Analyzing Baseball Data With R-Chapman and Hall - CRC (2013)

The R Series
Analyzing
Baseball Data
with R
Max Marchi
Jim Albert
Analyzing
Baseball Data
with R
Chapman & Hall/CRC
The R Series
Series Editors
John M. Chambers Torsten Hothorn
Department of Statistics Division of Biostatistics
Stanford University University of Zurich
Stanford, California, USA Switzerland
Duncan Temple Lang Hadley Wickham

Department of Statistics Department of Statistics
University of California, Davis Rice University
Davis, California, USA Houston, Texas, USA
Aims and Scope

This book series reflects the recent rapid growth in the development and application
of R, the programming language and software environment for statistical computing
and graphics. R is now widely used in academic research, education, and industry.
It is constantly growing, with new versions of the core software released regularly
and more than 4,000 packages available. It is difficult for the documentation to
keep pace with the expansion of the software, and this vital book series provides a
forum for the publication of books covering many aspects of the development and
application of R.
The scope of the series is wide, covering three main threads:
• Applications of R to specific disciplines such as biology, epidemiology,
genetics, engineering, finance, and the social sciences.
• Using R for the study of topics of statistical methodology, such as linear and
mixed modeling, time series, Bayesian methods, and missing data.
• The development of R, including programming, building packages, and
graphics.
The books will appeal to programmers and developers of R software, as well as
applied statisticians and data analysts in many fields. The books will feature
detailed worked examples and R code fully integrated into the text, ensuring their
usefulness to researchers, practitioners and students.
Published Titles
Analyzing Baseball Data with R, Max Marchi and Jim Albert
Customer and Business Analytics: Applied Data Mining for Business Decision
Making Using R, Daniel S. Putler and Robert E. Krider
Dynamic Documents with R and knitr, Yihui Xie
Event History Analysis with R, Göran Broström
Programming Graphical User Interfaces with R, Michael F. Lawrence and

John Verzani
R Graphics, Second Edition, Paul Murrell
Reproducible Research with R and RStudio, Christopher Gandrud
Statistical Computing in C++ and R, Randall L. Eubank and Ana Kupresanin

The R Series
Analyzing
Baseball Data
with R
Max Marchi
Jim Albert
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2014 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

