> table ( pima$BMI)

0 18.2 18.4 19.1 19.3 19.4 19.5 19.6 19.9 20

11 3 1 1 1 1 2 3 1 1
20.1 20.4 20.8 21 2 1 . 1 2 1 . 2 2 1 . 7 2 1 . 8 2 1 . 9 2 2 . 1
...

Here again, the 0s are clearly wrong. So, at the very least, let’s exclude
such data points:
> pima <− pima [ pima$BP > 0 & pima$BMI > 0 , ]
> dim( pima )
[ 1 ] 729 9

(We lost 38 cases.)

Now, for our analysis, start with fitting a logit model, then comparing to
k-NN. First, what value of k should we use?:
Let’s go with 50, and compare the parametric and nonparametric fits:
> kout <− k n n e s t ( pima$Diab , xdata , 5 0 )
> p a r v s n o n p a r p l o t ( glmout , kout )

The results of the plot are shown in Figure 6.10. There does appear to
be some overestimation by the logit at very high values of the regression
function, indeed all the range past 0.5. This can’t be explained by the fact,
noted before, that k-NN tends to underestimate at the high end.
Note carefully that if our goal is Prediction, it may not matter much at the
high end. Recall the discussion on classification contexts in Section 6.4.1. If
244 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

Figure 6.8: Outlier hunt (see color insert)

Figure 6.9: Best k for Pima

6.12. IMPROVING FIT 245

the true population regression value is 0.8 and we estimate it to be 0.88, we

still predict Y = 1, which is the same as what we would predict if we knew
the true population regression function. Similarly, if the true regression
value is 0.25 but we estimate it to be 0.28, we still make the proper guess
for Y .
For Description, though, we should consider a richer model. Running non-
parvsxplot() (not shown) suggests adding quadratic terms for the vari-
ables Gluc, Insul, Genet and especially Age. Adding these, and rerun-
ning knnest() with nearf = loclin to deal with k-NN’s high-end bias,11
our new parametric-vs.-nonparametric plot is shown in Figure 6.11.
The reader may ask if we now have underestimation by the parametric
model at the high end, but we must take into account the fact that with
nearf = loclin, we can get nonparametric estimates for the regression
function that are smaller than 0 or greater than 1, which is impossible in
the classification setting. The perceived “underestimation” actually occurs
at values at which the nonparametric figures are larger than 1.
In other words, we now seem to have a pretty good model.

6.12 Improving Fit

The currency example seemed so simple at first, with a very nice adjusted
R-squared value of 0.89, and with the yen seeming to have a clean linear
relation with the franc and the mark. And yet we later encountered some
troubling aspects to this data.
First we noticed that the adjusted R-squared value for the k-NN fit was
even better, at 0.98. Thus there is more to this data than simple linear
relationships. Later we found that the last 18 data points, possibly more,
have an inordinate influence on the βbj . This too could be a reflection of
nonlinear relationships between the currencies. The plots exhibited some
strange, even grotesque, relations.
So, let’s see what we might do to improve our parametric model.

6.12.1 Deleting Terms from the Model

Predictors with very little relation to the response variable may actually
degrade the fit, and we should consider deleting them. This topic is treated
11 This topic is covered in Section 11.1.
246 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

Figure 6.10: Estimation of regression values, two methods (I)

Figure 6.11: Estimation of regression values, two methods (II)

6.12. IMPROVING FIT 247

in depth in Chapter 9.

6.12.2 Adding Polynomial Terms

Our current model is linear in the variables. We might add second-degree

terms. Note that this means not only squares of the variables, but products
of pairs of them. The latter may be important, in view of our comment
in Section 6.7.3 that it might be useful to add a Canadian dollar/mark
interaction term to our model.

6.12.2.1 Example: Currency Data

Let’s add squared terms for each variable, and try the interaction term as
well. Here’s what we get:
> c u r r 2 <− c u r r 1
> c u r r 2 $C2 <− c u r r 2 $Canada ˆ2
> c u r r 2 $M2 <− c u r r 2 $Markˆ2
> c u r r 2 $F2 <− c u r r 2 $Franc ˆ2
> c u r r 2 $P2 <− c u r r 2 $Poundˆ2
> c u r r 2 $CM <− c u r r 2 $Canada∗ c u r r 2 $Mark
> summary(lm( Yen ∼ . , data=c u r r 2 ) )
...
Coefficients :
Estimate Std . E r r o r t v a l u e
( Intercept ) 223.575386 1.270220 176.013
Can −8.111223 1 . 5 4 0 2 9 1 −5.266
Mark 50.730731 1.804143 28.119
Franc −34.082155 2 . 5 4 3 6 3 9 −13.399
Pound −3.100987 1 . 6 9 9 2 8 9 −1.825
C2 −1.514778 0 . 8 4 8 2 4 0 −1.786
M2 −7.113813 1 . 1 7 5 1 6 1 −6.053
F2 11.182524 1.734476 6.447
P2 −1.182451 0 . 9 7 7 6 9 2 −1.209
CM 0.003089 1.432842 0.002
Pr ( >| t | )
( I n t e r c e p t ) < 2 e −16 ∗∗∗
Can 1 . 8 2 e −07 ∗∗∗
Mark < 2 e −16 ∗∗∗
Franc < 2 e −16 ∗∗∗
Pound 0.0684 .
248 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

C2 0.0745 .
M2 2 . 2 4 e −09 ∗∗∗
F2 2 . 0 4 e −10 ∗∗∗
P2 0.2269
CM 0.9983
−−−
...
M u l t i p l e R−s q u a r e d : 0 . 9 0 4 3 , Adj . R−s q u a r e d : 0 . 9 0 3 2

Adjusted R-squared increased only slightly. And this was despite the fact
that two of the squared-variable terms were “highly significant,” adorned
with three asterisks, showing how misleading significance testing can be.
The interaction term came out tiny, 0.003089. So, k-NN is still the winner
here.

6.12.2.2 Example: Census Data

Let’s take another look at the census data on programmers and engineers
in Silicon Valley, first introduced in Section 1.16.1.
We run
> data ( prgeng )
> pe <− prgeng # s e e ? k n n e s t
> # dummies f o r MS, PhD
> pe$ms <− as . integer ( pe$educ == 1 4 )
> pe$phd <− as . integer ( pe$educ == 1 6 )
> # computer o c c u p a t i o n s o n l y
> p e c s <− pe [ pe$ o c c >= 100 & pe$ o c c <= 1 0 9 , ]
> p e c s 1 <− p e c s [ , c ( 1 , 7 , 9 , 1 2 , 1 3 , 8 ) ]
> # p r e d i c t wage income from age , g e n d e r e t c .
> # p r e p a r e n e a r e s t −n e i g h b o r d a t a
> xdata <− p r e p r o c e s s x ( p e c s 1 [ , 1 : 5 ] , 1 5 0 )
> z o u t <− k n n e s t ( p e c s 1 [ , 6 ] , xdata , 5 )
> nonparvsxplot ( zout )

We find that the age variable, and possibly wkswrkd, seem to have a
quadratic relation to wageinc, as seen in Figures 6.12 and 6.13. So, let’s
try adding quadratic terms for those two variables. And, to assess how well
this works, let’s break the data into training and test sets:
> p e c s 2 <− p e c s 1
> p e c s 2 $ age2 <− p e c s 1 $ age ˆ2
6.12. IMPROVING FIT 249

Figure 6.12: Mean wage income vs. age

Figure 6.13: Mean wage income vs. weeks worked

250 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

> p e c s 2 $wks2 <− p e c s 1 $wkswrkd ˆ2

> n <− nrow( p e c s 1 )
> t r n i d x s <− sample ( 1 : n , 1 2 0 0 0 )
> p r e d i d x s <− s e t d i f f ( 1 : n , t r n i d x s )
> lmout1 <− lm( wageinc ∼ . , data=p e c s 1 [ t r n i d x s , ] )
> lmout2 <− lm( wageinc ∼ . , data=p e c s 2 [ t r n i d x s , ] )
> lmpred1 <− predict ( lmout1 , p e c s 1 [ p r e d i d x s , ] )
> lmpred2 <− predict ( lmout2 , p e c s 2 [ p r e d i d x s , ] )
> ypred <− p e c s 1 $ wageinc [ p r e d i d x s ]
> mean( abs ( ypred−lmpred1 ) )
[ 1 ] 25721.5
> mean( abs ( ypred−lmpred2 ) )
[ 1 ] 25381.08

So, adding the quadratic terms helped slightly, about a 1.3% improvement.
From a Prediction point of view, this is at best mild, There was also a
slight increase in adjusted R-squared, from 0.22 (not shown) to 0.23 (shown
below).
But for Description things are much more useful here:
> summary( lmout2 )
...
Coefficients :
Estimate Std . E r r o r t v a l u e
( I n t e r c e p t ) −63812.415 4 4 7 1 . 6 0 2 −14.271
age 3795.057 221.615 17.125
sex −10336.835 8 4 1 . 0 6 7 −12.290
wkswrkd 598.969 131.499 4.555
ms 14810.929 928.536 15.951
phd 20557.235 2197.921 9.353
age2 −39.833 2 . 6 0 8 −15.271
wks2 9.874 2.213 4.462
Pr ( >| t | )
( I n t e r c e p t ) < 2 e −16 ∗∗∗
age < 2 e −16 ∗∗∗
sex < 2 e −16 ∗∗∗
wkswrkd 5 . 2 9 e −06 ∗∗∗
ms < 2 e −16 ∗∗∗
phd < 2 e −16 ∗∗∗
age2 < 2 e −16 ∗∗∗
wks2 8 . 2 0 e −06 ∗∗∗
...
M u l t i p l e R−s q u a r e d : 0 . 2 3 8 5 , Adj . R−s q u a r e d : 0.2381
6.12. IMPROVING FIT 251

As usual, we should not make too much of the p-values, especially with a
sample size this large (16411 for pecs1). So, all those asterisks don’t tell
us too much. But a confidence interval computed from the standard error
shows that the absolute age-squared effect is at least about 34, far from 0,
and it does make a difference, say on the first person in the sample:
> predict ( lmout1 , p e c s 1 [ 1 , ] )
1
62406.48
> predict ( lmout2 , p e c s 2 [ 1 , ] )
1
63471.53

The more sophisticated model predicts about an extra $1,000 in wages for
this person.
Most important, the negative sign for the age-squared coefficient shows that
income tends to level off and even decline with age, something that could
be quite interesting in a Description-based analysis.
The positive sign for wkswrkd is likely due to the fact that full-time work-
ers tend to have better jobs.

6.12.3 Boosting

One of the techniques that has caused the most excitement in the machine
learning community is boosting, which in essence is a process of iteratively
refining, through reweighting, estimated regression and classification func-
tions (though it has primarily been applied to the latter).
This is a very complex topic, with many variations, and is basically beyond
the scope of this book. However, we will present an overview.

6.12.3.1 View from the 30,000-Foot Level

The main idea of boosting is to perform an iterative refitting of the model,

adjusting weights of the observations at each iteration. At any given step,
the observations predicted most poorly at the last step will now get larger
weights.
Berk [16] gives an excellent description of the basic philosophy of boost-
ing, by outlining a procedure that captures the method’s spirit. Here is a
modified version of his prescription, for the case of the linear model:
252 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

(a) Call lm() on the data (X1 , Y1 ), ..., (Xn , Yn ) as usual, yielding the
vector of estimated coefficients, βb(0) .

Repeat steps (b)-(c) for i = 1, ..., k:

(b) Form the residuals from the latest beta vector,

rj = Yj − (βb0 + βb1 Xi−1,1 + ...βbp(i−1) Xip−1, )

(i−1) (i−1)
(6.9)

Also compute

∑
n
di−1 = |rj | (6.10)
j=1

our total absolute prediction error.

(c) Run a weighted least squares analysis as in (3.3.2) (weights argument
in lm()), with

wj = |rj | (6.11)

yielding a new vector of estimated coefficients, βb(i) . We are giving the

observations with worse prediction errors more weight, in the hope of
getting better predictions at those points.

Finally:

(d) Compute dk as in (6.10), and set the final estimated coefficient vector
to

∑
k
βb = qs βb(s) (6.12)
s=0

where

1/ds
qs = ∑k (6.13)
t 1/dt

In other words, take a weighted average of the estimated coefficient

vectors we’ve computed, with the ones with better total prediction
error getting more weight.
6.12. IMPROVING FIT 253

The classification case is similar. In the 2-class setting, for instance, we

might set

rj = |Yj − Ybj | (6.14)

where Ybj is our predicted value for Yj , either 0 or 1. In step (d) we could use
“voting,” as with AVA in Section 5.5, but with the votes being weighted.
There are many, many variations — AdaBoost, gradient boosting and so
on — and their details are beyond the scope of this book. But the above
captures the essence of the method.
Why do all this? The key issue (often lost in the technical discussions) is
bias, in this case model bias, as follows. Putting aside issues of possible het-
eroscedasticity, ordinary (i.e., unweighted) least squares (OLS) estimation
is optimal for homoscedastic linear models: the Gauss-Markov Theorem
(Section 6.16.4) shows that OLS gives minimum variance among all un-
biased estimators. If we are in this setting, why use weights, especially
weights that come from such an involved process?
The answer is that the linear model is rarely if ever exactly correct. Thus
use of a linear model will result in bias; in some regions of X, the model will
overestimate, while in others it will underestimate — no matter how large
our sample is. We saw indications of this with our currency data earlier in
this chapter. It thus may be profitable to try to reduce bias in regions in
which our present predictions are very bad, at the hopefully small sacrifice
of some prediction accuracy in places where presently we are doing well.
The reweighting process is aimed at achieving a positive tradeoff of that
nature.
That tradeoff may be particularly useful in classification settings. As noted
in Section 6.4.1, in such settings, we can tolerate large errors in µb(t) on
the fringes of the dataset, so placing more weight in the middle, near the
classification boundary, could be a win.

6.12.3.2 Performance

Much has been made of the remark by the late statistician Leo Breiman,
that boosting is “the best off-the-shelf classifier in the world” [8], his term
off-the-shelf meaning that the given method can be used by nonspecialist
users without special tweaking. His statement has perhaps been overin-
terpreted (see Section 1.13), but many analysts have indeed reported that
some improvement (though not dramatic) results from the method. On the
254 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

other hand, it can also perform poorly relative to the nonboosting analysis
in some case. And it does seem to require good use of tuning parameters,
not “off the shelf” after all.
Note of course that improvement must be measured in terms of accuracy
in predicting new cases. Since boosting is aimed at reducing errors in
individual observations, there is a definite tendency toward overfitting.

6.13 A Tool to Aid Model Selection

In Section 5.5.3, we discussed the effects of nonlinear boundaries between

classes, and presented some possible solutions. In light of such situations,
a tool that helps us view those boundaries can be helpful. The regtools
package includes a plotting function, pwplot(), motivated by this consid-
eration. As our example, let’s again take the Pima diabetes data, Section
4.3.2.
We cannot view things in p dimensions, so pwplot() displays two predic-
tors, X (i) and X (j) at a time. Those two variables form the horizontal
and vertical axes, and knnest(), k-NN estimation, is applied, predicting
Y from X i) and X (j) . For each point the following computation is done:
We compare the estimated values of P (Y = 1 | X (i) , X (j) ) and P (Y = 1),
and then determine which is larger. The point here is to ask, Does knowing
X i and X j make Y = 1 more or less likely to occur, relative to having no
knowledge of those covariates? A ’1’ is plotted for each point for which the
answer is “more likely,” with ’0’ being plotted otherwise.
One of the function’s options is to plot only the points at which the es-
timated values of P (Y = 1 | X (i) , X (j) ) and P (Y = 1) are close to each
other. This produces a “contour” effect, indicating the boundary between
P (Y = 1 | X (i) , X (j) ) < P (Y = 1) and P (Y = 1 | X (i) , X (j) ) > P (Y = 1).
Let’s use this option, with our covariates being glucose and body mass:

> pwplot ( pima$ d i a b e t e s , pima [ , 1 : 8 ] , 2 5 ,

pairs=matrix ( c ( 2 , 6 ) , ncol =1) , cex =0.8 , band =0.05)

The result is shown in Figure 6.14. We see a snake-like contour, suggesting

that the boundary is quite nonlinear. If we are fitting a logistic model, we
might consider adding quadratic or even higher-order polynomial terms.
6.14. SPECIAL NOTE ON THE DESCRIPTION GOAL 255

Figure 6.14: Snakelike boundary

6.14 Special Note on the Description Goal

If we are unable to improve the fit of our parametric model in a setting

in which k-NN seems to give a substantially better fit, we should be quite
wary of placing too much emphasis on the values of the βj . As we saw
with the currency data, the estimated coefficients can be quite sensitive to
unusual observations and so on.
This is not to say the βj are useless in such settings. On the contrary, they
may be quite valuable. But they should be used with caution.

6.15 Computational Complements

6.15.1 Data Wrangling for the Currency Dataset

The file EXC.ASC has some nondata lines at the end, which need to be
removed before running the code below. We then read -it in, and do some
256 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

wrangling:
> c u r r <− read . table ( ’EXC. ASC ’ , h e a d e r=FALSE,
s t r i n g s A s F a c t o r s=FALSE)
> for ( i i n 1 : ncol ( c u r r ) ) c u r r [ , i ] <−
as . numeric ( c u r r [ , i ] )
Warning messages :
1 : NAs i n t r o d u c e d by c o e r c i o n
2 : NAs i n t r o d u c e d by c o e r c i o n
3 : NAs i n t r o d u c e d by c o e r c i o n
4 : NAs i n t r o d u c e d by c o e r c i o n
> colnames ( c u r r ) <−
c ( ’ Canada ’ , ’ Mark ’ , ’ Franc ’ , ’ Pound ’ , ’ Yen ’ )

What happened here? The above sequence is a little out of order, in the
sense that I ran it with prior knowledge of a certain problem, as follows.
The authors, in compiling this data file, decided to use ’.’ as their NA
code. R, in reading the file, then forced the numeric values, i.e., the bulk of
the data, to character strings for consistency. That in turn would have led
to each variable, i.e., each currency, being stored as an R factor. Having
discovered this earlier (not shown here), I added the argument stringsAs-
Factors = FALSE to my read call.
That still left me with character strings for my numeric values, so I ran
as.numeric() on each column. Finally, the original data set lacked names
for the columns, so I added some.

6.15.2 Data Wrangling for the Word Bank Dataset

There are a number of NA values in the data; let’s just look at complete
cases.
> # get variables of i n t e r e s t
> e n g l <− e n g l [ , c ( 2 , 5 : 8 , 1 0 ) ]
> # e x c l u d e c a s e s w i t h NAs
> e n c c <− e n g l [ complete . c a s e s ( e n g l ) , ]

One of the variables is birth order, an R factor. Let’s make it numeric:

> z <− e n c c $ b i r t h order
> class ( z )
[1] ” factor ”
> levels ( z )
6.16. MATHEMATICAL COMPLEMENTS 257

[ 1 ] ” Eighth ” ” F i f t h ” ” First ” ” Fourth ”

[ 5 ] ” Second ” ” Seventh ” ” S i x t h ” ” Third ”
> # c o n v e r t from F i r s t , Second , . . . t o 1 , 2 , . . .
> numcodes <− c ( 8 , 5 , 1 , 4 , 2 , 7 , 6 , 3 )
> zn <− as . numeric ( z )
> e n c c $ b i r t h order <− numcodes [ zn ]

Similarly, let’s convert the variable on the mother’s education, another R

factor, to a rough number of years of school, e.g., 10 years for the code
Some Secondary:
> z <− e n c c $mom ed
> levels ( z )
[ 1 ] ” College ” ” Graduate ”
[ 3 ] ” Primary ” ” Secondary ”
[ 5 ] ”Some C o l l e g e ” ”Some Graduate ”
[ 7 ] ”Some Secondary ”
> numcodes <− c ( 1 6 , 2 0 , 4 , 1 2 , 1 4 , 1 8 , 1 0 )
> e n c c $mom ed <− numcodes [ zn ]

Also, create the needed dummy variables, for gender and nonwhite cate-
gories:
> e n c c $male <− as . numeric ( e n c c $ s e x==’ Male ’ )
> e n c c $ s e x <− NULL
> e n c c $ a s i a n <− as . numeric ( e n c c $ e t h n i c i t y==’ Asian ’ )
> e n c c $ b l a c k <− as . numeric ( e n c c $ e t h n i c i t y==’ Black ’ )
> e n c c $ l a t i n o <− as . numeric ( e n c c $ e t h n i c i t y==’ H i s p a n i c ’ )
> e n c c $ o t h e r n o n w h i t e <− as . numeric ( e n c c $ e t h n i c i t y==’ Other ’ )
> e n c c $ e t h n i c i t y <− NULL

Note that a column in a data frame (or an element in any R list, of which
a data frame is a special case) can be removed by setting it to NULL.

6.16 Mathematical Complements

6.16.1 The Hat Matrix

We’ll use the notation of Section 2.4.2 here. The hat matrix is defined as
the n × n matrix

H = A(A′ A)−1 A′ (6.15)

258 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

The name stems from the fact that we use H to obtain “Y-hat,” the pre-
dicted values for the elements of D,

′
bi = µ
D fi ) βb
b(X (6.16)

Here is how the hat matrix comes in:

b = Aβb = A(A′ A)−1 A′ D = HD

D (6.17)

Using the famous relation (V W )′ = W ′ V ′ (Equation (A.14)), it is easily

verified that H is a symmetric matrix. Also, some easy algebra shows that
H is idempotent, i.e.,

H2 = H (6.18)

(The idempotency also follows from the fact that H is a projection operator;
once one projects, projecting the result won’t change it.)
This leads us directly to the residuals:

L = D − Aβb = D − HD = (I − H)D (6.19)

The diagonal elements

hii = Hii (6.20)

are known as the leverage values, another measure of influence like those in
Section 6.8.1, for the following reason. Looking at (6.17), we see that

b i = hii Di
D (6.21)

b i:
This shows us the effect of true value Di on the fitted value D

bi
∂D
hii = (6.22)
∂Di

So, hii can be viewed as a measure of how much influence observation i has
on its fitted value. A large value might thus raise concern — but how large
is “large”?
6.16. MATHEMATICAL COMPLEMENTS 259

Let wi denote row i of H, which is also column i since H is symmetric.

Then

∑
n
hii = Hii = (H 2 )ii = wi′ wi = h2ii + h2ij (6.23)
j=1,j̸=i

Since the far-right portion of the above equation is a sum of squares, This
directly tells us that hii ≥ 0. But it also tells us that hii ≥ h2ii , which forces
hii ≤ 1.
In other words,

0 ≤ hii ≤ 1 (6.24)

This will help assess whether a particular hii value is “large.”

6.16.2 Matrix Inverse Update

The famous Sherman-Morrison-Woodbury formula says that for an invert-

ible matrix B and vectors u and v

1
(B + uv ′ )−1 = B −1 − B −1 uv ′ B −1 (6.25)
1 + v ′ B −1 u

In other words, if we have already gone to the trouble of computing a matrix

inverse, and the matrix is then updated as above by adding uv ′ , then we
do not have to compute the new inverse from scratch; we need only modify
the old inverse, as specified above.
Let’s apply that to Section 6.8.1, where we discussed the effect of deleting
an observation from our data set. Write A in partioned form as in (2.19),
we have that

∑
n
A′ A = ei X
X e′ (6.26)
i
i=1

Thus the new version of A′ A after deleting observation i is

ei X
(A′ A)(−i) = A′ A − X ei′ (6.27)
260 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

This is then tailor-made for Sherman-Morrison-Woodbury! We set B =

A′ A. u = Xei and v = −Xei .

That will be our value for the left-hand side of (6.25). Look what happens
to the right-hand side:

ei′ (A′ A)−1 X

1 + v ′ B −1 u = 1 − X ei (6.28)

But that subtracted term is just hii ! Now that’s convenient.

It should be noted, however, that this approach has poor roundoff error
properties if A′ A is ill-conditioned, meaning that it is nearly singular. This
in turn can arise if some of the predictor variables are highly correlated
with each other, as we will see in Chapter 8.

6.16.3 The Median Minimizes Mean Absolute

Deviation

Let’s derive (6.7).

First, suppose a random variable W has a density fW . What value m
minimizes E[|W − m|]?

∫ ∞
E[|W − m|] = |t − m| fW (t)dt
−∞
∫ m ∫ ∞
= (m − t) fW (t)dt + (t − m) fW (t)dt
−∞ m
∫ m
= mP (W < m) − t fW (t)dt
−∞
∫ ∞
+ t fW (t)dt − mP (W > m) (6.29)
m

We have

mP (W < M ) − mP (W > m) = 2mFW (m) − m (6.30)

where FW is the cdf of W .

6.16. MATHEMATICAL COMPLEMENTS 261

Differentiating with respect to m and setting the result to 0, we have

0 = 2FW (m) − 1 (6.31)

In other words, m = median(W ).

The extension to conditional median then follows the same argument as in
Section 1.19.3.

6.16.4 The Gauss-Markov Theorem

The famous Gauss-Markov Theorem states that under the assumptions

of linearity of µ(t), homoscedasticity and independent observations, the
OLS is the Best Linear Unbiaed Estimator (BLUE) [116]. For any linear
combination c′ β of the elements of β, then c′ βb has minimum variance among
all unbiased estimators of β that consist of linear functions of the Yi . Note
that the normality assumption for the Yi is not needed.
Proofs of the theorem tend to take either an algebraic or an optimization
approach. We will follow the latter path here.

6.16.4.1 Lagrange Multipliers

A method for optimization under constraints is known as Lagrange multi-

pliers. The method will be useful for Gauss-Markov, where our constraint
is unbiasedness.
To see how the method works, say we wish to find the maximum value of
x3 + y subject to the constraint x2 + y 2 = 1. We set up the Lagrangian,

L = x3 + y + λ(x2 + y 2 − 1) (6.32)

We have introduced a number variable, λ, as an artifice to help us meet the

constraint. So, we set partial derivatives to 0:

∂L
0= = 3x2 + λ · 2x (6.33)
∂x

∂L
0= = 1 + λ · 2y (6.34)
∂y
262 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

∂L
0= = x2 + y 2 − 1 (6.35)
∂λ

The second equation gives us λ = −1/(2y), so that the first equation be-
comes

0 = 3x2 − 2x/(2y) (6.36)

Setting aside the case x = 0 for now, we have 0 = 3x − 1/y, i.e., 3xy = 1.
Using this and (6.35), we can solve for x and y (and finally dismiss x = 0).

6.16.4.2 Proof of Gauss-Markov

Let’s start with the case of p = 1, with a model with no intercept term.
The vector β now consists of a single element, and the matrix A in (2.28)
consists of a single column, a. The vector A′ D in that equation reduces to
a′ Y .
Consider any linear function of our Yi , say u′ Y . Its variance is

( )
∑
n
V ar(u′ Y ) = V ar u2i σ 2 (6.37)
i=1

The unbiasedness constraint is

β = E(u′ Y ) = u′ EY = u′ a′ β (6.38)

We are trying to find the BLUE u. Our Lagrangian is

∑
n
L = σ2 u2i + λ(u′ a′ β − β) (6.39)
i=1

Setting derivatives to 0, we have

∂L
0= = 2σ 2 ui + λai β (6.40)
∂ui

∂L
0= = u ′ a′ β − β (6.41)
∂λ
6.16. MATHEMATICAL COMPLEMENTS 263

From (6.40), we have

λai β
ui = − (6.42)
2σ 2

Equation (6.41) must hold for any β, so

∑
n
1= uj aj (6.43)
j=1

Substituting from (6.42) in (6.43), we have

λβ ∑ 2
n
1=− a (6.44)
2σ 2 j=1 j

i.e.,

λβ 1
− = ∑n 2 (6.45)
2σ 2 j=1 aj

Reusing that in (6.42), we finally have

ai
u i = ∑n (6.46)
j=1 a2j

Our BLUE estimator is then

∑
n
ui Yi (6.47)
i=1

The reader should check that this is exactly the OLS estimator (2.28): In
the latter, for instance,

∑
n
A′ A = a′ a = a2j (6.48)
j=1

For the general case, p > 1, one can actually use the same approach (Exer-
cise 12).
264 CHAPTER 6. MODEL FIT ASSESSMENT AND IMPROVEMENT

6.17 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. The contributors of the Forest Fire data set to the UCI Machine Learning
Repository, https://archive.ics.uci.edu/ml/datasets/Forest+Fires, describe
it as “a difficult regression problem.” Apply the methods of this chapter to
attempt to tame this data set.
2. Using the methods of this chapter, re-evaluate the two competing
Poisson-based analyses in Section 4.4.
3. Was logit a good model in the example in Section 4.5.3? Apply the
methods of this chapter to check.
4. Add an interaction term for age and gender in the linear model in Section
6.10. Interpret the results.
5. Apply parvsnonparplot() to the diabetes data in Section 6.11.1, and
discuss possible fit problems.
6. In the discussion of Figure 6.1, it was noted that in investigating possible
areas of poor fit for the parametric model, we should keep in mind possible
bias of the nonparametric model at the left and right ends of the figure. We
saw possible problems with the parametric model at both ends, but one of
them may be due in part to the bias issue. State which one, and explain
why.
Mini-CRAN and other computational problems:
7. Write an R function analogous to influence() for quantreg objects.
8. Recall that with Poisson regression we have (4.12), while with overdis-
persed models we have (4.13). Use knnest() to plot variance against mean
in the Pima data, Section 4.3.2, in order to partly assess whether a Poisson
model works there.
9. The knnest() function in the regtools package is quite versatile. Its
nearf argument allows us to apply general functions to Y values in a neigh-
borhood, rather than simply averaging them. Here we will apply that to
quantiles.
The γ quantile of a cdf F is defined to be a number d such that F (d) = γ.
For a continuous distribution, that number is unique, and the quantile
function is the inverse of the cdf.
6.17. DATA, CODE AND MATH PROBLEMS 265

But that is not true in the discrete case. The latter is especially problematic
in the case of finding sample quantiles. How, for instance, should we define
the sample median if the sample size n is an even number? This is such
a problem that R’s quantile() function actually offers the user 9 different
definitions of “quantile.”
We will be estimating conditional quantiles, defined by

q(t, γ) = Ft−1 (γ) (6.49)

where Ft (w) is defined to be P (Y ≤ w | X = t).

(a) Write an R function with call form

quanty ( predpt , nearxy ,gamma, type )

to be used as nearf in calling knnest(). Here type is as in quan-

tile().

(b) Apply this function to the baseball player data, regression weight
against height, and compare to the results of applying the quantreg
package.

10. Consider m-fold cross-validation, Section 6.6.2, in a k-NN context.

Change preprocessx() to allow this, so that the new call form will be
p r e p r o c e s s x ( x , kmax , x v a l=FALSE)

Math problems:
11. Say we use parallel coordinates (Section 6.7.3.1) to display some data
having p = 2. Say some of our points lie on a straight line in (X (1) , Xi(2)
space. Show that in the parallel coordinates plot, the lines corresponding
to these points will all intersect at a common point. (It might be helpful
to generate some data and form their parallel coordinates plot to help your
intuition.)
12. Prove the general case of the Gauss-Markov Theorem, showing that the
OLS estimator c′ βb is the BLUE of c′ β for any c. Follow the same pattern
as in Section 6.16.4.2, replacing β by c′ β and so on.
Chapter 7

Disaggregating Regressor
Effects

What does the above chapter title mean? It is, admittedly, rather overly
abstract, but it does capture the essense of this chapter, as follows.
Recall that a synonym for “predictor variable” is regressor. This chapter
is almost entirely focused on the Description goal, i.e., analysis of the in-
dividual effects of the regressors on the response variable Y . In a linear
model, those effects are measured by the βi . Well, then, what is meant by
the term disaggregating in our chapter title?
Recall the example in Section 1.11.1, regarding a study of the quality of
care given to heart attack patients in a certain hospital chain. The concern
was that one of the hospitals served a population with many elderly pa-
tients, thus raising the possibility that the analysis would unfairly present
this hospital as providing inferior care. We thus want to separate out —
disaggregate, as the economists say — the age effect, by modeling the prob-
ability of survival as a function of hospital branch and age. We could, for
instance, use a logistic model, with survival as the binary response variable,
and with age and dummy variables for the hospital branches as predictors.
The coefficients of the dummies would then assess the quality of care for
the various branches, independent of age issues.
Most of this chapter will be concerned with measuring effects of predictor
variables, with such disaggregation in mind. We will see, though, that
attaining this goal may require some subtle analysis. It will be especially
important to bear in mind the following principle:

267
268 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

The Predictor Effects Principle

The sign and magnitude of a regression coefficient (whether
sample estimate or population value) for one predictor variable
may depend on which other predictors are present.
This in turn is related to correlations among the predictors. In
a linear or generalized linear model, the coefficient for predictor
i may reflect not only the effect of that predictor, but may also
incorporate the effect of some predictor j that is correlated with
it.

In a Description context, this may be rather unsettling news, but it clearly

can’t be ignored. One must indeed be duly cautious in regression appli-
cations in which Description is the primary goal. One of the goals of this
chapter is to better understand this issue, and thus deal with it in practical
settings.
Re-aggregation
In addition, this chapter will also cover what might be termed re-aggregation
of the predictors. We use this term in the spirit of R’s aggregate() func-
tion, which computes summary statistics for subgroups of the data. As
noted above, a straight use of aggregate() in the hospital example would
be inappropriate; we need to do aggregation in a manner that takes co-
variates such as age into account, and that will be the focus of Section
7.5.
Our re-aggregation discussion here will also concern statistical methodology
known as small area estimation, which for example often arises in county
data. We may have data on aspects of interest for the larger counties of the
state, but have only limited data or none at all for some smaller ones. Small
area estimation involves using predictor variables to estimate the missing
values.

7.1 A Small Analytical Example

Before analyzing some real data, let’s get an overview of the situation, via
a small mathematical example. Our model will have two predictors,

E(Y |X (1) , X (2) ) = β0 + β1 X (1 + β2 X (2) (7.1)

7.1. A SMALL ANALYTICAL EXAMPLE 269

But what if we use only X (1) , i.e., we wish to work with

E(Y | X (1) ) (7.2)

Using the Tower Property for conditional expectation (stated in Section

1.19.5.4, proven in Section 7.8.1 in the Mathematical Complements section
at the end of this chapter), (7.2) is equal to the conditional expected value
of (7.1) with respect to X (1) , i.e.,

[ ]
E(Y | X (1) ) = E E(Y |X (1) , X (2) ) | X (1) (7.3)

This is very abstract, of course, but it merely says that the regression
function of Y on X (1) and X (2) , averaged over the values of X (2) is equal
to the regression function of Y on X (1) alone. If we take, say, the mean
weight of all people of a given height and age, and then average over the
values of age, we obtain mean weight of all people of a given height.
This gives us

E(Y | X (1) ) = β0 + β1 X (1) + β2 E(X (2) | X (1) ) (7.4)

Suppose the regression function of X (2) on X (1) is also linear (which for
example will be the case if the three variables have a trivariate normal
distribution):

E(X (2) | X (1) ) = γ0 + γ1 X (1) (7.5)

so that

E(Y | X (1) ) = β0 + β2 γ0 + (β1 + β2 γ1 )X (1) (7.6)

Here is the point: Say for convenience that β1 , β2 and γ1 are all positive.
Comparing (7.6) and (7.1), we see that if we use the two-predictor model
(7.1) instead of (7.2), the effect of X (1) on Y shrinks by the amount β2 γ1 .
Putting it more colloquially, adding the predictor X (2) “steals some of
X (1) ’s thunder.” So, the effect of X (1) on Y is smaller if we include X (2)
in our analysis.
270 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

On the other hand, suppose the βi are positive, but γ1 is so negative as to

make

β1 + β2 γ1 < 0 (7.7)

Then the sign of the X (1) effect changes from positive, β1 > 0, to negative,
(7.7).
Also, observe that if X (1) and X (2) are independent, or at least uncorre-
lated, then γ1 = 0, so that the coefficient of X (1) will be the same with or
without X (2) in our analysis.
This little example shows what is occurring in the background in the Pre-
dictor Effects Principle.

7.2 Example: Baseball Player Data

In Section 1.9.1.2, we found that the data indicated that older baseball
players — of the same height — tend to be heavier, with the difference
being about 1 pound gain per year of age. This finding may surprise some,
since athletes presumably go to great lengths to keep fit. Ah, so athletes
are similar to ordinary people after all.
We may then ask whether a baseball player’s weight is also related to the
position he plays. So, let’s now bring the Position variable in our data into
play. First, what is recorded for that variable?
> l e v e l s ( mlb$ P o s i t i o n )
[ 1 ] ” Catcher ” ” F i r s t Baseman”
[ 3 ] ” Outfielder ” ” Relief Pitcher ”
[ 5 ] ” Second Baseman” ” Shortstop ”
[ 7 ] ” Starting Pitcher ” ” Third Baseman”

So, all the outfield positions have been simply labeled “Outfielder,” though
pitchers have been separated into starters and relievers.
In order to have a handy basis of comparison below, let’s re-run the weight-
height-age analysis:
> summary(lm( Weight ∼ Height + Age , data=nondh ) )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
7.2. EXAMPLE: BASEBALL PLAYER DATA 271

( I n t e r c e p t ) −187.6382 17.9447 −10.46 < 2 e −16

Height 4.9236 0.2344 2 1 . 0 0 < 2 e −16
Age 0.9115 0.1257 7 . 2 5 8 . 2 5 e −13

( I n t e r c e p t ) ∗∗∗
Height ∗∗∗
Age ∗∗∗
...
M u l t i p l e R−s q u a r e d : 0.318 ,
Adjusted R−s q u a r e d : 0.3166

Now, for simplicity and also to guard against overfitting, let’s consolidate
into four kinds of positions: infielders, outfielders, catchers and pitchers.
That means we’ll need three dummy variables:
> pos <− mlb$ P o s i t i o n
> i n f l d <− as . integer ( pos %i n%
c ( ’ F i r s t Baseman ’ , ’ Second Baseman ’ , ’ S h o r t s t o p ’ ,
’ Third Baseman ’ ) )
> o u t f l d <− as . integer ( pos == ’ O u t f i e l d e r ’ )
> p i t c h e r <− as . integer ( pos %i n% c ( ’ R e l i e f P i t c h e r ’ ,
’ Starting Pitcher ’ ))

Again, remember that catchers are designated via the other three dummies
being 0.
So, let’s run the regression:
> lmpos <− lm( Weight ∼ Height + Age + i n f l d +
o u t f l d + p i t c h e r , data=mlb )
> summary( lmpos )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) −182.7216 1 8 . 3 2 4 1 −9.972 < 2 e −16
Height 4.9858 0 . 2 4 0 5 2 0 . 7 2 9 < 2 e −16
Age 0.8628 0.1241 6 . 9 5 2 6 . 4 5 e −12
infld −9.2075 1 . 6 8 3 6 −5.469 5 . 7 1 e −08
outfld −9.2061 1 . 7 8 5 6 −5.156 3 . 0 4 e −07
pitcher −10.0600 2 . 2 5 2 2 −4.467 8 . 8 4 e −06

( I n t e r c e p t ) ∗∗∗
Height ∗∗∗
Age ∗∗∗
272 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

infld ∗∗∗
outfld ∗∗∗
pitcher ∗∗∗
...
Mult . R−s q u a r e d : 0 . 3 4 0 4 , Adj . R−s q u a r e d : 0.3372
...

The estimated coefficients for the position variables are all negative. At
first, it might look like the town of Lake Wobegon in the radio show Prairie
Home Companion, “Where all children are above average.” Do the above
results say that players of all positions are below average?
No, not at all. Look at our model:

mean weight = (7.8)

β0 + β1 × height + β2 × age + β3 × infld + β4 × outfld + β5 × pitcher

Under this model, let’s find the difference in mean weight between two
subpopulations — 72-inches-tall, 30-year-old pitchers and catchers of the
same height and age. Keeping in mind that catchers are coded with 0s in
all three dummies, we see that the difference in mean weights is simply β5 !
In other words, β3 , β4 and β5 are mean weights relative to catchers. Thus
for example, the interpretation of the 10.06 figure is that, for a given height
and age, pitchers are on average about 10.06 pounds lighter than catchers
of the same height and age, while for outfielders the figure is about 9.2
pounds. An approximate 95% confidence interval for the population value
of the latter (population mean for outfielders minus population mean for
catchers) is

−9.2 ± 2 × 1.8 = (−12.8, −5.6) (7.9)

So, the image of the “beefy” catcher is borne out.

Note that the estimated coefficient for age shrank a little when we added
the position variables. In our original analysis, with just height and age
as predictors, it had been 0.9115,1 but now is only 0.8628. The associated
confidence interval, (0.61,1.11), still indicates weight increase with age, but
the effect is now smaller than before. This is an example of Predictor Effects
Principle, mentioned at the outset of this chapter, that the coefficient for
one predictor may depend on what other predictors are present.
1 This was the case even after removing the Designated Hitters, not shown here.
7.2. EXAMPLE: BASEBALL PLAYER DATA 273

It could be that this shrinkage arose because catchers are somewhat older
on average. In fact,
> mc <− [ mlb$PosCategory == ’ Catcher ’ , ]
> mean(mc$Age )
[ 1 ] 29.56368
> mean( mlb$Age )
[ 1 ] 28.70835

So the catchers “absorbed” some of that age effect, as explained in Section

7.1.
This also suggests that the age effect on weight is not uniform across playing
positions. To investigate this, let’s add interaction terms:
> summary(lm( Weight ∼ Height + Age +
infld + outfld + pitcher +
Age∗ i n f l d + Age∗ o u t f l d + Age∗ p i t c h e r , data=nondh ) )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) −168.5453 2 0 . 3 7 3 2 −8.273 4 . 1 1 e −16
Height 4.9854 0 . 2 4 0 7 2 0 . 7 1 4 < 2 e −16
Age 0.3837 0.3335 1.151 0.2501
infld −22.8916 1 1 . 2 4 2 9 −2.036 0.0420
outfld −27.9894 1 1 . 9 2 0 1 −2.348 0.0191
pitcher −31.9341 1 5 . 4 1 7 5 −2.071 0.0386
Age : i n f l d 0.4629 0.3792 1.221 0.2225
Age : o u t f l d 0.6416 0.4033 1.591 0.1120
Age : p i t c h e r 0.7467 0.5232 1.427 0.1539

( I n t e r c e p t ) ∗∗∗
Height ∗∗∗
Age
infld ∗
outfld ∗
pitcher ∗
Age : i n f l d
Age : o u t f l d
Age : p i t c h e r
...
Mult . R−s q u a r e d : 0.3424 , Adj . R−s q u a r e d : 0.3372

(Use of a colon in a regression formula, e.g., Age:infld here, means to

add an interaction term for the indicated pair of variables, which as we
274 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

have seen before is just their product. More on this in the Computational
Complements section at the end of this chapter.)
This doesn’t look helpful. Confidence intervals for the estimated interaction
coefficients are near 0,2 and equally important, are wide. Thus there could
be important interaction effects, or they could be tiny; we just don’t have
a large enough sample to say much.
Note that the coefficients for the position dummies changed quite a bit,
but this doesn’t mean we now think there is a larger discrepancy between
weights of catchers and the other players. For instance, for 30-year-old
players, the estimated difference in mean weight between infielders and
catchers of a given height is

−22.8916 + 30 × 0.4629 = −9.0046 (7.10)

similar to the -9.2075 figure we had before. Indeed, this is another indication
that interaction terms are not useful in this case.

7.3 Simpson’s Paradox

The famous Simpson’s Paradox should not be considered a paradox, when

viewed in the light of a central point we have been discussing in this chapter,
which we will state a little differently here:

The regression coefficient (sample or population) for a predictor

variable may change substantially when another predictor is
added. In particular, its sign may change, from positive to
negative or vice versa.

7.3.1 Example: UCB Admissions Data (Logit)

The most often-cited example, in a tabular context, is that of the UC Berke-

ley admissions data [20]. The issue at hand was whether the university had
been discriminating against women applicants for admission to graduate
school.
2 In
fact, they include 0, but as discussed before, we should not place any distinction
between near 0 and include 0.
7.3. SIMPSON’S PARADOX 275

On the surface, things looked bad for the school — 44.5% of the male appli-
cants had been admitted, compared to only 30.4% of the women. However,
upon closer inspection it was found that the seemingly-low female rate was
due to the fact that the women tended to apply to more selective academic
departments, compared to the men. After correcting for the Department
variable, it was found that rather than being victims of discrimination, the
women actually were slightly favored over men. There were six departments
in all, labeled A-F.
The data set is actually included in base R. As mentioned, it is stored in
the form of an R table:
> ucb <− UCBAdmissions
> c l a s s ( ucb )
[ 1 ] ” table ”
> ucb
, , Dept = A

Gender
Admit Male Female
Admitted 512 89
R e j e c t e d 313 19

, , Dept = B

Gender
Admit Male Female
Admitted 353 17
R e j e c t e d 207 8
...

In R, it is sometimes useful to convert a table to an artificial data frame,

which in this case would have as many rows as there were applicants in the
UCB study, 4526. The regtools function tbltofakedf() facilitates this:
> ucbdf <− t b l t o f a k e d f ( ucb )
> dim( ucbdf )
[ 1 ] 4526 3
> head ( ucbdf )
[ ,1] [ ,2] [ ,3]
[ 1 , ] ” Admitted ” ” Male ” ”A”
[ 2 , ] ” Admitted ” ” Male ” ”A”
[ 3 , ] ” Admitted ” ” Male ” ”A”
[ 4 , ] ” Admitted ” ” Male ” ”A”
276 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

[ 5 , ] ” Admitted ” ” Male ” ”A”

[ 6 , ] ” Admitted ” ” Male ” ”A”

The first six rows are the same, and in fact there will be 512 such rows,
since, as seen above, there were 512 male applicants who were admitted to
Department A.
Let’s analyze this data using logistic regression. With such coarsely discrete
data, this is not a typical approach,3 but it will illustrate the dynamics of
Simpson’s Paradox.
First, convert to usable form, not R factors. It will be convenient to use
the dummies package [26]:
> ucbdf $admit <− as . integer ( ucbdf [ , 1 ] == ’ Admitted ’ )
> ucbdf $male <− as . integer ( ucbdf [ , 2 ] == ’ Male ’ )
# s a v e work by u s i n g t h e ’ dummies ’ p a c k a g e
> l i b r a r y ( dummies )
> dept <− ucbdf [ , 3 ]
> deptdummies <− dummy( dept )
> head ( deptdummies )
deptA deptB deptC deptD deptE deptF
[1 ,] 1 0 0 0 0 0
[2 ,] 1 0 0 0 0 0
[3 ,] 1 0 0 0 0 0
[4 ,] 1 0 0 0 0 0
[5 ,] 1 0 0 0 0 0
[6 ,] 1 0 0 0 0 0
# o n l y 5 dummies
> ucbdf1 <− cbind ( ucbdf , deptdummies [ , − 6 ] ) [ , − ( 1 : 3 ) ]
> head ( ucbdf1 )
admit male deptA deptB deptC deptD deptE
1 1 1 1 0 0 0 0
2 1 1 1 0 0 0 0
3 1 1 1 0 0 0 0
4 1 1 1 0 0 0 0
5 1 1 1 0 0 0 0
6 1 1 1 0 0 0 0

Now run the logit, first only with the male predictor, then adding the
departments:
> glm( admit ∼ male , data=ucbdf1 , family=binomial )
3A popular method for tabular data is log-linear models [35] [2].
7.3. SIMPSON’S PARADOX 277

...
Coefficients :
( Intercept ) male
−0.8305 0.6104
...
> glm( admit ∼ . , data=ucbdf1 , family=binomial )
...
Coefficients :
( Intercept ) male deptA
−2.62456 −0.09987 3.30648
deptB deptC deptD
3.26308 2.04388 2.01187
deptE
1.56717
...

So the sign for the male variable switched from positive (men are favored)
to slightly negative (women have the advantage). Needless to say, this
analysis (again, in the original table form, not logit) caused quite a stir.
The evidence against the university had looked so strong, only to find later
that an overly simple statistical analysis had led to an invalid conclusion.
By the way, note that the coefficients for all five dummies were positive,
which reflects the fact that all the departments A-E had higher admissions
rates than department F:
> apply ( ucb , c ( 1 , 3 ) ,sum)
Dept
Admit A B C D E F
Admitted 601 370 322 269 147 46
R e j e c t e d 332 215 596 523 437 668

Let’s take one more look at this data, this time more explicitly taking the
selectivity of departments into account. We’ll create a new variable, finding
the acceptance rate for each department and then replacing each applicant’s
department information by the selectivity of that department:
> deptsums <− apply ( ucb , c ( 1 , 3 ) ,sum)
> d e p t r a t e s <− deptsums [ 1 , ] / colSums ( deptsums )
> deptsums <− apply ( ucb , c ( 1 , 3 ) ,sum)
> d e p t r a t e s <− deptsums [ 1 , ] / colSums ( deptsums )
> glm( admit ∼ male + d e p t r a t e , data=ucbdf )
...
Coefficients :
278 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

( Intercept ) male deptrate

0.003914 −0.016426 1.015092

Consistent with our earlier analysis, the coefficient for the male variable
is slightly negative. But we can also quantify our notion that the women
were applying to the more selective departments:
> tapply ( ucbdf $ d e p t r a t e , ucbdf $male , mean)
0 1
0.2951732 0.4508945

The mean admissions rate among departments applied to by women is much

lower than the corresponding figure for men.

7.3.2 The Verdict

Simpson’s is not really a paradox — let’s just call it Simpson’s Phenomenon

— but is crucially important to keep in mind in applications where Descrip-
tion is the goal. And the solution to the “paradox” is to think twice before
deleting any predictor variables.
Ironically, this last point is somewhat at odds with the theme of Chapter
9, in which we try to pare down the number of predictors. When we have
correlated variables, such as Gender and Department in the admissions
data, it might be tempting to delete one or more of them on the grounds
of “redundancy,” but we first should check the effects of deletion, e.g., sign
change.4
On the other hand, this is rather consistent with the method of ridge re-
gression in Chapter 8. That approach attempts to ameliorate the effects
of correlated predictor variables, rather than resorting to deleting some of
them.
Once again, we see that regression and classification methodology does not
always offer easy, pat solutions.

7.4 Unobserved Predictor Variables

In Statistical Heaven, we would have data on all the variables having sub-
stantial relations to the response variable. Reality is sadly different, and
4 In the admissions data, the correlation, though substantial, would probably not
warrant deletion in the first place, but the example does illustrate the dangers.
7.4. UNOBSERVED PREDICTOR VARIABLES 279

often we feel that our analyses are hampered for lack of data on crucial
variables.
Statisticians have actually developed methodology to deal with this prob-
lem, such as the famous Fay-Herriott model. Not surprisingly, the methods
have stringent assumptions, and they are hard to verify. This is especially
an issue in that many of the models used posit latent variables, meaning
ones that are unseen. But such methods should be part of the toolkit of
any data scientist, either to use where appropriate or at least understand
when presented with such analyses done by others. The following sections
will provide brief introductions to such methodology.

7.4.1 Instrumental Variables (IVs)

This one is quite controversial. (Or, at least the choice of one’s IV often
evokes controversy.) It’s primarily used by economists, but has become
increasingly popular in the social and life sciences. The goal is to solve the
problem of not being able to observe data on a variable that ideally we wish
to use as a predictor. We find a kind of proxy, known as an instrument.
The context is that of a Description goal. Suppose we are interested in the
relation between Y and two predictors, X (1) and X (2) , with a focus on the
former. What we would like to find is the coefficient of X (1) in the presence
of X (2) , i.e., estimate β1 in

E(Y | X (1) , X (2) ) = β0 + β1 X (1) + β2 X (2) (7.11)

But the problem at hand here is that we observe Y and X (1) but not X (2) .
We believe that the two population regression functions (one predictor vs.
two predictors) are well approximated by linear models:5

E(Y | X (1) ) = β01 + β11 X (1) (7.12)

E(Y | X (1) , X (2) ) = β02 + β12 X (1) + β22 X (2) (7.13)

Note that we are doubly-subscripting the β coefficients, since we have two

linear models.
5 The second model does not imply the first. What if X (2) ) = (X (1) )2 , for instance?
280 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

We are primarily interested in the role of X (1) , i.e., the value of β12 . How-
ever, as has been emphasized so often in this chapter, generally

β11 ̸= β12 (7.14)

Thus an analysis on our data that uses (7.12) — remember, we cannot use
(7.13), since we have no data on X (2) — may be seriously misleading.
A commonly offered example concerns a famous economic study regarding
the returns to education [31]. Here Y is weekly wage and X (1) is the number
of years of schooling. The concern was that this analysis doesn’t account
for “ability”; highly-able people (however defined) might pursue more years
of education, and thus get a good wage due to their ability, rather than the
education itself. If a measure of ability were included in our data, we could
simply use it as a covariate and fit the model (7.13), but no such measure
was included in the data.6
The instrumental variable (IV) approach involves using another variable,
observed, that is intended to remove from X (1) the portion of that variable
that involves ability. That variable — the instrument — works as a kind
of surrogate for the unobserved variable. If this works — a big “if” —
then we will be able to measure the effect of years of schooling without the
confounding effect of ability.
In the years-of-schooling example, the instrument proposed is distance from
a college. The rationale here is that, if there are no nearby postsecondary
institutions, the person will find it difficult to pursue a college education,
and may well decide to forego it — even if the person is of high ability. All
this will be quantified below, but keep this in mind as a running example.
Note that the study was based on data from 1980, when there were fewer
colleges in the U.S. than there are now. Thus this particular instrument
may be less useful today, but it was questioned even when first proposed.
As noted in the introduction to this section, the IV approach is quite con-
troversial.
Adding to the controversy is that different authors have defined the condi-
tions required for use of IVs differently. Furthermore, in some cases defini-
tions of IV have been insufficiently precise to determine whether they are
equivalent to others.
Nevertheless, the wide use of IV in certain disciplines warrants taking a
6 Of course, even with better data, “ability” would be hard to define. Does it mean
IQ (of which I am very skeptical), personal drive or what?
7.4. UNOBSERVED PREDICTOR VARIABLES 281

closer look, which we do below.

7.4.1.1 The IV Method

Let Z denote our instrument, i.e., an observed variable that we hope will
remove the effect of our unobserved variable X (2) . The instrument must
satisfy two conditions, to be described shortly. In preparation for this, set

ϵ = Y − E(Y |X (1) , X (2) ) (7.15)

= Y − (β02 + β12 X (1)
+ β22 X (2)
) (7.16)

Now, letting ρ(U, V ) denote the population correlation between and two
random variables U and V , the requirements for an IV Z are

(a) ρ(Z, X (1) ) ̸= 0

(b) ρ(Z, X (2) ) = 0

(c) ρ(Z, ϵ) = 0

In the education example, (a) means that distance is related to years of

schooling, as posited above, while (b) means that distance is unrelated to
ability. But what about (c)?
Recall that ϵ is called the “error term,” and is interpretable here as the
collective effect of all variables besides X (1) and X (2) that are related to Y
(Section 2.2). These other variables are assumed to have a causal relation
with Y (another controversy), as opposed to Z whose relation to Y is only
via X (1) . This motivates (c), which in the schooling example means that
not only is distance unrelated to ability, but also distance is unrelated to
any other variable having an impact on earnings.
Clearly, (c) is a very strong assumption, and one that is not assessable,
i.e., one cannot devise some check for this assumption as we did for various
assumptions in Chapter 6.
How are the above conditions used in the mathematics underlying the IV
method? Let’s perform a (population-level) calculation. Writing

Y = β02 + β12 X (1) + β22 X (2) + ϵ (7.17)

282 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

and using the linearity of covariance, we have

Cov(Z, Y ) = β12 Cov(Z, X (1) ) + β22 Cov(Z, X (2) ) + Cov(Z, ϵ)

= β12 Cov(Z, X (1) ) (7.18)

and thus

Cov(Z, Y )
β12 = (7.19)
Cov(Z, X (1) )

We then take

d
Cov(Z, Y)
βb12 = (7.20)
d
Cov(Z, X (1) )

where the estimated covariances come from the data, e.g., from the R cov()
function. We can thus estimate the parameter of interest, β12 — in spite
of not observing X (2) .
This is wonderful! Well, wait a minute...is it too good to be true? Well, as
noted, the assumptions are crucial, such as:

• We assume the linear models (7.12) and (7.13). The first can be
assessed from our data, but the second cannot.

• We assume condition (b) above, i.e., that our instrument is uncorre-

lated with our unseen variable. Often we are comfortable with that
assumption — e.g., that distance from a college is not related to abil-
ity — but again, it cannot easily be verified.7

• We assume condition (c), which as discussed earlier is quite strong.

• We need the instrument to have a fairly substantial correlation to the

observed predictor, i.e., ρ(Z, X (1) ) should be substantial. If it isn’t,
then we have a small or even tiny denominator in (7.20), so that the
sample variance of the quotient — and thus of our βb12 — will be
large, certainly not welcome news.
7 Sometests for this have been developed, but those have their own often-questionable
assumptions.
7.4. UNOBSERVED PREDICTOR VARIABLES 283

7.4.1.2 Two-Stage Least Squares:

Another way to look at the IV idea is Two-Stage Least Squares (2SLS), as

follows. Recall the phrasing used above, that the instrument

is a variable that is intended to remove from X (1) the portion

of that variable that involves ability.

That suggests regressing X (1) on Z.8

Let’s see what happens, again at the population level. Using (7.17), write

E(Y | Z) = β02 + β12 E(X (1) |Z) + β22 E(X (2) |Z) + E(ϵ|Z) (7.21)

By assumption, Z and X (2) are uncorrelated, as are Z and ϵ. If these

variables also have a multivariate normal distribution, then they are inde-
pendent. Assuming this, we have

E(X (2) |Z) = E[X (2) ] and E(ϵ|Z) = E[ϵ] = 0 (7.22)

In other words,

E(Y | Z) = c + β12 E(X (1) |Z) (7.23)

for a constant c = β02 + β22 E[X (2) ].

Now, remember, our goal is to estimate β12 , so its appearing in (7.23) is a
welcome sight! Then what does that equation suggest at the sample level?
Well, E(X (1) |Z) is the regression of X (1) on Z. In other words, the process
is as follows:

Two-Stage Least Squares:

• First regress X (1) on the instrument Z, to estimate E(X (1) |Z)

at our sample values Zi .
• In view of (7.23), we then treat these estimated regression
values as our new “predictor” values — which we use to
predict the Yi .
8 As noted before, the term “regress V on U ” means to model the regression function
of V given U .
284 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

• The resulting estimated slope will be βb12 , the estimate we

are seeking.

In other words, the IV method can be viewed as an application of 2SLS,

with the predictor variable in the first stage being our instrument.
In terms of R, this would mean
lmout <− lm( x1 ∼ z )
e s t r e g <− lmout$ f i t t e d . v a l u e s
b12hat <− coef (lm( y ∼ e s t r e g ) ) [ 2 ]

The purpose of this section was to explain the notion of IVs. It shows
directly where the name “Two-Stage” least squares comes from.
The above just gives us a point estimate of β12 . We need a standard error
as well. This can be derived using the Delta Method (Section 3.6.1), which
we explore in the exercises at the end of this chapter.
However, there are more sophisticated R packages for this, such as iv-
model, which give us all this and more, as seen in the next section.

7.4.1.3 Example: Years of Schooling

Data for the schooling example dicussed above is widely available. Here we
will use the set card.data, available for instance in the ivmodel package
[79].
There are many variables in the data set. Here we will just follow our earlier
example, analyzing the effect of years of schooling on wage (in cents per
hour), with nearness to a college as our instrument.
Let’s first do the computation “by hand,” as above:
> l i b r a r y ( ivmodel )
> data ( c a r d . data )
> sc h <− c a r d . data
# without the instrument
> lm( wage ∼ educ , data=s c h )
...
Coefficients :
( Intercept ) educ
183.95 29.66
# now w i t h t h e IV
> s t a g e 1 <− lm( educ ∼ nearc4 , data=s c h )
7.4. UNOBSERVED PREDICTOR VARIABLES 285

> lm( s c h $wage ∼ s t a g e 1 $ f i t t e d . v a l u e s )

...
Coefficients :
( Intercept ) stage1$fitted . values
−849.5 107.6

Here is how we can get this from the machinery in ivmodel:

> ivmodel (Y=s c h $wage ,D=s c h $educ , Z=s c h $ n e a r c 4 )

...
C o e f f i c i e n t s o f k−C l a s s E s t i m a t o r s :

k Estimate Std . E r r o r t v a l u e
OLS 0.0000 29.6554 1.7075 17.368
...
TSLS 1.0000 107.5723 15.3995 6.985
...

We see our IV result confirmed in the “TSLS” line, with the “OLS” line
restating our non-IV result.9 Note that the standard error increased quite
a bit.
Now, what does this tell us? On the one hand, a marginal increase, say 1
year, of schooling seems to pay off much more than the non-IV analysis had
indicated, about $1.08 per hour rather than $0.30. However, the original
analysis had a much higher estimated intercept term, $1.84 vs. -$8.50. Let’s
compute predicted wage for 12 years of school, for instance, under both
models:

> −8.50 + 12∗ 1 . 0 7

[ 1 ] 4.34
> 1 . 8 4 + 12∗ 0 . 3 0
[ 1 ] 5.44

That’s quite a discrepancy! We can do much better if we include the other

predictor variables available in the data (not shown), but we at least see
that blind use of IV models — and of course any models — can lead to
poor results.

9 Otherlines in the output, not shown here, show the results of applying other IV
methods, including those of the authors of the package.
286 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

7.4.1.4 The Verdict

In our years-of-schooling example (Section 7.4.1.1), it was mentioned that

the assumption that the distance variable was unreleated to ability was
debatable. For example, we might reason that able children come from
able parents, and able parents believe college is important enough that
they should live near one. This is an example of why the IV approach is so
controversial.
Nevertheless, the possible effect of unseen variables itself can make an anal-
ysis controversial. IVs may be used in an attempt to address such problems.
However, extra care is warranted if this method is used.

7.4.2 Random Effects Models

Continuing our theme here in Section 7.4 of approaches to account for

unseen variables, we now turn briefly to mixed effects models [77] [52].
Consider a usual linear regression model for one predictor X,

E(Y | X = t) = β0 + β1 t (7.24)

for unknown constants β0 and β1 that we estimated from the data.

Now alter the model so that β0 is random, each unit (e.g., each person)
having a different value, though all having a common value of β1 . We
might observe people over time, with X representing time and Y being
modeled as having a linear time trend. The slope β1 of that time trend is
assumed the same for all people, but the starting point β0 is not.
We might write our new model as

E(Y | X = t) = β0 + B + β1 t (7.25)

where B is a random variable having mean 0. Each person has a different

value of B, with the intercept for people now being a random variable with
2
mean β0 and variance σB .
It is more common to write

Y = β0 + α + β1 X + ϵ (7.26)
7.4. UNOBSERVED PREDICTOR VARIABLES 287

where α and ϵ have mean 0 and variances σa2 and σe2 . The population
values to be estimated from our data are β0 , β1 , σa2 and σe2 . Typically these
are estimated via Maximum Likelihood (with the assumptions that α and ϵ
have normal distributions, etc.), though the Method of Moments is possible
too.
The variables α and ϵ are called random effects (they are also called variance
components), while the β0 + β1 X portion of the model is called a fixed
effecs. This phrasing is taken from the term fixed-X regression, which we
saw in Section 2.3; actually, we could view this as a random-X setting, but
the point is that even β1 is fixed. Due to the presence of both fixed and
random effects, the term mixed-effects model is used.

7.4.2.1 Example: Movie Ratings Data

Consider again the MovieLens data introduced in Section 3.2.4. We’ll use
the 100,000-rating data here, which includes some demographic variables
for the users. The R package lme4 will be our estimation vehicle [9].
First we need to merge the ratings and demographic data. This entails use
of R’s merge() function, introduced in Section 3.5.1. See Section 7.7.1 for
details for our current setting. Our new data frame, after applying the code
in that section, is u.
We might speculate that older users are more lenient in their ratings. Let’s
take a look:
> z <− lmer ( r a t i n g ∼ age+g e n d e r +(1| usernum ) , data=u )
> summary( z )
...
Random e f f e c t s :
Groups Name V a r i a n c e Std . Dev .
usernum ( I n t e r c e p t ) 0 . 1 7 5 0.4183
Residual 1.073 1.0357
Number o f obs : 1 0 0 0 0 0 , g r o u p s : usernum , 943

Fixed e f f e c t s :
Estimate Std . E r r o r t v a l u e
( Intercept ) 3.469074 0.048085 72.14
age 0.003525 0.001184 2.98
genderM −0.002484 0.031795 −0.08

C o r r e l a t i o n o f Fixed E f f e c t s :
( I n t r ) age
288 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

age −0.829
genderM −0.461 −0.014

First, a word on syntax. Here our regression formula was

r a t i n g ∼ age + g e n d e r + ( 1 | usernum )

Most of this looks the same as what we are accustomed to in lm(), but the
last term indicates the random effect. In R formulas, ‘1’ is used to denote
a constant term in a regression equation (we write ‘-1’ in our formula if
we want no such term), and here ‘(1|usernum)’ specifies a random effects
intercept term that depends on usernum but is unobserved.
So, what is the answer to our speculation about age? Blind use of signif-
icance testing would mean announcing “Yes, there is a significant positive
relation between age and ratings.” But the effect is tiny; a 10-year differ-
ence in age would mean an average increase of only 0.03525, on a ratings
scale of 1 to 5. There doesn’t seem to be much difference between men and
women either.
The estimated variance of α, 0.175, is much smaller than that for ϵ, 1.073.
Of course, much more sophisticated analyses can be done, adding a variance
component for the movies, accounting for the different movie genres and so
on.

7.4.3 Multiple Random Effects

Of course, we can have more than one random effect. Consider the movie
data again, for instance (for simplicity, without the demographics). We
might model a movie rating Y as

Y =µ+γ+ν+ϵ (7.27)

where γ and ν are random effects for the user and for the movie.
The lme4 package can handle a very wide variety of such models, though
speficiation in the call to lmer() can become quite complex.

7.4.4 Why Use Random/Mixed Effects Models?

We may be interested in quantities such as σα2 for Description purposes,

especially relative to other variances, such as σϵ2 as in the example above.
7.5. REGRESSION FUNCTION AVERAGING 289

This is common in genetics applications, for instance, where we may wish

to compare within-family and between-family variation.
Random/mixed effects models may be used in Prediction contexts as well.
There is a rich set of methodology for this, concerning Best Linear Unbiased
Prediction (BLUP), unfortunately beyond the scope of this book.

7.5 Regression Function Averaging

Recall my old consulting problem from Section 1.11.1:

Long ago, when I was just finishing my doctoral study, I had

my first experience in statistical consulting. A chain of hos-
pitals was interested in comparing the levels of quality of care
given to heart attack patients at its various locations. A prob-
lem was noticed by the chain regarding straight comparison of
raw survival rates: One of the locations served a largely elderly
population, and since this demographic presumably has more
difficulty surviving a heart attack, this particular hospital may
misleadingly appear to be giving inferior care.

How do we deal with such situations? As mentioned in the introduction

to this chapter, my approach was to ask the question, How well would
Hospital 3 (the one with many elderly) do if it were to serve the populations
covered by all the hospitals, not just its own? This led to a method based
on computing the average value of an estimated regression function [106]
[107]. In subsequent years, this notion, which we’ll call regression function
averaging (RFA), has led not only to methods for handling situations like
the hospital example but also methodology for dealing with missing values
[34] [113].
Specifically, say we estimate some regression function µ(t), resulting in the
estimate µ b(t). RFA then averages the latter over some random variables
Q1 , ..., Qr , forming

1∑
r
b(Qi )
µ (7.28)
r i=1

Let’s see what can be done with this.

290 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

7.5.1 Estimating the Counterfactual

What an imposing word, counterfactual! It simply means, “What would

have happened if such-and-such had been different?” In our hospital ex-
ample above, the counterfactual was

How well would Hospital 3 (the one with many elderly) do if it

were to serve the populations covered by all the hospitals, not
just its own?

The problem there, in more specific statistical terms, was that the distri-
bution of one of the predictors, age, was different in Hospital 3 than for
the other hospitals. The field of propensity matching involves some rather
complex machinery designed to equilibrate away differences in predictor
variable distributions, by matching similar observations.
Roughly speaking, the idea is choose a subset of the patients at Hospital
3 who are similar to the patients at other hospitals. We can then fairly
compare the survival rate at Hospital 3 with those at the other institutions.
But we can do this more simply with RFA. The basic idea is to estimate
the regression function on the Hospital 3 data, then average that function
over the predictor variable data on all the hospitals. We would then have
an estimate of the overall survival rate if Hospital 3 had treated all the
patients. We could do this for each hospital, and thus compare them on a
“level playing field” basis.

7.5.1.1 Example: Job Training

The Lalonde data is an oft-used example in the propensity matching liter-

ature. It is available in the CRAN twang package [29]. The first column is
the treatment, 1 for participation in a training program and 0 if not in the
program.10 The outcome of interest is re78, earnings in 1978. We wish to
compare the training and nontraining groups with respect to earnings.
Let’s ask the question, “How much better would the nontraining group
have done if they had had the training?” We don’t want to simply do a
direct comparison of the two sample mean earnings, in the training and
nontraining groups, as the predictors likely have different distributions in
the two groups. Here is where RFA can be helpful.
10 Theterm “treatment” is not to be taken literally. It simply indicates a variable on
which we wish to compare groups, e.g., male vs. female.
7.5. REGRESSION FUNCTION AVERAGING 291

> l i b r a r y ( twang )
> data ( l a l o n d e )
> l l <− l a l o n d e
# s e p a r a t e d a t a frame i n t o
# t r a i n i n g and n o n t r a i n i n g g r p s
> t r t <− which ( l l $ t r e a t == 1 )
> l l . t <− l l [ t r t , −1]
> l l . nt <− l l [− t r t , −1]
# f i t r e g r e s s i o n on t r a i n i n g group
> lmout <− lm( r e 7 8 ∼ . , data= l l . t )
# f i n d and a v e r a g e t h e e s t r e g f t n
# v a l u e s on t h e n o n t r a i n i n g g r p
> c f y s <− predict ( lmout , l l . nt )
> mean( c f y s )
[ 1 ] 7812.121
# compare t o t h e i r a c t u a l e a r n i n g s
> mean( l l . nt$ r e 7 8 )
[ 1 ] 6984.17

So, by this analysis, had those in the nontraining group undergone training,
they would have earned about $800 more.
As with formal propensity analysis and many other statistical procedures,
there are important assumptions here. The analysis tacitly assumes that the
effects of the various predictors are the same for training and nontraining
populations, something to ponder seriously in making use of the above
figure.
Calculation of standard errors is discussed in Section 7.8.2 of the Mathe-
matical Complements section at the end of this chapter.

7.5.2 Small Area Estimation: “Borrowing from

Neighbors”

Mi casa es su casa [my house is your house] — Spanish saying

In this section we apply RFA to small area estimation. As noted in the
introductory section of this chapter, this term arises from the need to esti-
mate means, totals and other statistics in small subpopulations for which
very little data is available [117].
In other words, small-area estimation methods aim to bolster the statistical
accuracy of small samples. This is sometimes described as “borrowing” data
292 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

from other areas to improve estimation in a small area.

Models are typically rather elaborate, and depend heavily on certain pop-
ulation distributions being normal. Moreover, many such models involve
latent variables. But we can use RFA to do this more directly and with
fewer assumptions.
Let’s again look at the year 2000 census data on programmer and engineer
salaries, Section 6.12.2.2. Say we are interested in the mean wage of female
PhDs under age 35.11 Let’s find them in our sample:
> pe1doc <− pe1 [ which ( pe$phd == 1 ) , ]
> z <− pe1doc [ pe1doc$ age < 35 & pe1doc$ s e x == 2 , ]
> nrow( z )
[ 1 ] 23

Only 23 observations! Somewhat startling, since the full data set has over
20,000 observations, but we are dealing with a very narrow population here.
Let’s form a confidence interval for the mean wage:
> t . t e s t ( z$ wageinc )
...
95 p e r c e n t c o n f i d e n c e i n t e r v a l :
43875.90 70419.75
...

That is certainly a wide range, though again, not unexpectedly so. So, let’s
try RFA, as follows. (Justification for the intuitive analysis will be given
at the end.) In (7.28), the Qi are the Xi in the portion of our census data
corresponding to our condition (female PhDs under age 35).
So we fit a regression function model to our full data, PhDs and everyone,
b(Xi ). We then average those values:
resulting in fitted values µ

1 ∑
n
b(Xi )
µ (7.29)
N
Xi in A

where N is the number of female PhDs under 35.

We can call predict() to find the µ b(Xi ). (Or, we could use z$fitted.values;
this is left as an exercise for the reader.)
> lmout <− lm( wageinc ∼ . , data=pe1 )
11 Thisis a little different from typical SAE applications, where the grouping is geo-
graphical, but the principles are the same.
7.5. REGRESSION FUNCTION AVERAGING 293

> pr o u t <− predict ( lmout , z [ , − 6 ] )

> t . t e s t ( prout )
...
95 p e r c e n t c o n f i d e n c e i n t e r v a l :
58111.78 74948.87
mean o f x
66530.33

So, we indeed do obtain a narrower interval, i.e., have a more accurate

estimate of the mean wage in this particular subpopulation. Our point
estimate (7.29) is $66,530.
Note that we could do this even if we had no Yi information about the
subpopulation in question, i.e., if we did not know the salaries of those
femaler PhDs under 35, as long as we had their Xi values. This kind of
situation arises often in small-area applications.
It should be mentioned that there is a bit more to that RFA interval than
meets the eye. Technically, we have two sources of sampling variation:

• The Xi and Yi for this subpopulation, i.e., the 23 observations, are

considered a random sample from this subpopulation.

• The βbi coming from the call to lm() are also random, as they are
based on the 20,090 observations in our full data set, considered a
random sample from the full population.

The confidence interval computed above only takes into account variation
in that first bullet, not the second. We could use the methods developed
in Section 7.8.2, but actually the situation is simpler here. Owing to the
huge discrepancy between the two sample sizes, 23 versus 20090, we can to
a good approximation consider the βbi to be constants.
But there is more fundamental unfinished business to address. Does RFA
even work in this setting? Is it estimating the right thing? For instance, as
the sample size grows, does it produce a statistically consistent estimator
of the desired population quantity?
To answer those questions, let A denote the subpopulation region of inter-
est, such as female PhDs under age 35 in our example above. We are trying
to estimate

E(Y | X in A) (7.30)
294 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

the population mean wage income given that the person is a female PhD
under age 35. Let D denote the dummy variable indicating X in A. Then
(7.30) is equal to

E(DY )
(7.31)
P (X in A)

This makes intuitive sense. Since D is either 1 or 0, the numerator in (7.31)

is the population average value of Y among people for whom X is in A —
exactly (7.30) except for the scaling by P (X in A).
Now look at E(DY ). Using the Law of Total Expectation (1.58), write

E(DY ) = E [E(DY | X)] (7.32)

Now assume that D is a function of X. That is the case in our PhD

example above, where D is defined in terms of X. Then in the conditional
expectation in (7.32), D is a constant and

E(DY | X) = DE(Y |X) = Dµ(X) (7.33)

Combining this with (7.32), we have

E(DY ) = E[D µ(X)] (7.34)

so that (7.31) is

E[D µ(X)]
(7.35)
P (X in A)

The numerator here says, “Average the regression function over all people
in the population for whom X in A.” Now compare that to (7.29), which
we will rewrite as

∑
1
n Xi in A b(Xi )
µ
(7.36)
(N/n)

The numerator here is the sample analog of the numerator in (7.35), and
the denominator is the sample estimate of the denominator in (7.35). So,
(7.36) is exactly what we need, and it is exactly what we computed in the
R code for the PhD example above.
7.6. MULTIPLE INFERENCE 295

7.5.3 The Verdict

The RFA approaches introduced in the last two sections are appealing
in that they require many fewer assumptions than non-regression based
methodology for the same settings. Of course, the success of RFA relies on
having good predictor variables, at least collecitvely if not individually.

7.6 Multiple Inference

If you beat the data long enough, it will confess — old statistical joke
Events of small probability happen all the time, because there are so many
of them — the late Jim Sutton, economics professor and dean
We have already seen many examples of the use of lm() and glm(), and the
use of summary() to extract standard errors for the estimated regression
coefficients. We can form an approximate 95% confidence interval for each
population regression coefficient βi as

√
βbi ± 1.96 s.e.(βbi )/ n (7.37)

The problem is that the confidence intervals are only individually at the
95% level. Let’s look into this.

7.6.1 The Frequent Occurence of Extreme Events

Suppose we have a large collection of pennies. Although the minting process

results in some random imperfections, resulting in a slight variation in the
probability of heads from coin to coin, and though the design of a penny
inherently weights one side a bit more likely than the other, we will assume
here for simplicity that every penny has heads probability exactly 0.5.
Now suppose we give pennies to 500 people, and have each one toss his coin
100 times. Let’s find the probability that at least one of them obtains more
than 65 heads.
For a particular person, the probability of obtaining at most 65 heads is
> pbinom ( 6 5 , 1 0 0 , 0 . 5 )
[ 1 ] 0.999105
296 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

(See the Computational Complements section at the end of this chapter if

you are not familiar with this function.) So, the probability that a particular
person gets more than 65 heads is quite small:
> 1 − 0.999105
[ 1 ] 0.000895

But the probability that at least one of the 500 people obtains more than
65 heads is not small at all:
> 1 − 0.99910ˆ500
[ 1 ] 0.3609039

The point is that if one of the 500 pennies does yield more than 65 heads,
we should not say, “Oh, we’ve found a magic penny.” That penny would
be the same as the others. But given that we are tossing so many pennies,
there is a substantial probability that at least one of them yields a lot of
heads.
This issue arises often in statistics, as we will see in the next section. So,
let’s give it a name:

The Principle of Rare Events

Consider a set of independent events, each with a certain prob-
ability of occurrence. Even if the individual even probabilities
are small, there may still be a substantial probability that at
least one of them occurs.

7.6.2 Relation to Statistical Inference

In many statistical situations, when we form many confidence intervals,

we wish to state that they are collectively at the 95% level. (The reader
may substitute significance testing for confidence intervals in the following
discussion.) In the linear regression context, for instance, we wish to form
the intervals in such a way that there is a 95% chance that all of the
intervals do contain their respective βi . But that collective probability may
be considerably less than 0.95 if each of the intervals is at the 95% level.
In the context of the Principle of Frequent Occurrence of Rare Events, the
rare events at hand here consist of a confidence interval being incorrect,
i.e., failing to contain the parameter being estimated. Each event has only
a 0.95 probability, but there is a much larger chance than 0.05 that at least
one of the intervals is incorrect.
7.6. MULTIPLE INFERENCE 297

Why is this a problem? Say one is in some Description application, running

regression analysis with a large value of p, i.e., with many predictors. We
are interested in seeing which predictors have a substantial impact on Y ,
and we form a confidence interval for each βi . If a no-impact situation
corresponds to βi = 0, we are especially interested in which intervals are
substantially away from 0.12
The problem is that even if all the true βi are near 0, with large p the
chances are high that at least one interval will accidentally be far from 0.
Hence the joke quoted above — “beating the data long enough” alludes to
looking at a large number of predictors in hope of finding some with strong
impact, and “the data will confess” means at least one of the intervals will
(incorrectly) grant our wish.
This goal of attaining an overall 0.95 level for our intervals is known as the
multiple inference or simultaneous inference problem. Many methods have
been developed, enough to fill entire books, such as [73] [25]. Many of the
methods, though, are either tied exclusively to significance testing or rely
heavily on assumptions of normally-distributed populations, both of which
are restrictions this book aims to avoid.
Instead, we will focus here on two methods. The first, the Bonferroni
method is quite famous. The second, the Scheffe’s method, is well known in
its basic form, but little known in the generalized form presented here.

7.6.3 The Bonferroni Inequality

This one is simple. Say we form two 95% confidence intervals.13 Intuitively,
their overall confidence level will be only something like 90%. Formally, let
Ai , i = 1, 2 denote the events that the intervals fail to cover their respective
population quantities. Then

P (A1 or A2 ) =

P (A1 ) + P (A2 ) − P (A1 and A2 ) ≤ P (A1 ) + P (A2 ) (7.38)

In other words, the probability that at least one of the intervals fails is
at most 2 (0.05) = 0.10. If we desire an overall confidence level of 95%,
12 Note that if an interval merely excludes 0 but has both lower and upper bounds near

0, we would probably not judge this predictor as having an important impact on Y .

13 These typically will be approximate intervals, but for brevity this term will not be

repeatedly inserted here.

298 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

we can form two intervals of level 97.5%, and be assured that our overall
confidence is at least 95%.
Due to the qualifier at least above, we say that our intervals are conservative.
Using mathematical induction, this can be easily generalized to

∑
k
P (A1 or A2 or ... or Ak ) ≤ P (Ai ) (7.39)
i=1

Using this for forming simultaneously valid confidence intervals as above,

we have:

If we form k confidence intervals of confidence level 1 − α, the

overall confidence level is at least 1 − kα.
Thus, to achieve overall confidence level of at least 1 − α, we
must set the confidence level of each of our intervals to 1 − α/k.

7.6.4 Scheffe’s Method

Consider the setting of Section 2.8.4. There we found that the asymptotic
distribution of βb is multivariate normal with mean β and covariance matrix
as in (2.150). The results of Section 7.8.3 imply that the quantity

W = s2 (βb − β)′ (A′ A)−1 (βb − β) (7.40)

has an asymptotically chi-square distribution with p+1 degrees of freedom.

That implies that

P (W ≤ χ2α,p+1 ) ≈ 1 − α (7.41)

where χ2q,k denotes the upper-q quantile of the chi-square distribution with
k degrees of freedom.
This in turn implies that the set of all b such that

W = s2 (βb − b)′ (A′ A)−1 (βb − b) (7.42)

7.6. MULTIPLE INFERENCE 299

forms an approximate 1 − α confidence region for the population β in p + 1

dimensional space. This in itself is usually not very valuable, but it can be
used to show the following highly useful fact:

The confidence intervals

√
c′ βb ± s χ2α,p+1 c′ (A′ A)−1 c (7.43)

for the population quantities c′ β have overall confidence level

1 − α, where c ranges over all p + 1 dimensional vectors.

This may seem abstract, but for example consider a vector c consisting
of all 0s except for a 1 in position i. Then c′ βb = βbi and c′ β = βi . In
other words, (7.43) is giving us simultaneous confidence intervals for all the
coefficients βi .
Another common usage is to set c to a vector having all 0s except for a 1
at positions i and a -1 at j. This sets up a confidence interval for βi − βj ,
allowing us to compare coefficients.
Recall that the quantity

s2 (A′ A)−1 (7.44)

is given by R’s vcov() function. Denoting this matrix by V , (7.43) becomes

√
c′ βb ± χ2α,p+1 c′ V −1 c (7.45)

The above analysis extends to any asymptotically normal estimator:

Let θb denote an asymptotically r-variate normal estimator of a

d θ)
population vector θ, with a consistent estimator Cov( b of the
associated covariance matrix. Then the confidence intervals

√
c′ θb ± d θ)]
χ2α,p+1 c′ [Cov( b −1 c (7.46)

for the population quantities c′ θ have overall confidence level

1 − α, where c ranges over all r dimensional vectors.
300 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

This applies to logistic regression, for instance.

7.6.5 Example: MovieLens Data

Let’s predict movie rating from user age, gender and movie genres. The
latter are: Unknown, Action, Adventure, Animation, Children’s, Comedy,
Crime, Documentary, Drama, Fantasy, Film-Noir, Horror, Musical, Mys-
tery, Romance, Sci-Fi, Thriller, War and Western. In order to have inde-
pendent data points, we will proceed as in Section 3.2.4, where we computed
the average rating each user gave to the movies he/she reviewed. But how
should we handle genre, which is specific to each movie the user reviews?
In order to have independent data, we no longer are looking at the level of
each movie rated by the user.
We will handle this by calculating, for each user, the proportions of the
various genres in the movies rated by that user. Does a user rate a lot in
the Comedy genre, for instance? (Note that these proportions do not sum to
1.0, since a movie can have more than one genre.) In order to perform these
calculations, we will use R’s split() function, which is similar to tapply()
but more fundamental. (Actually, tapply() calls split().) Details are
shown in Section 7.7.2, where the final results are placed in a data frame
saout. Now let’s run the regression.
> sam <− as . matrix ( s a o u t )
> summary(lm( s a o u t [ , 3 ] ∼ sam [ , c ( 4 : 2 4 ) ] ) )

Note that we needed to convert saout from a data frame to a matrix, in

order to use column numbers in our call to lm(). Here is the result:
Coefficients :
Estima te Std . E r r o r t v a l u e
( Intercept ) 2.893495 0.325803 8.881
sam [ , c ( 4 : 2 4 ) ] age 0.001338 0.001312 1.020
sam [ , c ( 4 : 2 4 ) ] g e n d e r −0.008081 0 . 0 3 4 0 9 5 −0.237
sam [ , c ( 4 : 2 4 ) ] GN6 −0.625717 9 . 5 3 1 7 5 3 −0.066
sam [ , c ( 4 : 2 4 ) ] GN7 −0.717794 0 . 2 8 6 7 4 9 −2.503
sam [ , c ( 4 : 2 4 ) ] GN8 1.084081 0.427063 2.538
sam [ , c ( 4 : 2 4 ) ] GN9 1.009125 0.660956 1.527
sam [ , c ( 4 : 2 4 ) ] GN10 −0.481950 0 . 4 4 8 6 4 4 −1.074
sam [ , c ( 4 : 2 4 ) ] GN11 0.190537 0.278022 0.685
sam [ , c ( 4 : 2 4 ) ] GN12 0.083907 0.371962 0.226
sam [ , c ( 4 : 2 4 ) ] GN13 2.564606 1.272708 2.015
sam [ , c ( 4 : 2 4 ) ] GN14 0.377494 0.291400 1.295
7.6. MULTIPLE INFERENCE 301

sam [ , c ( 4 : 2 4 ) ] GN15 −0.576265 1.282073 −0.449

sam [ , c ( 4 : 2 4 ) ] GN16 1.268462 0.858390 1.478
sam [ , c ( 4 : 2 4 ) ] GN17 0.274987 0.313803 0.876
sam [ , c ( 4 : 2 4 ) ] GN18 −0.115167 0.577974 −0.199
sam [ , c ( 4 : 2 4 ) ] GN19 −0.578825 0.433863 −1.334
sam [ , c ( 4 : 2 4 ) ] GN20 0.631700 0.222688 2.837
sam [ , c ( 4 : 2 4 ) ] GN21 0.249433 0.345188 0.723
sam [ , c ( 4 : 2 4 ) ] GN22 0.577921 0.296188 1.951
sam [ , c ( 4 : 2 4 ) ] GN23 1.452932 0.356611 4.074
sam [ , c ( 4 : 2 4 ) ] GN24 1.821073 1.057404 1.722
Pr ( >| t | )
( Intercept ) < 2 e −16 ∗∗∗
sam [ , c ( 4 : 2 4 ) ] age 0.30788
sam [ , c ( 4 : 2 4 ) ] g e n d e r 0 . 8 1 2 7 0
sam [ , c ( 4 : 2 4 ) ] GN6 0.94767
sam [ , c ( 4 : 2 4 ) ] GN7 0.01248 ∗
sam [ , c ( 4 : 2 4 ) ] GN8 0.01130 ∗
sam [ , c ( 4 : 2 4 ) ] GN9 0.12716
sam [ , c ( 4 : 2 4 ) ] GN10 0.28300
sam [ , c ( 4 : 2 4 ) ] GN11 0.49331
sam [ , c ( 4 : 2 4 ) ] GN12 0.82158
sam [ , c ( 4 : 2 4 ) ] GN13 0.04419 ∗
sam [ , c ( 4 : 2 4 ) ] GN14 0.19549
sam [ , c ( 4 : 2 4 ) ] GN15 0.65319
sam [ , c ( 4 : 2 4 ) ] GN16 0.13982
sam [ , c ( 4 : 2 4 ) ] GN17 0.38109
sam [ , c ( 4 : 2 4 ) ] GN18 0.84210
sam [ , c ( 4 : 2 4 ) ] GN19 0.18249
sam [ , c ( 4 : 2 4 ) ] GN20 0 . 0 0 4 6 6 ∗∗
sam [ , c ( 4 : 2 4 ) ] GN21 0.47011
sam [ , c ( 4 : 2 4 ) ] GN22 0.05134 .
sam [ , c ( 4 : 2 4 ) ] GN23 5 . 0 1 e −05 ∗∗∗
sam [ , c ( 4 : 2 4 ) ] GN24 0.08537 .

It is interesting to compare this with the analysis in Section 3.2.4. There

age was found “significant” in the statistical sense but not in a practical
sense. Its estimated coefficient has now decreased even further, due to the
addition of the genre variables (see the “thunder stealing” remark in Section
7.1).
Instead, the relatively big effects here were those associated with genre
— but only relatively. One must keep in mind that the genre variables
are proportions. Consider for instance the Action genre, with estimated
302 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

coefficient -0.717794. If we have two users, one that has 20% of her ratings
on Action films, and another user for whom that figure is 30%, the estimated
impact on mean rating is only -0.07.
Only some of the genre variables came out “significant.” One might at first
think this odd, since we have 100,000 observations in our data, large enough
to pick up even small deviations from βi = 0. But that is 100,000 ratings,
and since we have collapsed our analysis to the person level, we must take
note of the fact that we have only 943 users here.
Now let’s apply multiple inference to the genre variables. Since there are
19 of them, in order to achieve a confidence level of (at least) 0.95, we need
to set the individual level for each interval at 0.951/19 . That means that
instead of using the standard 1.96 to compute the radius of our interval,
we need the value corresponding to this level:

> −qnorm( ( 1 − 0 . 9 9 7 3 0 4 ) / 2 )
[ 1 ] 3.000429

That means that each of the widths of the confidence intervals will increase
by a factor of 3.00/1.96, about 50%. That may be a worthwhile price to pay
for the ability to make intervals that hold jointly. But the price is rather
dramatic if one does significance testing. This book discourages the use of
testing, but it is instructive to look at the effect of multiple inference in a
testing context.
Here the “significant” genre variables are those whose entries in the ouptut
column labeled “t value” are greater than 3.00 in absolute value. Only
one genre, GN24, now qualifies, compared to five if no multiple inference is
performed.
What about Scheffe’ ? Here, instead of 1.96 we will use chi-square quantile
in (7.43):

> sqrt ( qchisq ( 0 . 9 5 , 1 9 ) )

[ 1 ] 5.490312

This is even larger than the Bonferroni value.

So, there is a price to be paid for multiple inference, but it does yield
protection that in many contexts is quite important.
7.7. COMPUTATIONAL COMPLEMENTS 303

7.6.6 The Verdict

As seen in the example above, Scheffe’s method is generally quite conserva-

tive. It is thus useful only if you wish to form a large number of confidence
intervals; otherwise, Bonferroni will be a better choice, among these two
methods. It should be noted, though, that both are relying on normal ap-
proximations, say of the βbi . These approximations are fine for most uses,
but in our call to qnorm() above, we are working far out in the right tail
of the distribution, where the normal approximation is questionable.
As noted earlier, there are many other methods for multiple inference. Most
of them have more restrictions than Bonferroni or Scheffe’, such as a popu-
lation normality assumption or being restricted to hypothesis testing, but
many analysts have found them useful. The interested reader should con-
sult the literature, such as the Hsu reference cited earlier.