Version Date: 20130916
International Standard Book Number-13: 978-1-4665-7023-8 (eBook - PDF)
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 stor-
age or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.copy-
right.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 pro-
vides licenses and registration for a variety of users. For organizations that have been granted a pho-
tocopy 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 xv
1 The Baseball Datasets 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 The Lahman Database: Season-by-Season Data . . . . . . . . 2
1.2.1 Bonds, Aaron, and Ruth home run trajectories . . . . 2
1.2.2 Obtaining the database . . . . . . . . . . . . . . . . . 4
1.2.3 The Master table . . . . . . . . . . . . . . . . . . . . . 4
1.2.4 The Batting table . . . . . . . . . . . . . . . . . . . . 6
1.2.5 The Pitching table . . . . . . . . . . . . . . . . . . . . 8
1.2.6 The Fielding table . . . . . . . . . . . . . . . . . . . . 11
1.2.7 The Teams table . . . . . . . . . . . . . . . . . . . . . 11
1.2.8 Baseball questions . . . . . . . . . . . . . . . . . . . . 13
1.3 Retrosheet Game-by-Game Data . . . . . . . . . . . . . . . . 14
1.3.1 The 1998 McGwire and Sosa home run race . . . . . . 14
1.3.2 Retrosheet . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.3 Game logs . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3.4 Obtaining the game logs from Retrosheet . . . . . . . 16
1.3.5 Game log example . . . . . . . . . . . . . . . . . . . . 16
1.4 Retrosheet Play-by-Play Data . . . . . . . . . . . . . . . . . 18
1.4.1 Event files . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.2 Event example . . . . . . . . . . . . . . . . . . . . . . 19
1.5 Pitch-by-Pitch Data . . . . . . . . . . . . . . . . . . . . . . . 21
1.5.1 MLBAM Gameday and PITCHf/x . . . . . . . . . . . 21
1.5.2 PITCHf/x Example . . . . . . . . . . . . . . . . . . . 22
1.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.7 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
vii
viii Contents
2 Introduction to R 29
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Installing R and RStudio . . . . . . . . . . . . . . . . . . . . 30
2.3 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.1 Career of Warren Spahn . . . . . . . . . . . . . . . . . 31
2.3.2 Vectors: defining and calculations . . . . . . . . . . . . 32
2.3.3 Vector functions . . . . . . . . . . . . . . . . . . . . . 34
2.3.4 Vector index and logical variables . . . . . . . . . . . . 35
2.4 Objects and Containers in R . . . . . . . . . . . . . . . . . . 36
2.4.1 Character data and matrices . . . . . . . . . . . . . . 37
2.4.2 Factors . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.3 Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.5 Collection of R Commands . . . . . . . . . . . . . . . . . . . 41
2.5.1 R scripts . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5.2 R functions . . . . . . . . . . . . . . . . . . . . . . . . 42
2.6 Reading and Writing Data in R . . . . . . . . . . . . . . . . 43
2.6.1 Importing data from a file . . . . . . . . . . . . . . . . 43
2.6.2 Saving datasets . . . . . . . . . . . . . . . . . . . . . . 45
2.7 Data Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 45
2.7.2 Manipulations with data frames . . . . . . . . . . . . 47
2.7.3 Merging and selecting from data frames . . . . . . . . 49
2.8 Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.9 Splitting, Applying, and Combining Data . . . . . . . . . . . 50
2.9.1 Using sapply . . . . . . . . . . . . . . . . . . . . . . . 51
2.9.2 Using ddply in the plyr package . . . . . . . . . . . . 52
2.10 Getting Help . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.11 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.12 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3 Traditional Graphics 59
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2 Factor Variable . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.1 A bar graph . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.2 Add axes labels and a title . . . . . . . . . . . . . . . 61
3.2.3 Other graphs of a factor . . . . . . . . . . . . . . . . . 62
3.3 Saving Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.4 Dot plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.5 Numeric Variable: Stripchart and Histogram . . . . . . . . . 65
3.6 Two Numeric Variables . . . . . . . . . . . . . . . . . . . . . 67
3.6.1 Scatterplot . . . . . . . . . . . . . . . . . . . . . . . . 67
3.6.2 Building a graph, step-by-step . . . . . . . . . . . . . 69
3.7 A Numeric Variable and a Factor Variable . . . . . . . . . . 73
Contents ix
3.7.1 Parallel stripcharts . . . . . . . . . . . . . . . . . . . . 74

3.7.2 Parallel boxplots . . . . . . . . . . . . . . . . . . . . . 74
3.8 Comparing Ruth, Aaron, Bonds, and A-Rod . . . . . . . . . 76
3.8.1 Getting the data . . . . . . . . . . . . . . . . . . . . . 76
3.8.2 Creating the player data frames . . . . . . . . . . . . . 78
3.8.3 Constructing the graph . . . . . . . . . . . . . . . . . 78
3.9 The 1998 Home Run Race . . . . . . . . . . . . . . . . . . . 79
3.9.1 Getting the data . . . . . . . . . . . . . . . . . . . . . 79
3.9.2 Extracting the variables . . . . . . . . . . . . . . . . . 81
3.9.3 Constructing the graph . . . . . . . . . . . . . . . . . 82
3.10 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4 The Relation Between Runs and Wins 87
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2 The Teams Table in Lahman’s Database . . . . . . . . . . . 88
4.3 Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 The Pythagorean Formula for Winning Percentage . . . . . . 93
4.5 The Exponent in the Pythagorean Formula . . . . . . . . . . 95
4.6 Good and Bad Predictions by the Pythagorean Formula . . . 96
4.7 How Many Runs for a Win? . . . . . . . . . . . . . . . . . . 99
4.8 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5 Value of Plays Using Run Expectancy 105
5.1 The Runs Expectancy Matrix . . . . . . . . . . . . . . . . . 105