7.7 Computational Complements

7.7.1 MovieLens Data Wrangling

Here are the details of the merge of the ratings data and demographics data
in Section 7.4.2.1.
As in Section 7.7.2, we need to use the R merge() function. Things are a
little trickier here, because that function relies on having a column of the
same name in the two data frames. Thus we need to assign our column
names carefully.
> r a t i n g s <− read . table ( ’ u . data ’ )
> names( r a t i n g s ) <− c ( ’ usernum ’ , ’ movienum ’ , ’ r a t i n g ’ , ’
transID ’ )
> demog <− read . table ( ’ u . u s e r ’ , s e p= ’ | ’ )
> names( demog ) <− c ( ’ usernum ’ , ’ age ’ , ’ g e n d e r ’ , ’ o c c ’ ,
’ ZIP ’ )
> u . b i g <− merge( r a t i n g s , demog , by . x=1,by . y=1)
> u <− u . b i g [ , c ( 1 , 3 , 5 , 6 ) ]

7.7.2 More Data Wrangling in the MovieLens Example

While tapply() partitions a data frame into groups and then applies some
summary function, split() does only the partitioning, returning an R list,
304 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

with each list element being the rows of the data frame corresponding to
one group. In our case here, we will group the user data by user ID.
> # read data , form column names
> ud <− read . table ( ’ u . data ’ , h e a d e r=F , s e p= ’ \ t ’ )
> uu <− read . table ( ’ u . u s e r ’ , h e a d e r=F , s e p= ’ | ’ )
> u i <− read . table ( ’ u . item ’ , h e a d e r=F , s e p= ’ | ’ )
> ud <− ud [ , − 4 ] # remove timestamp
> uu <− uu [ , 1 : 3 ] # user , age , g e n d e r
> u i <− u i [ , c ( 1 , 6 : 2 4 ) ] # item num, g e n r e s
> names( ud ) <− c ( ’ u s e r ’ , ’ item ’ , ’ r a t i n g ’ )
> names( uu ) <− c ( ’ u s e r ’ , ’ age ’ , ’ g e n d e r ’ )
> names( u i ) [ 1 ] <− ’ item ’
> names( u i ) [ − 1 ] <−
gsub ( ’V ’ , ’GN ’ ,names( u i ) [ − 1 ] ) # g e n r e s
> uu$ ge n d e r <− as . integer ( uu$ g e n d e r == ’M’ )
> # merge t h e 3 d f s
> u a l l <− merge( ud , uu )
> u a l l <− merge( u a l l , u i )
> # u a l l now i s ud+uu+u i ; now s p l i t by u s e r
> u s e r s <− s p l i t ( u a l l , u a l l $ u s e r )

At this point, for instance, users[[1]] will consist of all the rows in uall for
user ID 1:
> head ( u s e r s [ [ 1 ] ] , 3 )
item u s e r r a t i n g age g e n d e r GN6 GN7 GN8 GN9
1 1 1 5 24 1 0 0 0 1
510 2 1 3 24 1 0 1 1 0
637 3 1 4 24 1 0 0 0 0
GN10 GN11 GN12 GN13 GN14 GN15 GN16 GN17 GN18
1 1 1 0 0 0 0 0 0 0
510 0 0 0 0 0 0 0 0 0
637 0 0 0 0 0 0 0 0 0
GN19 GN20 GN21 GN22 GN23 GN24
1 0 0 0 0 0 0
510 0 0 0 1 0 0
637 0 0 0 1 0 0

Now, for each user, we need to find the mean rating and the proportions in
each genre. Since the genre variables are dummies, i.e., 0,1-valued, those
proportions are just their means. Here is the code:
> s a o u t <− sapply ( u s e r s , colMeans )
> s a o u t <− t ( s a o u t )
7.7. COMPUTATIONAL COMPLEMENTS 305

The R function sapply() applies a given function to each element of an R

list, producing a new list, and then attempting to collapse the latter into
a matrix or vector. In our case here, since each element of users is a data
frame, applying colMeans() to it produces a vector, after which sapply()
combines those vectors into a matrix.
As is typical with the output of sapply(), the row and colum arrangement
is not convenient, so we take the transpose. Here is the top of the final
matrix:

> head ( saout , 3 )

item u s e r r a t i n g age g e n d e r GN6
1 136.5000 1 3 . 6 1 0 2 9 4 24 1 0.003676471
2 249.5000 2 3 . 7 0 9 6 7 7 53 0 0.000000000
3 318.8148 3 2 . 7 9 6 2 9 6 23 1 0.000000000
GN7 GN8 GN9 GN10
1 0.2757353 0.15441176 0.04411765 0.09191176
2 0.1612903 0.04838710 0.01612903 0.06451613
3 0.2592593 0.07407407 0.00000000 0.00000000
GN11 GN12 GN13 GN14
1 0.3345588 0.09191176 0.01838235 0.3897059
2 0.2580645 0.14516129 0.00000000 0.5645161
3 0.2222222 0.18518519 0.01851852 0.4074074
GN15 GN16 GN17 GN18
1 0.007352941 0.003676471 0.04779412 0.04779412
2 0.016129032 0.032258065 0.03225806 0.01612903
3 0.000000000 0.037037037 0.09259259 0.03703704
GN19 GN20 GN21 GN22
1 0.01838235 0.16176471 0.15808824 0.1911765
2 0.06451613 0.25806452 0.06451613 0.1935484
3 0.20370370 0.09259259 0.14814815 0.3888889
GN23 GN24
1 0.09191176 0.02205882
2 0.04838710 0.00000000
3 0.09259259 0.00000000

Look at row 1, for instance. User 1 had a mean rating of 3.610294 in the
movies he rated. 27.57353% of the movies were of genre GN7 (Action), and
so on. Note that within each element of the list users, columns such as
age and gender are constant, so we have simply taken the average of those
constants, a bit wasteful but no problem.
306 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

7.8 Mathematical Complements

7.8.1 Iterated Projections

The Tower Property of conditional expectations states that

E [E(V |U1 , U2 ) | U1 ] = E(V | U1 ) (7.47)

This can be shown quite simply and elegantly using the vector space meth-
ods of Sections 1.19.5.5 and 2.12.8, as follows. (Readers may wish to review
Section 2.12.8 before continuing.) Our only assumption is that the variables
involved have finite variances, so that the vector space in 2.12.8 exists.
Let A and B denote the subspaces spanned by all functions of (U1 , U2 ) and
U1 , respectively. Denote the full space as C.
Write

u=V (7.48)

v = E(V |U1 , U2 ) (7.49)

w = E [E(V |U1 , U2 ) | U1 ] (7.50)

Note that B is a subspace of A. Indeed, w, a vector in B, is the projection

of v, a vector in A, onto B.
What we need to show is that w is also the projection of u onto B. Since
projections are unique, we will accomplish this if we show that

(w, u − w) = 0 (7.51)

Write

u − w = (u − v) + (v − w) (7.52)
7.8. MATHEMATICAL COMPLEMENTS 307

Now consider each of the two terms on the right, which we will show are
orthogonal to w. First,

(w, v − w) = 0 (7.53)

because w is the projection of v onto B. Second, since v is the projection

of u onto A, we know that

(q, u − v) = 0 (7.54)

for any q ϵ A, including the case q = w. Thus we have

(w, u − w) = 0 (7.55)

as desired, completing the proof.

7.8.2 Standard Errors for RFA

Recall that RFA is defined as

1∑
r
νb = b(Qi )
µ (7.56)
r i=1

We will assume that µ(t) is linear in t, with the notation of Section 2.4.2.
And for convenience, assume a model with no intercept term (Section 2.4.5).
Then (7.56) becomes

1 ∑ ′b
r
′
νb = Q β = Q βb (7.57)
r i=1 i

where Q is the (vector-valued) sample mean of the Qi .

As in Section 2.8, our standard errors will be conditional on the Xj . We
will also assume that the Qi are i.i.d. and independent of the (Yi , Xi ) from
which βb is computed.
308 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

To derive an estimate of V ar(b

ν ), we will use the Delta Method (Section
3.6.1),14 as follows.
Let Q denote a generic random variable having the distribution of the Qi .
From (7.57), our function f () in (3.26) is simply

f (s, t) = st (7.58)

with a = (EQ, β)′ . Using the reasoning in Section 3.6.1,

′
νb ≈ EQ′ βb + Q β = EQ′ βb + β ′ Q (7.59)

so that

AV ar(b b EQ + β ′ Cov(Q)β
ν ) = EQ′ Cov(β) (7.60)

b by (2.57), β by
To get standard errors, we then replace EQ by Q, Cov(β)
b
β and Cov(Q) by the sample covariance matrix of the Qi ,

1∑
r
(Qi − Q)(Qi − Q)′ (7.61)
r i=1

Here is code for the computation:

r f a s e <− function ( q , lmout ) {
quadf <− function ( u , v ) ( t ( u ) %∗% v %∗% u ) [ 1 ]
qbar <− colMeans ( q )
qbarcov <− cov ( q ) / nrow( q )
bhat <− coef ( lmout )
bhatcov <− vcov ( lmout )
sqrt ( quadf ( qbar , bhatcov ) + quadf ( bhat , qbarcov ) )
}

7.8.3 Asymptotic Chi-Square Distributions

The chi-square family of distributions has a single parameter, m, called

its degrees of freedom. The family is defined as follows. Let Z1 , ..., Zm be
14 In that section, we had n → ∞. Here we need r → ∞ too. For convenience take
r/n to be constant.
7.9. EXERCISES: DATA, CODE AND MATH PROBLEMS 309

independent random variables, each having a N(0,1) distribution, and set

Y = Z12 + ... + Zm
2
(7.62)

The chi-square distribution with m degrees of freedom is defined to be the

distribution of Y .
More generally, suppose V has an r-variate normal distribution with mean
vector ν and invertible covariance matrix Q. Then the diagonalizability of
Q can be used to show that the quantity

W = (V − ν)′ Q−1 (V − ν) (7.63)

has a chi-square distribution with r degrees of freedom (Exercise 6).

Now consider any asymptotically multivariate normal estimator θb of an r-

dimensional population vector θ. Let C denote the associated covariance
b denote a consistent estimate of C. Then the quantity
matrix, and let C

(θb − θ)′ C
b −1 (θb − θ) (7.64)

is then asymptotically chi-square distributed with r degrees of freedom.

7.9 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. Extend the analysis of the schooling data, Section 7.4.1.3, by incor-
porating other (observed) predictors besides educ, years of school. Pay
particular attention to fit, in view of the finding in that section that the
Two-Stage Least Squares fit was quite different from the OLS fit at the
point we tried, 12 years of education.
2. Form simultaneous confidence intervals for the 48 word variables in
Section 4.3.6. Try both the Bonferroni and Scheffe’ approaches.
3. In Chapter 1, we found that baseball players gain about a pound of
weight per year. We might ask whether there is a team effect. Add team
membership as a predictor, and investigate this possibility.
310 CHAPTER 7. DISAGGREGATING REGRESSOR EFFECTS

Mini-CRAN and other computational problems:

4. Write an R function of call form
s i m u l c i ( lmout , s i n g l e s=NULL, d i f f s =NULL,
a l p h =0.05 , method=c ( ’ bonf ’ , ’ s c h e f f e ’ )

that will form simultaneous confidence intervals, using the Bonferroni or

Scheffe’ method, of overall level 1 - alph, from the output lmout from
lm() or glm(). The intervals will be for the coefficients named in singles
and/or differences of coefficients named in diffs. The latter is a two-column
character matrix, one row per difference.
5. In the RFA computation in Section 7.5.2, it was suggested that fit-
ted.values be used insead of predict(), thereby saving a computational
step. Show how to do this. Hint: Use rownames().
Math problems:
6. Show that under the given assumptions, W in (7.63) has a chi-square
distribution as claimed.
7. Derive a closed-form expression for the Two-Stage Least Squares estima-
tor in the case in which the instrument Z is binary, i.e., a dummy variable.
Extend this to the case of any categorical variable (i.e. R factor) Z.
8. Use the Delta Method (Section 3.6.1) to derive the standard error for
βb12 in (7.20).
Chapter 8

Shrinkage Estimators

Suppose we are estimating a vector mean. Consider for instance the base-
ball player example, Section 1.6. Let the vector (H,W,A) denote the height,
weight and age of a player. Say we are interested in the population mean
vector

µ = (EH, EW, EA) (8.1)

The natural estimator would be the vector of sample means,

b = (H, W , A)
µ (8.2)

where H is the mean height of all players in our sample, and so on. Yet it
turns out that this “natural” estimator is not optimal in a certain theoreti-
cal sense. Instead, the theory suggests that it is actually better to “shrink”
(8.2) to a smaller size.
Theory is of course not a focus of this book. However, this particular
theoretical finding regarding “shrunken” estimators has had a major impact
on some of the current applied methodology in regression and classification.
Among the various methods developed for multivariate analysis in recent
years, many employ something called regularization. This technique shrinks
estimators, or equivalently, keeps them from getting too large. In particular,
the LASSO has become popular in machine learning circles and elsewhere.
The theoretical findings can help guide our intuition in practical settings,
so we will begin with a brief summary of the theory.

311
312 CHAPTER 8. SHRINKAGE ESTIMATORS

Shrinkage estimators will be the subject of this chapter, and will influence
Chapter 9 as well.

8.1 Relevance of James-Stein to Regression

Estimation

Recall that the very definition of the regression function is the conditional
mean,

µ(t) = E(Y | X = t) (8.3)

Since James-Stein theory says estimates of means should be shrunken, it is

not surprising that the same is true for estimates of regression functions,
at least say in linear models. In the latter context, µ becomes µ(t) and the
shrinking is applied to estimates of the vector β that parameterizes µ(t).
This illustrates the bias-variance tradeoff introduced in Section 1.11.1. In
the regression context, shrinking will likely produce bias (definitely so in
the case of the linear model), but by making the estimate smaller it should
reduce variance. If a good amount of shrinkage can be determined, it may
be a “win.”
Define the quantity

σ
v= (8.4)
b
||β||

where σ is as in (2.48). Then the discussion in Section 8.11.1.3, carried over

to our regression context, suggests the following about shrinking β: b

• For fixed p and v, the smaller our sample size n is, the more we need
to shrink.
• For fixed n and v, the more predictor variables we have, the more we
should shrink.
• For fixed n, p and β, the larger σ is, the more we need to shrink.

Again, it is questionable whether the assumptions of James-Stein theory

are directly relevant to real life. Nevertheless, the qualitative message of
that theory is highly relevant to actual practice.
8.2. MULTICOLLINEARITY 313

For instance, let’s consider what James-Stein might say about situations in
which our predictors are highly correlated with each other. Recall (2.54):

b = σ 2 (A′ A)−1
Cov(β) (8.5)

This quantity might be “large” (loosely defined) if the matrix inverse (A′ A)−1
is large, even if σ is small. The inverse will be large if A′ A is “small.” Again,
the latter term is loosely defined for now, but the main point that will
emerge below is that all this will occur if our predictors are multicollinear,
meaning that they are highly correlated.
Thus another bullet should be added to the rough guidelines above:

• For fixed values of n, p and v, the stronger the degree of multicollinear-

ity, the more the need for shrinkage.

What exactly is multicollinearity? We take this up next.

8.2 Multicollinearity

As we have been discussing in this chapter, the coefficient for a predictor

variable may be greatly affected by the presence or absence of other predic-
tors. This becomes a particular concern when there are strong correlations
among the predictors.

8.2.1 What’s All the Fuss About?

8.2.2 A Simple Guiding Model

To see the problem, first recall the R function scale(), introduced in Section
1.21. For each variable in the data it is applied to, this function centers
and scales the variable, i.e., subtracts the variable’s mean and divides by its
standard deviation. The resulting new versions of the variables now have
mean 0 and variance 1.
Suppose we apply scale() to our predictor variables. Also rewrite (2.28) as

1 1
βb = ( A′ A)−1 · A′ D (8.6)
n n
314 CHAPTER 8. SHRINKAGE ESTIMATORS

b but it will prove to be quite useful.

This won’t change the resulting β,
Due to the centering and scaling, we are in essence fitting a model without
a constant term β0 (Section 1.21), thus no 1s column in the matrix A. The
matrix n1 A′ A in (2.28) then becomes the correlation matrix of the predictors
(Exercise 4).
If we have just two predictors, for instance, we will have

( )
1 ′ 1 c
AA= (8.7)
n c 1

where c is the correlation between the two predictors. Then to get the
estimated regression coefficients, we will compute the inverse of n1 A′ A in
(2.28). That inverse, from (A.18), is

( )
1 1 −c
(8.8)
1 − c2 −c 1

Now we can see the problem arising if the two predictor variables are highly
correlated with each other: The value of c will be near 1, so the quantity
1/(1 − c2 ) may be huge. In that light, (2.54) tells us that our estimated
coefficients will have very large standard errors. This is bad — confidence
intervals will be very wide, and significance tests will have low power.

8.2.3 Checking for Multicollinearity

The previous section shows how to check for multicollinearity in the case
of just two predictors. How do we do this in general?

8.2.3.1 The Variance Inflation Factor

Noting the squared correlation in the demoninator of (8.8), a natural gen-

eralization would be to look at multiple R2 . For predictor variable X (i) ,
we would run lm(), predicting X (i) from all the other X (j) ; denoting the
resulting multiple R2 by Ri2 . To complete the analogy to (8.8), we define
the variance inflation factor (VIF) for X (i) by

1
V IFi = (8.9)
1 − Ri2
8.2. MULTICOLLINEARITY 315

A rule of thumb used by some analysts is to treat VIF values over 10 as

cause for concern, though setting this threshhold is up to each analyst.
We could calculate the VIF values ourselves, by repeated calls to lm(), but
the vif() function in the car package does this for us [49].

8.2.3.2 Example: Currency Data

Recall the currency data analysis in Section 6.8.1.1. We had regressed

the yen against the Canadian dollar, the mark, the franc and the pound,
assigning the output to fout1. We can input the latter object to vif():
> v i f ( fout1 )
Can Mark Franc Pound
5.575051 4.438270 14.119749 6.582210

Let’s check the Can case:

> tmp <− summary(lm( Can ∼Mark+Franc+Pound , data=c u r r 1 ) )
> 1 / ( 1 − tmp$ r . s q u a r e d )
[ 1 ] 5.575051

By the way, various definitions for VIF for the generalized linear model,
i.e., glm(), have been proposed. The vif() function in car applies one of
them.

8.2.4 What Can/Should One Do?

Once one has determined that the data has serious multicollinearity issues
— and note again, this is up to the individual analyst to decide whether it
is a problem — what are possible remedies?

8.2.4.1 Do Nothing

Multicollinearity can produce counterintuitive results in the signs of the

βbj . However, the analyst should be extremely reluctant to resort to anti-
multicollinearity measures shown below (or any others) such as ridge re-
gression and the LASSO, simply because he/she does not like the signs. On
the contrary, they may be the “real” signs in various senses, as discussed
in Chapter 7.
316 CHAPTER 8. SHRINKAGE ESTIMATORS

If our application involves the Description goal of regression analysis, we

may really want to know the effects of all the predictor variables. The
currency data analyzed above might be such a case. In this situation, we
might simply accept the fact that the standard errors of the coefficients are
large. Indeed, if the sample is sizable, then even those “large” standard
errors might be quite acceptable.

8.2.4.2 Eliminate Some Predictors

In principle, a variable with a high VIF is essentially redundant, and thus

not likely to add much predictive power. One might then remove it from
one’s set of predictor variables. Elimination of predictor variables is a major
issue in regression and classification analysis, and all of Chapter 9 will be
devoted to it.

8.2.4.3 Employ a Shrinkage Method

We turn to this very topic now.

8.3 Ridge Regression

Equation (8.8) suggested that multicollinearity results in ( n1 A′ A)−1 be-

coming “too large,” causing our estimated coefficient vector βb to be too
large. This fits in with the “too large, too often” intuitive motivation for
James-Stein theory, and thus the James-Stein crowd would react to the
multicollinearity problem by saying “Shrink it!” Two common methods for
this are ridge regression and the LASSO. The latter is more widely used,
but we can get better insight by considering the former first.
The approach involves adding a positive constant λ to the diagonal of A′ A,
to make the matrix “larger,” resulting in a “smaller” ( n1 A′ A)−1 . Here λ is
a tuning parameter, to be chosen by the analyst.
Equation (2.28) now becomes

βbridge = (A′ A + λI)−1 · A′ D (8.10)

In fact, shrinking does indeed occur, as we will see shortly.

8.3. RIDGE REGRESSION 317

A generalization would be to add different constants to different elements

of the diagonal, say because the predictors are of widely different scales.
But it is easier just to apply the R function scale() to the predictors, as
we saw earlier in this chapter.
To avoid confusion, let us denote the ordinary βb by βbOLS (Section 3.3.)

8.3.1 Alternate Definitions

There are other ways of defining ridge regression. These are not just math-
ematical parlor tricks; we will see later that they will be quite useful.
The first alternative definition is that for fixed λ, the ridge estimator is the
value of b that minimizes

∑
n
ei b)2 + λ||b||22
(Yi − X (8.11)
i=1

The idea here is that on the one hand we want to make the first term
(the sum of squares) small, so that b gives a good fit to our data, but on
the other hand we don’t want b to be too large. The second term acts
like a “governor” to try to prevent having too large a vector of estimated
regression coefficients; the larger we set λ, the more we penalize large values
of b.
Let’s see why this is consistent with (8.10). First, rewrite (8.11) as

(D − Ab)′ (D − Ab) + λb′ b (8.12)

in the notation of Section 2.4.2. Since we are discussing minimization of

(8.11), let’s take the derivative with respect to b, which, again following
Section 2.4.2, is

−2A′ (D − Ab) + 2λb (8.13)

Setting this to 0, and writing b = Ib, we have

(A′ A + λI)b = A′ D (8.14)

just as in (8.10).
318 CHAPTER 8. SHRINKAGE ESTIMATORS

This leads to our second alternative definition of ridge: It can be shown

that minimizing (8.12) is equivalent to minimizing

(D − Ab)′ (D − Ab) (8.15)

subject to a constraint of the form

||b||22 ≤ γ (8.16)

(γ and λ are not numerically equal, but are functions of each other.)

This really makes the point that we want to avoid letting βb get too large.
(See also Section 8.11.2 in the Mathematical Complements portion of this
chapter.)

8.3.2 Choosing the Value of λ

Well, then. Faced with actual data, what is the practioner to do? Using
ridge regression might solve the practioner’s multicollinearity problem, but
how is he to choose the value of λ?
One way to choose λ is visual: For each predictor variable, we draw a graph,
the ridge trace, that plots the associated estimated coefficient against the
value of λ. We choose the latter to be at the “knee” of the curve. The
function lm.ridge() in the MASS package (part of the base R distribution)
can be used for this, but here we will use the function ridgelm() from
regtools, due to its approach to scaling. Here is why:
As is standard, the ridgelm() function calls scale() on the predictors and
centers the response variable. But ridgelm() goes a little further, also
applying the 1/n scaling we used in Section 8.2.1. Equation (8.6) becomes

1 1
βb = ( A′ A + λI)−1 · A′ D (8.17)
n n

Here is the point: As noted earlier, n1 A′ A is the correlation matrix of our

predictor data. Thus this matrix has 1s on the diagonal, with the off-
diagonal elements being smaller than 1 in absolute value. This makes it
much easier to choose a range of values to try for λ, more or less indepen-
dent of what data we have. Thus the default value for the λ argument in
ridgelm() is (0.01,0.02,...,1.00), meaning that the function will try these
values of λ.
8.3. RIDGE REGRESSION 319

Figure 8.1: Ridge analysis of the currency data

The code for ridgeln() includes some noteworthy computational aspects,

and is discussed in the Computational Complements at the end of the
chapter.
Another approach to choosing λ is cross validation. We then choose λ to
be the value that predicts new cases best. The function ridge.cv() in the
parcor package does this [82].

8.3.3 Example: Currency Data

We continue from Section 8.2.3.2:

> library ( r e g t o o l s )
> plot ( r i d g e l m ( c u r r 1 ) )

The results are displayed in Figure 8.1. From top to bottom along the left
edge, the curves show the values of βM ark , βP ound , βCan and βF ranc . The
“knee” is visible for the franc and Canadian dollar at about 0.15 or 0.20,
though interestingly the curve for the mark continues to decline significantly
after that.
320 CHAPTER 8. SHRINKAGE ESTIMATORS

Ridge regression is thus rather subjective, without much “science” behind

it. Perhaps the most scientific approach is to use cross validation. Thus
let’s try ridge.cv():
> r i d g e . cv ( as . matrix ( c u r r 1 [ , − 5 ] ) , c u r r 1 [ , 5 ] )
$intercept

224.9451

$coefficients
XCanada XMark XFranc XPound
−5.786784 5 7 . 7 1 3 9 7 8 −34.399449 −5.394038

$lambda . opt
[ 1 ] 0.3168208

The recomended λ value here is about 0.32, rather larger than what we
might have chosen using the “knee” method. On the other hand, this
larger value makes sense in light of our earlier observation concerning the
mark.
Shrinkage did occur. Here are the OLS estimates:
> lm( Yen ∼ . , data=c u r r 1 )
...
Coefficients :
( Intercept ) Canada Mark Franc
224.945 −5.615 57.889 −34.703
Pound
−5.332

Ridge slightly reduced the absolute values of most of the coefficients, Canada
being the exception. The fact that the reductions were only slight should
not surprise us, given the rough guidelines in Section 8.11.1.3. The n/p
ratio is pretty large, and even the multicollinearity was mild according to
the generally used rule of thumb (Section 8.2.3.1).

8.4 The LASSO

Much of our material on the LASSO will appear in Chapters 9 and 12 but
we introduce it in this chapter due to its status as a shrinkage estimator. To
motivate this method, recall first that shrinkage estimators form another
8.4. THE LASSO 321

example of the bias-variance tradeoff. With ridge regression, for instance,

by shrinking β, b we are reducing its variance (actually, its covariance ma-
trix), at the expense of introducing some bias. If we can choose a good value
of λ, we can find a “sweet spot” in that tradeoff, and hopefully improve
predictive ability. This of course is the motivation for using cross-validation
to choose λ.
The Least Absolute Shrinkage and Selection Operator — the LASSO —
takes another approach to shrinking. As with ridge regression, the LASSO
actually has two equivalent formulations, which in rough terms are:

b
• Penalize large values of β.

b
• Place an explicit limit to the size of β.

We will begin with the first of these.

8.4.1 Definition

As noted in Section 8.2.4.2 and in earlier chapters, a more traditional way

than shrinkage to improve prediction error is subset selection, meaning to
pare down the set of predictor variables into a smaller but representative
set. As discussed earlier, this reduces variance, though again increasing
bias. One advantage of this approach is that it is appealing to deal with
just a small number of predictors, often termed a parsimonious model.
The LASSO was invented with the goal of combining the best aspects of
ridge regression on the one hand, and subset selection on the other. It
involves shrinkage, like ridge regression, but often results in a roundabout
way of doing subset selection.
So, how does the LASSO accomplish all this? The answer is remarkably
simple: In (8.11), simply replace ||b||22 by ||b||1 (see (A.2)). In other words,
the LASSO estimator is defined to be the value of b that minimizes

∑
n
ei b)2 + λ||b||1
(Yi − X (8.18)
i=1

Similar to the ridge case, one can show that an equivalent definition is that
322 CHAPTER 8. SHRINKAGE ESTIMATORS

the LASSO estimator is chosen to minimize

∑
n
ei b)2
(Yi − X (8.19)
i=1

subject to a constraint of the form

||b||1 ≤ γ (8.20)

Using the argument in Section 8.11.2, we see that the LASSO does produce
a shrinkage estimator. But it is designed so that typically many of the
estimated coefficients turn out to be 0, thus effecting subset selection, which
we will see in Section 9.7.7.1.

8.4.2 The lars Package

We’ll use the R package lars [63]. It starts with no predictors in the model,
then adds them (in some cases changing its mind and deleting some) one at
a time. At each step, the action is taken that is deemed to best improve the
model, as with forward stepwise regression, to be discussed in Chapter 9.
At each step, the LASSO is applied, with λ determined by cross-validation.
The lars package is quite versatile. Only its basic capabilities will be shown
here.

8.4.3 Example: Currency Data

As noted, the LASSO is commonly used as a method for variable selection,

the topic of Chapter 9. Since we have only p = 4 predictors, and more
than 700 observations, variable selection is not really an issue. But in this
chapter’s context of multicollinearity, it is of interest to see how much the
software decides to shrink.
> l a s s o u t <− l a r s ( as . matrix ( c u r r 1 [ , − 5 ] ) , c u r r 1 [ , 5 ] )
> lassout
...
R−s q u a r e d : 0 . 8 9 2
Sequence o f LASSO moves :
Canada Mark Pound Franc
Var 1 2 4 3
8.4. THE LASSO 323

Step 1 2 3 4

Note that lars requires the predictor values to be given as a matrix.

At Step 0, there are no predictors; it is a regression model with just a
constant term, so we are just predicting Y from its unconditional mean.
We see that at Step 1, lars() brought in the Canada predictor, then the
mark, then the pound and lastly, the franc.
Let’s take a closer look:

> summary( l a s s o u t )
LARS/LASSO
Call : l a r s ( x = as . matrix ( c u r r 1 [ , −5]) , y = c u r r 1 [ , 5 ] )
Df Rss Cp
0 1 2052191 6 2 6 3 . 5 0
1 2 2041869 6 2 3 0 . 1 8
2 3 392264 5 8 7 . 3 1
3 4 377574 5 3 9 . 0 4
4 5 220927 5.00

The Cp criterion is similar to adjusted-R2 , and will be discussed in full in

Chapter 9. The user may choose to use the Cp value as a guide as to which
model to use. In this case, that approach would choose the full model, with
all predictors, not surprising in this context of p << n.
Since the LASSO is mainly used for subset selection, the actual values of
the estimated coefficients are rather secondary, and not presented in the
output of summary(). But they are indeed accessible:

> l a s s o u t $beta
Canada Mark Franc Pound
0 0.0000000 0.00000 0.00000 0.000000
1 −0.2042481 0 . 0 0 0 0 0 0.00000 0.000000
2 −28.6567963 2 8 . 4 5 2 5 5 0.00000 0.000000
3 −28.1081479 2 9 . 6 1 3 5 0 0 . 0 0 0 0 0 −1.393401
4 −5.6151436 5 7 . 8 8 8 5 6 −34.70273 −5.331583
...

Again, this is presented in terms of the values at each step, and the 0s
show which predictors were not in the model as of that step. In our multi-
collinearity context in this chapter, we are interested in the final values, at
Step 4. They are seen to provide shrinkage similar to the mild amount we
saw in Section 8.3.3.
324 CHAPTER 8. SHRINKAGE ESTIMATORS

8.4.4 The Elastic Net

Although ridge regression had been around for many years, its popular-
ity was rather limited. But the introduction of the LASSO in the 1990s
(and some related methods proposed slightly prior to it) revived interest
in shrinkage estimators for regression contexts. A cottage industry among
statistics/machine learning researchers has thrived ever since then, includ-
ing various refinements of the LASSO idea.
One of those refinements is the elastic net, defined to be the value of b
minimizing

∑
n
ei b)2 + λ1 b1 + λ2 ||b||22
(Yi − X (8.21)
i=1

The idea behind this is that one might not be sure whether ridge or LASSO
would have better predictive power in a given context, so one can “hedge
one’s bets” by using both at once! Again, one could then rely on cross-
validation to choose the values of the λi .

8.5 Cases of Exact Multicollinearity,

Including p > n

These days it is common to have more predictors than observations, i.e., to

have p > n. Such a situation used to be dismissed as impossible, since the
matrix A in (2.28) would necessarily be of less than full rank, so that A′ A
would not be invertible.
However, today, people are more adventurous (some might say recklessly
so), and they hope to do regression and classification analysis in such situ-
ations. Shrinkage estimators provide a possible solution.

8.5.1 Why It May Work

This may be easily seen in the case of ridge regression. The key point is
that even if A′ A is not invertible, A′ A + λI will be invertible for any λ > 0).
(This follows from the analysis of Section 8.11.3 and the fact that the rank
of a matrix is the number of nonzero eigenvalues.) So, to many people,
there is hope for the case p > n!
8.5. CASES OF EXACT MULTICOLLINEARITY 325

8.5.2 Example: R mtcars Data

This is one of the famous built-in data sets in R, with n = 32 and p = 11.
So this is NOT an example of the p > n situation, but as will be seen, we
can use this data to show how ridge can resolve a situation in which A′ A
is not invertible.
The cyl column in this data set shows the number of cylinders in the car’s
engine, 4, 6 or 8. Let’s create dummy variables for each of these three
types:
> l i b r a r y ( dummies )
> dmy <− dummy( mtcars $ c y l )
> mtcars <− cbind ( mtcars , dmy)

Let’s predict mpg, gas mileage, just from the number of cylinders. Since
there are three categories of engine, we should only use two of those dummy
variables.1 But let’s see what happens if we retain all three dummies.
In the matrix A of predictor variables, our first column will consist of all 1s
as usual, but consider what happens when we add the vectors in columns
2, 3 and 4, where our dummies are: Their sum will be an all-1s vector, i.e.,
column 1! Thus one column of A will be equal to a linear combination of
some other columns (Section A.4), so A will be less than full rank. That
makes A′ A noninvertible.
Let’s use all three anyway:
> l i b r a r y ( dummies )
> d <− dummy( mtcars $ c y l )
> mtcars <− cbind ( mtcars , d )
> lm(mpg ∼ c y l 4+c y l 6+c y l 8 , data=mtcars )
...
Coefficients :
( Intercept ) cyl4 cyl6 cyl8
15.100 11.564 4.643 NA

R was smart enough to notice the column dependency, so it simply ignored

the cyl8 column. Indeed, if we omit that column ourselves, we get the
same result:

1 Readers familiar with Analysis of Variance (ANOVA) will recognize this as one-way

ANOVA. There the model is EYij = µ + αi + ϵij , where in our case i would equal 1,2,3.
∑
There again would be a redundancy, but it is handled with the constraint 3i=1 αi = 0.
326 CHAPTER 8. SHRINKAGE ESTIMATORS

> lm(mpg ∼ c y l 4+c y l 6 , data=mtcars )

...
Coefficients :
( Intercept ) cyl4 cyl6
15.100 11.564 4.643

We of course should have omitted that column in the first place. But
let’s see how ridge regression could serve as an alternate solution to the
dependency problem:
> head ( t ( r i d g e l m ( mtcars [ , c ( 1 2 : 1 3 , 1 ) ] ) $ b h a t s ) )
cyl4 cyl6
[ 1 , ] 11.40739 4.527374
[ 2 , ] 11.25571 4.416195
[ 3 , ] 11.10838 4.309103
[ 4 , ] 10.96522 4.205893
[ 5 , ] 10.82605 4.106375
[ 6 , ] 10.69068 4.010370

Recall that each row here is for a different value of λ. We see that shrinking
occurs, as anticipated, thus producing bias over repeated sampling, but use
of ridge indeed allowed us to overcome the column dependency problem.

8.5.2.1 Additional Motivation for the Elastic Net

In many senses, the elastic net was developed with the case p >> n in mind.
This occurs, for instance, in many genomics studies, with there being one
predictor for each gene under consideration. In this setting, Description,
not Prediction, is the main goal, as we wish to determine which genes affect
certain traits.
Though the LASSO would seem to have potential in such settings, it is
limited to finding p nonzero regression coefficients. This may be fine for
Prediction but problematic in, say, in genomics settings.
Similarly, the LASSO tends to weed out a predictor if it is highly correlated
with some other predictors. And this of course is exactly the issue of
concern, multicollinearity, in the early part of this chapter, and again, it
is fine for Prediction. But in the genomics setting, if a group of genes is
correlated but influential, we may wish to know about all of them.
The elastic net is motivated largely by these perceived drawbacks of the
LASSO. It is implemented in, for instance, the R package elasticvnet.
8.6. BIAS, STANDARD ERRORS AND SIGNFICANCE TESTS 327

8.6 Bias, Standard Errors and Signficance

Tests

Due to their bias, shrinkage estimators generally do not yield straightfor-

ward calculation of standard errors or p-values. Though some methods
have been proposed, mainly for p-values [27], they tend to have restrictive
assumptions and are not widely used and notably, software libraries for
shrinkage methods do not report standard errors.
Some analysts view the inclusion of a predictor by LASSO as akin to find-
ing that the predictor is “statistically significant.” This interpretation of
LASSO output is debatable, for the same reason as is the basic notion of
significance testing of coefficients: For large n (but fixed p), the LASSO
will include all the predictors.

8.7 Principal Components Analysis

The Variance Inflation Factor makes sense as a gauge of multicollinearity,

because it measures the possible deleterious impact of that problem. But
we might investigate multicollinearity in our data directly: Is one predictor
nearly equal to some linear combination of the others? The material in
Section A.6 is perfect for answering this question.
For convenience, suppose our predictor variables are centered, i.e., have
their means subtracted from them, so they have mean 0.
Here is our strategy: We will find p new predictor variables that are linear
combinations of our p original ones. Some of them will have very small
variance, which essentially makes them approximately constant. Since the
original predictors have mean 0, so will the new predictors, and thus in the
preceding sentence “approximately constant” will mean approximately 0 —
i.e., one of the original predictors will approximately be a linear combination
of the others.
In this way, we will identify multicollinearities. Note the plural in that
latter word, by the way. More than one of the new predictors could be
approximately 0.
The key equation is (A.21), which, to avoid a clash of symbols, we will
328 CHAPTER 8. SHRINKAGE ESTIMATORS

rewrite as

U ′ BU = G (8.22)

As in Section 2.4.2, let A denote the matrix of our predictor variables

(without a 1s column, due to centering of our data). In (8.22) take

B = A′ A (8.23)

i.e.,

G = U ′ A′ AU = (AU )′ AU (8.24)

from (A.14).
Column j of A is our data on predictor variable j. Now, in the product
AU , consider its first column. It will be A times the first column of U , and
will thus be a linear combination of the columns of A (the coefficients in
that linear combination are the elements of the first column of U ).
In other words, the first column of AU is a new variable, formed as a linear
combination of our original variables. This new variable is called a prin-
cipal component of A. The second column of AU will be a different linear
combination of the columns of A, forming another principal component,
and so on.
Also, Equation (2.79) implies2 that the covariance matrix of the new vari-
ables is G. Since G is diagonal, this means the new variables are uncor-
related. That will be useful later, but for now, the implication is that the
diagonal elements of G are the variances of the principal components.
And that is the point: If one of those diagonal elements is small, it cor-
responds to a linear combination with small variance. And in general, a
random variable with a small variance is close to constant. In other words,
multicollinearity!
“The bottom line,” is that we can identify multicollinearity by inspecting
the elements of G, which are the eigenvalues of A′ A. This can be done
using the R function prcomp().3
2 Again,
beware of the clash of symbols.
3 Wecould also use svd(), which computes the singular value decomposition of A. It
would probably be faster.
8.8. GENERALIZED LINEAR MODELS 329

This will be especially useful for generalized linear models, as VIF does not
directly apply. We will discuss this below.

8.8 Generalized Linear Models

Even though multicollinearity is typically discussed in the context of linear

models, it is certainly a concern from generalized linear models. The latter
“contain” linear models in their specification, and use them in their compu-
tation (Section 4.5). The shrinkage parameter λ is used as a counterweight
to extreme values of the log-likelihood function.
The R package glmnet can be used to compute shrinkage estimators in the
generalized linear models sense. Both the LASSO and the elastic net are
available as options in this package..

8.8.1 Example: Vertebrae Data

Consider the Vertebral Column data in Section 5.5.2. First, let’s check this
data for multicollinearity using diagonalization as discussed earlier:
> vx <− as . matrix ( v e r t [ , − 7 ] )
> prcomp ( vx )
Standard d e v i a t i o n s :
[ 1 ] 42.201442532 18.582872132 13.739609112 10.296340128
[ 5 ] 9.413356950 0.003085624

That’s quite a spread, with the standard deviation of one of the linear
combinations being especially small. That would suggest removing one of
the predictor variables. But let’s see what glmnet() does.
At first, let’s take the simple 2-class case, with classes DH and not-DH.
Continuing from Section 5.5, we have
> vy <− as . integer ( v e r t $V7 == ’DH ’ )
> coef ( glmnet ( vx , vy , family= ’ b i n o m i a l ’ ) )
V1 . . . .
V2 . . . .
V3 . . −0.0001184646 −0.002835562
V4 . −0.008013506 −0.0156299299 −0.020591892
V5 . . . .
V6 . . . .
330 CHAPTER 8. SHRINKAGE ESTIMATORS

V1 . . .
V2 . . .
V3 −0.005369667 −0.00775579 −0.01001947
V4 −0.025365930 −0.02996199 −0.03438807
V5 . . .
V6 . . .
...

i
The coefficients estimates from the first six iterations are shown, one column
per λ. The values of the latter were

> vout$lambda
[ 1 ] 0.1836231872 0.1673106093 0.1524471959
[ 4 ] 0.1389042072 0.1265643403 0.1153207131
...

There is as usual the question of which λ value is best. Unfortunately, it

is even less clear than the corresponding question for a linear model, and
we will not pursue it here. The analyst may wish to use the old standby,
cross-validation here.

8.9 Other Terminology

The notion of placing restraints on classical methods such as least-squares

is often called regularization. Ridge regression is called Tykhanov regu-
larization in the numerical analysis community. Another common term is
penalized methods, motivated by the fact that, for instance, the second term
in (8.18) “penalizes” us if we make b too large.

8.10 Further Reading

The classic paper on ridge regression is [68]. It’s well worth reading for the
rationale for the method.
A general treatment of the LASSO is [65].
8.11. MATHEMATICAL COMPLEMENTS 331

8.11 Mathematical Complements

8.11.1 James-Stein Theory

Consider further the height-weight-age example at the beginning of this

chapter, but with some other variables in there too, a total of p in all. We
have a random sample of size n, and are interested in the mean vector.
Assume that each of these variables has a normal distribution with vari-
ance σ 2 , with the latter value being known; these assumptions are rather
unrealistic, but our goal is to gain insight from the resulting theory, which
we will apply by extension to the regression case.

8.11.1.1 Definition

The James-Stein estimator [136] is defined to be

( )
(p − 2)σ 2 /n
1− b
µ (8.25)
||b
µ||2

Here || || denotes the Euclidean norm, the square root of the sums of squares
of the vector’s elements (Equation (A.1)).

8.11.1.2 Theoretical Properties

b in the sense
It can be shown that the James-Stein estimator outperforms µ
of Mean Squared Error,4 as long as p ≥ 3. This is pretty remarkable! In
b is nonoptimal in the height-
the ballplayer example above, it says that µ
weight-age case, but NOT if we are simply estimating mean height and
weight. Something changes when we go from two dimensions to three or
more.5

4 Defined µ − µ||2 ).
in the vector case as E(||b
5 Oddly, there is also a fundamental difference between one and two dimensions versus
three or more for random walk. It can be shown that a walker stepping in random
directions will definitely return to his starting point if he is wandering in one or two
dimensions, but not in three or more. In the latter case, there is a nonzero probability
that he will never return.
332 CHAPTER 8. SHRINKAGE ESTIMATORS

A key aspect of the theory is that if

(p − 2)σ 2 /n
<1 (8.26)
||b
µ||2

b toward 0.
then James-Stein shrinks µ

8.11.1.3 When Might Shrunken Estimators Be

Helpful?

Looking at (8.25), we can see the circumstances in which the James-Stein

(J-S) estimator will be substantially different from the ordinary one. We
can see immediately, for instance, that for large n, ceteris paribus, J-S will
essentially be the same as our ordinary sample-means vector µ b. J-S is
useful only for small samples. However, the factor p − 2 tells us that here
the definition of “small” depends on dimension; for larger values of p, J-S
may be helpful even in somewhat larger samples.
On a more subtle level, recall the quantity known as the coefficient of
variation from statistical practice, defined to be the ratio of a variable’s
standard deviation to its mean. We don’t know whether to consider a
standard deviation of say, 2.8, small or large, so we compare that 2.8 value
µ||, so J-S won’t be
to the mean of the variable. In (8.25), the quantity σ/||b
very useful if σ is small relative to ||b
µ||.

8.11.2 Yes, It Is Smaller

b as advertised? The answer is

So, does ridge regression actually shrink β,
yes, i.e.,

||βbridge ||2 ≤ ||βbOLS ||2 (8.27)

(In fact, the inequality above will be strict in any practical situation.) But
why?
There are intuitive answers to this question. We remarked earlier that
adding λI to A′ A would intuitively seem to make that matrix “larger,”
thus making βb smaller. And the constraint (8.16) would seem to imply
shrinkage, but such an argument is not airtight. What if, say, βb would
have satisified (8.16) anyway, without applying the ridge procedure?
8.11. MATHEMATICAL COMPLEMENTS 333

Instead, there is actually a very easy way to see that shrinking does indeed
occur. Let’s look again at (8.12), giving names to the various expressions:

f (b) = (D − Ab)′ (D − Ab) + λb′ b = g(b) + h(b) (8.28)

In minimizing that expression with respect to b, we have a tradeoff between

the two terms, g(b) and h(b). Since setting b = βbOLS minimizes g(b),
taking b = βbridge means accepting a larger value of g(b). But that must be
accompanied by a reduction in h(b); otherwise b = βbOLS would give us a
smaller f (b) than would b = βbridge , a contradiction, since the latter is the
b that minimizes f (b). So, h(b) does get smaller when we go from OLS to
ridge, and thus the ridge estimator is indeed shrunken.
More formally,

g(βbridge ) + h(βbridge ) = f (βbridge ) (8.29)

≤ f (βbOLS ) (8.30)
= g(βbOLS ) + h(βbOLS ) (8.31)
≤ g(βbridge ) + h(βbOLS ) (8.32)

In other words, it must be that

h(βbridge ) ≤ h(βbOLS ) (8.33)

i.e.,

||βbridge ||2 ≤ ||βbOLS ||2 (8.34)

Not only is this the simplest way to demonstrate mathematically that ridge
estimators are shrinkage estimators, the same argument above shows that
shrinkage occurs for any vector norm, not just l2 . The LASSO, to be
discussed later in this chapter, uses the l1 norm, so we immediately see
that the LASSO shrinks too.

8.11.3 Ridge Action Increases Eigenvalues

Let M = A′ A in the context of (2.19). It is nonnegative definite (Section

A.8).
334 CHAPTER 8. SHRINKAGE ESTIMATORS

Under ridge regression, a value λ is added to the diagonal of M , i.e., M is

replaced by

M + λI (8.35)

Now suppose ν is an eigenvalue of M , with eigenvector x. Then

(M + λI)x = M x + λx = νx + λx (8.36)

So, ν + λ is an eigenvalue of the “ridge-ized” matrix M + νI, again with

eigenvector x. Since λ is positive, applying the ridge action to M has
increased its eigenvalues.

8.12 Computational Complements

8.12.1 Code for ridgelm()

There are several noteworthy aspects. Here first is the code:

r i d g e l m <− function ( xy , lambda=seq ( 0 . 0 1 , 1 . 0 0 , 0 . 0 1 ) ,
mapback=TRUE) {
p <− ncol ( xy ) − 1 ; n <− nrow( xy )
x <− xy [ , 1 : p ]
y <− xy [ , p+1]
x <− s c a l e ( x ) ; y <− y − mean( y )
tx <− t ( x )
xpx <− tx %∗% x / n
xpy <− tx %∗% y / n
mapftn <− function ( lambval )
qr . solve ( xpx + lambval ∗diag ( p ) , xpy )
tmp <− Map( mapftn , lambda )
tmp <− Reduce ( cbind , tmp )
i f ( mapback ) {
s d s <− attr ( x , ’ s c a l e d : s c a l e ’ )
for ( i i n 1 : p ) tmp [ i , ] <− tmp [ i , ] / s d s [ i ]
}
r e s u l t <− l i s t ( b h a t s=tmp , lambda=lambda )
c l a s s ( r e s u l t ) <− ’ rlm ’
result
}
8.12. COMPUTATIONAL COMPLEMENTS 335

plot . rlm <− function ( r i d g e l m . out ) {

lamb <− r i d g e l m . out $lambda
bhs <− t ( r i d g e l m . out $ b h a t s )
matplot ( lamb , bhs , type= ’ l ’ , pch= ’ . ’ ,
x l a b= ’ lambda ’ , y l a b= ’ beta−hat ’ )
}

print . rlm <−

function ( r i d g e l m . out ) print ( t ( r i d g e l m . out $ b h a t s ) )

Note the use of the R functions Map() and Reduce(), borrowed from
functional languages such as LISP. The line

tmp <− Map( mapftn , lambda )

results in mapftn() being applied to each element of the vector lambda,

yielding a vector of βbi for each of those elements. Those vectors are returned
in an R list, and we wish to combine them into a matrix. The call

tmp <− Reduce ( cbind , tmp )

accomplishes this, by applying cbind() to all those vectors.6

Now look at this code excerpt:

x <− s c a l e ( x ) ; y <− y − mean( y )

...
i f ( mapback ) {
s d s <− attr ( x , ’ s c a l e d : s c a l e ’ )
for ( i i n 1 : p ) tmp [ i , ] <− tmp [ i , ] / s d s [ i ]
}

When we called scale() on x, R did as asked, and divided each of the

columns of x, i.e., each of the predictor variables, by its standard deviation.
That results in a corresponding increase in each βbi by the same factor. Thus,
if we wish to have the βbi on the original scale, we need to divide by those
standard deviations. Fortunately, they were saved for us — in x! Calling
scale() bestowed an R attribute on x, with information about the original
means and standard deviations of the columns of x. As seen above, we can
then retrieve the standard deviations via a call to attr().