5.2 Runs Scored in the Remainder of the Inning . . . . . . . . . 106
5.3 Creating the Matrix . . . . . . . . . . . . . . . . . . . . . . . 107
5.4 Measuring Success of a Batting Play . . . . . . . . . . . . . . 110
5.5 Albert Pujols . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.6 Opportunity and Success for All Hitters . . . . . . . . . . . . 114
5.7 Position in the Batting Lineup . . . . . . . . . . . . . . . . . 116
5.8 Run Values of Different Base Hits . . . . . . . . . . . . . . . 119
5.8.1 Value of a home run . . . . . . . . . . . . . . . . . . . 119
5.8.2 Value of a single . . . . . . . . . . . . . . . . . . . . . 121
5.9 Value of Base Stealing . . . . . . . . . . . . . . . . . . . . . . 123
5.10 Further Reading and Software . . . . . . . . . . . . . . . . . 126
5.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6 Advanced Graphics 129
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

6.2 The lattice Package . . . . . . . . . . . . . . . . . . . . . . 130
x Contents
6.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 130

6.2.2 The verlander dataset . . . . . . . . . . . . . . . . . 130
6.2.3 Basic plotting with lattice . . . . . . . . . . . . . . . 132
6.2.4 Multipanel conditioning . . . . . . . . . . . . . . . . . 133
6.2.5 Superposing group elements . . . . . . . . . . . . . . . 134
6.2.6 Scatterplots and dot plots . . . . . . . . . . . . . . . . 135
6.2.7 The panel function . . . . . . . . . . . . . . . . . . . 137
6.2.8 Building a graph, step-by-step . . . . . . . . . . . . . 139
6.3 The ggplot2 Package . . . . . . . . . . . . . . . . . . . . . . 144
6.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 144
6.3.2 The cabrera dataset . . . . . . . . . . . . . . . . . . . 145
6.3.3 The first layer . . . . . . . . . . . . . . . . . . . . . . . 146
6.3.4 Grouping factors . . . . . . . . . . . . . . . . . . . . . 148
6.3.5 Multipanel conditioning (faceting) . . . . . . . . . . . 149
6.3.6 Adding elements . . . . . . . . . . . . . . . . . . . . . 150
6.3.7 Combining information . . . . . . . . . . . . . . . . . 151
6.3.8 Adding a smooth line with error bands . . . . . . . . . 151
6.3.9 Dealing with cluttered charts . . . . . . . . . . . . . . 153
6.3.10 Adding a background image . . . . . . . . . . . . . . . 155
6.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7 Balls and Strikes Effects 161
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

7.2 Hitter’s Counts and Pitcher’s Counts . . . . . . . . . . . . . 162
7.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 162
7.2.2 An example for a single pitcher . . . . . . . . . . . . . 162
7.2.3 Pitch sequences on Retrosheet . . . . . . . . . . . . . 165
7.2.3.1 Functions for string manipulation . . . . . . 165
7.2.3.2 Finding plate appearances going through a
given count . . . . . . . . . . . . . . . . . . . 167
7.2.4 Expected run value by count . . . . . . . . . . . . . . 169
7.2.5 The importance of the previous count . . . . . . . . . 170
7.3 Behaviors by Count . . . . . . . . . . . . . . . . . . . . . . . 173
7.3.1 Swinging tendencies by count . . . . . . . . . . . . . . 173
7.3.1.1 Propensity to swing by location . . . . . . . 173
7.3.1.2 Effect of the ball/strike count . . . . . . . . . 176
7.3.2 Pitch selection by count . . . . . . . . . . . . . . . . . 178
7.3.3 Umpires’ behavior by count . . . . . . . . . . . . . . . 181
7.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 184
7.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Contents xi
8 Career Trajectories 187
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