6 The R function do.call() could also have been used here.

336 CHAPTER 8. SHRINKAGE ESTIMATORS

Finally, recall our brief discussion of R’s S3 classes in Section 1.20.4.7 How
does one actually create an object of a certain S3 class? This is illustrated
above. We first form an R list, containing the elements of our intended
class object, then set its class:
r e s u l t <− l i s t ( b h a t s=tmp , lambda=lambda )
c l a s s ( r e s u l t ) <− ’ rlm ’

The object result is now an object of S3 class ’rlm’.

We can then do function dispatch on objects of class ’rlm’, just as in Section
1.20.4. In particular, calling the generic function plot() on such objects
will result in the call being dispatched to plot.rlm(). The case of another
generic function, print(), is similar.
Note by the way the use of the R graphing function matplot(), which plots
multiple curves based on columns of matrix.

8.13 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. As discussed at various points in this book, one may improve a para-
metric model by adding squared and interaction terms (Section 1.16). In
principle, one could continue in that vein, adding cubic terms, quartic terms
and so on. However, in doing so, we are likely to quickly run into multi-
collinearity issues. Indeed, lm() may detect that the A′ A matrix of Section
2.4.2 is so close to singular that the function will refuse to do the compu-
tation.
Try this scheme on the baseball player data in Section 1.6, predicting weight
from height. Keep adding higher-degree powers of height until lm() com-
plains, or until at least one of the coefficients returned by the function is
the NA value.
Mini-CRAN and other computational problems:
2. A simple alternative approach to shrinkage estimators in the regression
context would be to do shrinking directly, as follows, in the LOOM context:
7R also features two other types of classes, S4 and reference classes.
8.13. DATA, CODE AND MATH PROBLEMS 337

We choose γ to minimize

∑
n
(Yi − γ X̃i βb−i )2 (8.37)
i=1

where βb−i is the βb vector resulting from fitting the model to all observations
but the ith . Write an R function to do this.
3. The output of summary.lars() shows Cp values. Alter that code so
that it also shows adjusted-R2 . You may wish to review Section 1.20.4 and
the fact that R treats functions as objects, i.e., they are mutable.
Math problems:
4. In Section 8.2.2, it was stated that if our predictor variables are centered
and scaled, then A′ A will become the correlation matrix of our predictors.
Derive this fact.
Note: Take the sample correlation between vectors U and V of length n to
be

∑n
1
n i=1 (Ui− U )(Vi − V )
(8.38)
sU sV

using the sample means and standard deviations.

5. Consider a fixed-X setting (Section 2.3), with an orthogonal design
(Section 2.3. Find a simple, closed-form expression for βbr , r = 1, ..., p
under the ridge method for a specified λ.
Chapter 9

Variable Selection and

Dimension Reduction

[Mathematical models] should be made as simple as possible, but not simpler.

— Albert Einstein
If I had more time, I would have written a shorter letter — 18th century
mathematician Blaise Pascal
For every complex problem there is an answer that is clear, simple, and
wrong — H.L. Mencken
In the long and continuing history of regression and classification methodol-
ogy, the burning question has always been, How should one select predictor
variables? We may have data on many potential predictors, but somehow
feel compelled to “thin them out.” Motivations generally include:

• Avoidance of overfitting, thus hoping for better estimation or predic-

tion.
• Parsimony — simple models are more appealing.
• Determination of the “important” predictors.

Denote the number of predictors used by p. Since we are working in p-

dimensional space (or p + 1, counting the response variable), some refer to
the variable selection process as dimension reduction. The term is especially
apt in view of approaches to the problem that are explicitly based on vector
subspace searches, to be covered in Section 9.9.

339
340 CHAPTER 9. VARIABLE SELECTION

In spite of the long quest for the Dimension Reduction Holy Grail, in many
settings there are good reasons NOT to delete any potential predictor vari-
able, such as:

• If one’s goal is Description, then variable selection generally renders

statistical inference techniques invalid. A nominal 95% confidence
interval no longer has that confidence level, because the distribution
of the data changes, once one takes into account the fact that that
distribution is now conditional on the variable-selection process. In
addition, there is the multiple inference problem, discussed in Section
7.6. (More on this in Section 9.8.)

• Again in the Description setting, a parsimonious model may be mis-

leading. We have already seen, in Chapter 7, that omission of some
predictors can substantially change the estimated coefficients of the
remaining predictors in a parametric model. This goes to the Mencken
quote at the top of this chapter.

• In the Prediction setting, an omitted variable, though not generally

very useful, might have good predictive power in some key subregion
of the predictor space.

On the other hand, with many modern data sets, the dimension reduction
issue cannot be ignored. In particular, it is now common for p to be much
larger than the number of observations n, a setting in which, for instance, a
classic linear parametric model cannot be fit. In such a situation, our hand
is forced; we have very few options other than to do variable selection.1
Much theoretical (and empirical) work has been done on this subject. This
is not a theoretical book, but at some points in this chapter we will discuss
in nontheoretical terms what the implications of the research work are for
the real-world practice of regression and classification analysis.
The aim of this chapter, then, is to investigate the dimension reduction
issue, both from why and how points of view: Why might it be desirable in
some situations, and how can one do it? The layout of the chapter is:

• Sections 9.1 through 9.2 will cover the why.

• Section 9.3 through 9.9 will handle the how.

1 Ridge regression is a possibility, but questionable if p is much larger than n.
9.1. A CLOSER LOOK AT UNDER/OVERFITTING 341

Note that polynomial and interaction terms (Section 1.16) are also predic-
tors. so, deciding on a model may also entail which quadratic terms to
include, for instance.
By the way, we will treat the case p >> n separately, in Chapter 12.

9.1 A Closer Look at Under/Overfitting

We introduced the notion of under/overfitting in Section 1.11, and will go

into more detail here, as it has strong implications for the methods to be
discussed in this chapter. (The reader may wish to review Section 1.11
before coninuing.)
b plays two roles:
Recall that our estimated regression function µ

• The quantity µb(t) is an estimate of µ(t), and in parametric models

it is thus an estimate of the coefficients. These are of interest in
Description contexts.

• In Prediction contexts, the quantity µ

b(Xnew ) is our predicted Y value
for a new data point with predictor vector Xnew .

b
The point is that, whether our interest is in Description or Prediction, µ
is central. But as an estimator, it is subject to both variance and bias
considerations. In particular, as noted in Section 1.11, we have that for
any t,

MSE(b
µ(t)) = V ar[b µ(t))]2
µ(t)] + [bias(b (9.1)

So, again we see the famous variance-bias tradeoff: Richer models, i.e.,
those with more predictors, have smaller bias but larger variance.
Equation (9.1) concerns estimation of the regression function at a single
point t. For the Prediction goal, we need to see how well we are doing over
a range of t. In Random-X contexts (Section 2.3), we typically have t range
according to the distribution of X, yielding the Mean Squared Prediction
Error (MSPE),

( )
E (V ar[b µ(X))]2
µ(X)]) + E [bias(b (9.2)
342 CHAPTER 9. VARIABLE SELECTION

9.1.1 A Simple Guiding Example

Here we present a simple toy example to help guide our intuition in under-
standing (9.1). Suppose we have the samples of boys’ and girls’ heights at
some age, say 10, X1 , ..., Xn and Y1 , ..., Yn . Assume for simplicity that the
variance of height is the same for each gender, σ 2 . The means of the two
populations are designated by µ1 and µ2 .
Say we wish to estimate µ1 and µ2 . The “obvious” estimators are

1∑
n
b1 =
µ Xi (9.3)
n i=1

and

1∑
n
b2 =
µ Yi (9.4)
n i=1

But at age 10, boys and girls tend to be about the same height. So if n is
small, we may wish to make the simplifying assumption that µ1 = µ2 , and
then just use the overall mean as our estimate of µ1 and µ2 :

1
µ̌i = (X + Y ), i = 1, 2 (9.5)
2

Each estimate is now based on 2n observations instead of just n, thus

reducing variance. On the other hand, if µ1 − µ2 is large — we don’t know
this, which is why we would be doing the estimation in the first place, but
we can still imagine the consequences of a large µ1 − µ2 — then we will
have introduced substantial bias, thus possibly negating the advantage of
using 2n data points for each estimator. Let’s investigate this precisely.
We’ll take as our criterion total MSE,

M SE(b
µ1 ) + M SE(b
µ2 ) (9.6)

and

M SE(µ̌1 ) + M SE(µ̌2 ) (9.7)

9.1. A CLOSER LOOK AT UNDER/OVERFITTING 343

The computations in the Mathematical Complements portion of this chap-

ter, Section 9.13.1, yield that (9.6) has the value

σ2 2σ 2
2( + 02 ) = (9.8)
n n

where the 0 quantity is the bias.

The computations for (9.7) yield

σ2 1
+ (µ1 − µ2 )2 (9.9)
n 2

So, let’s call our model

µ1 = µ2 (9.10)

the Lower-Dimensional Model, and call µ1 ̸= µ2 the Higher-Dimensional

Model, reflecting the fact that the latter has two parameters, µ1 and µ2 ,
while the former has only one, the presumed common value of µ1 and µ2 .
Comparing the two MSE values above, we see that using the Lower-Dimensional
Model will pay off if and only if

(µ1 − µ2 )2 < 2σ 2 /n (9.11)

Again, we don’t know the values of the µi and σ 2 , but we can ask “what
if” questions: The Lower-Dimensional Model will be a “win” if

• (µ1 − µ2 )2 is small, i.e., (9.10) is approximately true,

• n is small, i.e., we just don’t have enough data to estimate two sepa-
rate means, or

• σ 2 is large, which again amounts to not having enough data.

In such cases, the Higher-Dimensional Model is overfitting, i.e., is too rich

a model for our situation.
This, in a nutshell, is the essence of the notion of overfitting. We know that
technically (9.10) is incorrect, but if it is approximately correct, or if our
data are meager, it is helpful to make that assumption.
344 CHAPTER 9. VARIABLE SELECTION

The above example is trivially a regression problem. The Higher Dimen-

sional Model has p = 1, with our predictor X being a single dummy variable
for say, male. Our Lower Dimensional Model has p = 0, no predictors at
all, just our intercept term. (The latter is the common value of the µi .)
If we have some k-category variable, that would mean p = k − 1 predictors.
In this case, an assumption like, say,

µ1 = ..., = µk (9.12)

would be a huge dimension reduction, and of course there may be groupings

that might be considered for more moderate reductions.
The general regression/classification context is similar. Consider paramet-
ric models with p predictors on n observations. If we do not use all of our p
predictors, we have fewer βi to estimate, i.e., we have a lower-dimensional
model. If p is large and/or n is small, it may be desirable to use that
simplfiied model, say by omitting the predictors for which βi seems to be
small.

9.2 How Many Is Too Many?

To avoid overfitting, we should not have too many predictor variables p,

relative to our sample size n. So, how many is too many?
In Chapter 1, we set√a rough rule of thumb, due to Tukey, that one should
have no more than n predictors. Later work by Portnoy [115] produced
a similar result, and we will use this rule informally here. We will examine
it more closely in Chapter 12.

9.3 Fit Criteria

Let’s say we have m predictor variables in all, and we wish to choose a good
subset of them. Let p denote the size of the subset we ultimately choose.
Keep in mind that we typically do not choose p ourselves, but decide upon a
value based on one of the processes to be described shortly in this chapter,
or on some similar process.
9.3. FIT CRITERIA 345

9.3.1 Some Common Measures

In choosing a subset of predictor variables, we need a criterion to describe

how good a particular subset is. The criteria below (other than the first)
will be easier to explain if we assume a multivariate normal setting (Section
2.6.2). Let m denote the total number of variables we have available to serve
as predictors, so we are assuming that (X, Y ) has an (m+1) variate normal
distribution.
Suppose also that we have some algorithm by which to add predictors one
at a time (we will see such algorithms below), So we will only end up looking
at m + 1 sets of predictors (including the case of having none), not the 2m
possible sets we could examine.
In the multivariate normal setting, the regression function of Y against any
set of our predictors will be linear. Let σp2 denote the variance of Y given
our predictor set, p = 0, 1, ..., m.
Here are some of the most common criteria (the first two of which have
been covered earlier in this book):

• Re-predicting the Original Data

R2 is an example of this. Recall (Section 2.9.2) that it is the squared
sample correlation between the predicted and actual Y values, com-
puted by “predicting” Y in the data in which we did our model fitting.
In classification problems, we can compute the overall rate of correct
classification in the same manner.
As discussed earlier, e.g., in Section 6.6.1, this approach has a bias
problem. Since our fit procedures find the best fit in the particular
sample data at hand, and since the sample data will diverge somewhat
from the population distribution, the fitted model will probably not
be the best fit to the population. Thus R2 and the rate of correct
classification are generally too optimistic, i.e., are biased upward.
• Adjusted R2 : This was introduced in Section 2.9.4. Let’s take a
closer look. The quantity is defined to be
∑n b2
e ′ β)
1
n−p−1 i=1 (Yi −X i
2
Radj =1− ∑n (9.13)
1
n−1 i=1 (Yi − Y )2

ei is Xi with a 1 prepended.
Recall here that X
Note that there will be a different βb value for each subset of predictors
being used.
346 CHAPTER 9. VARIABLE SELECTION

Recall that the numerator and denominator here are unbiased esti-
mates of the values of σp2 and σ02 . Recall too from Section 2.9.4 that
both R2 and adjusted R2 , in their sample-based forms, are interpreted
as the estimated reduction in mean squared prediction error obtained
when one uses the p predictor variables, versus just predicting by a
constant.
As we add more and more predictors, R2 is guaranteed to increase,
since with more variables our minimization of (2.18) will be made
over a wider set of choices. By contrast, the adjusted R2 could go up
or down when we add a predictor.
We could take adjusted R2 as our stopping criterion: In adding our
first few predictors, adjusted R2 may increase, but eventually it might
start to come back down. We might choose our predictor set to be
the one that yields maximum adjusted R2 .
• Mallows’ Cp : This amounts to an alternative way to adjust R2 :

SSEp
Cp = 1 − n + 2p (9.14)
n−m+1 SSEm

where SSEp is

∑
n
(Yi − X b2
ei′ β) (9.15)
i=1

for the model with p predictors, p = 0, 1, ..., m.

Let’s see what motivates this definition. The reader should keep in
mind, by the way, that this is just a heuristic, so don’t worry much
about “leaps of faith” that may seem to occur.
Think of what would happen if a set of p predictors told the whole
story, i.e., were as good as using all m predictors. Then we would
have

σp2 = σp+1
2 2
= ... = σm (9.16)
As noted, the numerator in (2.71) is an unbiased estimate of σp2 , so
in (9.14), SSEp is approximately (n − p + 1)σp2 . The denominator is
2
approximately σm , no matter whether the set of p predictors tells the
whole story, but under our assumption it is equal to σp2 . Then (9.14)
is approximately

(n − p − 1)σp2
− n + 2p = p − 1 (9.17)
σp2
9.3. FIT CRITERIA 347

In other words, our stopping criterion might be to use the value of p

that achieves approximate equality in (9.17), with estimated values
for the σi2 .
For typical data sets, the graph of Cp against p is more or less con-
vex (concave-upward), possibly with some small deviations from that
shape. The reason for the convexity is that in (9.14), SSEp decreases
with p while of course −n + 2p is an increasing function of p (n is
fixed, as we are on the same data set). Moreover, if there is value of
p such that (9.16) holds, then in (9.14), SSEj will be approximately
constant for j = p, p + 1, ..., m, making it even more likely that the
graph will be increasing after j = p.
The bottom line, then, is that we might take as our stopping criterion
the model with minimum Cp .

• Akaike’s Information Criterion: This is basically

-2 log likelihood + 2 p (9.18)

In the linear regression model with the classic normality assumption,

this boils down to (except for unimportant constants)

n log(s2 ) + 2p (9.19)

where s2 is as in (2.55).
Note carefully that AIC assumes that the conditional distribution
of Y given X is known, in this case normal. Divergence from this
assumption has unknown impacts of the use of AIC for variable se-
lection.
AIC can also be computed for the logit model, though, and there we
are on safe ground, since by definition the conditional distribution in
question is Bernoulli/binomial.
Like Cp , AIC reflects a tradeoff between within-sample fit and number
of predictors. Using this criterion in choosing among a sequence of
models, we might choose the one with smallest AIC value.

• Cross-validation: As discussed earlier, here we split our data into

a training set and a test set. To compare several different predictor
sets, we fit each one to the training set, and see how well the resulting
model predicts the test set.
Consider a single partitioning of our data into r and n−r data points,
and consider the context mentioned above, in which we add predictors
348 CHAPTER 9. VARIABLE SELECTION

to our model one at a time, with p denoting the number of predictors

at any given step. For fixed n and fixed population distribution,
there is an optimal value of p, which we will call popt (n). Its value is
unknown, of course, but it does exist, and our model selection process
is, inter alia, estimating popt (n). Now consider the ramifications of
our choice for r:

– If we make r too small, we are estimating popt (r) instead of

popt (n), potentially a problem if r is much smaller than n.
– If we make r too large, our test set, consisting of n − r observa-
tions, will be small, and thus the measure of prediction accuracy
it gives us will have high variability.

This is the motivation for m-fold cross-validation: We randomly di-

vide our data into m subsets of about equal size, setting r to about
n/m. For rather small m, r will be close to n, solving the problem
in the first bullet above. But since we will look at m partitionings,
the variability problem in the second bullet may be ameliorated with
large m. But keep in mind that this is only a rough analysis, since
for large m, our estimates from each subset will have higher variance.

9.3.2 There Is No Panacea!

Choosing a subset of predictors on the basis of cross-validation, AIC and

so on is not foolproof by any means. Due to the Principle of Frequent
Occurrence of Rare Events (Section 7.6.1), some subsets may look very
good yet actually be artifacts.
On a theoretical basis, it has been shown that LOOM is not statistically
consistent [128]. That paper does find conditions under which one gets
consistency by leaving w out rather than 1, providing w/n → 1, but it is
not clear what the practical implications are. Empirical doubt on cross-
validation was cast in [118].

9.4 Variable Selection Methods

Alan Miller wrote a comprehensive account of the state of the art for vari-
able selection in 1990, and published a second edition in 2002 [110]. His
comment speaks volumes:
9.5. SIMPLE USE OF P-VALUES: PITFALLS 349

What has happened to the field since the first edition was first
published in 1990? The short answer is that there has been very
little progress.

The same statement applies today. The general issue of good variable
selection is still an unsolved problem. Nevertheless, a number of methods
have been developed that enjoy wide usage, and that many analysts have
found effective. We present some of them in the next few sections.

9.5 Simple Use of p-Values: Pitfalls

In applications in which Description is the primary goal, this method is the

most widespread, at least among nonstatisticians. The analyst simply reads
the output of, say, lm(), and decides that the predictors with asterisks are
the “important” ones, discarding the rest.
This book has discouraged such thinking (Section 2.10). The approach can
be very misleading, as shown in the MovieLens example in Section 3.2.4.
Moreover, if Prediction is the goal, it was disccovered long ago [14] that the
standard 0.05 cutoff for p-values may not be best, with values such as 0.25
and 0.35 giving better performance.2

9.6 Asking “What If ” Questions

An alternative is to form confidence intervals for the various βi , and gauge

the effects of the X (i) by comparing the locations of the intervals to the
general size of the Y variable. Examples of this were presented in Sections
2.10 and 3.2.4. Let’s review a finding in the latter example, involving the
possible impact of user age on movie ratings:

A 10-year difference in age only has an impact of about 0.03 on

ratings, which are on the scale 1 to 5.

Clearly, user age is not an important variable for predicting ratings, or for
analyzing the underlying process that affect ratings.
On the other hand, consider the baseball player example, results of which
are discussed in Section 1.9.1.2. The estimated regression coefficient for the
2 This is in the stepwise context to be discussed below.
350 CHAPTER 9. VARIABLE SELECTION

age variable was 0.9115, indicating that players do gain weight as they age,
almost a pound a year, in spite of needing to keep fit. Of course, one would
need to form a confidence interval for this, as it is only an estimate, but
the result is useful.
In other words, we would select predictors “by hand,” using common sense
and our knowledge of the setting in question. This is in contrast to using an
automated process such as selection by p-values or the stepwise methods to
be presented shortly. The drawback — to some people — is that one must
work harder this way, no automated system to make our decisions for us.
Nonetheless, this method has direct practical meaning, which the p-value
and other approaches do not.
In the case of something like a logit model, as noted in Section 4.3.3, the βi
are a little more difficult to interpret than in linear models. After forming
confidence intervals for the coefficients, how do we decide if they are “large”
or “small”? One quick way would be to compare them to the intercept term,
βb0 .
For example, consider the example of diabetes in indigenous Americans,
Section 4.3.2. The p-value for NPreg was tiny, generally considered “very
highly significant.” Yet the estimated effect of one extra pregnancy is only
0.1232, quite small compared to the intercept term, -8.4047. If our goal
is Description, the finding that having more pregancies puts a women at
greater risk for developing diabetes would be of interest, small but mean-
ingful. On the other hand, in a Prediction context, it is clear that NPreg
would not be of major help.
Note that we may not have a good domain-expert knowledge of the pre-
dictors. For instance we know, in the diabetes example above, what is
common for the number of pregancies a woman has had. But we may not
have similar knowledge of the triceps skin fold thickness variable. By look-
ing at that variable’s mean and standard deviation, we can attain at least
a first-pass understanding of its scale of variation, and thus gauge the size
of the coefficient in proper context. It gives us an idea of how much of a
change in that variable is typical, and we can multiply that value by the
coefficient to see how much µ b(t) changes.
For this reason, the analyst may find it useful, for instance, to routinely
run colMeans() on the predictor matrix/data frame after running lm()
or glm(). Similarly, we should run

apply ( dataname , 2 , sd )
9.7. STEPWISE SELECTION 351

where dataname is our matrix/data frame. Of course, a more direct way

would be to apply scale() to the predictors before running
On the other hand, with a small sample (especially relative to the size of
p), we have the opposite danger: There may be some good predictors that
a simple p-value analysis may overlook. It may pay to take a good look
at the “non-signicant” variables, especially if we are armed with domain
expertise.
Again, this approach will not satisfy those who “want the computer to make
their decisions for them.” But placing the decision-making into the hands
of the actual user may yield a better result in the end. (Recall Section
1.18.)

9.7 Stepwise Selection

When the use of computers first became common in statistical applications,

stepwise methods were the rage. The idea is quite straightforward.

9.7.1 Basic Notion

We start with no predictors in the model, and test the hypothesis

H0 : βi = 0 (9.20)

for each predictor X (i) . Whichever one yields the smallest p-value, that
predictor is added to the model. We then determine which predictor to
add to the model next in the same manner, via (9.20). We stop when we
no longer have any p-values below a cutoff value, which classically has been
the usual 0.05 for stepwise methods.3
This is called forward selection; backward selection begins with all the pre-
dictors in the model, and removes them one by one, again using (9.20).
The original forms of these methods are now considered out of date by
many, but the notion of stepwise adding/deleting predictors is still quite in
favor.
3 As noted earlier, research has shown that a much larger cutoff tends to work better.

An interesting related fact is that including a predictor if the Z-score in Equation (2.56) is
greater than 1.0 in absolute value — a p-value of about 0.32 — is equivalent to including
the predictor if it raises the Adjusted R-squared value [58].
352 CHAPTER 9. VARIABLE SELECTION

9.7.2 Forward vs. Backward Selection

Forward selection seems especially appealing. Since our goal is to find

a parsimonious model, it is natural to start with the most parsimonious
model of all — the one without any predictors at all — and then slowly
add more predictors. One setting in which forward selection is virtually
forced upon us is that in which p > n, making linear or other parametric
models impossible to fit (without resorting to ridge regression or the like).
On the other hand, many analysts believe that forward selection may under-
fit, in the sense that it may miss a group of predictors in which each one is
not a strong predictor but collectively the group has substantial predictive
power.
Consider the MovieLens data (Section 3.2.4). After calculating the Ti as
our Y variable here, we added age and gender predictors. We might also
add dummy variables for the individual movies. Presumably each one will
contribute only a negligible amount to prediction, but collectively they
could be rather powerful. We will investigate this in Chapter 12.
We can also try both forward and backward stepping. In the forward case,
for instance, after adding a few predictors it may be that one of the ones
added early in the process is now not very helpful.

9.7.3 R Functions for Stepwise Regression

There are many such functions. Here, we will use stepAIC() for the linear
and logit models. It is part of the MASS package included in the base
R distribution. It uses AIC for its fit criterion, so that for instance in
forward selection, the predictor that is added will be the one that brings
about the largest drop in AIC. The argument direction allows the user
to specify forward or backward selection, or even both. In the latter case,
which is the default value, backward elimination is used but variables can
be re-added to the model at various stages of the process.
By the way, lars(), which we will use for LASSO below, also does offer
other options, including one for stepwise selection.

9.7.4 Example: Bodyfat Data

This data set was introduced in Section 1.2, with n = 252 observations
on p = 13 predictor variables. The role of prediction for that data was
9.7. STEPWISE SELECTION 353

explained:

Body fat is expensive and unwieldy to measure directly, as it

involves underwater weighing. Thus it would be highly desirable
to “predict” that quantity from easily measurable variables such
as height, age, weight, abdomen circumference and so on.

The first column of the data set is the case number. If the numbering is
sequential in time, this might be useful for investigating time trends, but
we will omit it.
The next three columns all involve the underwater weighing, something
we are trying to avoid. We wish to see how various body circumference
methods can predict body fat, so we will use only one of the available
measures, say the first:
> l i b r a r y ( mfp )
> data ( b o d y f a t )
> b o d y f a t <− b o d y f a t [ , − c ( 1 , 3 : 4 ) ]

Now let’s run lm():

> l m a l l <− lm( b r o z e k ∼ . , data=b o d y f a t )
> summary( l m a l l )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) −15.29255 1 6 . 0 6 9 9 2 −0.952 0 . 3 4 2 2 5
age 0.05679 0.02996 1.895 0.05929 .
weight −0.08031 0 . 0 4 9 5 8 −1.620 0 . 1 0 6 6 0
height −0.06460 0 . 0 8 8 9 3 −0.726 0 . 4 6 8 3 0
neck −0.43754 0 . 2 1 5 3 3 −2.032 0 . 0 4 3 2 7 ∗
chest −0.02360 0 . 0 9 1 8 4 −0.257 0 . 7 9 7 4 0
abdomen 0.88543 0 . 0 8 0 0 8 1 1 . 0 5 7 < 2 e −16 ∗∗∗
hip −0.19842 0 . 1 3 5 1 6 −1.468 0 . 1 4 3 4 1
thigh 0.23190 0.13372 1.734 0.08418 .
knee −0.01168 0 . 2 2 4 1 4 −0.052 0 . 9 5 8 5 0
ankle 0.16354 0.20514 0.797 0.42614
biceps 0.15280 0.15851 0.964 0.33605
forearm 0.43049 0.18445 2.334 0.02044 ∗
wrist −1.47654 0 . 4 9 5 5 2 −2.980 0 . 0 0 3 1 8 ∗∗
...
M u l t i p l e R−s q u a r e d : 0 . 7 4 9 , Adj . R−s q u a r e d : 0 . 7 3 5 3
354 CHAPTER 9. VARIABLE SELECTION

This is rather good fit, with a straightforward analysis with all predictor
variables present: Adjusted R2 was 0.7353. Why, then, should we do vari-
able selection?
Again, the answer is expense. The point of predicting bodyfat in the first
place was to save on cost, and since collecting data on the predictor variables
entails labor and thus costs, it would be nice if we found a parsimonious
subset. Let’s see what stepAIC() does with it.
The function stepAIC() needs the full model to be fit first, using it to
acquire information about the data, model and so on. We have already
done that, so now we run the analysis, showing the output of the first step
here:
> l i b r a r y (MASS)
> s t e p o u t <− stepAIC ( l m a l l )
S t a r t : AIC=710.77
b r o z e k ∼ age + w e i g h t + h e i g h t + neck + c h e s t +
abdomen + h i p + t h i g h + knee + a n k l e + b i c e p s +
forearm + w r i s t

Df Sum o f Sq RSS AIC

− knee 1 0.04 3785.2 708.77
− chest 1 1.05 3786.2 708.84
− height 1 8.39 3793.5 709.33
− ankle 1 10.11 3795.2 709.44
− biceps 1 14.78 3799.9 709.75
<none> 3785.1 710.77
− hip 1 34.28 3819.4 711.04
− weight 1 41.73 3826.9 711.53
− thigh 1 47.83 3833.0 711.94
− age 1 57.12 3842.3 712.55
− neck 1 65.66 3850.8 713.10
− forearm 1 86.63 3871.8 714.47
− wrist 1 141.21 3926.3 718.00
− abdomen 1 1944.46 5729.6 813.24

The initial AIC value, i.e., with all predictors present, was 710.77. The
function then entertained removal of the various predictors, one by one.
Removing the knee measurement, for instance, would reduce AIC to 708.77,
while removing chest would achieve 708.84, and so on.
By contrast, removing the hip measurement would be worse than doing
nothing, actually increasing AIC to 711.04. (The <none> line separates
the variables whose removal decreases AIC from those that increase it.
9.7. STEPWISE SELECTION 355

The difference between the knee and chest variables is negligible, but the
function decides to remove the knee variable, as seen in the model after one
step,
Step : AIC=708.77
b r o z e k ∼ age + w e i g h t + h e i g h t + neck + c h e s t +
abdomen + h i p + t h i g h + a n k l e + b i c e p s +
forearm + w r i s t

At the second step, the algorithm was faced with the following choices:
Df Sum o f Sq RSS AIC
− ankle 1 11.20 3805.1 704.09
− biceps 1 16.21 3810.1 704.43
− hip 1 28.16 3822.0 705.22
<none> 3793.9 705.35
− thigh 1 63.66 3857.5 707.55
− neck 1 65.45 3859.3 707.66
− age 1 66.23 3860.1 707.71
− forearm 1 88.14 3882.0 709.14
− weight 1 102.94 3896.8 710.10
− wrist 1 151.52 3945.4 713.22
− abdomen 1 2737.19 6531.1 840.23

It then chose to remove the ankle variable.

Eventually, all variables fall below the <none> line,
Step : AIC=703.08
b r o z e k ∼ age + w e i g h t + neck +
abdomen + h i p + t h i g h + f o r e a r m + w r i s t

Df Sum o f Sq RSS AIC

<none> 3820.0 703.08
− hip 1 33.23 3853.2 703.26
− neck 1 67.79 3887.8 705.51
− age 1 67.88 3887.9 705.52
− weight 1 81.50 3901.5 706.40
− thigh 1 90.34 3910.3 706.97
− forearm 1 122.99 3943.0 709.07
− wrist 1 139.46 3959.4 710.12
− abdomen 1 2726.49 6546.5 836.83

and the final model is chosen:

356 CHAPTER 9. VARIABLE SELECTION

Coefficients :
( Intercept ) age weight
−20.06213 0.05922 −0.08414
neck abdomen hip
−0.43189 0.87721 −0.18641
thigh forearm wrist
0.28644 0.48255 −1.40487

So, how does this reduced model fare in terms of predictive ability?
> summary( s t e p o u t )
...
Coefficients :
Estimate Std . E r r o r t v a l u e Pr ( >| t | )
( I n t e r c e p t ) −20.06213 1 0 . 8 4 6 5 4 −1.850 0 . 0 6 5 5 8
age 0.05922 0.02850 2.078 0.03876
weight −0.08414 0 . 0 3 6 9 5 −2.277 0 . 0 2 3 6 6
neck −0.43189 0 . 2 0 7 9 9 −2.077 0 . 0 3 8 8 9
abdomen 0.87721 0 . 0 6 6 6 1 1 3 . 1 7 0 < 2 e −16
hip −0.18641 0 . 1 2 8 2 1 −1.454 0 . 1 4 7 2 7
thigh 0.28644 0.11949 2.397 0.01727
forearm 0.48255 0.17251 2.797 0.00557
wrist −1.40487 0 . 4 7 1 6 7 −2.978 0 . 0 0 3 1 9

( Intercept ) .
age ∗
weight ∗
neck ∗
abdomen ∗∗∗
hip
thigh ∗
forearm ∗∗
wrist ∗∗
...
M u l t i p l e R−s q u a r e d : 0 . 7 4 6 7 ,
Adjusted R−s q u a r e d : 0 . 7 3 8 3

The two R-squared values are quite close to those of the full model. In
other words, our pared-down predictor set has about the same predictive
power as the full model, but at a much lower data collection cost.
Note, though, that the variable selection process changes all the distribu-
tions. The p-values are overly optimistic, as is adjusted R-squared. Nev-
ertheless, use of the more restrictive predictor set, with the attendant cost
9.7. STEPWISE SELECTION 357

savings, does seem to be a safe bet. As is often the case, it would be wise
to check this with a domain expert.

9.7.5 Classification Settings

Dimension reduction is of course an important issue in classification set-

tings. Continuing our use of stepAIC() as our vehicle for stepwise predic-
tor selection, let’s take a look at what can be done.

9.7.5.1 Example: Bank Marketing Data

This data set is one of the best known in the UCI collection. It consists
of data on a marketing campaign by a Portuguese bank, with the goal of
predicting whether a customer would open a new term deposit account.
The latter is indicated by the y column in the data frame:
# n o t e t h e ’ ; ’ s e p a r a t o r symbol , not commas
> bank <− read . csv ( ’ bank . c s v ’ , s e p= ’ ; ’ )
> head ( bank )
age j o b m a r i t a l e d u c a t i o n default
1 30 unemployed m a r r i e d primary no
2 33 s e r v i c e s married secondary no
3 35 management s i n g l e t e r t i a r y no
4 30 management m a r r i e d t e r t i a r y no
5 59 blue −c o l l a r m a r r i e d s e c o n d a r y no
6 35 management s i n g l e t e r t i a r y no
b a l a n c e h o u s i n g l o a n c o n t a c t day month
1 1787 no no c e l l u l a r 19 oct
2 4789 y e s y e s c e l l u l a r 11 may
3 1350 yes no c e l l u l a r 16 apr
4 1476 y e s y e s unknown 3 jun
5 0 yes no unknown 5 may
6 747 no no c e l l u l a r 23 feb
d u r a t i o n campaign pdays p r e v i o u s poutcome y
1 79 1 −1 0 unknown no
2 220 1 339 4 failure no
3 185 1 330 1 failure no
4 199 4 −1 0 unknown no
5 226 1 −1 0 unknown no
6 141 2 176 3 failure no
> dim( bank )
358 CHAPTER 9. VARIABLE SELECTION

[ 1 ] 4521 17

So we have n = 4521 and a nominal value p = 16. The true value of p is

much larger, due to the presence of R factors, which correspond to dummy
variables; more on this below. Another point to keep in mind is that this
is actually the smaller data set in this bank data package.
Now, what about those dummy variables? In R, a popular way to store
a categorical variable, say eye color, is as a factor, essentially a numeric
vector that is normally dealt with by the names of the values. See the
Computational Complements section on factors at the end of this chapter
for details, but for now the point is that with factors we can avoid forming
dummy variables; R does this work for us:
> g l o u t <− glm( y ∼ . , data=bank , family=binomial )
> summary( g l o u t )
Coefficients :
Esti mate Std . E r r o r z v a l u e
( Intercept ) −2.462 e+00 6 . 0 3 8 e −01 −4.077
age −4.232 e −03 7 . 1 2 5 e −03 −0.594
j o b b l u e −c o l l a r −3.924 e −01 2 . 4 2 0 e −01 −1.621
jobentrepreneur −2.498 e −01 3 . 8 1 1 e −01 −0.655
jobhousemaid −3.530 e −01 4 . 1 7 6 e −01 −0.845
jobmanagement −7.302 e −02 2 . 4 0 7 e −01 −0.303
jobretired 6 . 3 1 5 e −01 3 . 1 1 2 e −01 2.029
j o b s e l f −employed −1.812 e −01 3 . 5 3 3 e −01 −0.513
jobservices −1.457 e −01 2 . 7 2 9 e −01 −0.534
jobstudent 3 . 7 8 4 e −01 3 . 7 5 0 e −01 1.009
jobtechnician −1.926 e −01 2 . 3 0 1 e −01 −0.837
jobunemployed −6.395 e −01 4 . 2 1 4 e −01 −1.518
jobunknown 5 . 2 0 7 e −01 5 . 8 5 3 e −01 0.890
maritalmarried −4.696 e −01 1 . 7 4 3 e −01 −2.694
maritalsingle −3.051 e −01 2 . 0 3 8 e −01 −1.497
...
Pr ( >| z | )
( Intercept ) 4 . 5 5 e −05 ∗∗∗
age 0.552537
j o b b l u e −c o l l a r 0.104937
jobentrepreneur 0.512199
jobhousemaid 0.398000
jobmanagement 0.761602
jobretired 0.042454 ∗
j o b s e l f −employed 0.608167
jobservices 0.593542
9.7. STEPWISE SELECTION 359

jobstudent 0.312958
jobtechnician 0.402496
jobunemployed 0.129138
jobunknown 0.373669
maritalmarried 0 . 0 0 7 0 5 8 ∗∗
maritalsingle 0.134354

Notice that R not only created dummies, but gave them names according
to the levels; for instance, R noticed that one of the levels of job was blue-
collar, and thus named the dummy jobblue-collar. If we want to know
which levels R chose for constructing the dummies, we can examine the
xlevels component:
glout$ xlevels
$job
[ 1 ] ”admin . ” ” blue −c o l l a r ”
[ 3 ] ” entrepreneur ” ” housemaid ”
[ 5 ] ”management” ” retired ”
[ 7 ] ” s e l f −employed ” ” services ”
[ 9 ] ” student ” ” technician ”
[ 1 1 ] ” unemployed ” ”unknown”

$marital
[ 1 ] ” d i v o r c e d ” ” m ar r i e d ” ” single ”

$education
[ 1 ] ” primary ” ” secondary ” ” t e r t i a r y ”
[ 4 ] ”unknown”
...

We can see that the marital column in the data frame has levels “divorced,”
“married” and “single,” yet lm() only produced coefficients for the latter
two. So those coefficients are relative to the “divorced” level. Since they
are both negative, it seems that the divorced customers are more likely to
open the new account.
It’s important to know what the real value of p is (as opposed to the nom-
inal value 16 we mentioned earlier), as the larger p is, the more we risk
overfitting. As already discussed, there is no magic rule for determining if
p is too large, but we should at least know what value of p we have:
> length ( coef ( g l o u t ) ) − 1 # exclude intercept
[ 1 ] 42

Let’s apply the stepwise selection process:

360 CHAPTER 9. VARIABLE SELECTION

> s t e p o u t <− stepAIC ( g l o u t )

> stepout
...
Coefficients :
( Intercept ) maritalmarried
−2.731855 −0.510829
m a r i t a l s i n g l e educationsecondary
−0.318135 0.126807
educationtertiary educationunknown
0.381100 −0.283123
housingyes loanyes
−0.343417 −0.643090
contacttelephone contactunknown
−0.007710 −1.457777
day monthaug
0.016295 −0.318045
monthdec monthfeb
0.202759 0.139813
monthjan monthjul
−1.085164 −0.731248
monthjun monthmar
0.560045 1.626269
monthmay monthnov
−0.473499 −0.861876
monthoct monthsep
1.410274 0.755895
duration campaign
0.004184 −0.072216
poutcomeother poutcomesuccess
0.481385 2.414545
poutcomeunknown
−0.093110

How much was trimmed?

> length ( coef ( s t e p o u t ) )
[ 1 ] 27

So, 15 of the original 42 predictors were removed.

Let’s see whether this resulted in much compromise in predictive power.
Again, this would be more accurately assessed using cross-validation, but
here is a quick assessment:
> y <− as . integer ( bank$y ) − 1
9.7. STEPWISE SELECTION 361

> yhat <− round ( g l o u t $ f i t t e d . v a l u e s )

> mean( y )
[ 1 ] 0.11524
> mean( yhat == y )
[ 1 ] 0.9051095

So, without covariate information, we would also guess y to be ’no’, and

would be correct about 1 - 0.11524 = 0.88476 of the time. With the covari-
ates we would do slightly better, 0.9051095. Let’s compare the latter figure
with the model yielded by stepAIC():
> y h a t s t e p <− round ( mnout$ f i t t e d . v a l u e s )
> mean( y h a t s t e p == y )
[ 1 ] 0.9046671

Good, almost identical.

As mentioned, the data package for this bank marketing data contains two
versions of the data, first the one we used above, bank.csv, and a much
larger one, bank-full.csv. After going through the same computations
as above for this large set (not shown), it turns out that only 2 of the
42 predictors are removed, compared to 15 for the smaller data set. This
makes sense; as discussed before, the larger n is, the more predictors the
model can handle well.

9.7.5.2 Example: Vertebrae Data

As a multiclass example, let’s again use the vertebrae data from Section
5.5.2:
> mnout <− multinom (V7 ∼ . , data=v e r t )
...
> s t e p o u t <− stepAIC ( mnout )
...
> summary( s t e p o u t )
Call :
multinom ( formula = V7 ∼ V1 + V2 + V5 + V6 , data = v e r t )

Coefficients :
( Intercept ) V1 V2 V5
1 −20.85643 0 . 1 8 2 7 6 3 8 −0.2634111 0 . 1 3 5 9 2 4 7 0
2 −21.37424 0 . 2 2 0 2 8 8 6 −0.2183829 0 . 0 7 7 1 6 2 2 4
V6
362 CHAPTER 9. VARIABLE SELECTION

1 −0.000133527
2 0.311163955

Std . E r r o r s :
( Intercept ) V1 V2 V5
1 4.249310 0.03310576 0.04888643 0.02873804
2 5.239924 0.05140014 0.08326775 0.03095909
V6
1 0.03862044
2 0.05794419
...

So, the predictors V3 and V4 were dropped.

9.7.6 Nonparametric Settings

So, how does all this relate to nonparametric regression methods (including
classification, of course), say k-nearest neighbor? There are actually two
questions here:

• Is dimension reduction even an issue in the nonparametric case?

• If so, how might this goal be accomplished?

9.7.6.1 Is Dimension Reduction Important in

the Nonparametric Setting?

The answer to this question is indeed yes. To see why, let’s look again at
the vertebrae data, estimating µ(t) for t equal to the first observation in
the data set, which for convenience we will call the prediction point.
To find the nearest neighbors of the prediction point, we’ll use the function
get.knnx() from the FNN package on CRAN [88] (which is also used in
our regtools package [97]). The call form is
get . knnx ( dataframe , t v a l u e )

where dataframe is our data frame of predictor values, and tvalue is our
prediction point. This call returns an R list with components nn.index
and nn.dist.
9.7. STEPWISE SELECTION 363

p=1 p=2 p=3

245 1 1
1 254 236
23 308 175
109 23 308
201 17 119
153 201 120
260 231 4
197 236 240
254 119 286
17 144 22

Table 9.1: Indices of Nearest Neighbors

To gauge the effects of different values of the number of predictors p, we’ll

consider p = 1, 2, 3, so that for instance p = 2 means that we do prediction
based on the variables V 1 and V 2. Here is the code:
t e s t <− v e r t [ 1 , , drop=FALSE ]
get . knnx ( v e r t [ , 1 ] , t e s t [ , 1 ] ) $nn . index
get . knnx ( v e r t [ , 1 : 2 ] , t e s t [ , 1 : 2 ] ) $nn . index
get . knnx ( v e r t [ , 1 : 3 ] , t e s t [ , 1 : 3 ] ) $nn . index

The results are shown in Table 9.1. The indices of the 10 closest neighbors
to the prediction point are shown, one per row, for each value of p.
Suppose that, unknown to the analyst, the regression function µ(t1 , t2 , ...)
depends only on ti , i.e., only the first predictor has impact on the response
variable Y . Then (again, unknown to the analyst) the nearest-neighbor
finding process need only consider the first predictor, V1. In that case, it
turns out that observation number 245 is the closest. Yet that observation
doesn’t make the nearest-10 list at all for the cases p = 2 and p = 3. And
though there is some commonality among the three columns of the table, it
is clear that generally the nearest neighbors of a point for p = 1 will differ
from those for the other two values of p.
What are the implications of this? Recall the bias-variance tradeoff issue in
under/overfitting (Section 9.1). The more distant an observation from the
prediction point, the more the bias. So, making a “mistake” in choosing
the nearest neighbors will generally give us more-distant neighbors, and
364 CHAPTER 9. VARIABLE SELECTION

increase prediction error.

In other words, yes, dimension reduction is just as much an issue in the
nonparametric setting as in the parametric one.

9.7.7 The LASSO

One of the reasons for the popularity of the LASSO is that it does automatic
variable selection. We will take a closer look at LASSO methods in this
section.

9.7.7.1 Why the LASSO Often Performs Subsetting

First, similar to the ridge case, minimizing (8.18) is equivalent to minimiz-

ing4

∑
n
q(b) = ei b)2
(Yi − X (9.21)
i=1

subject to the constraint

||b||1 ≤ λ (9.22)

This motivates Figure 9.1.

The figure is for the case of p = 2 predictors (for simplicity, we assume
there is no constant term β0 ). Writing b = (b1 , b2 )′ , then the horizontal
and vertical axes are for b1 and b2 , as shown. The corners of the diamond
are at (λ, 0), (0, λ) and so on. Due to the constraint (9.22), our LASSO
estimator βbl must be somewhere within the diamond.
What about the ellipses? They are contours of q: For a given value of
q, say c, then the locus of points b for which q(b) = c takes the form of
an ellipse. Each value of c gives us a different ellipse; two of them, out of
infinitely many, are shown in the figure, with the smaller one corresponding
to a smaller value of c.
But remember, we are trying to minimize q, so we want c to be as small as
possible, i.e., we want the countour curve to be small — but our constraint
4 The computational details of the minimization process are beyond the scope of this
book.
9.7. STEPWISE SELECTION 365

Figure 9.1: Subsetting nature of the LASSO

requires that the curve must include at least one point within the diamond.
In our figure here, this implies that we must choose c so that the ellipse is
barely touching the diamond, as the larger ellipse does.
Now, here is the key point: The point at which the ellipse barely touches
the diamond will typically be one of the four corners of the diamond. And
at each of those corners, either b1 or b2 is 0 — i.e., βbl has selected a subset
of the predictors, in this case a subset of size 1.
The same geometric argument works in higher dimensions, and this is then
the appeal of the LASSO for many analysts:

The LASSO often does automatic subset selection. The analyst

need only use the predictors X (i) for which βbi ̸= 0.

We say that the LASSO tends to produce a sparse estimator of β. Needless

to say, though this is indeed an appealing property, there is no guarantee
that this produces a “good” set of predictors.
Suppose in the figure, the inner ellipse corresponds to the ordinary estima-
366 CHAPTER 9. VARIABLE SELECTION

tor b = βbOLS , i.e.,

c = q(βbOLS ) (9.23)

In order to satisfy the LASSO constraint, we needed to accept a larger

b This
value of q, corresponding to the outer ellipse, and thus a smaller β.
illustrates the shrinkage nature of the LASSO.
On the other hand, the ellipse corresponding to OLS might already dip into
the diamond. In this case,

βbl = βbOLS (9.24)

So, it is not guaranteed that the LASSO will choose a sparse β. b As was
noted earlier for shrinkage estimators in general, for fixed p, the larger n is,
the less need for shrinkage, and the above situation may occur.
There is of course the matter of choosing the value of λ. Our old friend,
cross-validation, is an obvious approach to this, and others have been pro-
posed as well. The lars package includes a function cv.lars() to do k-fold
cross-validation.

9.7.7.2 Example: Bodyfat Data

Let’s continue the example of Section 9.7.4. Let’s see what lars finds here.
> library ( l a r s )
> l a r s o u t <− l a r s ( as . matrix ( b o d y f a t [ , − 1 ] ) , b o d y f a t [ , 1 ] )
> larsout
Call :
l a r s ( x = as . matrix ( b o d y f a t [ , −1]) , y = b o d y f a t [ , 1 ] )
R−s q u a r e d : 0 . 7 4 9
Sequence o f LASSO moves :
abdomen h e i g h t age w r i s t neck f o r e a r m h i p
Var 6 3 1 13 4 12 7
Step 1 2 3 4 5 6 7
w e i g h t b i c e p s t h i g h a n k l e c h e s t knee
Var 2 11 8 10 5 9
Step 8 9 10 11 12 13

So, at Step 1, the abdomen predictor was brought in, then height at Step
2, and so on. Now look further:
9.8. POST-SELECTION INFERENCE 367

> summary( l a r s o u t )
...
Df Rss Cp
0 1 15079.0 698.131
1 2 5423.4 93.012
2 3 5230.7 82.893
3 4 4914.9 65.038
4 5 4333.6 30.484
5 6 4313.5 31.225
6 7 4101.8 19.910
7 8 4090.5 21.202
8 9 4006.5 17.919
9 10 3 9 8 0 . 0 1 8 . 2 5 2
10 11 3 8 5 9 . 5 1 2 . 6 7 9
11 12 3 7 9 3 . 0 1 0 . 4 9 5
12 13 3 7 8 6 . 0 1 2 . 0 5 7
13 14 3 7 8 5 . 1 1 4 . 0 0 0

Based on the Cp value, we might stop after Step 11, right after the ankle
variable is brought in. The resulting model would consist of predictors
abdomen, height, age, wrist, neck, forearm, hip, weight, biceps, thigh and
ankle.
By contrast, if one takes the traditional approach and selects the variables
on the basis of p-values, as discussed in Section 9.5, only 4 predictors would
be chosen (see output in Section 9.7.4), rather than 9 as above.
We can also determine what λ values were used:
> l a r s o u t $lambda
[ 1 ] 99.9203960 18.1246879 15.5110550 10.7746865
[ 5 ] 4.8247693 4.5923026 2.6282871 2.5472757
[ 9 ] 1.9518718 1.7731184 1.0385786 0.3162681
[ 1 3 ] 0.1051796

9.8 Post-Selection Inference

Stepwise predictor selection is an adaptive technique. This refers to any

statistical method that works in stages, with the outcome of any stage de-
termining what action is taken at the next stage. The problem with this is
that a proper statistical analysis would be based on the conditional distribu-
tion in the later stage, given the earlier stage, rather than the unconditional
368 CHAPTER 9. VARIABLE SELECTION

distribution. If for instance we wish to form a confidence interval for β5

after removing the predictor X (2) , the distribution of βb5 is no longer given
by the material in Section 2.8. If the latter is used, then inferences will be
incorrect. Following is an intuitive derivation of that property.
For concreteness, let’s consider the following simple example. We have
p = 2 predictors, and use a linear model without an intercept term,

E(Y |X (1) = t1 , X (2) = t2 ) = β1 t1 + β2 t2 (9.25)

Consider a classical variable selection approach, in which the variable to

delete is chosen by p-values (Section 9.7.1). The latter are calculated using
(2.56), which is based on the standard unconditional distribution assump-
tions.
Under those assumptions, the pair (βb1 , βb2 )′ has a bivariate normal distri-
bution with mean vector (β1 , β2 )′ (Section 2.8.4). For simplicity, suppose
in actuality β1 = β2 = 0. Also, suppose the standard errors of the βbi are
about 1.0.
In the forward stepwise process, we first test the hypothesis H0 : β1 = 0;
using the standard 0.05 significance level, we will reject the hypothesis if
and only if

|βb1 | > 1.96 (9.26)

If that hypothesis is rejected, then we retain X (1) as a predictor, and move

on to X (2) , retaining it if and only if

|βb2 | > 1.96 (9.27)

But (9.27) is based on the unconditional distribution of βb2 . Here, under

H0 : β2 = 0,

P (|βb2 | > 1.96) = 0.05 (9.28)

in the setting β2 = 0. But the proper probability to use would be

P (|βb2 | > 1.96 | |βb1 | > 1.96) (9.29)

9.9. DIRECT METHODS FOR DIMENSION REDUCTION 369

and this could be quite different from 0.05. Indeed, it could be near 1.0!
Here is why:
Recall the analysis in Section 8.2.1, especially (8.8). In that setting, βb1 and
βb2 are negatively/positively correlated if X (1) and X (2) are positively/neg-
atively correlated, i.e., c > 0 versus c < 0 in (8.8). Moreover, if |c| is near
1.0 there, the correlation is very near -1.0 and 1.0, respectively.
So, when c is near -1.0, for example, βb1 and βb2 will be highly positively
correlated. In fact, given that they have the same mean and variance, the
two quantities will be very close to identical, with high probability (Exercise
11, Chapter 2). In other words, (9.29) will be near 1.0, not the 0.05 value
we desire.
The ramification of this is that any calculation of confidence intervals or
p-values made on the final chosen model after stepwise selection cannot be
taken too much at face value, and indeed could be way off the mark. Sim-
ilarly, quantities such as the adjusted R-squared value may not be reliable
after variable selection.
Some research has been done to develop adaptive methods for such settings,
termed post-selection inference, but they tend to be hypothesis-testing ori-
ented and difficult to convert to confidence intervals, as well as having
restrictive assumptions. See for instance [59] [17]. There is no “free lunch”
here.

9.9 Direct Methods for Dimension Reduction

Given this chapter’s title, “Dimension Reduction,” we should discuss direct

methods for that goal.
Typically the methods are applied to X, the vector of predictor variables,
excluding Y . The oldest method is the classical statistical technique of
principal components analysis (PCA), but a number of others have been
devised in recent years. Here we will cover PCA and another method now
very popular in classification contexts, nonnegative matrix factorization, as
well as the parallel coordinates graphical approach we touched on in Section
6.7.3.

9.9.1 Informal Nature

Black cat, white cat, it doesn’t matter as long as it catches mice — former
370 CHAPTER 9. VARIABLE SELECTION

Chinese leader Deng Xiaoping

Though statisticians tend to view things through the lens of exact models,
proper computation of standard errors and so on, much usage of regression
models is much less formal, especially in the machine learning community.
Research in the area is often presented as intuitive ad hoc models that
seem to do well on some data sets. Since these models would typically
be extremely difficult to analyze mathematically, empirical study is all we
have.
Nevertheless, some methods have been found to work well over the years,
and must be part of the analyst’s toolkit, without worrying too much about
assumptions. For example, PCA was originally intended for use in multi-
variate normal distributions, but one might apply it to dummy variables.

9.9.2 Role in Regression Analysis

The basic approach for using these methods in regression is as follows:

• Center and scale the data.

• Find new predictors, typically linear combinations of the original pre-
dictor variables.
• Discard some of the new variables, on the grounds that they are
largely redundant.
• Then fit whatever regression model is desired — linear/nonlinear
parametric, or even nonparametric — on this new lower-dimensional
predictor space, i.e., using the remaining new variables.

9.9.3 PCA

We’ll start with Principal Components Analysis. This is a time-honored

statistical method, closely related to Singular Value Decomposition (SVD);
see for example [84] [44]. PCA was discussed in Section 8.7. Recall from
there that we create new variables, called principal components, which are
linear combinations of the original predictor variables. They are deter-
mined through a matrix diagonalization process (Section A.6) applied to
the sample covariance matrix S of the original predictors.
We then use these new variables as our predictors, discarding any having a
small variance, or equivalently, corresponding to a small eigenvalue.
9.9. DIRECT METHODS FOR DIMENSION REDUCTION 371

9.9.3.1 Issues

Though PCA regression would seem to be a natural solution to the dimen-

sion reduction problem, of course, this approach is no panacea:

• The sample covariance matrix is subject to sampling error, and thus

the ordering of the principal components with respect to their vari-
ances may be incorrect, resulting in some components being discarded
when they should be retained and vice versa. The coefficients in the
linear combinations may be similarly unstable. And the uncorrelated
nature of the principal components only holds for the sample, and
substantial correlations may exist at the population level.
• Even though one of the new predictors has a small variance, it still
could have a strong correlation with Y . Indeed, the problem is that
the entire dimension reduction process using PCA is oblivious to Y .
• As usual, we are faced with choosing the value of a tuning parameter,
the number of principal components to retain, k. A rule of thumb
used by many is to choose k so that the variances of the retained
variances sum to, say, 90% of the total. Of course, one might choose
k via cross-validation on the regression analysis.

9.9.3.2 Example: Bodyfat Data

Let’s apply PCA to the bodyfat data:

> l i b r a r y ( mfp )
Loading r e q u i r e d package : s u r v i v a l
Loading r e q u i r e d package : s p l i n e s
> data ( b o d y f a t )
> b o d y f a t <− b o d y f a t [ , − c ( 1 , 3 : 4 ) ]
> p r c <− prcomp ( b o d y f a t [ , − 1 ] , c e n t e r=TRUE, s c a l e=TRUE)
> s t r ( prc )
List of 5
$ sdev : num [ 1 : 1 3 ] 2 . 8 3 6 1 . 1 6 4 1 . 0 0 1 0 . 8 1 7 0 . 7 7 4 . . .
$ r o t a t i o n : num [ 1 : 1 3 , 1 : 1 3 ] 0 . 0 0 9 8 5 0 . 3 4 4 5 4
0.10114 0.30559 0.31614 . . .
...

R’s prcomp() function does PCA analysis, returning an object of class

’prcomp’. Whenever working with a new R class, it’s always useful to
get a quick overview, via the str() (“structure”) function. The full results
372 CHAPTER 9. VARIABLE SELECTION

are not shown here, but they show that prc includes various components,
notably sdev, the standard deviations of the principal components, and
rotation, the matrix of principal components themselves. Let’s look a
little more closely at the latter.
For instance, rotation[1,] will be the coefficients, called loadings, in the
linear combination of X that comprises the first principal component (which
is the first row of M ):
> prc$ r o t a t i o n [ , 1 ]
age weight height neck
0.009847707 0.344543793 0.101142423 0.305593448
chest abdomen hip thigh
0.316137873 0.311798252 0.325857835 0.310088798
knee ankle biceps forearm
0.308297081 0.230527382 0.299337590 0.249740971
wrist
0.279127655

In other words, the first principal component is

0.009847707 × age + 0.344543793 × weight + ... + 0.279127655 × wrist

(9.30)
Keep in mind that we have centered and scaled the original predictors, so
that for instance age has been so transformed. For example,
> prc$scale [ 1 ]
age
12.60204
> var ( b o d y f a t [ , − 1 ] $ age )
[ 1 ] 158.8114
> var ( b o d y f a t [ , − 1 ] $ age / p r c $ s c a l e [ 1 ] )
[1] 1

We could transform (9.30) to the original scaling by multiplying the coeffi-

cients by prc$scale. On the other hand, prc$x contains the new version
of our data, expressed in terms of the principal components. Our original
data, bodyfat[,1], contained 252 observations on 13 variables, and prc$x
is a matrix of those same dimensions:
> head ( p r c $x )
PC1 PC2 PC3 PC4
1 −2.2196368 1 . 2 3 6 6 1 2 −1.49623685 −0.2846021
9.9. DIRECT METHODS FOR DIMENSION REDUCTION 373

2 −0.8861482 2 . 0 0 9 7 1 8 −0.03778771 −0.2279745

3 −2.3610815 1 . 2 1 9 9 7 4 −2.19356148 1 . 9 1 4 2 0 2 7
4 −0.1158392 1 . 5 4 8 6 4 3 −0.31048859 −0.5128123
5 0 . 3 2 1 5 9 4 7 1 . 6 8 7 2 3 1 −1.21348165 1 . 4 0 7 0 3 6 5
6 3 . 0 6 9 2 5 9 4 2 . 3 6 4 8 9 6 −0.12285297 0 . 3 1 7 7 9 9 2
PC5 PC6 PC7 PC8
1 0 . 1 6 3 0 5 0 7 −0.2355776 −0.2260290 0 . 2 6 2 8 6 9 3 5
...

Whoever was the first person in our dataset has a value of -2.2196368 for
the principal component and so on.
Now, which principal components should we retain?
> summary( p r c )
Importance o f components :
PC1 PC2 PC3
Standard d e v i a t i o n 2.8355 1.1641 1.0012
Proportion of Variance 0.6185 0.1042 0.0771
Cumulative P r o p o r t i o n 0 . 6 1 8 5 0 . 7 2 2 7 0 . 7 9 9 8
PC4 PC5 PC6
Standard d e v i a t i o n 0.81700 0.77383 0.56014
Proportion of Variance 0.05135 0.04606 0.02413
Cumulative P r o p o r t i o n 0 . 8 5 1 1 6 0 . 8 9 7 2 2 0 . 9 2 1 3 6
PC7 PC8 PC9
Standard d e v i a t i o n 0.53495 0.51079 0.42776
Proportion of Variance 0.02201 0.02007 0.01408
Cumulative P r o p o r t i o n 0 . 9 4 3 3 7 0 . 9 6 3 4 4 0 . 9 7 7 5 1
PC10 PC11 PC12
Standard d e v i a t i o n 0.36627 0.27855 0.23866
Proportion of Variance 0.01032 0.00597 0.00438
Cumulative P r o p o r t i o n 0 . 9 8 7 8 3 0 . 9 9 3 8 0 0 . 9 9 8 1 8
PC13
Standard d e v i a t i o n 0.15364
Proportion of Variance 0.00182
Cumulative P r o p o r t i o n 1 . 0 0 0 0 0

Applying the 90% rule of thumb (again, cross-validation might be bet-

ter), we would use only the first 6 principal components. Let’s take that
as our example. Note that this means that our new predictor dataset is
prc$x[,1:6]. We can now use that new data in lm():
> lmout <− lm( b o d y f a t [ , 1 ] ∼ p r c $x [ , 1 : 6 ] )
> summary( lmout )
374 CHAPTER 9. VARIABLE SELECTION

...
Coefficients :
Estimate Std . E r r o r t v a l u e
( Intercept ) 18.9385 0.3018 62.751
p r c $x [ , 1 : 6 ] PC1 1.6991 0.1066 15.932
p r c $x [ , 1 : 6 ] PC2 −2.6145 0 . 2 5 9 8 −10.064
p r c $x [ , 1 : 6 ] PC3 −1.5999 0 . 3 0 2 1 −5.297
p r c $x [ , 1 : 6 ] PC4 0.5104 0.3701 1.379
p r c $x [ , 1 : 6 ] PC5 1.3987 0.3908 3.579
p r c $x [ , 1 : 6 ] PC6 2.0243 0.5399 3.750
Pr ( >| t | )
( Intercept ) < 2 e −16 ∗∗∗
p r c $x [ , 1 : 6 ] PC1 < 2 e −16 ∗∗∗
p r c $x [ , 1 : 6 ] PC2 < 2 e −16 ∗∗∗
p r c $x [ , 1 : 6 ] PC3 2 . 6 2 e −07 ∗∗∗
p r c $x [ , 1 : 6 ] PC4 0.169140
p r c $x [ , 1 : 6 ] PC5 0.000415 ∗∗∗
p r c $x [ , 1 : 6 ] PC6 0.000221 ∗∗∗
...
M u l t i p l e R−s q u a r e d : 0 . 6 2 7 1 , Adj . R−s q u a r e d : 0.6179
...

The adjusted R-squared value, about 0.62, is considerably less than what
we obtained from stepwise regression earlier, or for that matter, than what
the full model gave us. This suggests that we have used too few principal
components. (Note that if we had used all of them, we would get the full
model back again, albeit with transformed variables.)
As mentioned, we could choose the number of principal components via
cross-validation: We would break the data into training and test sets, then
apply lm() to the training set p times, once with just one component, then
with two and so on. We would then see how well each of these fits predicts
in the test set.
However, in doing so, we would lose one of the advantages of the PCA
approach, which is that it does predictor selection independent of the Y
values. Our selection process with PCA does not suffer from the problems
cited in Section 9.8.
On the other hand, in situations with very large values of p, say in the
hundreds or even more, PCA provides a handy way to cut things down to
size. On that scale, it also may pay to use a sparse version of PCA [138].
9.9. DIRECT METHODS FOR DIMENSION REDUCTION 375

9.9.3.3 Example: Instructor Evaluations

This dataset, another from the UCI repository,5 involves student evalua-
tions of instructors in Turkey. It consists of four questions about the student
and so on, and student ratings of the instructors on 28 different aspects,
such as “The quizzes, assignments, projects and exams contributed to help-
ing the learning.” The student gives a rating of 1 to 5 on each question.
We might be interested in regressing Question 9, which measures student
enthusiasm for the instructor against the other variables, including the
difficulty variable (“Level of difficulty of the course as perceived by the
student”). We might ask, for instance, how much of a factor is that variable,
when other variables, such as “quizzes helpful for learning,” are adjusted
for.
It would be nice to reduce those 28 rating variables to just a few. Let’s try
PCA:
> t u r k <−
read . csv ( ’ t u r k i y e −s t u d e n t −e v a l u a t i o n g e n e r i c . c s v ’ )
> t p c a <− prcomp ( t u r k [ , − ( 1 : 2 ) ] )
> summary( t p c a )
Importance o f components :
PC1 PC2 PC3
Standard d e v i a t i o n 6.1372 1.70133 1.40887
Proportion of Variance 0.7535 0.05791 0.03971
Cumulative P r o p o r t i o n 0 . 7 5 3 5 0 . 8 1 1 4 3 0 . 8 5 1 1 4
PC4 PC5 PC6
Standard d e v i a t i o n 1.05886 0.81337 0.75777
Proportion of Variance 0.02243 0.01323 0.01149
Cumulative P r o p o r t i o n 0 . 8 7 3 5 7 0 . 8 8 6 8 1 0 . 8 9 8 3 0
...

So, the first principal component already has about 75% of the total varia-
tion of the data, rather remarkable since there are 32 variables. Moreover,
the 28 ratings all have about the same coefficients:
> tpca$ r o t a t i o n [ , 1 ]
instr class nb . repeat
> tpca$ r o t a t i o n [ , 1 ]
nb . repeat attendance difficulty
0 . 0 0 3 5 7 1 0 4 7 −0.048347081 −0.019218696
Q1 Q2 Q3
5 https://archive.ics.uci.edu/ml/datasets/Turkiye+Student+Evaluation.
376 CHAPTER 9. VARIABLE SELECTION

−0.178343832 −0.186652683 −0.181963142

Q4 Q5 Q6
−0.183842387 −0.189891414 −0.186746480
Q7 Q8 Q9
−0.187475846 −0.186424737 −0.182169214
Q10 Q11 Q12
−0.192004876 −0.186461239 −0.185863361
Q13 Q14 Q15
−0.192045861 −0.191005635 −0.190046305
Q16 Q17 Q18
−0.195944361 −0.180852697 −0.193325354
Q19 Q20 Q21
−0.192520512 −0.193006276 −0.190947008
Q22 Q23 Q24
−0.190710860 −0.194580328 −0.192843769
Q25 Q26 Q27
−0.188742048 −0.190603260 −0.189426706
Q28
−0.188508546

In other words, the school administrators seem to be wasting the students’

time with all those questions! One could capture most of the content of the
data with just one of the questions — any of them. To be fair, the second
component shows somewhat more variation about the student questions.

> tpca$ r o t a t i o n [ , 2 ]
nb . repeat attendance difficulty
−0.0007927081 −0.7309762335 −0.6074205968
Q1 Q2 Q3
0.1124617340 0.0666808104 0.0229400512
Q4 Q5 Q6
0.0745878231 0.0661087383 0.0695127306
Q7 Q8 Q9
0.0843229105 0.0869128408 0.0303247621
...

Also interesting is that the difficulty variable is basically missing from the
first component, but is there with a large coefficient in the second one.
9.9. DIRECT METHODS FOR DIMENSION REDUCTION 377

9.9.4 Nonnegative Matrix Factorization (NMF)

Nonnegative matrix factorization (NMF) is a popular tool in many appli-

cations, such as image recognition and text classification.

9.9.4.1 Overview

Given an u × v matrix A with nonnegative elements, we wish to find non-

negative, rank-k matrices6 W (u × k) and H (k × v) such that

A ≈ WH (9.31)

The larger the rank, the better our approximation in (9.31). But we typi-
cally hope that a good approximation can be achieved with

k ≪ rank(A) (9.32)

The rank here is analogous to the choice of number of components in PCA.

However, there may be other criteria for choosing k.
One of the popular methods for computing NMF actually uses regression
analysis in a creative manner. See Section 9.12.1.

9.9.4.2 Interpretation

The rows and columns of A above will correspond to some quantities of

interest, and as will be seen here, the rows of H are viewed as “typical”
patterns of the ways the column variables interact. To make this concrete,
consider the text classification example in Section 4.3.6.
In that example, the matrix A had 4601 rows, corresponding to the 4601
e-mail messages in our data. There were 48 columns of word frequency
data, corresponding to the 48 words tabulated.
The matrix H will then have dimensions k × 48. Its rows can then be
thought of as k synthetic “typical” e-mail messages. And row i of W will
give us the coefficients in the linear combination of those synthetic messages
that approximates message i in our data.
6 Recallthat the rank of a matrix is the maximal number of linearly independent rows
or columns.
378 CHAPTER 9. VARIABLE SELECTION

How are these synthetic messages then used to do machine classification of

new messages, to determine spam/no spam status? One can then use these
variables as inputs to a logit model or whatever method we wish.
Take for instance our e-mail example above. Say we choose k = 3. We now
have p = 3 in, say, our logit model. The matrix W , of size 4061 × 3, will
take the place of A as our matrix of predictor values.
Note that the factorization (9.31) is not unique. For any invertible k × k
matrix R with a nonnegative inverse, then setting W̃ = W R−1 and H̃ =
RH will produce the same product. But this should not be a problem
(other than possible computational convergence issues); any set of synthetic
typical messages in the above sense should work, provided we have a good
value of k.

9.9.4.3 Sum-of-Parts Property

Up there in the sky.

Don’t you see him?
No, not the moon.
The Man in the Moon. — William Joyce
“Sum-of-Parts” is a hoped-for-property, not guaranteed to arise, and pos-
sibly in the eye of the beholder. Due to the nonnegativity of the matrices
involved, it is hoped that the matrix H will be sparse. In that case, the
locations of the nonzero elements may be directly meaningful to the given
application [86]. In facial recognition, for instance, we may find that rows
of H correspond to the forehead, eyes, nose and so on.
As noted above, the NMF factorization is not unique. Thus if a nicely
interpretable factorization is found, changing the factorization using a ma-
trix R as above will likely destroy that interpretability. Thus interpretation
should be done with care, with review by domain experts.

9.9.4.4 Example: Spam Detection

Let’s again use the spam dataset from Section 4.3.6.

> l i b r a r y ( ElemStatLearn )
> data ( spam )
> spam$spam <− as . integer ( spam$spam == ’ spam ’ )
> l i b r a r y (NMF)
> spam48 <− spam [ , 1 : 4 8 ]
9.9. DIRECT METHODS FOR DIMENSION REDUCTION 379

> rowsums <− apply ( spam48 , 1 ,sum)

> spam48a <− spam48 [ rowsums > 0 , ]
> nmfout <− nmf ( spam48a , 1 0 )
> h <− nmfout@fit@H

In our first try at using NMF on this dataset (not shown), it turned out
that some rows of spam48 consisted of all 0s, preventing the computation
from being done. Thus we removed the offending rows.
The choice of 10 for the rank was rather abitrary. It needs to be less than
or equal to the minimum of the numbers of rows and columns in the input
matrix, in this case 4601 and 48, preferably much less. Again, we might
choose the rank via cross-validation.
It turns out that this data set (as is typical in the text classification case)
does yield the sum-of-parts property. Here is the first row of H:

> h[1 ,]
A. 1 A. 2 A. 3 A. 4
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 5 A. 6 A. 7 A. 8
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 9 A. 1 0 A. 1 1 A. 1 2
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 1 3 A. 1 4 A. 1 5 A. 1 6
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 1 7 A. 1 8 A. 1 9 A. 2 0
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 2 1 A. 2 2 A. 2 3 A. 2 4
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 2 5 A. 2 6 A. 2 7 A. 2 8
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 2 9 A. 3 0 A. 3 1 A. 3 2
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 3 3 A. 3 4 A. 3 5 A. 3 6
2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
A. 3 7 A. 3 8 A. 3 9 A. 4 0
6 . 5 2 3 5 8 8 e −03 2 . 2 2 0 4 4 6 e −16 4 . 3 0 9 1 9 6 e −03 2 . 2 2 0 4 4 6 e −16
A. 4 1 A. 4 2 A. 4 3 A. 4 4
2 . 3 9 3 1 2 4 e −03 2 . 2 2 0 4 4 6 e −16 2 . 3 0 7 6 1 4 e −03 2 . 2 2 0 4 4 6 e −16
A. 4 5 A. 4 6 A. 4 7 A. 4 8
1 . 6 5 0 8 4 0 e −02 9 . 8 5 5 1 5 4 e −03 2 . 2 2 0 4 4 6 e −16 2 . 2 2 0 4 4 6 e −16
380 CHAPTER 9. VARIABLE SELECTION

The entries 2.220446e-16 are basically 0s, so the nonzero entries are for
words A.37, ’1999’, A.39, ’pm’, A.41, ’cs’, a.43, ’original’, A.45, ’re’ (i.e., re-
garding) and A.46, ’edu’. Since the dataset consists of messages received by
the person who compiled the data, a Silicon Valley engineer, this row seems
to correspond to the messages about meetings, possibly with acdemia. The
second row (not shown) has just two nonzero entries, A.47, ’table’, and
A.48, ’conference’. Some rows, such as row 4, seem to be flagging spam
messages, containing words like ’credit.’
As in the PCA case, we could use these 10 variables for prediction, instead
of the original 48. (We would still use the other variables, e.g., related to
long all-capital-letter words.)
Note the specific form of these new variables. We see from h[1,] above, for
instance, that the first new variable is
0 . 0 0 6 5 ∗ A. 3 7 + 0 . 0 0 4 3 ∗ A39 + 0 . 0 0 2 4 ∗ A. 4 1 +
0 . 0 0 2 3 ∗ A. 4 3 + 0 . 0 1 6 5 A. 4 5 + 0 . 0 0 9 9 ∗ A. 4 6

9.9.5 Use of freqparcoord for Dimension Reduction

The freqparcoord package provides us with another tool for dimension

reduction.

9.9.5.1 Example: Student Evaluations of Instructors

Consider the Turkish instructor evaluations again. Let’s run freqparco-

ord:
> library ( freqparcoord )
> t u r k <−
read . csv ( ’ t u r k i y e −s t u d e n t −e v a l u a t i o n g e n e r i c . c s v ’ )
> f r e q p a r c o o r d ( j i t t e r ( as . matrix ( t u r k [ , − ( 1 : 4 ) ] ) ) ,m=5)

The R jitter() function adds a small amount of random noise to data. We

needed it here for technical reasons; freqparcoord() cannot operate if the
data includes too many tied values.7
The result, shown in Figure 9.2, is rather striking. We seem to be working
with just 3 kinds of instructors: those who consistently get high ratings to
7 The function uses k-NN density estimation, Section 5.10.1.1. With lots of ties and

moderate k, the (multidimensional analog of the) quantity h will be 0, causing a divide-

by-0 problem.
9.9. DIRECT METHODS FOR DIMENSION REDUCTION 381

Figure 9.2: Turkish student evaluations

the various questions; those who consistently get medium-high ratings; and
those who consistently get low ratings. (In viewing the vertical axis, recall
that the data are centered and scaled.) This suggests removing many of
the questions from our analysis, so that we can better estimate the effect
of the difficulty variable.
Of course, given what we learned about this data through PCA above, the
results here are not entirely unexpected. But freqparcoord is giving us
further insight, showing three groups.

9.9.5.2 Dimension Reduction for Dummy/R Factor Variables

As discussed in Section 1.16, in many cases, in order to obtain a good fit in a

parametric model, we need to include interaction terms. This is true for any
382 CHAPTER 9. VARIABLE SELECTION

kinds of predictors, including dummy variables (or R factors). To include

second-order interactions, for example, we would form new predictors from
products of pairs of the original dummies; for three-way interactions, we
would add in products of triples and so on.
Such product terms can be numerous. In the bank marketing example
in Section 9.7.5.1, there are 12 kinds of jobs, 3 types of marital status, 4
levels of education etc. For three-way interactions, that would already be
11 × 2 × 3 = 66 new terms. In other words, the value of p would increase
by 66 just from those added terms, and there would be many more.
If, say, job type and education were independent random variables, we
would not need pairs of that type. But of course there is some dependency
there, and elsewhere among the dummies for this data set. Thus there may
be a real need for dimension reduction.
This need can be met via log-linear models [35], which are used to analyze
tabular data. Though this topic is beyond the scope of this book, the basic
idea is to write the log of the cell probabilities in the given table as a sum
of main effects, first-order interactions and so on. Readers who are familiar
with analysis of variance (ANOVA) will find the structure to be similar.
Computations can be done with for instance, the built-in loglin().

9.10 The Verdict

So, what predictor selection method should one use? Though many empir-
ical studies have been conducted on real and simulated data, there simply
is no good answer to this question. The issue of dimension reduction is an
unsolved problem.
The classical version of stepwise regression described at the beginning of
Section 9.7.1, in which a hypothesis test is run to decide whether to include
a predictor, has been the subject of a great deal of research and even more
criticism. Sophisticated techniques have been developed (see for example
[110] and [112]), but they have stringent assumptions, such as normality,
homoscedasticity and exact validity of the linear model. Again, this is an
unsettled question.
Nevertheless, a wide variety of methods have been developed for approach-
ing this problem, allowing analysts to experiment and find ones they believe
work reasonably well.
We will return to this problem in Chapter 12 to discuss the case p >> n,
9.11. FURTHER READING 383

where it is especially important.

9.11 Further Reading

Alan Miller’s book [110] is a tour de force on the topic of predictor variable
selection.
There is much interesting material on the use of PCA, NMF and so on in
classification problems in [44].
As noted earlier, [65] is a comprehensive treatment of the LASSO and
related estimators. The fundamental paper presenting the theory behind
the lars package, notably the relation of LASSO and stepwise regression
to another technique called least-angle regression is [42]. See [28] for a
theoretical treatment.

9.12 Computational Complements

9.12.1 Computation for NMF

The matrices W and H are calculated iteratively, with one of the major
methods being regression. (There are other methods, such as a multiplica-
tive update method; see [90].) Here is how:
We make initial guesses for W and H, say with random numbers. Now
consider an odd-numbered iteration. Suppose just for a moment that we
know the exact value of W , with H unknown. Then for each j we could
“predict” column j of A from the columns of W . The coefficient vector
returned by lm() will become column j of H. (One must specify a model
without an intercept term, which is signaled via a -1 in the predictor list;
see Section 2.4.5.) We do this for j = 1, 2, ..., v.
In even-numbered iterations, suppose we know H but not W . We could
take transposes,

A′ = H ′ W ′ (9.33)

and then just interchange the roles of W and H above. Here a call to lm()
gives us a row of W , and we do this for all rows.
384 CHAPTER 9. VARIABLE SELECTION

Figure 9.3: Mt. Rushmore

R’s NMF package [54] for NMF computation is quite versatile, with many,
many options. In its simplest form, though, it is quite easy to use. For a
matrix a and desired rank k, we simply run
> nout <− nmf ( a , k )

The factors are then in nout@fit@W and nout@fit@H.

Let’s illustrate it in an image context, using the image in Figure 9.3.
Though typically NMF is used for image classification, with input data
consisting of many images,8 here we have only one image, and we’ll use
NMF to compress it, not do classification. We first obtain A:
> l i b r a r y ( pixmap )
> mtr <−
read . pnm( ’ MtRush . pgm ’ ) # s e e t h i s book ’ s w e b s i t e
> a <− mtr@grey

Now, perform NMF, find the approximation to A, and display it, as seen
in Figure 9.4:
8 Each image is stored linearly in one column of the matrix A.
9.12. COMPUTATIONAL COMPLEMENTS 385

Figure 9.4: Mt. Rushmore, compressed image

> aout <− nmf ( a , 5 0 )

> w <− aout@fit@W
> h <− aout@fit@H
> approxa <− w %∗% h
# b r i g h t n e s s v a l u e s must be i n [ 0 , 1 ]
> approxa <− pmin( approxa , 1 )
> mtrnew <− mtr
> mtrnew@grey <− approxa
> plot ( mtrnew )

This is understandably blurry. The original matrix has dimension 194×259,

and thus presumably has rank 194.9 We’ve approximated the matrix by
9 This is confirmed by running the function rankMatrix() in the Matrix package
[10].
386 CHAPTER 9. VARIABLE SELECTION

one of rank only 50, with a 75% storage savings. This is not important for
one small p;cture, but possibly worthwhile if we have many large ones. The
approximation is not bad in that light, and may be good enough for image
recognition or other applications.
Indeed, in many if not most applications of NMF, we need to worry about
overfitting. As you will see later, overfitting in this context amounts to
using too high a value for our rank, something to be avoided.

9.13 Mathematical Complements

9.13.1 MSEs for the Simple Example

Here are the details of the MSE computations in Section 9.1.1. From Section
1.19.2 we know that

M SE = variance + (bias)2 (9.34)

In the case of (9.6), we know that each µ b1 is an unbiased estimator of µi ,

with variance σ 2 /n. So (9.6) has the value 2σ 2 /n.
Things are a little more complicated for (9.7). First, the variance term is

1 1 1
V ar[ (X + Y )] = 2V ar(X) = σ 2 /n (9.35)
2 4 2

so that the variance portion of (9.7), σ 2 /n, is smaller than that of (9.7).
This of course is the goal of using the µ̌i . But that improvement is offset
by the nonzero bias. How large is it?

1 1
bias = E(µ̌i ) − µi = E[ (X + Y )] − µi = (µ1 + µ2 ) − µi (9.36)
2 2

so that

1
bias2 = (µ1 − µ2 )2 (9.37)
4
9.14. EXERCISES: DATA, CODE AND MATH PROBLEMS 387

In other words, the total value of (9.7) is

1
σ 2 /n + (µ1 − µ2 )2 (9.38)
2

9.14 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. Apply lars() to the bodyfat data, as in Section 9.7.7.2, this time setting
type = ’stepwise’, and compare the results.
2. Apply principal component regression to the letters recognition data in
Section 5.5.4.
3. Take the NMF factorization in Section 9.9.4.4, and use logit to predict,
as in Section 4.3.6. Try various values of the rank k. Compare to the logit
full model in Section 4.3.6.
4. In our analysis of evaluations of Turkish instructors in Section 9.9.5,
we found that there were three main groups of students; in one group the
students gave instructors consistently high evaluations, and so on. The
lme4 package [9] also contains a data set of this kind, InstEval. Explore
that data, to see whether a similar pattern emerges.
Mini-CRAN and other computational problems:
5. Edit R’s summary.lm so that in addition to the information printed
out by the ordinary version, it also reports Cp .
6. Edit R’s summary.lars so that in addition to the information printed
out by the ordinary version, it prints out R2 .
7. Write a function stepAR2() that works similarly to stepAIC(), except
that this new function uses adjusted R2 as its criterion for adding or deleting
a predictor. The call form will be
stepAR2 ( lmobj , dir= ’ fwd ’ , n s t e p s=ncol ( lmobj $model) −1)

where: lmobj is an object of class ’lm’; dir is ’fwd’ or ’back’; and nsteps
is a positive integer.
388 CHAPTER 9. VARIABLE SELECTION

Predictors will be added/deleted one at a time, over the course of nsteps

models, according to which one maximizes adjusted R2 (even if that value
is lower than the present one).
The return value will be an S3 object of type ’stepr2’. with sole component
a data frame of βbi values (0s meaning the predictor is not currently in the
prediction equation), one row per model. There will also be an R2 column.
Write a summary() function for this class, that shows the actions taken
at each step of the process.
8. In principal components regression analysis, we transform the original p
predictor variables X (i) to p new ones that are linear combinations of the
originals, then choose q ≤ p of the new ones to use as our actual predictors.
Let W j , j = 1, ..., q denote those new predictors. We have

∑
p
Wj = wji X (i) (9.39)
i=1

Our regression model is now

E(Y | W ) = γ0 + γ1 W (1) + ... + γq W (q) (9.40)

But after fitting that model, we could transform back to X, substituting

with (9.39). In this way, we could obtain an estimated regression function
of Y on the original X, which might have useful interpretability.
Write an R function with call form
xformback ( p r c )

that takes the argument prc, which is output from prcomp(), and returns
the vector of estimated regression coefficients with respect to X derived as
above.
9. Consider the example in Section 9.8. Write an R function that will
evaluate (9.29) for the value of the argument c. Call the function for var-
ious values in (-1,1) and plot the results. Assume that the distribution of
Y given X is normal. Calculate the bivariate normal probabilities using
pmvnorm() in the mvtnorm() package [70].
10. With stepwise variable selection procedures, a central question is of
course what stopping rule to use, i.e., what policy to use. One approach, say
for the forward direction setting, has been to add p artificial noise variables
to the predictor set, and then stop the stepwise process the first time an
9.14. DATA, CODE AND MATH PROBLEMS 389

artificial variable is entered by the stepwise procedure [111]. Candes et al

[30] suggest that the artificial variables, called knockoffs, have a distribution
matching the distribution of the (real) predictor vector X in certain senses,
but we will assume i.i.d. knockoffs here.

(a) Write an R function with call form

k o f i l t e r ( dframe , yname , d i r e c t i o n )

that does the following: It first adds knockoffs, then calls lm() on the
data frame dframe with yname being the name of the response vari-
able. It then applies stepAIC() in the specified direction, stopping
according to the above prescription. either stopping the first time the
procedure attempts to enter a knockoff variable (forward direction)
or the first time all knockoffs are removed (backward direction).

(b) Apply this function to the bodyfat data, comparing the results to the
ones obtained in this chapter.

Math problems:
11. In this problem we will extend the analysis of Section 9.1.1.
Suppose we have a categorical variable X with k categories (rather than
just 2 categories as in Section 9.1.1). Let µi denote the mean of Y in the
subpopulation defined by X = i, and suppose µi − µi−1 = d, i = 2, ..., k.
Derive a variance-bias tradeoff analysis like that of (9.1.1).
12. Consider the following simple regression model with p = 2:

µ(t1 , t2 ) = β1 t1 + β2 t2 , V ar(Y |X = (t1 , t2 )) = σ 2 (9.41)

Assume we have an orthogonal design (Section 2.3). Derive a relation

similar to (9.11). As in Section 2.8.4, consider the conditional distribution
of Y given X in MSE computation.
13. For any k-dimensional vector u = (u1 , ..., uk )′ , its lq norm is defined to
be

∑k
||u||q = ( uqi )1/q (9.42)
i=1
390 CHAPTER 9. VARIABLE SELECTION

The familiar cases are q = 2, yielding the standard Euclidean norm in (A.1),
and q = 1, the latter playing a central role in the LASSO, Equation (8.18).
But it is defined for general q > 0.
In fact, an important special case is q = ∞. If one lets q → ∞ in (9.42), it
can be shown that one gets

||u||∞ = max |ui | (9.43)

i

Now, look at Figure 9.1. Since this was for the LASSO, it was based on the
use of the l1 norm in (8.18). Discuss how the figure would change for other
lq norms, in particular the cases q = 2, q = ∞ and 0 < q. In particular,
argue that the latter case would also lead to a variable selection process.
14. Fill in the gaps in the intuitive derivation in Section 9.8 to make it a
careful proof. You may find Exercise 11 of Chapter 2 helpful.
Chapter 10

Partition-Based Methods

These methods partition the X or feature space,1 space into rectangular

subregions, sub-subregions and so on. To predict a new case, we determine
which region its X value belongs to, and then predict from the Y values in
that region.
This approach is similar to k-NN methods, in that we are in essence finding
neighbors of the new case, but with a quite different way to find those
neighbors. The method was invented by statisticians and is popular in that
community, but are widely used in the machine learning community.
Here is a sneak preview, using the Letters Recognition data from Section
5.8.3. The code
> l r <− L e t t e r R e c o g n i t i o n # l e a v e l e t t r as f a c t o r !
> library ( rpart )
> l i b r a r y ( r p a r t . plot )
> r p l r <− r p a r t ( l e t t r ∼ . , data=l r , method= ’ c l a s s ’ )
> prp ( r p l r )

produces the “flow chart” in Figure 10.1. For instance, suppose we have a
certain letter image for which x2ybr is less than 2.5, but for which y2bar
is greater than or equal to 3.5. Then we predict this to be the letter ’L’.
Given the tree-like structure in Figure 10.1, it is not surprising that it is
referred to as a “tree.” This kind of approach is very appealing. It is easy to
1 The predictors are called features in the machine learning literature, inherited from

the electrical engineering community. In a classification problem, the names of the classes
are called labels, and a point with unknown class, to be predicted, is termed unlabeled.

391
392 CHAPTER 10. PARTITION-BASED METHODS

implement and even easier to explain to nonspecialists. Computationally,

it can handle large numbers of predictors, works well with dummy variables
and so on.
In the next section we will discuss the methodology behind this, known as
CART. A number of questions arise, such as how far to take the partitioning
process.
We then move on to the refinement, ensembles of trees. The idea behind
forming an ensemble of trees is to generate many different trees from the
same data, and combine the results. The combining is done by averaging
in the regression case, and by “voting” (as in the AVA approach) in the
classification setting.
The idea of random forests is to address the “discreteness” of CART. If an
observation lies near the boundary of some subregion, this produces some
instability: If this data point had been just slightly different, the entire
portion of the tree below it could change. We address this by looking at
random subsets of the data, producing the ensembles used for averaging or
voting.
The approach of boosted trees again generates an ensemble of trees, using
boosting, a technique we discussed briefly in Section 6.12.3. At any given
step in the process, we assign weights to each of our n data points. At
step i, we update the weights that were used in generating tree i − 1, with
greater weight placed on data points that were not predictd well. We then
generate tree i from these new weights.

10.1 CART

Classification and Regression Trees (CART) were developed by Breiman,

Friedman, Olshen and Stone [24], building on preliminary ideas from various
researchers (see [71] for a history). The basic concept is simple: Build a
“flow chart” as in the letter recognition data above.
In the leaf nodes, the estimated µb(t) is recorded if we have a regression
problem. For classification problems, the most likely class for that node is
recorded.
We will use the rpart package [132], as well as the rpart.plot package for
plotting [109]. (For larger graphs, use prp() from the same package.)
10.1. CART 393

Figure 10.1: Letter recognition flow chart (see color insert)

394 CHAPTER 10. PARTITION-BASED METHODS

10.2 Example: Vertebral Column Data

Let’s try CART on the dataset in Section 5.5:

> v e r t <− read . table ( ’ column 3C. dat ’ , h e a d e r=FALSE)
> library ( rpart )
> l i b r a r y ( r p a r t . plot )
> r p v e r t <− r p a r t (V7 ∼ . , data=v e r t , method= ’ c l a s s ’ )

Note that we needed to inform rpart() that this was a classification prob-
lem, flagged by method=’class’.
Now plot the result:
> r p a r t . plot ( r p v e r t )

The result is shown in Figure 10.2. We see that the first split is based on
the predictor V6. If that variable is greater than or equal to 16 (actually
16.08), we would guess V 7 = 2. Otherwise, we check whether V4 is less
than 28; if so, our prediction is V 7 = 0, and so on.
Figure 10.3 shows the partioning of the predictor space we would have if
we were to just use the variables V6 and V4. We have divided the space
into three rectangles, corresponding to the conditions in Figure 10.2. If we
were to add V5 to our predictor set, the rectangle labeled “V6 < 16, V4
>= 28” would be further split, according to the condition V5 < 117, and
so on.
Figure 10.3 also shows that CART is similar to a locally adaptive k-NN
procedure. In the k-NN context, that term means that the number of
nearest neighbors used can differ at different locations in the X space.
Here, each rectangle plays the role of a neighborhood.
Our estimated class probabilities for any point in a certain rectangle are
the proportions of observations in that rectangle in the various classes, just
as they would be in a neighborhood with k-NN. For instance, consider the
tall, narrow rectangle in the
> vgt16 <− v e r t [ v e r t $V6 > 1 6 . 0 8 , ]
> mean( vgt16 $V7 == ’ SL ’ )
[ 1 ] 0.9797297

shown as 0.98 in the rightmost leaf, i.e., terminal node.

Now take a look inside the returned object, which gives us more detailed
information:
10.2. EXAMPLE: VERTEBRAL COLUMN DATA 395

Figure 10.2: Flow chart for vertebral column data (see color insert)
396 CHAPTER 10. PARTITION-BASED METHODS

Figure 10.3: Partitioning for vertebral column data

10.2. EXAMPLE: VERTEBRAL COLUMN DATA 397

> rpvert
n= 310

node ) , s p l i t , n , l o s s , yval , ( yprob )

∗ d e n o t e s t e r m i n a l node

1 ) r o o t 310 160 SL ( 0 . 1 9 3 5 4 8 3 9 0 . 3 2 2 5 8 0 6 5 0 . 4 8 3 8 7 0 9 7 )
2 ) V6< 1 6 . 0 8 162 65 NO
(0.37037037 0.59876543 0.03086420)
4 ) V4< 2 8 . 1 3 5 35 9 DH
(0.74285714 0.25714286 0.00000000) ∗
5 ) V4>=28.135 127 39 NO
(0.26771654 0.69291339 0.03937008)
1 0 ) V5< 1 1 7 . 3 6 47 23 DH
(0.51063830 0.40425532 0.08510638)
2 0 ) V4< 4 6 . 6 4 5 34 10 DH
(0.70588235 0.26470588 0.02941176) ∗
2 1 ) V4>=46.645 13 3 NO
(0.00000000 0.76923077 0.23076923) ∗
1 1 ) V5>=117.36 80 11 NO
(0.12500000 0.86250000 0.01250000) ∗
3 ) V6>=16.08 148 3 SL
(0.00000000 0.02027027 0.97972973) ∗

The line labeled “2),” for instance, tells us that its decision rule is V 6 <
16.08; that there are 65 such cases; that we would guess V 7 = N O if we
were to stop there; and so on. Leaf nodes are designated with asterisks.
Prediction is done in the usual way. For instance, let’s re-predict the first
case:
z <− v e r t [ 1 , ]
> predict ( r p v e r t , z )
DH NO SL
1 0.7058824 0.2647059 0.02941176
> predict ( r p v e r t , z , type= ’ c l a s s ’ )
1
DH
L e v e l s : DH NO SL

The conditional probabilities that V7 is DH, NO or SL for this case are

0.71 etc., and we would guess V 7 = SL.
So, how well did we predict (on the original set, no cross-validation)?
398 CHAPTER 10. PARTITION-BASED METHODS

> rpypred <− predict ( r p v e r t , type= ’ c l a s s ’ )

> mean( rpypred == v e r t $V7)
[ 1 ] 0.883871

10.3 Technical Details

The devil is in the details — old saying

CART has been the subject of much research over the years, reflected in
a number of algorithms and options in the implementation software. The
details matter a lot.
The full details of CART are complex and beyond the scope of this book.
(And we have used default values for the various optional arguments in
rpart().) See [64] [92] for further information on general techniques, and
[131] for the details on rpart in particular.
However, we briefly discuss implementation in this section.