8.2 Mickey Mantle’s Batting Trajectory . . . . . . . . . . . . . . 188
8.3 Comparing Trajectories . . . . . . . . . . . . . . . . . . . . . 192
8.3.1 Some preliminary work . . . . . . . . . . . . . . . . . 192
8.3.2 Computing career statistics . . . . . . . . . . . . . . . 194
8.3.3 Computing similarity scores . . . . . . . . . . . . . . . 195
8.3.4 Defining age, OBP, SLG, and OPS variables . . . . . . 197
8.3.5 Fitting and plotting trajectories . . . . . . . . . . . . 198
8.4 General Patterns of Peak Ages . . . . . . . . . . . . . . . . . 202
8.4.1 Computing all fitted trajectories . . . . . . . . . . . . 202
8.4.2 Patterns of peak age over time . . . . . . . . . . . . . 203
8.4.3 Peak age and career at-bats . . . . . . . . . . . . . . . 204
8.5 Trajectories and Fielding Position . . . . . . . . . . . . . . . 205
8.6 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
9 Simulation 211
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

9.2 Simulating a Half Inning . . . . . . . . . . . . . . . . . . . . 212
9.2.1 Markov chains . . . . . . . . . . . . . . . . . . . . . . 212
9.2.2 Review of work in runs expectancy . . . . . . . . . . . 213
9.2.3 Computing the transition probabilities . . . . . . . . . 215
9.2.4 Simulating the Markov chain . . . . . . . . . . . . . . 216
9.2.5 Beyond runs expectancy . . . . . . . . . . . . . . . . . 219
9.2.6 Transition probabilities for individual teams . . . . . . 220
9.3 Simulating a Baseball Season . . . . . . . . . . . . . . . . . . 223
9.3.1 The Bradley-Terry model . . . . . . . . . . . . . . . . 223
9.3.2 Making up a schedule . . . . . . . . . . . . . . . . . . 224
9.3.3 Simulating talents and computing win probabilities . . 225
9.3.4 Simulating the regular season . . . . . . . . . . . . . . 225
9.3.5 Simulating the post-season . . . . . . . . . . . . . . . 226
9.3.6 Function to simulate one season . . . . . . . . . . . . 227
9.3.7 Simulating many seasons . . . . . . . . . . . . . . . . 228
9.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 231
9.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
10 Exploring Streaky Performances 237
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

10.2 The Great Streak . . . . . . . . . . . . . . . . . . . . . . . . 238
10.2.1 Finding game hitting streaks . . . . . . . . . . . . . . 238
10.2.2 Moving batting averages . . . . . . . . . . . . . . . . . 240
xii Contents
10.3 Streaks in Individual At-Bats . . . . . . . . . . . . . . . . . . 242

10.3.1 Streaks of hits and outs . . . . . . . . . . . . . . . . . 242
10.3.2 Moving batting averages . . . . . . . . . . . . . . . . . 243
10.3.3 Finding hitting slumps for all players . . . . . . . . . . 243
10.3.4 Were Suzuki and Ibanez unusually streaky? . . . . . . 246
10.4 Local Patterns of Weighted On-Base Average . . . . . . . . . 249
10.5 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 255
10.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
11 Learning About Park Effects by Database Management

Tools 259
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

11.2 Installing MySQL and Creating a Database . . . . . . . . . . 260
11.3 Connecting R to MySQL . . . . . . . . . . . . . . . . . . . . 262
11.3.1 Connecting using package RMySQL . . . . . . . . . . . . 262
11.3.2 Connecting using Package RODBC . . . . . . . . . . . . 263
11.4 Filling a MySQL Game Log Database from R . . . . . . . . 264
11.4.1 From Retrosheet to R . . . . . . . . . . . . . . . . . . 265
11.4.2 From R to MySQL . . . . . . . . . . . . . . . . . . . . 265
11.5 Querying Data from R . . . . . . . . . . . . . . . . . . . . . 268
11.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 268
11.5.2 Coors Field and run scoring . . . . . . . . . . . . . . . 271
11.6 Baseball Data as MySQL Dumps . . . . . . . . . . . . . . . 273
11.6.1 Lahman’s database . . . . . . . . . . . . . . . . . . . . 273
11.6.2 Retrosheet database . . . . . . . . . . . . . . . . . . . 274
11.6.3 PITCHf/x database . . . . . . . . . . . . . . . . . . . 274
11.7 Calculating Basic Park Factors . . . . . . . . . . . . . . . . . 275
11.7.1 Loading the data in R . . . . . . . . . . . . . . . . . . 275
11.7.2 Home run park factor . . . . . . . . . . . . . . . . . . 276
11.7.3 Assumptions of the proposed approach . . . . . . . . . 277
11.7.4 Applying park factors . . . . . . . . . . . . . . . . . . 278
11.8 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 279
11.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
12 Exploring Fielding Metrics with Contributed R Packages 283
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