10.3.1 Split Criterion

At each step, CART must decide (a) whether to make further splits below
the given node, (b) if so, on which predictor it should make its next split,
and (c) the split itself, i.e., the cutoff value to use for this predictor. How
are these decisions made?
One method is roughly as follows. For a candidate split, we test whether
the mean Y values on the two sides of the split are statistically significantly
different. If so, we take the split with the smallest p-value; if not, we don’t
split.

10.3.2 Predictor Reuse and Statistical Consistency

Details of some theoretical properties are in [36]. However, the issue of sta-
tistical consistency (Section 2.7.3) should be discussed here. As presented
above, CART is not consistent. After all, with p predictors, the greatest
number of rectangles we can get is 2p , no matter how large the sample size
n is. Therefore the rectangles cannot get smaller and smaller as n grows,
which they do in k-NN. Thus consistency is impossible.
10.4. TUNING PARAMETERS 399

However, this is remedied by allowing the same predictor to enter into the
process multiple times. In Figure 10.3, for instance, V6 might come in
again at the node involving V6 and V6 at the lower left of the figure.
Here is a rough argument as to why statistical consistency can then be
achieved. Consider the case of p = 1, a single predictor, and think of
what happens at the first splitting decision. Suppose we take a simple split
criterion based on significance testing, as mentioned above. Then, as long
as the regression function is nonconstant, we will indeed find a split for
sufficiently large n. The same is true for the second split, and so on, so we
are indeed obtaining smaller and smaller rectangles as in k-NN.

10.4 Tuning Parameters

Almost any nonparametric regression method has one or more tuning pa-
rameters. For CART, the obvious parameter is the size of the tree. Too
large a tree, for example, would mean too few observations in each rectan-
gle, causing a poor estimate, just like having too few neighbors with k-NN.
Statistically, this dearth of data corresponds to a high sampling variance,
so once again we see the variance-bias tradeoff at work. The rpart() func-
tion has various ways to tune this, such as setting the argument minsplit,
which specifies a lower bound on how many points can fall into a leaf node.
Note that the same predictor may be used in more than one node, a key issue
that we will return to later. Also, a reminder: Though tuning parameters
can be chosen by cross-validation, keep in mind that cross-validation itself
can be subject to major problems (Sections 9.3.2 and 1.13).

10.5 Random Forests

CART attracted much attention after the publication of [24], but the au-
thors and other researchers continued to refine it. While engaged in a
consulting project, one of the original CART authors discovered a rather
troubling aspect, finding that the process was rather unstable [130]. If the
value of a variable in a single data point were to be perturbed slightly, the
cutoff point at that node could change, with that change possibly propa-
gating down the tree. In statistical terms, this means the variances of key
quantities might be large. The solution was to randomize, as follows.
400 CHAPTER 10. PARTITION-BASED METHODS

10.5.1 Bagging

To set up this notion, let’s first discuss bagging, where “bag” stands for
“bootstrap aggregation.”
The bootstrap [39] is a resampling procedure, a general statistical term
(not just for regression contexts) whose name alludes to the fact that one
randomly chooses samples from our sample! This may seem odd, but the
motivation is easy to understand. Say we have some population parameter
θ that we estimate by θ, b but we have no formula for the standard error of
the latter. We can generate many subsamples (typically with replacement),
calculate θb on each one, and compute the standard deviation of those values;
b
this becomes our standard error for θ.
We could then apply this idea to CART. We resample many times, con-
structing a tree each time. We then average or vote, depending on whether
we are doing regression or classification. The resampling acts to “smooth
out the rough edges,” solving the problem of an observation coming near a
boundary.
The method of random forests then tweaks the idea of bagging. Rather than
just resampling the data, we also take random subsets of our predictors,
and finding splits using them.
Thus many trees are generated, a “forest,” after which averaging or voting
is done as mentioned above. The first work in this direction was that of
Ho [67], and the seminal paper was by Breiman [22].
A popular package for this method is randomForest [89].

10.5.2 Example: Vertebrae Data

Let’s apply random forests to the vertebrae data. Continuing with the data
frame vert in Section 10.2, we have
> l i b r a r y ( randomForest )
> r f v e r t <− randomForest (V7 ∼ . , data=v e r t )

Prediction follows the usual format. Let’s look at five random rows:
> predict ( r f v e r t , v e r t [ sample ( 1 : 3 1 0 , 5 ) , ] )
267 205 247 64 211
NO SL NO SL NO
L e v e l s : DH NO SL
10.5. RANDOM FORESTS 401

So, cases 267, 205 and so on are predicted to have V7 equal to NO, SL etc.
Let’s check our prediction accuracy:
> r f y p r e d <− predict ( r f v e r t )
> mean( r f y p r e d == v e r t $V7)
[ 1 ] 0.8451613

This is actually a bit less than what we got from CART, though the differ-
ence is probably commensurate with sampling error.
The randomForest() function has many, many options, far more than we
can discuss here. The same is true for the return value, of class ’random-
Forest’ of course. One we might discuss is the votes component:
> r f v e r t $votes
DH NO SL
1 0.539682540 0.343915344 0.116402116
2 0.746268657 0.248756219 0.004975124
3 0.422222222 0.483333333 0.094444444
4 0.451977401 0.254237288 0.293785311
5 0.597883598 0.333333333 0.068783069
6 0.350877193 0.649122807 0.000000000
...

So, of the 500 trees (this value is queriable in rfvert$ntree), observation 1

was predicted to be DH by 0.54 of the trees, while 0.34 of them predicted
NO and so on. These are also the conditional probabilities of the various
classes.

10.5.3 Example: Letter Recognition

Let’s first apply CART:

> r p l r <− r p a r t ( l e t t r ∼ . , data=l r , method= ’ c l a s s ’ )
> rpypred <− predict ( r p l r , type= ’ c l a s s ’ )
> mean( rpypred == l r $ l e t t r )
[ 1 ] 0.4799

Oh, no! This number, about 48% is far below what we obtained with our
logit model in Section 5.5.4, even without adding quadratic terms. But
recall that the reason for adding those terms is that we suspected that the
regression functions are not monotonic in some of the predictors. That
could well be the case here. But if so, perhaps random forests can remedy
that problem. Let’s check:
402 CHAPTER 10. PARTITION-BASED METHODS

> r f l r <− randomForest ( l e t t r ∼ . , data=l r )

> r f y p r e d <− predict ( r f l r )
> mean( r f y p r e d == l r $ l e t t r )
[ 1 ] 0.96875

Ah, very nice. So use of random Forests led to a dramatic improvement

over CART in this case.

10.6 Other Implementations of CART

A number of authors have written software implementing CART (and other

partitioning-based methods). One alternative package to rpart worth try-
ing is partykit [72], with its main single-tree function being ctree(). It
uses p-value-based splitting and stopping rules. You may find that it pro-
duces better accuracy than rpart, possibly at the expense of slower run
time.
Here is how it does on the letter recognition data:
> library ( p a r t y k i t )
Loading r e q u i r e d package : grid
> l i b r a r y ( mlbench )
> data ( L e t t e r R e c o g n i t i o n )
> l r <− L e t t e r R e c o g n i t i o n
> c t o u t <− c t r e e ( l e t t r ∼ . , data=l r )
> c t p r e d <− predict ( c t o u t , l r )
> mean( c t p r e d == l r $ l e t t r )
[ 1 ] 0.8552

This is much better than what we got from rpart, though still somewhat
below random forests. (The partykit version, cforest() is still experimen-
tal as of this writing.)
Given that, let’s take another look at rpart. By inspecting the tree gen-
erated earlier, we see that no predictor was split more than once. This
could be a real problem with non-monotonic data, and may be caused by
premature stopping of the tree-building process.
With that in mind, let’s try a smaller value of the cp argument, which is
a cutoff value for split/no split, relative to the ratio of the before-and-after
split criterion (default is 0.01),
> r p l r <− r p a r t ( l e t t r ∼ . , data=l r , method= ’ c l a s s ’ ,
10.7. EXERCISES: DATA, CODE AND MATH PROBLEMS 403

cp =0.00005)
> rpypred <− predict ( r p l r , type= ’ c l a s s ’ )
> mean( rpypred == l r $ l e t t r )
[ 1 ] 0.8806

Great improvement! It may be the case that this is generally true for non-
monotonic data.

10.7 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. Fill in the remainder of Figure 10.3.
2. Not surprising in view of the ’R’ in CART, the latter can indeed be used
for regression problems, not just classification. In rpart(), this is specified
via method = ’anova’. In this problem, you will apply this to the bodyfat
data (Section 9.7.4).
Fit a CART model, and compare to our previous results by calculating R2
in both cases, using the code in Problem 4.
3. Download the New York City taxi data (or possibly just get one of
the files), http://www.andresmh.com/nyctaxitrips/ Predict trip time from
other variables of your choice, using CART.
Mini-CRAN and other computational problems:
4. As noted in Section 2.9.2, R2 , originally defined for the classic linear
regression model, can be applied to general regression methodology, as it
is the squared correlation between Y and predicted-Y values. Write an R
function with call form
r p a r t r 2 ( r p a r t o u t , newdata=NULL, type= ’ c l a s s ’ )

that returns the R2 value in a CART setting. Here rpartout is an object

of class ’rpart’, output of a call to rpart(); newdata is a data frame on
which prediction is to be done, with that being taken as the original data
set if this argument is NULL; and type is as in predict.rpart().
5. It would be interesting, in problems with many predictors, to see which
ones are chosen by rpart(). Write an R function with call form
404 CHAPTER 10. PARTITION-BASED METHODS

s p l i t s v a r s ( rpartout )

that returns a vector of the (unique) predictor variable names involved in

the splits reported in rpartout, an object of class ’rpart’.
6. In classification problems, it would be useful to identify the hard-to-
classify cases. We might do this, for instance, by finding the cases in which
the two largest probabilities in the votes component of an object of class
’randomForest’ are very close to each other. Write a function with call
form
hardcases ( rfobj , to l )

that reports the indices of such cases, with the argument tol being the
percentage difference between the two largest probabilities.
Chapter 11

Semi-Linear Methods

Back in Chapter 1, we presented the contrasting approaches to estimating

a regression function:

• Nonparametric methods, such as k-Nearest Neighbors, have the ad-

vantage of being model-free, and thus generally applicable.
• Parametric models, if reasonably close to reality, are more powerful,
generally yielding more accurate estimates of µ(t).

In this chapter, we try to “have our cake and eat it too.” We present
methods that are still model-free, but make use of linearity in various ways.
For that reason, they have the potential of being more accurate than the
unenhanced model-free methods presented earlier. We’ll call these methods
semi-linear.
Here is an overview of the techniques we’ll cover, each using this theme of
a “semi-linear” approach::

• k-NN with Local Linearization: Instead of estimating µ(t) using

the average Y value in a neighborhood of a given point t, fit a linear
regression model to the neighborhood and predict Y at t using that
model. The idea here is that near t, µ(t1 , ..., tp ) should be monotonic
in each tj , indeed approximately linear in those variables,1 and this
approach takes advantage of that.2
1 Think of the tangent plane at the regression surface at t.
2 The reader may notice some similarity to the loess() function in R, which in turn
is based on work of Cleveland [37].

405
406 CHAPTER 11. SEMI-LINEAR METHODS

• Support Vector Machines: Consider for example the two-class

classification problem, and think of the boundary between the two
classes in the predictor space — on one side of the boundary we guess
Y = 1 while on the other side we guess Y = 0. We noted in Section
4.3.7 that under the logit model, this boundary is linear, i.e., a line
in two dimensions, a plane in three dimensions, and a hyperplane in
higher dimensions. SVMs drop the assumption of logit form for µ(t)
but retain the assumption that the boundary is linear.

In addition, recall that if we believe the effect of a particular predic-

tor X on Y is nonlinear, we might add an X 2 term to our paramet-
ric model, and possibly interaction terms (Section 1.16). SVM does
something like this too.

After the hyperplanes are found, OVA or AVA is used (Section 5.5);
[33], for instance, uses AVA.

• Neural networks: This too is based on modeling the class-separating

hypersurfaces as linear, and again input variables are put through
nonlinear transformations. The difference is that we go through a se-
ries of stages in which new variables are formed from old ones in this
manner. We first form new variables as linear functions of the orig-
inal inputs, and transform those new variables. We then find linear
functions of those transformed variables, and pass those new variables
through a transformation, and so on. Each step takes the variables
in one layer, and forms the next layer from them. The final layer is
used for prediction.

The latter two methods above are mainly used in classification settings,
though they can be adapted to regression. The basic idea behind all of
these methods is:

By making use of monotonicity and so on, these methods “should”

work better than ordinary k-NN or random forests.

But as the use of quotation marks here implies, one should not count on
these things.
11.1. K-NN WITH LINEAR SMOOTHING 407

11.1 k-NN with Linear Smoothing

Consider the code below, which simulates a setting with

E(Y | X = t) = t (11.1)

n <− 1000
z <− matrix ( runif ( 2 ∗n ) , ncol =2)
x <− z [ , 1 ]
y <− z %∗% c ( 1 , 0 . 5 )
xd <− p r e p r o c e s s x ( x , 1 0 0 )
kout <− k n n e s t ( y , xd , 1 0 0 )
plot ( x , kout$ r e g e s t )

The plot is shown in Figure 11.1. Most of it looks like a 45-degree line, as
it should, but near 0 and 1 the curve is flat. Here is why:
Think of a point very close to 1. Most or all of its neighbors will be to its
left, so their µ(t) values will be lower than for our given point. So, averaging
them, as straight k-NN does, will produce a downward bias. Similarly, a
point near 0 will experience an upward bias, hence the flattening of µ b(t) for
values of t near 0 and 1.

11.1.1 Extrapolation Via lm()

One solution to this is to fit a linear regression to the points in a neigh-

borhood, rather than just taking the average. This should reduce the bias
near 0 and 1, and may produce smoother estimates even in interior points.
In regtools, this can be specified by setting nearf = loclin in the call to
knnest():

> k o u t l l <− k n n e s t ( y , xd , 1 0 0 , n e a r f=l o c l i n )

> plot ( x , k o u t l l $ r e g e s t )

The result, shown in Figure 11.2, looks much better.

Of course, one could even try a quadratic approximation scheme. This is
explored in Exercise 3 at the end of this chapter.
408 CHAPTER 11. SEMI-LINEAR METHODS

Figure 11.1: k-NN bias on the edges of the data

Figure 11.2: Reducing k-NN bias on the edges of the data

11.1. K-NN WITH LINEAR SMOOTHING 409

11.1.2 Multicollinearity Issues

One issue to keep in mind here is multicollinearity. Recall the comment near
the end of Section 8.11.1.3, which in essence said that multicollinearity is
exacerbated in small samples. This is relevant in our context here: If our
dataset as a whole suffers from multicollinearity, then the problem would
be accentuated in the neighborhoods, since “n” is small for them. In such
cases, lm() may compute some of its coefficients as NA values, or possibly
even refuse to do the computation.
Remedies are as before: We can simply use fewer predictors, say choosing
them via PCA, or might even try a ridge approach.

11.1.3 Example: Bodyfat Data

Let’s apply these ideas to the bodyfat data of Section 9.7.4. Recognizing
the possible multicollinearity issues as above, we will only use a few of the
predictors, basically the first few found by stepAIC() in Section 9.7.4:

> b f <− b o d y f a t [ , c ( 1 , 2 , 3 , 5 , 7 ) ]
> xd <− p r e p r o c e s s x ( b f [ , − 1 ] , 1 0 )
> kout <− k n n e s t ( b f [ , 1 ] , xd , 1 0 )
> mean( abs ( b f [ , 1 ] − kout$ r e g e s t ) )
[ 1 ] 3.468016
> k o u t l l <− k n n e s t ( b f [ , 1 ] , xd , 1 0 , n e a r f=l o c l i n )
> mean( abs ( b f [ , 1 ] − k o u t l l $ r e g e s t ) )
[ 1 ] 2.453403

That’s quite an improvement. Of course, with n being only 252, part of the
difference may be due to sampling error, but the result does make sense.
This being data on humans, there are likely some individuals who are on
the fringes of the data, say people who are exceptionally thin. Use of the
local linear method may help predict such people more accurately.

11.1.4 Tuning Parameter

As before, the tuning parameter here is k, the number of nearest neighbors.

410 CHAPTER 11. SEMI-LINEAR METHODS

11.2 Linear Approximation of Class

Boundaries

In Section 4.3.7, it was noted that if a logistic model holds, then class
boundaries — say, on one side, we predict Class A and on the other side we
predict B — are linear: With one predictor, the boundary is a single point
value, with two it is a line, with three it is a plane, and with more than
three predictors it is a hyperplane. (As noted earlier, for the multiclass
case, one can apply OVA or AVA to the methods below. We will stay with
the two-class case here.)
The remaining two methods to be described in this chapter, Support Vec-
tor Machines (SVMs) and Neural Networks (NNs), seek to estimate class
boundaries, using linearity assumptions in indirect ways. (Both methods
can also be used for general regression purposes, but we will not pursue
that aspect much here.) In this sense they differ from the other methods
in this book, which involve estimation of conditional class probabilities,
P (Y = i | X = t).3 SVMs and NNs are much more “machine learning-ish”
than k-NN or random forests, and it is no coincidence that the latter two
were developed in the statistics community while the former are from ML.
This section is devoted to those ML methods.
A word on notation: In ML, a class membership variable Y is coded as +1
or -1, rather than 1 or 0 as in statistics.
Note: SVMs and NNs are highly complex methods, with many variations
and tuning parameters. This chapter can only scratch the surface. For
further details, see for instance [64] [83] [129]. Also, though cross-validation
can be used to choose the values of the tuning parameters, it must be once
again pointed out that cross-validation itself has problems (Section 9.3.2).

11.2.1 SVMs

SVM methodology [38] is used in a wide variety of applications, with a rich

theoretical foundation having been developed. We’ll introduce the subject
here.
We will assume the two-class classification problem, which can then be ap-
plied to multiclass settings via OVA or AVA. As noted above, the main mo-
tivation of SVM stems from the idea of a linear boundary between classes,
3 NNs can estimate those probabilities, as will be seen below, but tend to be thought
of as class predictors rather than probability estimators.
11.2. LINEAR APPROXIMATION OF CLASS BOUNDARIES 411

as with for instance the logistic model. But the linear-boundary nature of
the logit stems from a model on the regression function,

µ(t) = P (Y = 1 | X = t) (11.2)

whereas SVM makes NO assumption on the nature of that function; SVM

just models the interclass boundary.
How, then, might we estimate the parameters in that linear boundary?4
Lacking much probabilistic structure, we turn to geometry.5

11.2.1.1 Geometric Motivation

Let’s generate some artificial data, using this code:

li b r a r y ( mvtnorm )
cv <− rbind ( c ( 1 , − 0 . 1 ) , c ( − 0 . 1 , 1 ) )
set . s e e d ( 9 9 9 9 9 )
m <− 3 . 5
means <− rbind ( c ( 0 , 0 ) , c (m,m) )
n <− 7
p t s 0 <− rmvnorm ( n , means [ 1 , ] , cv )
p t s 1 <− rmvnorm ( n , means [ 2 , ] , cv )
plot ( x=NULL, y=NULL, xlim=c ( −2 ,m+3) , ylim=c ( −2 ,m+3) ,
x l a b= ’X1 ’ , y l a b= ’X2 ’ )
points ( pts0 , pch= ’ x ’ )
points ( pts1 , pch= ’ o ’ )

Our two predictor variables are X1 and X2, so a hyperplane is simply a

b Again,
line. Let’s refer to the population line as ℓ and our estimate as ℓ.
the question is, how should we choose ℓ?b

As you can see in Figure 11.3. there is quite a separation between the x
and o data, meant to be classes 1 and 0 in this simulation, so much so
that there are many lines we might take as our ℓb (all of which would give
a “perfect” fit, a matter that will come into play later). SVM makes that
choice in a manner that should warm the hearts of geometers.
4 SVM was invented in the machine learning community. As mentioned in this book’s

Preface, that community typically doesn’t think in terms of samples from a population.
We take the statistical view here.
5 This idea is certainly not limited to the ML community. After all, the origin of

least-squares regression was geometric. A modern example is concave density estimation,

which fits concave curves without other assumptions [124].
412 CHAPTER 11. SEMI-LINEAR METHODS

Figure 11.3: Artificial data

To explain, we need to define a term, convex hull. A geometric set C is

called convex if the line segment connecting any two points in the set itself
lies in the set. More formally,

x, y  C =⇒ tx + (1 − t)y  C for all 0 ≤ t ≤ 1 (11.3)

The convex hull of a collection of points is the smallest set that contains
the given points. The convex hulls of the x and o points are shown in
Figure 11.4. A common way to describe a convex hull is to imagine a string
wrapped tightly around the points, as can be seen in the picture.
I used R’s chull function to compute those convex hulls:
> ch0 <− chull ( p t s 0 )
> l i n e s ( p t s 0 [ c ( ch0 , ch0 [ 1 ] ) , ] )
# ch0 [ 1 ] needed f o r f i n a l segment
> ch1 <− chull ( p t s 1 )
> l i n e s ( p t s 1 [ c ( ch1 , ch1 [ 1 ] ) , ] )

With the convex hulls added to the picture, it is clearer what our choices
 but there are still infinitely many choices. We
are to choose the line ,
choose one as follows.
11.2. LINEAR APPROXIMATION OF CLASS BOUNDARIES 413

Figure 11.4: Convex hulls

First, we find the points closest to each other in the two sets. If such a
point is not a vertex of its set, it will at least be on one of the edges of the
set, and we record the vertices at the ends of that edge. In our case here,
that gives us vertices labeled SV01, SV11 and SV12 in Figure 11.5. These
are called the support vectors.
We then drop a line segment between the support vectors, in this case
a perpendicular line from SV01 to the line connecting SV11 and SV12.
Finally, we take  to be the line that perpendicularly bisects this dropped
line segment. The result is shown in Figure 11.6. The line  extends between
the top left and the bottom right.

11.2.1.2 Reduced convex hulls

Of course, the most salient “toy” aspect of the above example is that, as
mentioned, the data has been constructed to be separable, meaning that
a line can be drawn that fully separates the points of the two classes. In
reality, though, our data will not be separable. There will be overlap be-
tween the two point clouds, making the convex hulls overlap, and the above
formulation fails.
414 CHAPTER 11. SEMI-LINEAR METHODS

Figure 11.5: Support vectors

Figure 11.6: Support vectors

11.2. LINEAR APPROXIMATION OF CLASS BOUNDARIES 415

One obvious way to solve the problem of overlapping convex hulls is to

make them smaller, so that some points are not included. This is done via
reduced convex hulls (RCHs), defined as follows.
For a set of points W = {p1 , ..., pm }, the convex hull of W is defined
algebraically as all convex combinations of the pi

a1 p1 + ... + am pm (11.4)

with

a1 + ... + am = 1, ai ≥ 0 for all i (11.5)

In the reduced case, we set a cost c and impose the additional constraint

ai ≤ c for all i (11.6)

For any vertex i in the original convex hull, we would need αi = 1 so

any c value smaller than 1 would eliminate pi in the reduced convex hull.
This results in smaller convex hulls, and for sufficiently small c, they will be
nonoverlapping. The above method can then be used to find the separating
hyperplane.
There is an equivalent algebraic dual formulation of SVM, outlined in Sec-
tion 11.4.2 in the Mathematical Complements section at the end of this
chapter. There our parameter c above is replaced by a related value, C,
termed the cost.

11.2.1.3 Tuning Parameter

The smaller we set c, the fewer points there are of the form (11.4) that
satisfy (11.6). In other words, the smaller c, the smaller our RCH. The
quantity c is then our tuning parameter.
Let’s consider the variance-bias tradeoff (Section 1.11) in this setting. First
note that in the end, ℓb depends only on the support vectors. This makes
the solution ℓb sensitive to perturbations in the support vectors. In Figure
11.6, suppose the point SV12 were to be moved straight downward a bit.
This would force ℓb to move downward as well, i.e., have a more negative
slope.
416 CHAPTER 11. SEMI-LINEAR METHODS

However, suppose the two convex hulls in the picture were of the same
shape, orientation with respect to each other and so on, but much further
apart than in the picture. The same amount of downward motion of SV12
would have a smaller impact on ℓb in this case.
The point is that perturbations of, say, SV12 correspond to sampling vari-
ation, i.e., from one sample to another. The above thought experiment
shows that sampling variation corresponds to variation in ℓ. b We knew that
from general statistical principles, of course, but the implication is this:

A small value of c increases the distance (called the margin)

between the two RCHs, and thus reduces the variance of (the
b
coefficients in) ℓ.

b as it
On the other hand, if we set c too small, we are introducing bias in ℓ,
is ignoring the vital region near the boundary, basing our estimation only
on points at the edges of our combined data. In other words, we have a
variance-bias tradeoff, with smaller c giving us smaller variation but larger
bias, and the opposite for larger c.
But it’s not quite so simple as that. We could argue that with very small
c, our ℓb will be based on two RCHs that each contain a very small number
of points. That should increase variance.
So, we have competing intuitive arguments. Which one is correct? In view
of this conflict, it is not surprising that [133] found that increasing bias
does not necessarily decrease variance, and vice versa. Thus there is no
clear answer. Of course, we can still use cross-validation to choose c, but
perhaps less confidently than in other situations.

11.2.1.4 Nonlinear Boundaries

In our discussions of parametric models, we have frequently explored form-

ing squared and interaction terms from our predictor variables (Section
1.16). It would be natural, then, to do so with SVM, and that is indeed
what is commonly done, though with a computation-saving extra aspect
known as the kernel trick.
The details are presented in Section 11.4.3 in the Mathematical Comple-
ments section at the end of this chapter. The bottom line, though, is that
one can model nonlinearity of ℓ by transforming our X values. The as-
sumption is that in the transformed space, ℓ becomes linear.
11.2. LINEAR APPROXIMATION OF CLASS BOUNDARIES 417

The type of transformation is called the kernel, and most SVM software
packages offer the user a choice of kernels.

11.2.1.5 Statistical Consistency

Under the proper conditions, if we use techniques such as k-NN or random

forests, the estimates of our regression function µ(t) converge to the true
values as the sample size n goes to infinity. This will not quite be the case
with SVM. Here is the intuition:
First, as noted, we are not really estimating a regression function (though
regression versions of SVM do exist). We are estimating a class boundary.
But more important, we are assuming that that boundary has a linear
form (possibly after a kernel transform). If that assumption is true, then
with proper choice of c (or the cost C in Section 11.4.2), one could show
statistical consistency. The situation is similar to that of ordinary linear
models.
Of course, one can try using kernels of higher and higher order of complexity,
say polynomial kernels of higher and higher degree. Theoretically this would
give statistically consistent estimation for any interclass boundary surface.
But as noted in Section 1.16.4 for the linear-model case, this is not practical.

11.2.1.6 Example: Letter Recognition Data

Here we will apply SVM to the Letter Recognition data analyzed in Sections
5.5.4 and 10.5.3, using the svm() function from the popular e1071 pacakge.
As explained in the Preface to this book, we use the default values. notably
for the choice of kernel (radial basis) and cost C (1). With n = 20000 and
only 16 predictors, we will not bother with cross validation.
> l i b r a r y ( mlbench )
> data ( L e t t e r R e c o g n i t i o n )
> l r <− L e t t e r R e c o g n i t i o n
> l i b r a r y ( e1071 )
> e o u t <− svm ( l e t t r ∼ . , data=l r )
> svmpred <− predict ( eout , data=l r )
> mean( svmpred == l r $ l e t t r )
[ 1 ] 0.9624

This is similar to what we previously obtained with random forests, and

possibly slightly better than the result with k-NN.
418 CHAPTER 11. SEMI-LINEAR METHODS

11.2.2 Neural Networks

The term neural network (NN) alludes to a machine learning method that
is inspired by the biology of human thought. In a two-class classification
problem, for instance, the predictor variables serve as inputs to a “neuron,”
with output 1 or 0, with 1 meaning that the neuron “fires” and we decide
Class 1. NNs of several hidden layers, in which the outputs of one layer of
neurons are fed into the next layer and so on, until the process reaches the
final output layer, were also given biological interpretation.
The method was later generalized, using activation functions with outputs
more general than just 1 and 0, and allowing backward feedback from later
layers to earlier ones. This led development of the field somewhat away from
the biological motivation, and some questioned the biological intepretation
anyway, but NNs have a strong appeal for many in the machine learning
community. Indeed, well-publicized large projects using deep learning have
revitalized interest in NNs.

11.2.2.1 Example: Vertebrae Data

Let us once again consider the vertebrae data. We’ll use the neuralnet
package, available from CRAN.
> library ( neuralnet )
> v e r t <− read . table ( ’ column 3C. dat ’ , h e a d e r=FALSE)
> l i b r a r y ( dummies )
> ys <− dummy( v e r t $V7)
> v e r t <− cbind ( v e r t [ , 1 : 6 ] , ys )
> names( v e r t ) [ 7 : 9 ] <− c ( ’DH ’ , ’NO ’ , ’ SL ’ )
> set . s e e d ( 9 9 9 9 )
> nnout <− n e u r a l n e t (DH+NO+SL ∼ V1+V2+V3+V4+V5+V6 ,
data=v e r t , hidden =3, l i n e a r . output=FALSE)
> plot ( nnout )

Note that we needed to create dummy variables for each of the three classes.
Also, neuralnet()’s computations involve some randomness, so for the sake
of reproducibility, we’ve called set.seed().
As usual in this book, we are using the default values for the many possible
arguments, including using the logistic function for activation.

1
g(t) = (11.7)
1 + e−t
11.2. LINEAR APPROXIMATION OF CLASS BOUNDARIES 419

However, we have set a nondefault value of hidden = 3. That argument

is a vector through which the user states the number of hidden layers and
the number of neurons in each layer. By setting the value 3 here, we are
specifying three nodes in just one hidden layer, to keep the example simple.
We also had to specify linear.output = FALSE, to indicate that this is
a classification problem rather than regression.
The plot is shown in Figure 11.7, which will serve as our introduction to
NNs, as follows. The circles represent neurons. Values coming in from the
left of a neuron are fed into the activation function. Lines coming out of the
right of a neuron represent inputs going in to the next layer on the right.
The leftmost column of circles represents transforming our original input
variables via the activation function; the center column of three circles
represents our single hidden layer; and the rightmost column produces the
final estimated regression function values, in this case the values of P(V7 |
V1,...,V6).
The several lines coming into a neuron from the left are labeled with num-
bers representing weights. The final value fed into that neuron is the
weighted sum of the “ancestor” neurons on the left. That weighted sum is
input to the activation function, with the value of that function then being
the output value at the neuron.
Weights are computed on the outputs of all circles except the rightmost.
For instance, it turns out that the weight of logit(V1) as input to the second
circle in the center column is 0.64057.
For any circle except those in the leftmost column, a weighted sum of
outputs from the previous column is input. For example, the input to the
first circle, center column is 1.0841 · 1 + 0.84311 V1 + 0.49439 V2 + ...
The weighted sum is then fed into the logit function, with the final output
being the estimated conditional probability.
So, how do we predict with this? Instead of having a generic function for
predict, here the function is compute(). Let’s use it to go back and
re-predict our data:

> p r e d i n f o <− compute ( nnout , v e r t [ , 1 : 6 ] )

> ypred <− p r e d i n f o $ n e t . r e s u l t
> ypred [ 5 6 , ]
[ 1 ] 0.37218683338 0.60285701068 0.02462869721

So, for the 56th observation, our estimated probabilities for DH, NO and
SL are about 0.37, 0.60 and 0.02.
420 CHAPTER 11. SEMI-LINEAR METHODS

Figure 11.7: Verbebrae NN, 1 hidden layer (see color insert)

How about our overall accuracy?

> mean( apply ( ypred , 1 , which .max) ==
apply ( v e r t [ , 7 : 9 ] , 1 , which .max) )
[ 1 ] 0.783870967

This is not quite what we attained with a parametric logistic model in

Section 5.5.2.6 For a more careful analysis, we should use cross validation
as we did there.

11.2.2.2 Tuning Parameters and Other Technical Details

Clearly, the number of hidden layers, as well as the number of neurons per
layer, are tuning parameters. But there are more of them, involving things
such as to what degree iterative feedback (back propagation) is used.
The weights are typically calculated via least-squares minimization (in one
6 Of course, the logit model there should not be confused with our use of logit as our
activation function here.
11.2. LINEAR APPROXIMATION OF CLASS BOUNDARIES 421

variant or another), possibly with a LASSO-type regularization parameter.

These and other details are beyond the scope of this book.

11.2.2.3 Dimension Reduction

Say we have ℓ layers and d nodes per layer. The number of weights is then
potentially ℓd2 , which could be tremendous in large problems. The danger
of overfitting is thus quite grave, even though NNs are typically used with
very large n.
The solution is to somehow bar certain connections among the neurons in
one layer to those of the next. In other words, we impose structural 0s into
certain types of weights. This can be application-specific. For instance,
convolutional neural networks [114] are geared to image classification prob-
lems, as follows.
Images are naturally thought of as two-dimensional arrays, but are stored
as one-dimensional arrays, say in column-major form: First the top col-
umn is stored, then the second column and so on. This destroys the two-
dimensional locality of an image — points near a pixel are likely to have
similar values to it — but we can restore that locality by requiring the
weights in our NN to favor neural connections involving neighboring pixels.

11.2.2.4 Why Does It Work (If It Does)?

Hopefully the reader is not the type who is satisfied with “black box”
techniques, so this section presents an outline of why all this may actually
work!
As with SVM, a geometric view can be very helpful. Toward that end,
think of the case p = 2, and suppose we use an “ideal” neuron activation
functiuon a(s) equal to 1 for s > 0 and 0 otherwise. We can interpret
any set of weights coming in to a neuron in the first hidden layer as being
represented by a line in the (t1 , t2 ) plane,

w1 t1 + w2 t2 − c (11.8)

(We can pick up the c term by, for instance, allowing a 1 input to the
network, in addition to the predictors X (1) and X (2) .) The neuron receiving
these inputs fires if and only if we are on one particular side of the line. We
have as many lines as there are neurons in the first hidden layer.
422 CHAPTER 11. SEMI-LINEAR METHODS

Let’s do a simple “What if?” analysis. Consider a toy example with p = 2

and two classes, in which our regression function is

P [class 1 | (X (1) , X (2) ) = (t1 , t2 )] = µ(t1 , t2 ) =

{
1, if 0.4 < t1 < 0.6 and 0.4 < t2 < 0.6
(11.9)
0, otherwise

So, ideally we should predict class 1 if (X (1) , X (2) ) falls in this little square
in the center of the space, and predict class 0 otherwise.
If we knew the above — again, this is a “What if?” analysis — we could
in principle predict perfectly with a three-layer network, as follows. (This
will be an outline; the reader is invited to fill in the details.)
We would have three inputs, X (1) , X (2) and 1, to the left layer, and we
would have two neurons in the rightmost layer, for our two classes. Fol-
lowing up on our geometric view above, note that for instance, for the
constraint

0.4 < t1 (11.10)

the weights for X (1) , X (2) and 1 would be 1, 0 and -0.4. From (11.9), we
see we need eight such lines, thus eight neurons in the middle layer.
The weights coming out of the second layer into the top neuron of the
rightmost layer would all be 1/8, so that that neuron would fire if and only
if all of the eight constraints in (11.9) are satisfied. The lower neuron would
do the opposite.
The word all above is key. It basically says that we have formed an “AND”
network. But what if in (11.9) there were two square regions in which
µ(t) = 1, rather than just one? Then we sould need to effect an “OR”
operation. We could do this by having two AND nets, with a second hidden
layer playing the role of OR.
Finally, note that any general µ(t) could be approximated by having many
little squares like this, with the value of µ(t) now being more general than
just 0 and 1. We could still use a four-layer AND-OR network, with some
modification to account for µ(t) now having values between 0 and 1.
Now, let’s come down to Earth a bit. The above assumes that µ(t) is
known, which generally is not the case; it must be estimated from our data,
11.3. THE VERDICT 423

typically with p (and n) very large. It is hoped that, since the weights are
computed by least-squares fits to our data, plus iterative techniques such
as back propagation — adjusting our earlier iterates by feedback from the
prediction accuracy of the final layer — eventually “it all comes out in the
wash.” Hopefully in the end we obtain a network that works well. Of
course, there are no guarantees.
Note too that this shows that in principle we need only two hidden layers
(actually even one, by modifying this analysis), no matter how many pre-
dictor variables we have. However, in practice, we may find it easier to use
more.
You can see that the above intuition could be the basis for a proof of
statistical consistency. Some researchers have developed complicated tech-
nical conditions under which NNs can be shown to yield statistical consis-
tency [69]. which says that, given enough neurons, NNs can approximate
any smooth regression function. Showing statistical consistency then be-
comes a matter of determining how fast the number of neurons can grow
with n. The Stone-Weierstrass Theorem (Section 1.16.4), which states that
we approximate any continuous regression function by polynomials, is used
in some of this theory.

11.3 The Verdict

The reader, having come this far in the book, is now armed with a number
of techniques, linear/nonlinear and parametric/nonparametric. Which is
best? There is no good answer to this, and though many research papers
or books will say something like “Method A is better in such-and-such
settings, while Method B is better in some other situations, etc.”, the reader
is advised to retain a healthy skepticism.
SVMs and NNs were developed in the machine learning community, and
have attracted much attention in the press. These methods, especially
NNs, have generated some highly impressive example applications [85], but
they have also generated controversy [15]. There has been concern that the
science fiction-like names of the methods are overinterpreted as implying
that these methods somehow have special powers. As remarked in [64]:

There has been a great deal of hype surrounding neural net-

works, making them seem magical and mysterious. As we make
clear in this section, they are just nonlinear statistical models...
424 CHAPTER 11. SEMI-LINEAR METHODS

Again, none of these methods, or any other, is guaranteed to work.

There are always challenges that the analyst must face. But these are good
tools available to try. Furthemore, keep in mind that SVM and NN, which
make use of linear functions, may be more powerful, since in many applied
problems there µ(t) is monotonic in the ti .

11.4 Mathematical Complements

11.4.1 Edge Bias in Nonparametric Regression

It is important to keep in mind that k-NN and other nonparametric methods

are subject to bias near the boundaries of our data. It will be easier to
explain this in the density estimation context (Section 5.10.1.1). We’ll use
(5.30), which for convenience is duplicated here:

#(t − h, t + h)
fb(t) = (11.11)
2hn

Let R denote the variable whose density is of interest. Suppose the true
population density is fR (t) = 4t3 for t in (0,1), 0 elsewhere. The quantity
in the numerator has a binomial distribution with n trials and probability
of success per trial

∫ t+h
p = P (t − h < R < t + h) = 4u3 du = (t + h)4 − (t − h)4 = 8t3 h + 8th3
t−h
(11.12)
By the binomial property, the numerator of (11.11) has expected value np,
and thus

np
E[fc
R (t)] = = 4t3 + 4th2 (11.13)
2nh

Subtracting fR (t), we have

bias[fc
R (t)] = 4th
2
(11.14)

So, the smaller we set h, the smaller the bias, consistent with intuition. But
note too the source of the bias: Since the density is increasing, we are likely
11.4. MATHEMATICAL COMPLEMENTS 425

to have more neighbors on the right side of t than the left, thus biasing our
density estimate upward.
How about the variance? Again using the binomial property, the variance
of the numerator of (11.11) is np(1 − p), so that

np(1 − p) np 1 − p 1−p
V ar[[fc
R (t)] = = · = (4t3 + 4th2 ) · (11.15)
(2nh)2 2nh 2nh 2nh

This matches intuition too: On the one hand, for fixed h, the larger n is, the
smaller the variance of our estimator — i.e., larger samples are better, as
expected. On the other hand, the smaller we set h, the larger the variance,
because with small h there just won’t be many Ri falling into our interval
(t − h, t + h).
So, you can really see the bias-variance tradeoff here, in terms of what value
we choose for h.
The nonparametric regression case is similar. For p = 1, the numerator of
(11.11) now becomes the sum of all Yi for which Xi is in (t − h, t + h). The
expected value of the numerator is now

E(Y |t − h < X < t + h) (11.16)

The relevant density is then the conditional density of X, given X is in

(t−h, t+h). We can then compute the bias and variance as above (Exercise
7).

11.4.2 Dual Formulation for SVM

It can be shown that the vector w of coefficients of ℓb can be computed in

the separable case as follows.

Find w and a scalar b that minimize

1
||w||2 (11.17)
2

subject to
426 CHAPTER 11. SEMI-LINEAR METHODS

Yi (w′ Xi + b) ≥ 1, i = 1, ..., n (11.18)

The intuition is this: Look at Figure 11.6. The separating line is mathe-
matically

w′ t + b = 0 (11.19)

with the value b chosen so that this is the case. Thus we are on one side of
the line if

w′ t + b > 0 (11.20)

and are on the other side if

w′ t + b < 0 (11.21)

There will be two supporting hyperplanes on the edges of the two convex
hulls. In Figure 11.6, one of those hyperplanes is the line through SV11
and SV12, and the other is the parallel line through SV01. Recall that the
margin is the distance between the two supporting hyperplanes. We want
to maximize the distance between them, that is, separate the classes as
much as possible. Simple geometric calculation shows that the margin is
equal to 2/||w||2 . We want to maximize the margin, thus minimize (11.17).
Recall that Yi = ±1. So, in the class having all Yi = 1, we will want our
prediction to be at least 1, i.e.,

w′ Xi + b ≥ c, i = 1, ..., n (11.22)

with equality for the support vectors, while in the other class we will want

w′ Xi + b ≤ −c, i = 1, ..., n (11.23)

The scaling of w is defined so that c = 1.

Again since Yi = ±1, (11.18) neatly captures both cases.
In the general, nonseparable case, for cost C (not numerically equal to c
above), the problem becomes:
11.4. MATHEMATICAL COMPLEMENTS 427

Find w, a scalar b, and nonnegative quantities ξi that minimize

1 ∑
||w||2 + C ξi (11.24)
2

subject to

Yi (w′ Xi + b) ≥ 1 − ξi , i = 1, ..., n (11.25)

There is an equivalent formulation (in mathematical terminology, the dual

problem):7

Find quantities αi that minimize

1 ∑∑ ∑
n n n
Yi Yj αi αj Xi′ Xj − αi (11.26)
2 i=1 j=1 i=1

such that

∑
n
Yi αi = 0 (11.27)
i=1

and 0 ≤ αi ≤ C. Then

∑
n
w= αi Yi Xi (11.28)
i=1

and

b = Yj − Xj′ w (11.29)
7 For
readers with background in Lagrange multipliers, that is the technique used here.
The variables αi are the Lagrange variables.
428 CHAPTER 11. SEMI-LINEAR METHODS

for any j having αj > 0, i.e., any support vector.

The guess for Y for a new observation with predictor vector X
is

( )
∑
n
sign αi Yi Xi′ X +b (11.30)
i=1

11.4.3 The Kernel Trick

Take the case p = 2, for instance. We would replace X = (X1 , X2 )′ by

h(X), where

h(t1 , t2 ) = (t1 , t21 , t2 , t22 , t1 t2 )′ (11.31)

as our new predictor vector. Then instead of minimizing

1 ∑∑ ∑
m m n
Yi Yj αi αj Xi′ Xj − αi (11.32)
2 i=1 j=1 i=1

we would minimize

1 ∑∑ ∑
m m n
Yi Yj αi αj h(Xi )′ h(Xj ) − αi (11.33)
2 i=1 j=1 i=1

In SVM, one can reduce computational cost by changing this a bit, mini-
mizing

1 ∑∑ ∑
m m n
Yi Yj αi αj K(Xi , Xj ) − αi (11.34)
2 i=1 j=1 i=1

where the function K() is a kernel, meaning that it must satisfy certain
mathematical properties, basically that it is an inner product in some space.
A few such functions have been found to be useful and are incorporated
into SVM software packages. One of them is

K(u, v) = (1 + u′ v)2 (11.35)

11.5. FURTHER READING 429

This expression is quadratic in the ui and vi , so it achieves our goal of a

second-degree polynomial transformation as posed above. And, we could
take the third power in (11.35) to get a cubic transformation and so on.
Another kernel often found in software packages for SVM is the radial basis
kernel,

K(w, x) = e−γ||w−x||
2
(11.36)

11.5 Further Reading

Much theoretical (though nonstatistical) work has been done on SVMs. See
for instance [38] [36].
For a statistical view of neural networks, see [120].

11.6 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. In the discussion of Figure 6.1, it was noted that ordinary k-NN is
biased at the edges of the data, and that this might be remedied by use of
local-linear smoothing, a topic treated here in this chapter, Section 11.1.
Re-run Figure 6.1 using local-linear smoothing, and comment on what
changes, if any, emerge.
2. Write code to generate nreps samples, say 10000, in the example in
Section 11.1, and compute the bias for ordinary vs. local-linear k-NN at r
equally-spaced points in (0,1). In your experiment, vary k, n and r.
Mini-CRAN and other computational problems:
3. Write a function with call form
l o c q u a d ( predpt , nearxy , prodterms=FALSE)

analogous to regtools’ loclin(), except that it is locally quadratic instead

of locally linear. The function will add squared terms for the predictors,
and if prodterms is TRUE, then the interactions will be added as well.
430 CHAPTER 11. SEMI-LINEAR METHODS

Math problems:
4. Given points p1 , ..., pm in k-dimensional space Rk , let CHc denote the
reduced convex hull for cost c. For any c ≥ 1, this becomes the original
convex hull of the points. Show that CHc is indeed a convex set.
5. Following up on the discussion in Section 11.2.2.4, show by construction
of the actual weights that any second-degree polynomial function can be
reproduced exactly.
6. Show that the kernel in (11.35) is indeed an inner product, i.e.,

K(u, v) = q(u)′ q(v) (11.37)

for

q((s, t)′ ) = (as2 , bt2 , cst, ds, et, 1)′ (11.38)

with some suitable a, b, c.