12.2 A Motivating Example: Comparing Fielding Metrics . . . . . 284
12.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 284
12.2.2 The fielding metrics . . . . . . . . . . . . . . . . . . . 285
12.2.3 Reading an Excel spreadsheet (XLConnect) . . . . . . 286
12.2.4 Summarizing multiple columns (doBy) . . . . . . . . . 287
12.2.5 Finding the most similar string (stringdist) . . . . . 288
12.2.6 Applying a function on multiple columns (plyr) . . . 291
Contents xiii
12.2.7 Weighted correlations (weights) . . . . . . . . . . . . 291

12.2.8 Displaying correlation matrices (ellipse) . . . . . . . 292
12.2.9 Evaluating the fielding metrics (psych) . . . . . . . . 293
12.3 Comparing Two Shortstops . . . . . . . . . . . . . . . . . . . 294
12.3.1 Reshaping the data (reshape2) . . . . . . . . . . . . . 296
12.3.2 Plotting the data (ggplot2 and directlabels) . . . . 296
12.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 297
12.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
A Retrosheet Files Reference 301
A.1 Downloading Play-by-Play Files . . . . . . . . . . . . . . . . 301

A.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 301
A.1.2 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
A.1.3 Using a special function for a particular season . . . . 302
A.1.4 Reading the files into R . . . . . . . . . . . . . . . . . 302
A.1.5 The function parse.retrosheet.pbp . . . . . . . . . 302
A.2 Retrosheet Event Files: a Short Reference . . . . . . . . . . . 304
A.2.1 Game and event identifiers . . . . . . . . . . . . . . . 304
A.2.2 The state of the game . . . . . . . . . . . . . . . . . . 305
A.3 Parsing Retrosheet Pitch Sequences . . . . . . . . . . . . . . 306
A.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 306
A.3.2 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
A.3.3 Evaluating every count . . . . . . . . . . . . . . . . . . 307
B Accessing and Using MLBAM Gameday and PITCHf/x

Data 311
B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

B.2 Where are the Data Stored? . . . . . . . . . . . . . . . . . . 312
B.3 Suitable Formats for PITCHf/x Data . . . . . . . . . . . . . 314
B.3.1 Obtaining data from on-line resources . . . . . . . . . 314
B.3.2 Parsing in R . . . . . . . . . . . . . . . . . . . . . . . 314
B.3.2.1 A wrapper function . . . . . . . . . . . . . . 315
B.4 Details on the Data . . . . . . . . . . . . . . . . . . . . . . . 316
B.4.1 atbat attributes . . . . . . . . . . . . . . . . . . . . . 316
B.4.2 pitch attributes . . . . . . . . . . . . . . . . . . . . . 317
B.4.3 hip attributes (hit locations data) . . . . . . . . . . . 318
B.5 Special Notes About the Gameday and PITCHf/x Data . . . 319
B.6 Miscellanea . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
B.6.1 Calculating the pitch trajectory . . . . . . . . . . . . . 320
B.6.2 An R package for getting and visualizing PITCHf/x
data: pitchRx . . . . . . . . . . . . . . . . . . . . . . 321
B.6.3 Cross-referencing with other data sources . . . . . . . 323
B.6.4 Online resources . . . . . . . . . . . . . . . . . . . . . 323
xiv Contents
Bibliography 325
Index 329
Preface
Baseball has always had a fascination with statistics. Schwarz (2005) docu-
ments the quantitative measurements of teams and players since the begin-
ning of professional baseball history in the 19th century. Since the foundation
of the Society of Baseball Research in 1971, an explosion of new measures
have been developed for understanding offensive and defensive contributions
of players. One can learn much about the current developments in sabermet-
rics by viewing articles at websites such as www.baseballprospectus.com,
www.hardballtimes.com, and www.fangraphs.com.
The quantity and detail of baseball data has exhibited remarkable growth
since the birth of the Internet. The first data were collected for players and
teams for individual seasons – this type of data were what would be dis-
played on the back side of a Topps baseball card. The volunteer-run Project
Scoresheet organized the collection of play-by-play game data, and these
type of data are currently freely available at the Retrosheet organization at
www.retrosheet.org/. Since 2006, PITCHf/x data have been measuring the
speeds and trajectories of every pitched ball, and newer types of data are col-
lecting the speeds and locations of batted balls and the locations and move-
ments of fielders.
The ready availability of these large baseball datasets has led to challenges
for the baseball enthusiast interested in answering baseball questions with
these data. It can be problematic to download and organize the data. Stan-
dard statistical software packages may be well-suited for working with small
datasets of a specific format, but they are less helpful in merging datasets of
different types or performing particular types of analyses, say contour graphs
of pitch locations, that are helpful for PITCHf/x data.
Fortunately, a new open-source statistical computing environment, R, has
experienced increasing popularity among the statistical and computer science
community. R is a system for statistical computation and graphics, and it is
a computer language designed for typical and possibly specialized statistical
and graphical applications. The software is available for Unix, Windows, and
Macintosh platforms and is available from www.r-project.org.
The public availability of baseball data and the open-source R software is
an attractive marriage. R provides a large range of tools for importing, ar-
ranging, and organizing large datasets. By the use of built-in functions and
collections of packages from the R user-community, one can perform various
data and graphical analyses, and communicate this work easily to other base-
xv
xvi Preface
ball enthusiasts over the Internet. One of us recently asked a number of MLB
team analytics groups about their use of R and here are some responses:
• “We use: R, MySQL / Oracle, perl, php.”