7. Find the bias and variance in the regression example in Section 11.4.1,
taking in that example µ(s) = s2 .
Chapter 12

Regression and
Classification in Big Data

Gone are the old days of textbook examples consisting of n = 25 obser-

vation. Automatic data collection by large companies has led to data sets
on the terabyte scale or more. This has enabled applications of previ-
ously unimaginable power, some noble such as geographical prediction of
epidemics and crime incidents, others that create disturbing invasions of
privacy. For good or bad, though, Big Data is here to stay.
All this draws renewed attention on the predictive methods covered in this
book, but with a new set of problems in the case of Big Data:

• Computational: Typically any of the methods in this book will have

very long run times, hours, days or even more in some cases. Equally
important, the data may not fit into available memory, and thus be-
come impossible to run.

• Statistical: In many applications we have p >> n, many more pre-

dictors than observations. This is common in genomics problems, for
instance. Since we need p < n to avoid indeterminacy in a linear
model, and need p << n to avoid overfitting, this presents a real
challenge.

This chapter is devoted to addressing such issues. As in [102], we will

distinguish here betwen Big-n and Big-p.

431
432 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

12.1 Solving the Big-n Problem

Consider the case of n large, p moderate. Here there is a simple solution to

the computational problem, which I call Software Alchemy. (See [103] for
the various authors who have independently developed this idea, and for
details on the method.)

12.1.1 Software Alchemy

Let’s illustrate this notion with lm(). The procedure is as follows:

• Break the rows of the data matrix into chunks, say r of them.

• Apply the statistical procedure, say lm(), to each chunk.

• Average the r βb vectors to obtain the overall β.

b

Since each chunk is much smaller than the full data set, the run time for
each chunk is smaller as well. And since the chunks are run in parallel,
a substantial speedup can be attained. An analysis of the types of appli-
cations in which speedups are possible is presented in the Mathematical
Complements section at the end of this chapter.
It is easy to prove that this does work. One does not obtain the same value
of βb that would be computed from the entire data set, but the result is just
as good — the chunked and the full estimator have the same asymptotic
distribution.
It is easy to apply this to nonparametric models. With k-NN for instance,
we would compute µ b(t) on each chunk, and average the resulting values. For
nonparametric classification methods that normally return just a predicted
class rather than computing µ b(t), say CART, we can “vote” among the r
predicted classes, taking our guessed class to be the one that gets the most
votes, as with AVA. Similarly, with dimension reduction via PCA or NMF,
we cannot average the bases, but we can average the µ b(t) values or vote
among the predicted classes.
Software Alchemy is implemented in the partools package [96].1 The main
work is done in the function cabase(). There are various wrappers for
that function, such as calm(), which applies Software Alchemy to lm().2
1 Version 1.1.5 or higher is needed below.
2 The prefix ‘ca’ stands for “chunk averaging.”
12.1. SOLVING THE BIG-N PROBLEM 433

Some details on the innards of partools are given in the Computational

Complements section at the end this chapter.

12.1.2 Example: Flight Delay Data

Yet another famous dataset is that of airline flight data [6]. There are
records for all U.S. flights between 1987 and 2008, with the focus on arrival
and departure delay, ArrDelay and DepDelay.
In order to facilitate efforts by readers to replicate our analysis here, we will
focus on just one year, 2008, which is already “big enough,” over 7 million
records.
> library ( p a r t o o l s )
> c l s <− makeCluster ( 1 6 )
> setclsinfo ( cls )
> y2008 <− read . csv ( ’ y2008 ’ , h e a d e r=TRUE)
> mnthnames <−
c ( ’ Jan ’ , ’ Feb ’ , ’ Mar ’ , ’ Apr ’ , ’May ’ , ’ Jun ’ ,
’ J u l ’ , ’ Aug ’ , ’ Sep ’ , ’ Oct ’ , ’ Nov ’ , ’ Dec ’ )
> mnth <− mnthnames [ y2008$Month ]
> daynames <−
c ( ’ Sun ’ , ’Mon ’ , ’ Tue ’ , ’Wed ’ , ’Thu ’ , ’ F r i ’ , ’ Sat ’ )
> day <− daynames [ y2008$DayOfWeek ]
> y2008$Month <− as . factor ( mnth )
> y2008$DayOfWeek <− as . factor ( day )
> system . time ( calmout <− calm ( c l s , ’ ArrDelay ∼
DepDelay+D i s t a n c e+TaxiOut+U n i q u e C a r r i e r+Month+
DayOfWeek , data=y2008 ’ ) )
u s e r system e l a p s e d
40.788 3.748 50.040
> system . time ( lmout <− lm( ArrDelay ∼
DepDelay+D i s t a n c e+TaxiOut+U n i q u e C a r r i e r+Month+
DayOfWeek , data=y2008 ) )
u s e r system e l a p s e d
74.720 2.508 77.376
> d i s t r i b s p l i t ( c l s , ’ y2008 ’ , s c r a m b l e=TRUE)

Here we regress arrival delay against departure delay, distance and so on

(UniqueCarrier is the airline, e.g., DL for Delta). Note that we needed
to convert the month and day variables from numeric to factor, to avoid
having, say, March, count 3 times as much as January, which would be
meaningless.
434 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

We set up a (virtual) cluster of 16 nodes, which in this case run on 16 cores

(the machine had 32, counting hyperthreading); we could run on 16 nodes
of a physical cluster and so on. Each node is running its own invocation of
R, in communication with the invocation running on the parent node.
We split the dataset y2008 across the 16 nodes, so that each node had about
1/16th of the data. With much less data to work with, this potentially
enables speedier computation, since the nodes run in parallel. Though
we cannot expect to attain a 16-fold speedup, for reasons given in the
Computational Complements section of this chapter, we did get a speedup
of about 60%.3
Note the argument scramble = TRUE in the call to distribsplit().
Many datasets are ordered on one or more of the variables, in this case
ordered by date. A straightforward splitting of the data would mean that
some nodes may receive data for only one month, making it impossible to
compute a Month coefficient in the regression. (Also, Software Alchemy
assumes i.i.d. data.) The random permutation solves that problem.
So, how does the Software Alchemy output compare with that of the anal-
ysis on the full data?
> calmout $ t h t
( Intercept ) DepDelay Distance
−16.230510414 0.991774602 −0.002382778
TaxiOut UniqueCarrierAA UniqueCarrierAQ
0.826768548 4.516806436 8.953951702
UniqueCarrierAS UniqueCarrierB6 UniqueCarrierCO
4.030485591 −0.367642225 0.194950670
UniqueCarrierDL UniqueCarrierEV UniqueCarrierF9
1.581873640 2.056932472 6.766798981
UniqueCarrierFL UniqueCarrierHA UniqueCarrierMQ
5.670985961 10.692474278 3.469432309
UniqueCarrierNW UniqueCarrierOH UniqueCarrierOO
3.449247970 −0.072744312 3.842304091
UniqueCarrierUA UniqueCarrierUS UniqueCarrierWN
1.681950224 0.441686744 4.140866177
UniqueCarrierXE UniqueCarrierYV MonthAug
2.791769787 2.928448428 −1.400718871
MonthDec MonthFeb MonthJan
−0.038449264 −0.017799483 −0.404931144
MonthJul MonthJun MonthMar
3 There was overhead in splitting the data. As noted in Section 12.4.1, the real

advantage of partools is accrued when many operations are done on the distributed
data after a split.
12.1. SOLVING THE BIG-N PROBLEM 435

−0.951371174 0.275108880 −0.107499646

MonthMay MonthNov MonthOct
−0.019260538 −1.767687711 −1.510315535
MonthSep DayOfWeekMon DayOfWeekSat
−1.534679083 0.581085330 0.320872459
DayOfWeekSun DayOfWeekThu DayOfWeekTue
0.289664672 0.673834948 0.356366605
DayOfWeekWed
0.581757662

> coef ( lmout )

( Intercept ) DepDelay Distance
−16.230600841 0.991777979 −0.002382658
TaxiOut UniqueCarrierAA UniqueCarrierAQ
0.826780859 4.516897691 8.957282637
UniqueCarrierAS UniqueCarrierB6 UniqueCarrierCO
4.030860520 −0.368104197 0.195027499
UniqueCarrierDL UniqueCarrierEV UniqueCarrierF9
1.581651876 2.057249536 6.766480218
UniqueCarrierFL UniqueCarrierHA UniqueCarrierMQ
5.671174474 10.691920438 3.469761618
UniqueCarrierNW UniqueCarrierOH UniqueCarrierOO
3.448967869 −0.072846826 3.842374665
UniqueCarrierUA UniqueCarrierUS UniqueCarrierWN
1.681752229 0.441697940 4.141035514
UniqueCarrierXE UniqueCarrierYV MonthAug
2.791406582 2.928769069 −1.400844315
MonthDec MonthFeb MonthJan
−0.038074477 −0.017720576 −0.405148865
MonthJul MonthJun MonthMar
−0.951546924 0.274937723 −0.107709304
MonthMay MonthNov MonthOct
−0.019137396 −1.767617225 −1.510315831
MonthSep DayOfWeekMon DayOfWeekSat
−1.534714190 0.581024609 0.320610432
DayOfWeekSun DayOfWeekThu DayOfWeekTue
0.289584205 0.673469157 0.356256017
DayOfWeekWed
0.581504814

The results are essentially identical. Note very carefully again, though, that
the chunked estimator has the same asymptotic variance as the original one,
so any discrepancies that may occur between them should not be interpreted
436 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

as meaning that one is better than the other. In other words, we saved some
computation time here with results of equal quality.
Let’s try k-NN on this data. Continuing from our computation above, we
have:
> y2008mini <− y2008 [ , c ( 1 5 , 1 6 , 1 9 , 2 1 ) ]
# can ’ t have NAs
> y2008mini <− y2008mini [ complete . c a s e s ( y2008mini ) , ]
> d i s t r i b s p l i t ( c l s , ’ y2008mini ’ , s c r a m b l e=TRUE)
> system . time ( caknn ( c l s , ’ y2008mini [ , 1 ] ’ , 5 0 ,
xname= ’ y2008mini [ , − 1 ] ’ ) )
u s e r system e l a p s e d
35.552 3.792 113.514
> library ( r e g t o o l s )
> system . time ( xd <− p r e p r o c e s s x ( y2008mini [ , − 1 ] , 5 0 )
u s e r system e l a p s e d
303.516 4.400 308.068
> system . time ( kout <− k n n e s t ( y2008mini [ , 1 ] , xd , 5 0 ) )
u s e r system e l a p s e d
701.832 37.720 740.036

The speedup here in the fitting stage is large, 113.514 seconds vs. 308.068
+ 740.036, almost 10-fold.
Note that caknn() has the call form
caknn ( c l s , yname , k , xname = ’ ’ )

The optional argument indicates whether preprocessx() has already been

called at the cluster nodes. A nonblank value indicates that this is not the
case, and contains the expression for the desired matrix/data frame of X
values.

12.1.3 More on the Insufficient Memory Issue

As noted earlier, use of parallel computation in large problems is not just

a matter of reducing run time, but is also in many cases a solution to lack
of sufficient memory. Indeed, the problem may be too large for R’s address
space.
Software Alchemy can solve that problem, by breaking the task into chunks,
each of which can fit into memory. This would be feasible if the task is run
12.1. SOLVING THE BIG-N PROBLEM 437

on a cluster of machines, rather than one multicore one. In the multicore

case, we could run the chunks one at a time.
A subtle problem may lurk here, though. Software Alchemy relies on the
Central Limit Theorem, in multivariate form, and the convergence rate
depends on the dimension [19]. So, if p/n is not small, Software Alchemy
may not be appropriate. How, then, can we do computation on the full
data set while circumventing the memory problem?
For a linear or generalized linear model, a solution is available in packages
such as biglm [94], and the package biganalytics [46], which provides a
wrapper to biglm. (The biganalytics package may also provide a solution
to memory space problems.)
The partools package includes some file-based functions, such as file-
split(), counterpart to distribsplit(), another way to avoid memory re-
strictions.

12.1.4 Deceivingly “Big” n

Even with millions of observations, n may not be large enough to estimate

what we want. Consider the airline data, for instance. Suppose there is
some small, obsolescent airport that is rarely used, so rarely that there are
only, say, 5 records in the dataset involving this airport. Then clearly we
have insufficient data to be able to say anything about this airport. If for
example we have as a predictor a dummy variable for this airport and use a
linear model, the resulting estimated coefficient will have a large standard
error, despite the huge n.

12.1.5 The Independence Assumption in Big-n Data

Many instances of Big Data arise with multiple observations on the same
unit, say the same person. This raises issues with assumptions of statistical
independence, the core of most statistical methods.
However, with Big Data, people are generally not interested in statistical
inference, since the standard errors will typically be tiny (though not in
the situations discussed in the last section). Since in parametric models
the assumption of independent observations enters in mainly in deriving
standard errors, lack of independence is typically not an issue.
438 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

12.2 Addressing Big-p

We now turn to the case of large p, including but not limited to p >> n.
Note the subdued term in the title here, addressing rather than solving as
in Section 12.1. Fully satisfactory methods have not been developed for
this situation.

12.2.1 How Many Is Too Many?

We know that overfitting can be a problem. In terms of our number of

observation n and number of variable p, we risk overfitting if p/n is not
small. But what is “small”? — not an easy question at all. This question
is the subject of dozens of methods developed over the decades, some of
which will be discussed in this chapter. None of them completely settles
the issue, but the point here is that that issue cannot be addressed without
understanding the dynamics underlying it.
From earlier discussions in this book, the reader is familiar with the notion
of examining what happens as the number of data points n goes to infinity,
as for instance in the Central Limit Theorem, whose implication was that
the estimated coefficient vector βb in a linear model has an approximately
multivariate normal distribution, important for forming confidence intervals
and finding p-values.
But in the context of variable selection, we have an additional aspect to
consider, as seen in our example in Section (9.1.1). The latter suggests
that the larger n is, the more predictors p our model can handle well. For
this reason, researchers in this field typically not only model n as going
to infinity, but also have p do so. A question then arises as to whether p
should go to infinity more slowly than n does.
To nonmathematicians, this concept may seem odd, maybe even surreal.
After all, in practice one just has a certain sample size and a certain number
of predictors to choose from, so why care about those quantities going to
infinity? This is a valid concern, but the researchers’ findings are useful
in a “cultural” way, in the sense that the results give us a general feeling
about the interplay between p and n. So, in this section we will take a brief
look at the nature of the results researchers have obtained.
12.2. ADDRESSING BIG-P 439

12.2.1.1 Toy Model

As a warmup, consider a very simple example, in the fixed-X setting with

an orthogonal design (Section 2.3).
Again, in the fixed-X case, the X values are chosen beforehand by the
experimenter. We will assume that she has done so in a manner in which
none of the values is very big, which in our simple model here we will take
to mean that

∑
n
(j)
(Xi )2 = n, j = 1, ..., p (12.1)
i=1

We will assume a linear model with no intercept term,

∑
n
µ(t) = βj tj (12.2)
j=1

In our setting here of an orthogonal design, the factor A′ A in (2.28) will

have 0s for its off-diagonal elements, with the j th diagonal entry being

∑
n
(j)
(Xi )2 (12.3)
i=1

which by our assumption has the value n. Then (2.28) has the easy closed-
form solution

∑n (j)
Xi Yi
βbj = i=1
, j = 1, ..., p (12.4)
n

We’ll also assume homoscedasticity, which the reader will recall means

V ar(Y | X = (t1 , ..., tp )) = σ 2 (12.5)

regardless of the values of the tj . (All calculations here will be conditional

on the X values, since they are fixed constants chosen by the experimenter,
but we will not explicitly write the | symbol.) Thus we have

(j) (j)
V ar(Xi Yi ) = (Xi )2 σ 2 (12.6)
440 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

and thus from (12.4), we see that

1
V ar(βbj ) = σ 2 , j = 1, ..., p (12.7)
n

Now finally, say we wish to estimate µ(1, ..., 1), i.e., enable prediction for
the case where all the predictors have the value 1. Our estimate will be

∑
p
βbj · 1 (12.8)
j=1

and its variance will then be

pσ 2
V ar[b
µ(1, , ..., 1)] = (12.9)
n

Now, what does this have to do with the question, “How many predictors
is too many?” The point is that since we want our estimator to be more
and more accurate as the sample size grows, we need (12.9) to go to 0 as
n goes to infinity. We see that even if at the same time we have more and
more predictors p, the variance will go to 0 as long as

p
→0 (12.10)
n

Alas, this still doesn’t fully answer the question of how many predictors we
can afford to use for the specific value of n we have in our particular data
set. But it does give us some insight, in that we see that the variance of
our estimator is being inflated by a factor of p if we use p predictors. This
is a warning not to use too many of them, and the simple model shows that
we will have “too many” if p/n is large. This is the line of thought taken
by theoretical research in the field.

12.2.1.2 Results from the Research Literature

The orthogonal nature of the design in the example in the last section is
not very general. Let’s see what theoretical research has yielded.
Stephen Portnoy of the University of Illinois proved some results for general
Maximum Likelihood Estimators [115], not just in the regression/classifi-
cation context, though restricted to distributions in an exponential family
12.2. ADDRESSING BIG-P 441

(Section 4.2.1). We won’t dwell on the technical definition of that term

here, but just note again that it includes many of the famous distribution
families, such as normal, exponential, gamma and Poisson. Portnoy proved
that we obtain consistent estimates as long as

p
√ →0 (12.11)
n

This is more conservative than what we found in the last section, but much
more broadly applicable.
We might thus
√ take as a rough rule of thumb that we are not overfitting as
long as p < n. Strictly speaking, this really doesn’t follow from Portnoy’s
result, but some analysts use this as a rough guideline. As noted earlier,
this was a recommendation of the late John Tukey, one of the pioneers of
modern statistics.
There is theoretical work on p, n consistency of LASSO estimators, CART
and so on, but they are too complex to discuss here. The interested reader
is referred to, for instance, [28] and [66].

12.2.1.3 A Much Simpler and More Direct Approach

The discussion in the last couple of subsections at least shows that the key
to the overfitting issue is the size of p relative to n. The question is where
to draw the dividing line between overfitting and a “safe” value of p. The
answer to this question has never been settled, but we do have one very
handy measure that is pretty reasonable in the case of linear models: the
adjusted R2 value. If p < n, this will provide some rough guidance.
As you will recall (Section 2.9.4), ordinary R2 is biased upward, due to
overfitting (even if only slight). Adjusted R2 is a much less biased version of
R2 , so a substantial discrepancy between the ordinary and adjusted versions
of R2 may be a good indication of overfitting. Note, though, biases in almost
any quantity can become severe in adaptive methods such as the various
stepwise techniques described in Chapter 9.

12.2.1.4 Nonparametric Case

The problems occur with nonparametric methods too.

k-NN:
442 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

With p predictor variables, think of breaking the predictor space into little
cubes of dimension p, each side of length d. This is similar to a k-NN
setting. In estimating µ(t) for a particular point t, consider the density
of X there, f (t). For small d, the density will be approximately constant
throughout the cube. Then the probability of an observation falling into
the cube is about dp f (t), where the volume of the cube is dp . Thus the
expected number of observations in the cube is approximately

ndp f (t) (12.12)

Now, think of what happens as n → ∞. On one hand, we need the bias of

b(t) to go to 0, so we need d → 0. But on the other hand, we also need the
µ
b(t) to go to 0. So we need the number of observations in the
variance of µ
cube to go to infinity, which from (12.12) means

log n + p log d → ∞ (12.13)

as n → ∞.
Since d will be smaller than 1, its log is negative, so for fixed d in (12.13),
the larger p is, the slower (12.13) will go to infinity. If p is also going to
infinity, it will need to do so more slowly than log n. This informal result
is consistent with that of [135].
As noted, the implications of the asymptotic analysis for the particular
values of n and p in the data we have at hand are unclear.
√ But comparison
to the parametric case above, with p growing like n, does suggest that
nonparametric methods are less forgiving of a large value of p than are
parametric methods.
There is also the related issue of the Curse of Dimensionality, to be dis-
cussed shortly.
NNs:
Say we have ℓ layers, with d nodes per layer. Then from one layer to the
next, we will need d2 weights, thus ℓd2 weights in all. Since the weights are
calculated using least squares (albeit in complicated ways), we might think
of this as a situation with

p = ℓd2 (12.14)

Somewhat recklessly applying the Tukey/Portnoy finding, that in the para-

12.2. ADDRESSING BIG-P 443

√
metric case we need p grow more slowly than n, that would mean that,
say, d should grow more slowly than n1/4 for fixed ℓ.

12.2.1.5 The Curse of Dimensionality

The term Curse of Dimensionality (COD) was introduced by Richard Bell-

man, back in 1953 [13]. Roughly speaking it says that the difficulty of
estimating a statistical model increases exponentially with the dimension
of the problem. In our regression/classification context, that would imply
that if our number of predictors p is very large, we may need an enormous
number of data points n to develop an effective model. This is especially
troubling in today’s era of Big Data, in which p may be on the order of
hundreds, thousands or even more.
One popular interpretation of the COD is that in very high dimensions the
data is quite sparse. With k-NN, for instance, one can show that with high
values of p, neighboring observations tend to be far from a given one [18],
say a point at which prediction is to be done. The neighbors thus tend to
be “unrepresentative” of the given point, especially at the edges of the data
space. This arguably is a cause of what we found earlier about the need for
p to grow slowly when we use nonparametric regression methods.
Perhaps this could be ameliorated using linear smoothing, as in Section
11.1. The general effect should be similar, though.

12.2.2 Example: Currency Data

Here we return to the currency data (Section 6.5). A favorite investigatory

approach by researchers of p >> n situations is to add noise variables to a
real data set, and then see whether a given method correctly ignores those
variables.
Let’s try that here, starting with the LASSO. The dataset has n = 761,
and we will add 1500 noise variables:
> c u r r <− read . table ( ’EXC. ASC ’ , h e a d e r=TRUE)
> u <− matrix (rnorm( 7 6 1 ∗ 1 5 0 0 ) ,nrow=761)
> c u r r u <− cbind ( c u r r , u )
> library ( l a r s )
> l a r s o u t <−
l a r s ( as . matrix ( c u r r u [ , − 5 ] ) , c u r r u [ , 5 ] ,
n o r m a l i z e=FALSE, u s e . Gram=FALSE)
> larsout
444 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

Sequence o f LASSO moves :

Mark Franc 88 855 1308 1466 310 611 159 59
Var 2 3 92 859 1312 1470 314 615 163 63
Step 1 2 3 4 5 6 7 8 9 10
...

This is not a good start! The algorithm has already gone 10 steps, yet has
not added in the variables for the Canadian dollar and the pound, while
adding in 8 of the noise variables. Trying the same thing with stepwise
regression, calling lars() with the argument type = ’stepwise’, produces
similar results.
Let’s try CART as well:
> library ( rpart )
> r p a r t o u t <− r p a r t ( Yen ∼ . , data=c u r r u )
> rpartout
n= 761

node ) , s p l i t , n , deviance , y v a l
∗ d e n o t e s t e r m i n a l node

1 ) r o o t 761 2 0 5 2 1 9 1 . 0 0 2 2 4 . 9 4 5 1
2 ) Mark< 2 . 2 7 5 2 356 6 2 6 2 0 3 . 0 0 1 8 5 . 3 6 3 1
4 ) Franc >=5.32 170 1 0 5 1 1 2 . 1 0 1 4 7 . 9 4 6 0
8 ) Mark< 2 . 2 0 6 3 5 148 25880.71 140.4329 ∗
9 ) Mark>=2.20635 22 14676.99 198.4886 ∗
5 ) Franc< 5 . 3 2 186 65550.84 219.5616 ∗
3 ) Mark>=2.2752 405 3 7 7 9 5 6 . 0 0 2 5 9 . 7 3 8 1
6 ) Canada >=1.08 242 68360.23 237.7370
1 2 ) Canada >=1.38845 14 5393.71 199.2500 ∗
1 3 ) Canada< 1 . 3 8 8 4 5 228 40955.69 240.1002 ∗
7 ) Canada< 1 . 0 8 163 18541.41 292.4025 ∗

Much better. None of the noise variables was selected. One disappointment
is that the pound was not chosen. Running the analysis with ctree() did
pick up all four currencies, and again did not choose any noise variables.

12.2.3 Example: Quiz Documents

I give weekly quizzes in my courses. Here I have 143 such documents, in

LaTeX, for five courses. We will try to classify documents into their courses,
based on frequencies of various words.
12.2. ADDRESSING BIG-P 445

There should be a considerable amount of overlap between the courses,

making classification something of a challenge. For example, two of the
courses, one undergraduate and the other graduate-level, are statistical in
nature, thus possibly difficult to distinguish from each other. Among the
other three, two are in the computer systems area, thus with common termi-
nology such as cache, causing the same problem in distinguishing between
them. The fifth is on scripting languages (Python and R).
There are quite a few R packages available for text mining. The tm package
provides an extensive set of utilities, such as removal of stop words, very
common words such as a and the that usually have no predictive power. I
used this package (details are presented in Section 12.4.2), and I removed
the LaTeX keywords using the detex utility.
I ended up with my term-document matrix nmtd, with n = 143 rows and
p = 4670 columns, the latter figure reflecting the fact that 4670 distinct
words were found in the entire document set. So, we certainly have p >> n.
The element in row i, column j is the number of occurrences of word j in
document i. (Various other formulations are possible.)
The distribution of classes was as follows:
> table ( l a b e l s ) /sum( table ( l a b e l s ) )
labels
ECS132 ECS145 ECS158 ECS256
0.41958042 0.10489510 0.25874126 0.02797203
ECS50
0.18881119

Once again, we will run our packages naively, just using default values.
Here are the results for SVM:
> l i b r a r y (tm)
> l i b r a r y ( SnowballC ) # s u p p l e m e n t t o tm
> # tm o p e r a t i o n s not shown h e r e
> nmtd <− as . matrix ( nmtd )
> # ’ labels ’ is the vector of class l a b e l s
> d f w h o l e <−
as . data . frame ( cbind ( labels , as . data . frame ( nmtd ) ) )
> # c r o s s −v a l i d a t i o n
> t r a i n i d x s <− sample ( 1 : 1 4 3 , 7 2 )
> d f t r n <− d f w h o l e [ t r a i n i d x s , ]
> d f t s t <− d f w h o l e [− t r a i n i d x s , ]
> l i b r a r y ( e1071 )
> svmout <− svm ( l a b e l s ∼ . , data=d f t r n )
446 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

> ypredsvm <− predict ( svmout , d f t s t [ , − 1 ] )

[ 1 ] 0.5352113

Without the word information, we would always guess the course ECS 132,
and would be right about 42% of the time, as seen above. So, a 54% rate
does represent substantial improvement.
Let’s try CART:
> library ( rpart )
> r p a r t o u t <−
r p a r t ( l a b e l s ∼ . , data=d f t r n , method= ’ c l a s s ’ )
> y p r e d r p a r t <−
predict ( r p a r t o u t , d f t s t [ , − 1 ] , type= ’ c l a s s ’ )
> mean( y p r e d r p a r t == d f t s t [ , 1 ] )
[ 1 ] 0.5774648

About 58%, pretty good.

12.2.4 The Verdict

In Chapter 9, it was repeatedly emphasized that there is no good solution

to the problem of selecting predictor variables. That statement holds even
more strongly for the situation p >> n.
In particular, no matter which method we use, the Principle of Rare Events
(Section 7.6.1), for p large enough, there is a high probability that some
predictor variable with little or no relation to our response variable will be
chosen by the method.
It is instructive to consider two conflicting statements by prominent re-
searchers. On the one hand, Hastie et al [65, p. 86] remark (citing Chapter
3 of [64]),

Forward stepwise methods...are hard to beat in terms of finding

good, sparse sets of variables.

On the other hand, Gelman [55] writes, in a blog post provocatively titled,
“Why We Hate Stepwise Regression,”

To address the issue more directly: the motivation behind step-

wise regression is that you have a lot of potential predictors but
12.3. MATHEMATICAL COMPLEMENTS 447

not enough data to estimate their coefficients in any meaningful

way. This sort of problem comes up all the time, for example
heres an example from my research, a meta-analysis of the ef-
fects of incentives in sample surveys.
The trouble with stepwise regression is that, at any given step,
the model is fit using unconstrained least squares. I prefer meth-
ods such as factor analysis or lasso that group or constrain the
coefficient estimates in some way.

Though Hastie et al likely have in mind a more sophisticated use of stepwise

regression than Gelman, the latter’s s last statement is a little ironic, in view
of the leading role that Hastie et al have played in the development of the
LASSO and related estimation methods.
It must be stated one more time, then, that the analyst needs to keep in
mind that all these techniques are merely tools available for possible use.
One should also use them in conjunction with cross-validation, though even
then one must heed the warning in Section 9.3.2.
These are powerful methods, to be sure, but should be used with care.

12.3 Mathematical Complements

12.3.1 Speedup from Software Alchemy

(Readers who do not have previous background in “big oh” notation should
review Section 5.10.2 before continuing.)
Some back-of-the envelope analysis illustrates what kinds of applications
can benefit greatly from Software Alchemy (SA), and to what kinds SA
might bring only modest benefits.
Consider statistical methods needing time O(nc ). For instance, matrix
multiplication, say with both factors having size n × n, has an O(n3 ) time
complexity.
If we have r processes, say running on r cores of a multicore machine,
then SA assigns about n/r data points to each process, each with run time
O((n/r)c ). Since the processes run independently in parallel, SA would
reduce the run time from O(nc ) to O((n/r)c ) = O(nc /rc ), a speedup of rc .
The larger the exponent c is, the greater the speedup.
Moreover, suppose we run the r chunks sequentially, i.e., one at a time,
448 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

rather than in parallel. We may do this, for instance, because we cannot

fit more than one chunk in memory at a time. Then instead of O((n/r)c )
time, we would have O(r(n/r)c ). That would yield a speedup of O(rc−1 ),
so even in this case SA gives us faster runs if c > 1.
On the other hand, consider a linear regression computation. In the first
phase, we need O(np2 ) time to compute A′ A, and then in the second phase
we must compute (A′ A)−1 or equivalent. A straightforward computation
for the latter takes O(p3 ) time, or possibly O(p2 ) time with QR, depending
on what information is desired. Alas, though SA helps in the first phase,
it is of no help in that second phase. Thus SA will typically yield limited
benefits here.

12.4 Computational Complements

12.4.1 The partools Package

This package consists of a number of functions for parallel computation and

data wrangling. Similar to Hadoop and Spark, it is centered on the notion
of distributed files and data. In the file case, for instance, this means that
a nominal file x is actually composed of many files, say x.001, x.002 and
so on. But unlike Hadoop and Spark, partools does not use MapReduce,
which is very constraining, e.g., due to requiring a sort operation whether
needed or not.
The partools package is itself based on R’s parallel package. With that
package one creates a virtual cluster, running an invocation of R on each
node. The virtual cluster could be a real cluster of machines, the cores on
a multicore machine, or a combination of both. The original R process, the
manager, invokes tasks at the created R processes, the workers, and collects
the results.
Here is an example, involving finding the largest row sum of a matrix:

> library ( p a r a l l e l )
> c l s <− makeCluster ( 2 )
> x <− matrix ( runif ( 1 0 0 ) , nrow=10)
> maxsum <− function (m) max( apply (m, 1 ,sum) )
# d i s t r i b u t e t h e t o p and bottom h a l v e s o f x t o t h e
# nodes , # have them each c a l l maxsum on t h e i r h a l v e s ,
# and c o l l e c t t h e r e s u l t s
> tmp <−
12.4. COMPUTATIONAL COMPLEMENTS 449

c l u s t e r A p p l y ( c l s , l i s t ( x [ 1 : 5 , ] , x [ 6 : 1 0 , ] ) , maxsum)
> tmp
[[1]]
[ 1 ] 6.052948
[[2]]
[ 1 ] 5.906015
> Reduce (max, tmp )
[ 1 ] 6.052948
# check
> maxsum( x )
[ 1 ] 6.052948

Of course, for this to pay off, one must (a) have much larger scale and (b)
use the distributed data set repeatedly, in various operations.
For more information on parallel computation, see [101].

12.4.2 Use of the tm Package

Here is the code I used with the tm package on the quiz documents:
li b r a r y (tm)
li b r a r y ( SnowballC )
nm <− Corpus ( D i r S o u r c e ( ’ MyQuizTexts ’ ) )
nm <− tm map(nm, t o l o w e r )
nm <− tm map(nm, removePunctuation )
nm <− tm map(nm, removeNumbers )
nm <− tm map(nm, removeWords , s to p w or d s ( ” e n g l i s h ” ) )
nm <− tm map(nm, stemDocument , l a n g u a g e = ” e n g l i s h ” )
nm <− tm map(nm, PlainTextDocument )
nmtd <− DocumentTermMatrix (nm)
nmtd <− as . matrix ( nmtd )

The call to DirSource() expects to see a number of documents in the spec-

ified directory, where I had my quizzes. There are various other methods
in tm to input data as well.
The function tolower() changes all letters to lower-case, and removePunc-
tuation() and removeNumbers() are self-explanatory. As noted earlier,
we also wish to remove stop words such as “a” and “the.”
Finally, DocumentTermMatrix() creates the matrix, though in tm()’s
form, hence the call to as.matrix().
450 CHAPTER 12. REGRESSION/CLASSIFICATION IN BIG DATA

12.5 Exercises: Data, Code and Math Prob-

lems

Data problems:
1. The dataset used in the example in Section 12.2.3 is available in the
regtools package. Try alternative analyses using the LASSO, say from the
package glmnet and NMF.
2. Analyze the airline dataset as to the possible difference among the
various destination airports, regarding arrival delay.
3. Download the New York City taxi data (or possibly just get one of
the files), http://www.andresmh.com/nyctaxitrips/. Predict trip time from
other variables of your choice, but instead of using an ordinary linear model,
fit median regression, using the quantreg package (Section 6.9.1). Com-
pare run times of Software Alchemy vs. a direct fit to the full data.
Mini-CRAN and other computational problems:
4. In the example in Section 12.2.3, we performed various text operations
such as removal of stop words, but could have gone further. One way to do
this would be to restrict the analysis only to the most frequently appearing
words, rather than say all 4670 in the example.
Write an R function with call form
c u l l w o r d s ( td , howmany=0.80)

aimed in this direction. Here td is the term-document matrix, which was

nmtd in our example. The argument howmany specifies how many words
we want to choose. If it is a value less than 1.0, the meaning is that we
wish to extract the top 80% of the words in terms of frequency. If instead it
is an integer, say 500, it means extract the 500 most-frequent words. Have
the function return the column numbers in td of the extracted words.
5. Write an R function, svplot(), modeled after pwplot() in the regtools
package (Section 6.13), with call form
s v p l o t ( svmout , pairs=combn ( length ( svmout$terms −1) , 2 ) )

that will plot the support vectors, two predictors at a time.

Appendix A

Matrix Algebra

This book assumes the reader has background in matrix algebra, or is

willing to learn the material as part of reading this book. This appendix
is intended as a review for the former group, or a quick treatment for the
latter group. Even the latter group may find some of the topics new.
It is recommended that both groups of readers skim through this appendix
for now, and then consult it for details later as the need arises.

A.1 Terminology and Notation

A matrix is a rectangular array of numbers. A vector is a matrix with

only one row (a row vector or only one column (a column vector).
The expression, “the (i,j) element of a matrix A,” will mean its element in
row i, column j, denoted Aij .
If A is a square matrix, i.e., one with equal numbers n of rows and columns,
then its diagonal elements are aii , i = 1,...,n.
A square matrix is called upper-triangular if aij = 0 whenever i > j,
with a corresponding definition for lower-triangular matrices.
The norm (or length) of an n-element vector X is

v
u n
u∑
||X||2 = t x2i (A.1)
i=1

451
452 APPENDIX A. MATRIX REVIEW

The norm is thus nonnegative.

This is actually one of many possible norms, called the Euclidean norm.
Another commonly used norm is the l1 norm,

∑
n
||X||1 = |x|i (A.2)
i=1

A.2 Matrix Addition and Multiplication

• For two matrices having the same numbers of rows and same numbers
of columns, addition is defined elementwise, e.g.,

     
1 5 6 2 7 7
 0 3 + 0 1 = 0 4  (A.3)
4 8 4 0 8 8

• Multiplication of a matrix by a scalar, i.e., a number, is also defined

elementwise, e.g.,

   
7 7 2.8 2.8
0.4  0 4  =  0 1.6  (A.4)
8 8 3.2 3.2

• The inner product or dot product of equal-length vectors X and

Y is defined to be

∑
n
xk yk (A.5)
k=1

• The product of matrices A and B is defined if the number of rows

of B equals the number of columns of A (A and B are said to be
conformable). In that case, the (i,j) element of the product C is
defined to be

∑
n
cij = aik bkj (A.6)
k=1
A.3. MATRIX TRANSPOSE 453

For instance,
   
7 6 ( ) 19 66
 0 4  1 6
=  8 16  (A.7)
2 4
8 8 24 80

It is helpful to visualize cij as the inner product of row i of A and

column j of B, e.g., as shown in bold face here:
   
7 6 ( ) 19 66
 0 1 6
4  =  8 16  (A.8)
2 4
8 8 24 80

• Matrix multiplication is associative and distributive, but in general

not commutative:

A(BC) = (AB)C (A.9)

A(B + C) = AB + AC (A.10)

AB ̸= BA (A.11)

A.3 Matrix Transpose

• The transpose of a matrix A, denoted A′ or AT , is obtained by ex-

changing the rows and columns of A, e.g.,
 ′
7 70 ( )
 8 16  = 7 8 8
(A.12)
70 16 80
8 80

• If A + B is defined, then

(A + B)′ = A′ + B ′ (A.13)

• If A and B are conformable, then

(AB)′ = B ′ A′ (A.14)
454 APPENDIX A. MATRIX REVIEW

• A square matrix A is said to be symmetric if A′ = A.

• Note that for a vector x, written as a one-column matrix,

∑
n
x′ x =∥ x ∥1 = |x|i (A.15)
i=1

A.4 Linear Independence

Vectors X1 ,...,Xk are said to be linearly independent if it is impossible

for

a1 X1 + ... + ak Xk = 0 (A.16)

unless all the ai are 0.

By the way, an expression of the form on the left side of (A.16) is called a
linear combination of the Xi .

A.5 Matrix Inverse

• The identity matrix I of size n has 1s in all of its diagonal elements

but 0s in all off-diagonal elements. It has the property that AI = A
and IA = A whenever those products are defined.
• If A is a square matrix and AB = I, then B is said to be the inverse
of A, denoted A−1 . Then BA = I will hold as well.
• A−1 exists if and only if its rows (or columns) are linearly indepen-
dent.
• A−1 exists if and only if det(A) ̸= 0.
• If A and B are square, conformable and invertible, then AB is also
invertible, and

(AB)−1 = B −1 A−1 (A.17)

• A matrix U is said to be orthogonal if its rows each have norm 1

and are orthogonal to each other, i.e., their inner product is 0. U thus
has the property that U U ′ = I, i.e., U −1 = U .
A.6. EIGENVALUES AND EIGENVECTORS 455

A special case that is used in this book to create simple examples of various
phenomena is the inverse of a 2 × 2 matrix:

( )−1 ( )
a b 1 d −b
= (A.18)
c d ad − bc −c a

providing ad − because ̸= 0. The reader should verify that the product of

the two matrices is indeed the 2 × 2 identity matrix.
Typically one does not compute matrix inverses directly. A common al-
ternative is the QR decomposition: For a matrix A, matrices Q and R
are calculated so that A = QR, where Q is an orthogonal matrix and R is
upper-triangular.
If A is square and invertible, A−1 is easily found:

A−1 = (QR)−1 = R−1 Q′ (A.19)

Again, though, in some cases A is part of a more complex system, and the
inverse is not explicitly computed.

A.6 Eigenvalues and Eigenvectors

Let A be a square matrix.1

• A scalar λ and a nonzero vector X that satisfy

AX = λX (A.20)

are called an eigenvalue and eigenvector of A, respectively.

• If A is symmetric and real, then it is diagonalizable, i.e., there exists
an orthogonal matrix U such that

U ′ AU = D (A.21)

for a diagonal matrix D. The elements of D are the eigenvalues of A,

and the columns of U are the eigenvectors of A (scaled to have length
1).
1 For
nonsquare matrices, the discussion here would generalize to the topic of singular
value decomposition.
456 APPENDIX A. MATRIX REVIEW

A different sufficient condition for A.21 is that the eigenvalues of A

are distinct. In this case, U will not necessarily be orthogonal.
By the way, this latter sufficient condition shows that “most” square
matrices are diagonalizable, if we treat their entries as continuous ran-
dom variables. Under such a circumstance, the probability of having
repeated eigenvalues would be 0.

A.7 Rank of a Matrix

Definition: The rank of a matrix A is the maximal number of linearly

independent columns in A.
Rank has the following properties:

• rank(A′ ) = rank(A)

• Thus the rank of A is also the maximal number of linearly independent

rows in A.

• Let A be r × s. Then

rank(A) ≤ min(r, s) (A.22)

• If A is n×n, it is invertible if and only if it has full rank, i.e. rank(A) =

n.

• The rank of a square matrix is equal to the number of nonzero eigen-

values.

A.8 Matrices of the Form B’B

Let B be any rectangular matrix, and write C = B ′ B. Then C has the

following properties:

• C is symmetric (follows from (A.14)).

• The rank of C is that of B.

A.9. PARTITIONED MATRICES 457

• C is nonnegative definite: For any conformable vector x,

x′ Cx ≥ 0 (A.23)

(follows from writing x′ Cx = (x′ B ′ )(Bx) and noting that the latter
is the squared norm of Bx, thus nonnegative).

A.9 Partitioned Matrices

In certain senses, we can treat submatrices in a matrix almost like scalars,

through a process called partitioning. This turns out to be very useful.
For example, let

 
1 5 12
A= 0 3 6  (A.24)
4 8 2

and

 
0 2 5
B= 0 9 10  , (A.25)
1 1 2

so that

 
12 59 79
C = AB =  6 33 42  . (A.26)
2 82 104

We could partition A as

( )
A00 A01
A= , (A.27)
A10 A11

where

( )
1 5
A00 = , (A.28)
0 3
458 APPENDIX A. MATRIX REVIEW

( )
12
A01 = , (A.29)
6

( )
A10 = 4 8 (A.30)

and

( )
A11 = 2 . (A.31)

Similarly we would partition B and C into blocks of a compatible size to A,

( )
B00 B01
B= (A.32)
B10 B11

and

( )
C00 C01
C= , (A.33)
C10 C11

so that for example

( )
B10 = 1 1 . (A.34)

The key point is that multiplication still works if we pretend that those
submatrices are numbers! For example, pretending like that would give the
relation

C00 = A00 B00 + A01 B10 , (A.35)

which the reader should verify really is correct as matrices, i.e. the compu-
tation on the right side really does yield a matrix equal to C00 .

A.10 Matrix Derivatives

There is an entire body of formulas for taking derivatives of matrix-valued

expressions. One of particular importance to us is for the vector of deriva-
A.11. MATRIX ALGEBRA IN R 459

tives

dg(u)
(A.36)
du

for a vector u of length k. This is the gradient of g(u), i.e. the vector

∂g(u) ∂g(u) ′
( , ..., ) (A.37)
∂u1 ∂uk

A bit of calculus shows that the gradient can be represented compactly. in

some cases, such as

d
(M u + w) = M ′ (A.38)
du

for a matrix M and vector w that do not depend on u. The reader should
verify this by looking at the individual ∂g(u)
∂ui .

Another example is the quadratic form

d ′
u Qu = 2Qu (A.39)
du

for a symmetric matrix Q and a vector u.

And there is a Chain Rule. For example if u = M v + w, then

∂ ′
u u = 2M ′ u (A.40)
∂v

A.11 Matrix Algebra in R

The R programming language has extensive facilities for matrix algebra,

introduced here. Note by the way that R uses column-major order.
A linear algebra vector can be formed as an R vector, or as a one-row or
one-column matrix.
> # constructing matrices
> a <− rbind ( 1 : 3 , 1 0 : 1 2 )
> a
460 APPENDIX A. MATRIX REVIEW

[ ,1] [ ,2] [ ,3]

[1 ,] 1 2 3
[2 ,] 10 11 12
> b <− matrix ( 1 : 9 , ncol =3)
> b
[ ,1] [ ,2] [ ,3]
[1 ,] 1 4 7
[2 ,] 2 5 8
[3 ,] 3 6 9
# multiplication , etc .
> c <− a %∗% b ; c + matrix ( c ( 1 , − 1 , 0 , 0 , 3 , 8 ) ,nrow=2)
[ ,1] [ ,2] [ ,3]
[1 ,] 15 32 53
[2 ,] 67 167 274
> c %∗% c ( 1 , 5 , 6 ) # n o t e 2 d i f f e r e n t c ’ s
[ ,1]
[ 1 , ] 474
> # transpose , inverse
> t (a) # transpose
[ ,1] [ ,2]
[1 ,] 1 10
[2 ,] 2 11
[3 ,] 3 12
> u <− matrix ( runif ( 9 ) , nrow=3)
> u
[ ,1] [ ,2] [ ,3]
[ 1 , ] 0.08446154 0.86335270 0.6962092
[ 2 , ] 0.31174324 0.35352138 0.7310355
[ 3 , ] 0.56182226 0.02375487 0.2950227
> uinv <− solve ( u )
> uinv
[ ,1] [ ,2] [ ,3]
[ 1 , ] 0 . 5 8 1 8 4 8 2 −1.594123 2 . 5 7 6 9 9 5
[ 2 , ] 2 . 1 3 3 3 9 6 5 −2.451237 1 . 0 3 9 4 1 5
[ 3 , ] −1.2798127 3 . 2 3 3 1 1 5 −1.601586
> u %∗% uinv # n o t e r o u n d o f f e r r o r
[ ,1] [ ,2] [ ,3]
[ 1 , ] 1 . 0 0 0 0 0 0 e+00 −1.680513 e −16 −2.283330 e −16
[ 2 , ] 6 . 6 5 1 5 8 0 e −17 1 . 0 0 0 0 0 0 e+00 4 . 4 1 2 7 0 3 e −17
[ 3 , ] 2 . 2 8 7 6 6 7 e −17 −3.539920 e −17 1 . 0 0 0 0 0 0 e+00
> # e i g e n v a l u e s and e i g e n v e c t o r s
> eigen ( u )
$values
A.11. MATRIX ALGEBRA IN R 461

[ 1 ] 1 . 2 4 5 6 2 2 0 + 0 . 0 0 0 0 0 0 0 i −0.2563082+0.2329172 i
−0.2563082 −0.2329172 i

$vectors
[ ,1] [ ,2]
[ ,3]
[ 1 , ] −0.6901599+0 i −0.6537478+0.0000000 i
−0.6537478+0.0000000 i
[ 2 , ] −0.5874584+0 i −0.1989163 −0.3827132 i
−0.1989163+0.3827132 i
[ 3 , ] −0.4225778+0 i 0 . 5 6 6 6 5 7 9 + 0 . 2 5 5 8 8 2 0 i
0.5666579 −0.2558820 i
> # d i a g o n a l m a t r i c e s ( o f f −d i a g o n a l s 0)
> diag ( 3 )
[ ,1] [ ,2] [ ,3]
[1 ,] 1 0 0
[2 ,] 0 1 0
[3 ,] 0 0 1
> diag ( ( c ( 5 , 1 2 , 1 3 ) ) )
[ ,1] [ ,2] [ ,3]
[1 ,] 5 0 0
[2 ,] 0 12 0
[3 ,] 0 0 13

We can obtain matrix inverse using solve(), e.g.,

> m <− rbind ( 1 : 2 , 3 : 4 )
> m
[ ,1] [ ,2]
[1 ,] 1 2
[2 ,] 3 4
> minv <− solve (m)
> minv
[ ,1] [ ,2]
[ 1 , ] −2.0 1 . 0
[2 ,] 1 . 5 −0.5
> m %∗% minv # s h o u l d g e t I b a c k
[ ,1] [ ,2]
[1 ,] 1 1 . 1 1 0 2 2 3 e −16
[2 ,] 0 1 . 0 0 0 0 0 0 e+00

Note the roundoff error, even with this small matrix. We can try the QR
method, provided to us in R via qr(). In fact, if we just want the inverse,
462 APPENDIX A. MATRIX REVIEW

qr.solve() will compute (A.19) for us.

We can in principle obtain rank from, for example, the rank component
from the output of qr(). Note however that although rank is clearly defined
in theory, the presence of roundoff error in computation may make rank
difficult to determine reliably.

A.12 Further Reading

A number of excellent, statistics-oriented books are available on these top-

ics. Among them are [62] [56] [7] on the mathematics, and [134] for usage
in R.
Bibliography

[1] Aggarwal, C. Recommender Systems: The Textbook. Springer

International Publishing, 2016.

[2] Agresti, A. Categorical Data Analysis. Wiley Series in Probability

and Statistics. Wiley, 2014.

[3] Agresti, A. Foundations of Linear and Generalized Linear Models.

Wiley Series in Probability and Statistics. Wiley, 2015.

[4] Alfons, A. cvtools: Cross-validation tools for regression

models. https://cran.r-project.org/web/packages/cvTools/
index.html.

[5] Ambler, G., et al. mfp: Multivariable fractional polynomials.

https://cran.r-project.org/web/packages/mpf/index.html.

[6] ASA. Data expo ’09. http://stat-computing.org/dataexpo/2009/the-

data.html, 2009.

[7] Banerjee, S., and Roy, A. Linear Algebra and Matrix Analysis
for Statistics. Chapman & Hall/CRC Texts in Statistical Science.
Taylor & Francis, 2014.

[8] Bartlett, P. L., and Traskin, M. Adaboost is consistent. Jour-

nal of Machine Learning Research 8 (2007), 2347–2368.

[9] Bates, D., et al. lme4: Linear mixed-effects models using

’eigen’ and s4. https://cran.r-project.org/web/packages/lme4/
index.html.

[10] Bates, D., and Maechler, M. Matrix: Sparse and dense

matrix classes and methods. https://cran.r-project.org/web/
packages/Matrix/index.html.

463
464 BIBLIOGRAPHY

[11] Baty, F., et al. nlstools: Tools for nonlinear regression

analysis. https://cran.r-project.org/web/packages/nlstools/
index.html.
[12] Becker, R. The variance drain and Jensen’s inequality. AEPR
Working Paper 2012-004, Indiana University, 2010.
[13] Bellman, R. Computational Problems in the Theory of Dynamic
Programming. Rand paper series. Rand Corporation, 1953.
[14] Bendel, R. B., and Afifi, A. A. Comparison of stopping rules
in forward stepwise regression. Journal of the American Statistical
Association 72, 357 (1977), 46–53.
[15] Bennett, K. P., and Campbell, C. Support vector machines:
hype or hallelujah? ACM SIGKDD Explorations Newsletter 2, 2
(2000), 1–13.
[16] Berk, R. Statistical Learning from a Regression Perspective.
Springer Texts in Statistics. Springer International Publishing, 2016.
[17] Berk, R., Brown, L., Buja, A., Zhang, K., and Zhao, L. Valid
post-selection inference. Ann. Statist. 41, 2 (04 2013), 802–837.
[18] Beyer, K. S., Goldstein, J., Ramakrishnan, R., and Shaft,
U. When is ”nearest neighbor” meaningful? In Proceedings of the 7th
International Conference on Database Theory (London, UK, 1999),
ICDT ’99, Springer-Verlag, pp. 217–235.
[19] Bhattacharya, R., and Holmes, S. An exposition of Götze’s
estimation of the rate of convergence in the multivariate central limit
theorem. arXiv 1003.4254, 2010.
[20] Bickel, P. J., Hammel, E. A., and O’Connell, J. W. Sex Bias
in Graduate Admissions: Data from Berkeley. Science 187, 4175
(1975), 398–404.
[21] Braun, W., and Murdoch, D. A First Course in Statistical Pro-
gramming with R. Cambridge University Press, 2016.
[22] Breiman, L. Random forests. Machine Learning 45, 1 (2001), 5–32.
[23] Breiman, L. Statistical modeling: The two cultures (with comments
and a rejoinder by the author). Statist. Sci. 16, 3 (08 2001), 199–231.
[24] Breiman, L., Friedman, J., Stone, C., and Olshen, R. Clas-
sification and Regression Trees. The Wadsworth and Brooks-Cole
statistics-probability series. Taylor & Francis, 1984.
BIBLIOGRAPHY 465

[25] Bretz, F., Hothorn, T., and Westfall, P. Multiple Compar-

isons Using R. CRC Press, 2016.

[26] Brown, C. dummies: Create dummy/indicator variables flexi-

bly and efficiently. https://cran.r-project.org/web/packages/
dummies/index.html.

[27] Buhlmann, P. Statistical significance in high-dimensional linear

models. Bernoulli 19, 4 (09 2013), 1212–1242.

[28] Bühlmann, P., and van de Geer, S. Statistics for High-

Dimensional Data: Methods, Theory and Applications. Springer Se-
ries in Statistics. Springer Berlin Heidelberg, 2011.

[29] Burgette, L., et al. twang: Toolkit for weighting and anal-
ysis of nonequivalent groups. https://cran.r-project.org/web/
packages/twang/index.html.

[30] Candes, E., Fan, Y., Janson, L., and Lv, J. Panning for gold:
Model-free knockoffs for high-dimensional controlled variable selec-
tion. arXiv 1610.02351, 2016.

[31] Card, D. Estimating the Return to Schooling: Progress on Some

Persistent Econometric Problems. Econometrica 69, 5 (2001), 1127–
1160.

[32] Carroll, R., and Ruppert, D. Transformation and Weighting

in Regression. Chapman & Hall/CRC Monographs on Statistics &
Applied Probability. Taylor & Francis, 1988.

[33] Chang, C.-C., and Lin, C.-J. Libsvm: A library for support vector
machines. ACM Trans. Intell. Syst. Technol. 2, 3 (May 2011), 27:1–
27:27.

[34] Chen, S. X., Qin, J., and Tang, C. Y. Mann-Whitney test with
adjustments to pretreatment variables for missing values and obser-
vational study. Journal of the Royal Statistical Society: Series B
(Statistical Methodology) 75, 1 (2013), 81–102.

[35] Christensen, R. Log-linear models and logistic regression, 2nd ed.

Springer texts in statistics. Springer, New York, c1997. Earlier ed.
published under title: Log-linear models. 1990.

[36] Clarke, B., Fokoue, E., and Zhang, H. Principles and Theory
for Data Mining and Machine Learning. Springer Series in Statistics.
Springer New York, 2009.
466 BIBLIOGRAPHY

[37] Cleveland, W. S., and Loader, C. Smoothing by Local Regres-

sion: Principles and Methods. Physica-Verlag HD, Heidelberg, 1996,
pp. 10–49.
[38] Cristianini, N., and Shawe-Taylor, J. An Introduction to Sup-
port Vector Machines: And Other Kernel-based Learning Methods.
Cambridge University Press, New York, NY, USA, 2000.
[39] Davison, A., and Hinkley, D. Bootstrap Methods and Their Ap-
plication. Cambridge Series in Statistical and Probabilistic Mathe-
matics. Cambridge University Press, 1997.
[40] Deng, H., and Wickham, H. Densty estimation in R.
http://vita.had.co.nz/papers/density-estimation.pdf, 2011.
[41] Dobson, A., and Barnett, A. An Introduction to Generalized
Linear Models, Third Edition. Chapman & Hall/CRC Texts in Sta-
tistical Science. CRC Press, 2008.
[42] Efron, B., Hastie, T., Johnstone, I., and Tibshirani, R. Least
angle regression. Ann. Statist. 32, 2 (04 2004), 407–499.
[43] Eicker, F. Limit theorems for regressions with unequal and de-
pendent errors. In Proceedings of the Fifth Berkeley Symposium on
Mathematical Statistics and Probability, Volume 1: Statistics. Uni-
versity of California Press, Berkeley, Calif., 1967, pp. 59–82.
[44] Eldén, L. Matrix Methods in Data Mining and Pattern Recogni-
tion. Fundamentals of Algorithms. Society for Industrial and Applied
MathematicsS, 2007.
[45] Elzhov, T., et al. minpack.lm: R interface to the Levenberg-
Marquardt nonlinear least-squares algorithm found in minpack, plus
support for bounds. https://cran.r-project.org/web/packages/
minpack.lm/index.html.
[46] Emerson, J., and Kane, M. biganalytics: Utilities for ’big.matrix’
objects from package ’bigmemory’. https://cran.r-project.org/
web/packages/biganalytics/index.html.
[47] Faraway, J. Extending the Linear Model with R: Generalized Linear,
Mixed Effects and Nonparametric Regression Models. Chapman &
Hall/CRC Texts in Statistical Science. CRC Press, 2016.
[48] Fong, W. M., and Ouliaris, S. Spectral tests of the martingale
hypothesis for exchange rates. Journal of Applied Econometrics 10,
3 (1995), 255–71.
BIBLIOGRAPHY 467

[49] Fox, J., et al. car: Companion to applied regression. https:

//cran.r-project.org/web/packages/car/index.html.

[50] Fox, J., and Weisberg, S. An R Companion to Applied Regression.

SAGE Publications, 2011.

[51] Friedman, J. Another approach to polychotomous classification.

Tech. rep., Stanford University, 1996.

[52] Galecki, A., and Burzykowski, T. Linear Mixed-Effects Mod-

els Using R: A Step-by-Step Approach. Springer Texts in Statistics.
Springer New York, 2013.

[53] Galilli, T. Is it harder to advertise to the more educated? https:

//www.r-statistics.com/tag/data-set/, 2010.

[54] Gaujoux, R., and Seoighe, C. Nmf: Algorithms and frame-

work for nonnegative matrix factorization (nmf). https://cran.
r-project.org/web/packages/NMF/index.html.

[55] Gelman, A. Why we hate stepwise regression. https://www.

r-statistics.com/tag/data-set/, 2010.

[56] Gentle, J. Matrix Algebra: Theory, Computations, and Applica-

tions in Statistics. Springer Texts in Statistics. Springer New York,
2010.

[57] Gilbert, P., and Varadhan, R. numderiv: Accurate nu-

merical derivatives. https://cran.r-project.org/web/packages/
numDeriv/index.html.

[58] Giles, D. The adjusted r-squared, again.

http://davegiles.blogspot.com/2013/07/the-adjusted-r-squared-
again.html, 2013.

[59] Giles, J. A., and Giles, D. Pre-test estimation and testing in

econometrics: Recent developments. Journal of Economic Surveys 7,
2 (1993), 145–97.

[60] Goldenberg, A., Zheng, A., Fienberg, S., and Airoldi, E. A

Survey of Statistical Network Models. Foundations and Trends(r) in
Machine Learning. Now Publishers, 2010.

[61] Harrell, F. Regression Modeling Strategies: With Applications to

Linear Models, Logistic and Ordinal Regression, and Survival Anal-
ysis. Springer Series in Statistics. Springer International Publishing,
2015.
468 BIBLIOGRAPHY

[62] Harville, D. Matrix Algebra From a Statistician’s Perspective.

Springer New York, 2008.

[63] Hastie, T., and Efron, B. lars: Least angle regression, lasso and
forward stagewise? https://cran.r-project.org/web/packages/
lars/index.html.

[64] Hastie, T., Tibshirani, R., and Friedman, J. The Elements of

Statistical Learning. Springer Series in Statistics. Springer New York
Inc., New York, NY, USA, 2001.

[65] Hastie, T., Tibshirani, R., and Wainwright, M. Statistical

Learning with Sparsity: The Lasso and Generalizations. Chapman
& Hall/CRC Monographs on Statistics & Applied Probability. CRC
Press, 2015.

[66] Hastie, T., Tibshirani, R., and Wainwright, M. Statistical

Learning with Sparsity: The Lasso and Generalizations. Chapman &
Hall. Taylor & Francis, 2015.

[67] Ho, T. K. The random subspace method for constructing decision

forests. IEEE Transactions on Pattern Analysis and Machine Intel-
ligence 20, 8 (1998), 832–844.

[68] Hoerl, A. E., and Kennard, R. W. Ridge regression: Biased

estimation for nonorthogonal problems. Technometrics 42, 1 (Feb.
2000), 80–86.

[69] Hornik, K., Stinchcombe, M., and White, H. Multilayer feed-

forward networks are universal approximators. Neural Netw. 2, 5
(July 1989), 359–366.

[70] Hothorn, T., et al. mvtnorm: Multivariate normal and t distri-

butions. https://cran.r-project.org/web/packages/mvtnorm/
index.html.

[71] Hothorn, T., Hornik, K., and Zeileis, A. Unbiased recursive

partitioning: A conditional inference framework. Journal of Compu-
tational and Graphical Statistics 15, 3 (2006), 651–674.

[72] Hothorn, T., and Zeileis, A. partykit: A toolkit for recur-

sive partytioning. https://cran.r-project.org/web/packages/
partykit/index.html.

[73] Hsu, J. Multiple Comparisons: Theory and Methods. Taylor & Fran-
cis, 1996.
BIBLIOGRAPHY 469

[74] Inselberg, A. Parallel Coordinates: Visual Multidimensional Ge-

ometry and Its Applications. Advanced series in agricultural sciences.
Springer New York, 2009.

[75] Jackson, M. Social and Economic Networks. Princeton University

Press. Princeton University Press, 2010.

[76] James, G., Witten, D., Hastie, T., and Tibshirani, R. An

Introduction to Statistical Learning: with Applications in R. Springer
Texts in Statistics. Springer New York, 2013.

[77] Jiang, J. Linear and Generalized Linear Mixed Models and Their
Applications. Springer Series in Statistics. Springer, Dordrecht, 2007.

[78] Jordan, M. I. Leo Breiman. Ann. Appl. Stat. 4, 4 (12 2010), 1642–
1643.

[79] Kang, H., et al. ivmodel: Statistical inference and sensitivity anal-
ysis for instrumental variables model. https://cran.r-project.
org/web/packages/ivmodel/index.html.

[80] Kaye, D., and Freedman, D. Reference guide on statistics. https:

//www.nap.edu/read/13163/chapter/7.

[81] Koenker, R., et al. quantreg: Quantile regression. https://

cran.r-project.org/web/packages/quantreg/index.html.

[82] Kraemer, N., and Schaefer, J. parcor: Regularized estimation

of partial correlation matrices. https://cran.r-project.org/web/
packages/parcor/index.html.

[83] Kuhn, M., and Johnson, K. Applied Predictive Modeling. Springer

New York, 2013.

[84] Lange, K. Numerical Analysis for Statisticians. Springer, 2001.

[85] Le, Q., Ranzato, M., Monga, R., Devin, M., Chen, K., Cor-
rado, G., Dean, J., and Ng, A. Building high-level features using
large scale unsupervised learning. In International Conference in Ma-
chine Learning (2012).

[86] Lee, D. D., and Seung, H. S. Learning the parts of objects by

nonnegative matrix factorization. Nature 401 (1999), 788–791.

[87] Leisch, F., and Dimitriadou, E. mlbench: Machine learn-

ing benchmark problems. https://cran.r-project.org/web/
packages/mlbench/index.html.
470 BIBLIOGRAPHY

[88] Li, S., et al. Fnn: Fast nearest neighbor search algorithms and
applications. https://cran.r-project.org/web/packages/FNN/
index.html.

[89] Liaw, A. randomForest: Breiman and Cutler’s random forests for

classification and regression. https://cran.r-project.org/web/
packages/randomForest/index.html.

[90] Lin, C.-J. On the convergence of multiplicative update algorithms

for nonnegative matrix factorization. Trans. Neur. Netw. 18, 6 (Nov.
2007), 1589–1596.

[91] Little, R., and Rubin, D. Statistical Analysis with Missing Data.
Wiley Series in Probability and Statistics. Wiley, 2014.

[92] Loh, W.-Y. Classification and regression trees. Data Mining and
Knowledge Discovery 1 (2011), 14–23.

[93] Luke, D. A Users Guide to Network Analysis in R. Use R! Springer

International Publishing, 2015.

[94] Lumley, T. biglm: bounded memory linear and generalized linear

models. https://cran.r-project.org/web/packages/sandwich/
index.html.

[95] Lumley, T., and Zeileis, A. sandwich: Robust covariance

matrix estimators. https://cran.r-project.org/web/packages/
sandwich/index.html.

[96] Matloff, N. partools: Tools for the ’parallel’ package. https:

//cran.r-project.org/web/packages/partools/index.html.

[97] Matloff, N. regtools: Regression tools. https://cran.

r-project.org/web/packages/regtools/index.html.

[98] Matloff, N. Estimation of internet file-access/modification rates

from indirect data. ACM Trans. Model. Comput. Simul. 15, 3 (July
2005), 233–253.

[99] Matloff, N. From Algorithms to Z Scores: Prob-

abilistic and Statistical Modeling in Computer Science.
http://heather.cs.ucdavis.edu/probstatbook, 2010.

[100] Matloff, N. The Art of R Programming: A Tour of Statistical

Software Design. No Starch Press Series. No Starch Press, 2011.
BIBLIOGRAPHY 471

[101] Matloff, N. Parallel Computing for Data Science: With Examples

in R, C++ and CUDA. Chapman & Hall/CRC The R Series. CRC
Press, 2015.
[102] Matloff, N. Big-n and big-p in big data. In Handbook of Big Data,
P. Bühlmann, P. Drineas, M. Kane, and M. J. van der Laan, Eds.
CRC, 2016.
[103] Matloff, N. Software alchemy: Turning complex statistical compu-
tations into embrassingly-parallel ones. Journal of Statistical Software
71, 1 (2016).
[104] Matloff, N., and Xie, Y. freqparcoord: Novel methods for par-
allel coordinates. https://cran.r-project.org/web/packages/
freqparcoord/index.html.
[105] Matloff, N. S. Ergodicity conditions for a dissonant voting model.
Ann. Probab. 5, 3 (06 1977), 371–386.
[106] Matloff, N. S. Use of regression functions for improved estimation
of means. Biometrika 68, 3 (Dec. 1981), 685–689.
[107] Matloff, N. S. Use of covariates in randomized response settings.
Statistics & Probability Letters 2, 1 (1984), 31–34.
[108] McCann, L. Written Testimony of Laurie A. McCann, AARP Foun-
dation Litigation, EEOC. 2016.
[109] Milborrow, S. rpart.plot: Plot ’rpart’ models: An enhanced ver-
sion of ’plot.rpart’. https://cran.r-project.org/web/packages/
rpart.plot/index.html.
[110] Miller, A. Subset Selection in Regression. Chapman & Hall/CRC
Monographs on Statistics & Applied Probability. CRC Press, 2002.
[111] Miller, A. J. Selection of subsets of regression variables. Journal
of the Royal Statistical Society. Series A (General) (1984), 389–425.
[112] Mizon, G. Inferential procedures in nonlinear models: An applica-
tion in a UK industrial cross section study of factor substitution and
returns to scale. Econometrica 45, 5 (1977), 1221–42.
[113] Müller, U. U. Estimating linear functionals in nonlinear regression
with responses missing at random. Ann. Statist. 37, 5A (10 2009),
2245–2277.
[114] Murphy, K. Machine Learning: A Probabilistic Perspective. Adap-
tive computation and machine learning series. MIT Press, 2012.
472 BIBLIOGRAPHY

[115] Portnoy, S. Asymptotic behavior of m-estimators of p regression

parameters when p2 /n is large. Ann. Statist. 12, 4 (12 1984), 1298–
1309.

[116] Rao, C., and Toutenburg, H. Linear Models: Least Squares and
Alternatives. Springer, 1999.

[117] Rao, J., and Molina, I. Small Area Estimation. Wiley Series in
Survey Methodology. Wiley, 2015.

[118] Rao, R. B., and Fung, G. On the dangers of cross-validation. An

experimental evaluation. SIAM, pp. 588–596.

[119] Rifkin, R., and Klautau, A. In defense of one-vs-all classification.

J. Mach. Learn. Res. 5 (Dec. 2004), 101–141.

[120] Ripley, B. Pattern Recognition and Neural Networks. Cambridge

University Press, 2007.

[121] Ripley, B., and Venables, W. nnet: Feed-forward neu-

ral networks and multinomial log-linear models. https://cran.
r-project.org/web/packages/nnet/index.html.

[122] Rose, R. Nonparametric Estimation of Weights in Least-Squares Re-

gression Analysis. PhD thesis, University of California, Davis, 1978.

[123] Rudin, W. Principles of Mathematical Analysis. McGraw-Hill, 1964.

[124] Rufibach, K., and Duembgen, L. logcondens: Estimate a log-

concave probability density from iid observations. https://cran.
r-project.org/web/packages/logcondens/index.html.

[125] Sarkar, D. lattice: Trellis graphics for r. https://cran.

r-project.org/web/packages/lattice/index.html.

[126] Schoerker, B., et al. Ggally: Extension to ’ggplot2’. https:

//cran.r-project.org/web/packages/GGally/index.html.

[127] Scott, D. Multivariate Density Estimation: Theory, Practice, and

Visualization. Wiley Series in Probability and Statistics. Wiley, 2015.

[128] Shao, J. Linear model selection by cross-validation. J. of the Amer-

ican Statistical Association 88 (1993), 486–494.

[129] Shawe-Taylor, J., and Cristianini, N. Kernel Methods for Pat-

tern Analysis. Cambridge University Press, 2004.

[130] Stone, C. Private communication, 2016.

BIBLIOGRAPHY 473

[131] Therneau, T., et al. Introduction to recursive partitioning using

the rpart routines. https://cran.r-project.org/web/packages/
rpart/vignettes/longintro.pdf.

[132] Therneau, T., et al. rpart: Recursive partitioning and regression

trees. https://cran.r-project.org/web/packages/rpart/index.
html.
[133] Valentini, G., Dietterich, T. G., and Cristianini, N. Bias-
variance analysis of support vector machines for the development of
svm-based ensemble methods. Journal of Machine Learning Research
5 (2004), 725–775.
[134] Vinod, H. Hands-on Matrix Algebra Using R: Active and Motivated
Learning with Applications. World Scientific, 2011.
[135] Wasserman, L. All of Nonparametric Statistics. Springer Texts in
Statistics. Springer New York, 2006.
[136] Wasserman, L. All of Statistics: A Concise Course in Statistical
Inference. Springer Texts in Statistics. Springer New York, 2013.
[137] Zeileis, A. Object-oriented computation of sandwich estimators.
Journal of Statistical Software 16, 1 (2006), 1–16.
[138] Zou, H., Hastie, T., and Tibshirani, R. Sparse principal com-
ponent analysis. Journal of Computational and Graphical Statistics
15 (2004), 2006.
Index

A example (quiz docu-

ments), 444–446
Akaike’s Information Criterion how many is too many,
(AIC), 347 438–443
All vs. All (AVA), 183 nonparametric case, 441–
Amazon recommender system 3 443
results from research liter-
B ature, 440–441
simpler and more direct
Baseball player approach, 441
data, 22, 270–274 toy model, 439–440
description goal, 6–15 verdict, 446–447
Best Linear Unbiased Estimator computational complements,
(BLUE), 261 448–449
Big Data, regression and classifi- partools package, 448–449
cation in, 431–450 tm package, 449
Big-n, 432–437 exercises, 450
deceivingly “Big” n, 437 mathematical complements,
example (flight delay 447–448
data), 433–436 Bike sharing
independence assumption activity, predicting, 1–2
in Big-n data, 437 data, 32–38
insufficient memory issue, classification, 45–47
436–437 inference under ho-
Software Alchemy, 432– moscedasticity, 86–88
433 least-squares computa-
Big-p, 438–447 tion, 171–172
Curse of Dimensionality, linear modeling of µ(t),
443 33–38
example (currency data), nonparametric analysis,
443–444 38

475
476 INDEX

BLUE, see Best Linear Unbiased innards of the k-NN code,

Estimator 58–59
Bodyfat data linear regression models,
dimension reduction, 371– 115–120
374 model fit, 255–257
k-NN with linear smoothing, multiclass classification
409 problems, 207–211
stepwise variable selection, shrinkage estimators, 334–
352–357 336
Bonferroni inequality, 297–298 variable selection and dimen-
Boosting (model fit), 251–254 sion reduction, 383–386
Conditional expectation
C
geometry of, 108
Cauchy-Schwartz Inequality, 109 properties of, 52–54
Central Limit Theorem (CLT), 82 geometric view, 54
regression case 83 law of total expectation,
statement of, 104 53
Chi-square distributions, 308–309 law of total variance, 54
Classification and Regression random variable, condi-
Trees (CART), 392 tional expectation as,
Classification techniques, 44–47 52–53
example (bike-sharing data), Tower Property, 54, 269,
45–47 306
regression problem, 44–45 Confusion matrix, 202
Click-through rate (CTR), 3 CRAN packages, 55–56
CLT, see Central Limit Theorem Cross-validation, 26–31, 347–348
COD, see Curse of Dimensional- choosing the partition sizes,
ity 30–31
Coefficient of variation, 332 k-NN case, 29–30
Column vector, 451 linear model case, 28–29
Computational complements, 55–
CTR, see Click-through rate
60
Currency data
Big Data, 448–449
classification, 207–211 Big-p, 443–444
CRAN packages, 55–56 data wrangling, 255–256
disaggregating regressor ef- LASSO, 322–323
fects, 303–305 model fit assessment and im-
function dispatch, 59–60 provement, 233, 238
function tapply() and its multicollinearity, 315
cousins, 56–58 ridge regression, 319–320
GLM, 173–175 Curse of Dimensionality (COD),
homoscedasticity, 140–141 443
INDEX 477

D (“borrowing from neigh-

bors”), 291–294
Degrees of freedom, 308 verdict, 295
Delta Method, 141–142, 168 Simpson’s paradox, 274–278
Description goal, 6–15, 255 small analytical example,
Disaggregating regressor effects, 268–270
267–310 small area estimation, 268
computational complements, unobserved predictor vari-
303–305 ables, 278–289
example (baseball player example (movie ratings
data), 270–274 data), 287–288
exercises, 309–310 fixed effects, 287
mathematical complements, instrumental variables
306–309 (IVs), 279–286
asymptotic chi-square dis- mixed-effects model, 287
tributions, 308–309 multiple random effects,
degrees of freedom, 308 288
iterated projections, 306– random effects models,
307 286–288
standard errors for RFA, reason for using ran-
307–308 dom/mixed effects mod-
multiple inference, 295–303 els, 288–289
Bonferroni inequality, Dummy variables, 35
297–298
example (MovieLens E
data), 300–302
frequent occurrence of ex- Eigenvalues and eigenvectors,
treme events, 295–296 455–456
principle of rare events, Enzyme kinetics model, 148–150
296 Euclidean distance, 20
relation to statistical infer- Euclidean norm, 452
ence, 296–297 Exercises
Scheffe’s method, 298–300 Big Data, regression and
verdict, 303 classification in, 450
predictor effects principle, disaggregating regressor ef-
268 fects, 309–310
regression function averag- generalized linear and non-
ing, 289–295 linear models, 176–177
estimating the counterfac- homoscedasticity and other
tual, 290–291 assumptions in practice,
example (job training), 143–145
290–291 linear regression models,
small area estimation 120–122
478 INDEX

model fit assessment and im- definition, 150–151

provement, 264–265 exponential families, 152–
multiclass classification 153
problems, 211–213 link function, 151
partition-based methods,
overdispersion, 152
403–404
Poisson regression, 151–
semi-linear methods, 429–
152
430
shrinkage estimators, 336– R’s glm() function, 153–
337 154
variable selection and dimen- GLM (logistic model), 154–
sion reduction, 387–390 164
example (Pima diabetes
F data), 158–159
example (predicting spam
Fisher linear discriminant analy-
e-mail), 163–164
sis, 188–191
Fixed effects, 287 interpretation of coeffi-
Fixed-X setting, 67 cients, 159–161
Forward stepwise regression, 322 linear boundary, 164
FPR, see Rate of false positives motivation, 155–158
freqparcoord package, 229–232 overall prediction accu-
for dimension reduction, racy, 162
380–382 predict() function, 161
for outlier detection, 235–236
text classification, 163
G GLM (Poisson regression
model), 165–166
Gauss-Markov theorem, 261–263 least-squares computation,
Best Linear Unbiased Esti- 166–173
mator, 261 Eicker-White asymptotic
Lagrange multipliers, 261– standard errors, 168–
262 171
proof, 262–263
“elephant in the room”
Generalized linear and nonlinear
(convergence issues),
models, 147–177
172–173
computational complements,
example (bike sharing
173–175
data), 171–172
example (enzyme kinetics
model), 148–150 Gauss-Newton method,
exercises, 176–177 166–168
generalized linear model mathematical complements,
(GLM), 150–154, 329– 175
330 R factors, 174–175
INDEX 479

H example (MovieLens
data), 127–130
Hat matrix, 224, 257–259 inference on linear regres-
Holdout method, 222 sion coefficients, 126
Homoscedasticity, inference un- what can be done, 126–127
der (linear regression join operation, 140
models ), 81–88 mathematical complements,
example (bike-sharing data), 141–143
86–88 Delta Method, 141–142
normal distribution model, distortion due to transfor-
82–83 mation, 142–143
regression case, 83–86 normality assumption, 124–
review (classical inference on 125
single mean), 81
Slutsky’s Theorem, 82 I
standard error, concept of,
83 Identity matrix, 454
Homoscedasticity and other as- Indicator variables, 35
sumptions in practice, Instrumental variables (IVs),
123–145 279–286
computational complements, example (years of schooling),
140–141 284–285
dropping the homoscedastic- method, 281–282
ity assumption, 130–139 two-stage least squares, 283–
example (female wages), 284
136 verdict, 286
methodology, 135–136 Iteratively reweighted least
procedure for valid infer- squares, 174
ence, 135
robustness of the ho- J
moscedasticity assump-
tion, 131–133 James-Stein theory, 331–332
simulation test, 137 Join operation, 140
variance-stabilizing trans-
formations, 137–139 K
verdict, 139
weighted least squares, Kernel trick, 416, 428–429
133–135 k-fold cross-validation, 95
exercises, 143–145 k-Nearest Neighbor (k-NN), 21
independence assumption, code, innards of, 58–59
125–130 cross-validation, 29–30
estimation of a single on letter recognition data,
mean, 125–126 187–188
480 INDEX

with linear smoothing, 407– multiclass classification

409 problems, 186–187
LOOM for, 95 random forests, 401–402
nonparametric regression es- SVMs, 417
timation, 21 Leverage values, 258
Knockoffs, 389 Linear discriminant analysis
(LDA), 188
L Linear models, see General-
ized linear and nonlinear
models
Lagrange multipliers, 261–262
Linear regression models, 65–122
LASSO, 320–324, 364–367
assumptions, 75–78
definition, 321–322
classical, 75–76
elastic net, 324 motivation, 76–78
example (bodyfat data), closer look at lm() output,
366–367 73–75
example (currency data), collective predictive strength
322–323 of X(j), 88–96
lars package, 322 adjusted-R2, 92–94
stepwise variable selection, basic properties, 88–89
364–367 bias issues, 91–92
subsetting, 364–366 definition of R2, 90–91
Law of total expectation, 53 extensions of LOOM, 95
Law of total variance, 54 k-fold cross-validation, 95
LDA, see Linear discriminant “leaving-one-out method,”
analysis 94–95
Least-angle regression, 383 LOOM for k-NN, 95
Least-squares estimation, 68–73 computational complements,
consistency, 106–108 115–120
distribution, 83, 124 details of the computation,
matrix formulations, 70 115–116
matrix operations to mini- ’lm’ objects, details of,
mize, 71–72 118–120
matrix terms, 71 R functions for multivari-
models without intercept ate normal distribution
term, 72–73 family, 116–117
motivation, 68–70 “error term,” 67
property, 80 exercises, 120–122
unbiasedness, 105–106 inference under homoscedas-
Leaving-one-out method ticity, 81–88
(LOOM), 94–95, 348 example (bike-sharing
Letter recognition data, 186–187 data), 86–88
k-NN, 187–188 normal distribution
INDEX 481

model, 82–83 unbiasedness and consis-

regression case, 83–86 tency, 79–81
review (classical inference Link function, 151
on single mean), 81 LOOM, see Leaving-one-out
Slutsky’s Theorem, 82 method
standard error, concept of,
83 M
least-squares estimation, 68–
73 Major concepts, 1–63
mathematical complements, after fitting a model, how we
101–115 use it for prediction, 22–
biased nature of S, 108 24
Cauchy-Schwartz Inequal- centering and scaling, 60–61
ity, 109 classification techniques, 44–
Central Limit Theorem, 47
104 example (bike-sharing
conditional expectation, data), 45–47
geometry of, 108–110 regression problem, 44–45
covariance matrices, 101– computational complements,
103 55–60
least-squares estimator, CRAN packages, 55–56
consistency of, 106–108 function dispatch, 59–60
least-squares estimator, function tapply() and its
unbiasedness of, 105– cousins, 56–58
106 innards of the k-NN code,
models without a constant 58–59
term, details on, 104– cross-validation, 26–31
105 E (), samples and popula-
multivariate normal distri- tions, 5–6
bution family, 103–104 examples
projections, 109 bike-sharing data, 32–38
random variables as inner click-through rate, 15–17
product spaces, 108–109 description goal (baseball
missing values, 100–101 players), 6–15
notation, 65–66 predicting bike-sharing ac-
p-values, 96 tivity, 1–2
misleadingly large p- prediction goal (body fat),
values, 99–100 2
misleadingly small p- Who clicks Web ads?, 3–4
values, 97–99 exercises, 61–63
verdict, 100 generic predict() function,
random- vs. fixed-X cases, 23–24
67–68 indicator variables, 35
482 INDEX

interaction terms, including Mathematical complements, 48–

quadratics, 38–43 54
example (salaries of female basic concepts, 48–54
programmers and engi- Big Data, 447–448
neers), 39–42 classification, 203–206
fitting separate models, disaggregating regressor ef-
42–43 fects, 306–309
higher-order polynomial homoscedasticity, 141–143
models, 43 linear regression models,
saving your work, 43 101–115
intuition, 24–26 model fit, 257–263
mathematical complements, multiclass classification
48–54 problems, 203–206
conditional expectation, semi-linear methods, 424–
properties of, 52–54 429
indicator random vari- shrinkage estimators, 331–
ables, 48 334
mean squared error of an statistical estimation, 175
estimator, 48–49 variable selection and dimen-
µ(t) minimizes mean sion reduction, 386–387
squared prediction er- Matrix algebra, 451–462
ror, 49–50 eigenvalues and eigenvectors,
µ(t) minimizes misclassifi- 455–456
cation rate, 50–52 linear independence, 454
model bias, 25 matrices of the form B’B,
nonparametric settings, 23 456–457
overfitting and the variance- matrix addition and multi-
bias tradeoff, 24–26 plication, 452–453
parametric settings, 22–23 matrix algebra in R, 459–462
parametric vs. nonparamet- matrix derivatives, 458–459
ric models, 15 matrix inverse, 454–455
participation, 47–48 matrix transpose, 453–454
prediction, approach to, 4 partitioned matrices, 457–
predictor variables, 17–22 458
multipredictor linear mod- rank of a matrix, 456
els, 18–19 terminology and notation,
nonparametric regression 451–452
estimation (k-NN), 19– Maximum Likelihood Estimators
22 (MLEs), 175
rough rule of thumb, 32 Median regression, 236
training data, 23 Mixed-effects model, 287
tuning parameters, impor- Model bias, 25
tant note on, 31–32 Model fit assessment and im-
INDEX 483

provement, 215–265 timator, 261

automated outlier resistance, Gauss-Markov theorem,
236–238 261–263
example (currency data), hat matrix, 257–259
238 Lagrange multipliers, 261–
median regression, 236– 262
238 leverage values, 258
robust regression, 236 matrix inverse update,
classification settings, 241– 259–260
245 median minimizes mean
computational complements, absolute deviation, 260–
255–257 261
description goal, 255 proof, 262–263
diagnostics related to indi- methods, 216
vidual predictors, 224– notation, 216
232 overall measures of model fit,
freqparcoord package, 220–224
229–232 cross-validation, revisited,
partial residual plots, 225– 222
227 hat matrix, 224
plotting nonparametric fit holdout method, 222
against each predictor, plotting parametric fit
227–228 against a nonparamet-
example (currency data), ric one, 222–223
219–220 residuals vs. smoothing,
example (vocabulary acquisi- 223–224
tion), 238–241 R-squared, revisited, 221–
exercises, 264–265 222
goals of model fit-checking, tool to aid model selection,
217–218 254
center vs. fringes of the unusual observations, effects
data set, 218 of, 232–236
description context, 218 example (currency data),
prediction context, 217 233
improving fit, 245–254 influence() function, 232–
adding polynomial terms, 235
247–251 outliers, 232
boosting, 251–254 use of freqparcoord for
deleting terms from the outlier detection, 235–
model, 245 236
mathematical complements, MovieLens data
257–263 homoscedasticity, 127–130
Best Linear Unbiased Es- multiple inference, 300–302
484 INDEX

wrangling, 303–305 191–193

Multiclass classification prob- example (vertebrae data),
lems, 179–213 192–193
computational complements, model, 191
207–211 software, 192
OVA and AVA logit analy- one vs. all or all vs. all, 183–
sis, R code for, 207–210 188
ROC code, 210–211 better choice, 184
estimating the functions example (k-NN on letter
µi (t), 182 recognition data), 187–
exercises, 211–213 188
Fisher linear discriminant example (letter recogni-
analysis, 188–191 tion data), 186–187
background, 189 example (vertebrae data),
derivation, 189–190 184–185
example (vertebrae data), intuition, 185
190 verdict, 188
LDA code and results, “unbalanced” (and bal-
190–191 anced) data, 193–200
quadratic discriminant crucial sampling issue,
analysis, 190 195–198
going beyond using the 0.5 example (letter recogni-
threshold, 200–206 tion), 198–200
confusion matrix and ROC reason for concern regard-
curve, 202–203 ing balance, 194–195
unbalanced data, 201–202 Multiple inference problem, 297
unequal misclassification
costs, 200–201 N
how we use models for pre-
diction, 182 Neural networks, 406, 418–423
key equations, 180–181 dimension reduction, 421
key notation, 179–180 example (vertebrae data),
mathematical complements, 418–420
203–206 tuning parameters, 420–421
classification via density why it works, 421–423
estimation, 203–205 Nonlinear model, see General-
optimal classification rule ized linear and nonlinear
for unequal error costs, models
206 Nonnegative matrix factorization
time complexity compari- (NMF), 377–380
son, OVA vs. AVA, 205– computation for, 383–386
206 example (spam detection),
multinomial logistic model, 378–380
INDEX 485

interpretation, 377–378 predictor reuse and statis-

overview, 377 tical consistency, 398–
sum-of-parts property, 378 399
Nonparametric regression estima- split criterion, 398
tion (k-NN), 19–22 tuning parameters, 399
example (baseball player Partitioning, 457
data), 22 PCA, see Principal components
k-NN method and tuning pa- analysis
rameters, 21 Pima diabetes data
looking at nearby points, 20 classification, 242–245
measures of nearness, 20–21 generalized linear model,
nearest-neighbor analysis in 158–159
the regtools package, model fit assessment and im-
21–22 provement, 242–245
Post-selection inference, 367–369
O Prediction goal (example), 2
Predictor effects principle, 268
One vs. All (OVA), 183
Predictor variables, 17–22
Ordinary least-squares (OLS) es-
bodyfat data, 352
timator, 130, 253
dimension reduction, 369
Outliers, 232
elimination of, 316
Overdispersion, 152
fit criteria, 344
Overfitting, variance-bias trade-
linear discriminant analysis,
off and, 24–26
188
example (student evalua-
matrix formulations, 70
tions of instructors), 26
model fit, 220
intuition, 24–26
multipredictor linear models,
P 18–19
neural networks, 418
Parsimonious model, 321 nonlinear boundaries, 416
Partition-based methods, 391– nonparametric regression es-
404 timation (k-NN), 19–22
CART, 392 OVA vs. AVA, 205
example (vertebral column principal components analy-
data), 394–398 sis, 327
exercises, 403–404 unobserved, 278–289
random forests, 399–402 Principal components analysis
bagging, 400 (PCA), 327, 369, 370–
example (letter recogni- 376
tion), 401–402 Principle of Frequent Occurrence
example (vertebrae data), of Rare Events, 348
400–401 Principle of rare events, 296
technical details, 398–399 p-values, 96
486 INDEX

misleadingly large p-values, RFA, see Regression function av-

99–100 eraging
misleadingly small p-values, Ridge regression, 316–320
97–99 Robust regression, 236
simple uses of (pitfalls), 349 ROC curve, see Receiver Operat-
verdict, 100 ing Characteristic curve
Pythagorean Theorem, 54, 109
S
Q
Scheffe’s method, 298–300
Quadratic discriminant analysis, Semi-linear methods, 405–430
190 exercises, 429–430
Quadratics, 38–43 k-NN with linear smoothing,
Quiz documents, Big-p, 444–446 407–409
example (bodyfat data),
R 409
extrapolation via lm(), 407
Random effects models, 286–288 multicollinearity issues,
Random variable, conditional ex- 409
pectation as, 52–53 tuning parameter, 409
Random-X setting, 67 linear approximation of class
Rare events, principle of, 296, 348 boundaries, 410–423
Rate of false positives (FPR), 202 neural networks, 418–423
Rate of true positives (TPR), 202 SVMs, 410–417
Receiver Operating Characteris- mathematical complements,
tic (ROC) curve, 202– 424–429
203 dual formulation for SVM,
Reduced convex hulls (RCHs), 425–428
413–415 edge bias in nonparamet-
Regression function averaging ric regression, 424–425
(RFA), 289–295 kernel trick, 428–429
estimating the counterfac- verdict, 423–424
tual, 290–291 Shrinkage estimators, 311–337
example (job training), 290– bias, standard errors and sig-
291 nificance tests, 327
small area estimation (“bor- computational complements,
rowing from neigh- 334–336
bors”), 291–294 exact multicollinearity, cases
standard errors for, 307–308 of, 324–326
verdict, 295 exercises, 336–337
regtools package, 21–22 generalized linear models,
Regularization, 311, 330 329–330
Residual values, 224 LASSO, 320–324
INDEX 487

definition, 321–322 Slutsky’s Theorem, 82

elastic net, 324 Small area estimation, 268
example (currency data), Software Alchemy, 432–433, 447–
322–323 448
lars package, 322 Square matrix, 451
mathematical complements, Standard error, concept of, 83
331–334 Strong Law of Large Numbers
James-Stein theory, 331– (SLLN), 106
332 Student evaluations of instruc-
ridge action increases tors, 26
eigenvalues, 333–334 Support vector machines
multicollinearity, 313–316 (SVMs), 406, 410–417
checking for, 314–315 example (letter recognition
example (currency data), data), 417
315 geometric motivation, 411–
guiding model, 313–314 413
variance inflation factor, nonlinear boundaries, 416–
314–315 417
what one can/should do, reduced convex hulls, 413–
315–316 415
penalized methods, 330 statistical consistency, 417
principal components analy- tuning parameter, 415–416
sis, 327–329 SVD, see Singular Value Decom-
regularization, 311, 330 position
relevance of James-Stein to
regression estimation, T
312–313
ridge regression, 316–320 Test set, 26
alternate definitions, 317– Text classification, 163
318 Total expectation, law of, 53
choosing the value of , Total variance, law of, 54
318–319 Tower Property, 54, 269, 306
example (currency data), TPR, see Rate of true positives
319–320 Tuning parameters, 21
Tykhanov regularization, important note on, 31–32
330 k-NN with linear smoothing,
Simpson’s paradox, 274–278 409
Simultaneous inference problem, neural networks, 420–421
297 nonparametric regression es-
Singular Value Decomposition timation, 21
(SVD), 370 partition-based methods,
SLLN, see Strong Law of Large 399
Numbers SVMs, 415–416
488 INDEX

Two-Stage Least Squares, 283 cross-validation, 347–348

Tykhanov regularization, 330 how many is too many, 344
knockoffs, 389
U least-angle regression, 383
mathematical complements,
Unbalanced data, 193–200
386–387
Unbiasedness, 79
post-selection inference,
Upper-triangular matrix, 451
367–369
V p-values, simple uses of (pit-
falls), 349
Validation set, 26 stepwise selection, 351–367
Variable selection and dimension backward selection, 351,
reduction, 339–390 352
asking “what if” questions, basic notion, 351
349–351 classification settings,
computational complements, 357–362
383–386 example (bodyfat data),
dimension reduction, direct 352–357, 366–367
methods for, 369–382 forward selection, 351, 352
dummy/R factor vari- LASSO, 364–367
ables, 381–382 nonparametric settings,
example (bodyfat data), 362–364
371–374 R functions for, 352
example (instructor evalu- subsetting, 364–366
ations), 375–376 under/overfitting, 341–344
freqparcoord for dimen- variable selection methods,
sion reduction, 380–382 348–349
informal nature, 369–370 verdict, 382
loadings, 372 Variance
nonnegative matrix factor- -bias tradeoff, overfitting
ization, 377–380 and, 24–26
PCA, 370–376 components, 287
role in regression analysis, inflation factor, 314–315
370 Vertebrae data
Singular Value Decompo- Fisher linear discriminant
sition, 370 analysis, 190
sum-of-parts property, 378 multiclass classification
exercises, 387–390 problems, 184–185
fit criteria, 344–348 neural networks, 418–420
Akaike’s Information Cri- random forests, 400–401
terion, 347 Vocabulary acquisition, model
common measures, 345– fit assessment and im-
348 provement, 238–241
INDEX 489

W
Web ads, who clicks on, 3–4
Weighted least squares (WLS),
130, 133–135
Word bank dataset, data wran-
gling for, 256–257
 Fig. 1.3, Turkish student evaluations

0.015
z

0.010

0.005

−10

−5

10
0
5
s

5 0
t
−5
10 −10

Fig. 2.1, Bivariate normal density, ρ = 0.2

0.025
0.020
0.015
z

0.010
0.005
0.000
−10

−5

10
0
5
s

5 0
t
−5
10 −10

Fig. 2.2, Bivariate normal density, ρ = 0.8

Fig. 6.3, Partial residuals plot, currency data

Fig. 6.6, Freqparcoord plot, currency data

Fig. 6.8, Outlier hunt

Fig. 10.2, Flow chart for vertebral column data

Fig. 11.7, Vertebrae NN, 1 hidden layer