• “We do use R extensively, and it is our primary statistical package. The
only other major tool we use is probably Excel.”
• “We do use R here. It is our primary statistical package for projects that
need something more than the statistical functions in Excel.”
• “With the occasional exception of Python+NumPy, (R) is the only sta-
tistical programming language or package we use.”
• “We do use R, its used in conjunction with Excel for analysis.”
It is clear that R is a major tool for the analytical work of MLB teams.
The purpose of this book is to introduce R to sabermetricians, baseball
enthusiasts, and students interested in exploring baseball data. Chapter 1 pro-
vides an overview of the publicly available baseball datasets and Chapter 2
gives a gentle introduction to the type of data structures and exploratory and
data management capabilities of R. One of the strongest features of R is its
graphics capabilities – Chapter 3 provides an overview of the traditional graph-
ics functions available in the base package and Chapter 6 introduces more
sophisticated graphical displays available through the lattice and ggplot2
packages.
The remainder of the book illustrates the use of R in exploring a number
of popular topics in sabermetrics. Two fundamental ideas in sabermetrics are
the relationship between runs and wins, and the measurement of the value of
baseball events by runs. Chapter 4 explores the famous Pythagorean formula
derived by Bill James and Chapters 5 and 7 describe the value of plays and
pitch sequences using run expectancy. It is fascinating to explore career per-
formance trajectories of ballplayers and Chapter 8 illustrates the use of R to
fit quadratic models to player trajectories. Chapter 9 illustrates the use of R
simulation functions to simulate a game of baseball by a Markov chain model,
and simulate a season of baseball competition. Baseball fans are interested in
streaky patterns of performance of teams and players and Chapter 10 explores
methods of describing and understanding the significance of streaky patterns
of hitting. Given the large size of baseball datasets, it may be more convenient
to work with a database and Chapter 11 illustrates the application of several
R packages to interface with a MySQL database. Chapter 12 describes the
usefulness of several R packages for exploring fielding statistics. The datafiles
available through Retrosheet and MLBAM Gameway and PITCHf/x are rela-
tively sophisticated and the appendix material provides detailed descriptions
on downloading and reading this data into R.
The reader is encouraged to work on the book datasets and try out
the presented R code as the chapters are read. All of the data files and
Preface xvii
R code sections used in the book are available at the GitHub repository
at github.com/maxtoki/baseball_R. In addition, there is a book blog at
baseballwithr.wordpress.com where the authors will provide advice on us-
ing R in sabermetrics research and keep the reader informed on new develop-
ments in R software and baseball datasets.
The authors are very grateful for the efforts of our editor, John Kimmel,
who played an important role in our collaboration and provided us with timely
reviews that led to significant improvements of the manuscript. We wish to
thank Anne and Ramona for encouragement and inspiration. Although the
two of us live thousands of miles apart, we share a passion both for statistics,
baseball, and the knowledge that one can learn about the game through the
exploration of data.
1
The Baseball Datasets
CONTENTS
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 The Lahman Database: Season-by-Season Data . . . . . . . . . . . . . . . . . 2
1.2.1 Bonds, Aaron, and Ruth home run trajectories . . . . . . . . . 2
1.2.2 Obtaining the database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.3 The Master table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.4 The Batting table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.5 The Pitching table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.6 The Fielding table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.7 The Teams table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.8 Baseball questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Retrosheet Game-by-Game Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.1 The 1998 McGwire and Sosa home run race . . . . . . . . . . . . 14
1.3.2 Retrosheet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.3 Game logs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3.4 Obtaining the game logs from Retrosheet . . . . . . . . . . . . . . . 16
1.3.5 Game log example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 Retrosheet Play-by-Play Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.1 Event files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.2 Event example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5 Pitch-by-Pitch Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.5.1 MLBAM Gameday and PITCHf/x . . . . . . . . . . . . . . . . . . . . . . 21
1.5.2 PITCHf/x Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.7 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1
2 Analyzing Baseball Data with R
1.1 Introduction
Baseball’s marriage with numbers goes back to the origins of the sport. The
pioneers of the game in the 1840s had not yet decided the ultimate distance
between the pitcher’s rubber and home plate, nor the number of balls needed
to be awarded a base, when the first box scores and the first stats appeared
in newspapers.
This chapter introduces three sources of freely available data, the Lahman
database, Retrosheet data, and PITCHf/x data. Baseball records from these
sources have a growing level of detail, from seasonal stats available since the
1871 season, to box score data for individual games, to play-by-play accounts
covering most games since 1945, to extremely detailed pitch-by-pitch data
recorded for nearly all the pitches thrown in MLB parks since 2008. Examples
throughout this book will predominately use subsets of data coming from
these three sources.
1.2 The Lahman Database: Season-by-Season Data

1.2.1 Bonds, Aaron, and Ruth home run trajectories
In the 2007 baseball season Barry Bonds became the new home run king
surpassing Hank Aaron’s record of 755 career home runs. Aaron had held the
throne since 1974, when he had moved past the legendary Babe Ruth with his
715th home run. Figure 1.1 plots the cumulative home runs of Bonds, Aaron,
and Ruth as a function of age. It is clear from the graph that the home run
careers of the three sluggers have followed different paths. Aaron was the clear
home run leader until age 32 and Aaron and Ruth had similar career home
run paths until retirement. Bonds was far behind Aaron and Ruth in career
home runs in his 30s, but narrowed the gap and overtook the two sluggers in
his 40s.
Babe Ruth began his career as a teenage pitcher for the Boston Red Sox
in the so-called Deadball Era when home runs were rare. Ruth’s home run
impact was not felt until his sixth season, when he began sending the ball out
of the park with regularity and outslug nearly every other American League
team with 29 home runs. Given his late start, his career line is S-shaped due
to his slow start and inevitable decline at the end of his career.
Hank Aaron also made his MLB debut at a very young age and shows a
nearly straight line in the graph for the best part of his career. His pattern
of hitting home runs was marked by consistency as he hit between 30 and
50 home runs for most seasons of his career. Similar to the Babe, Aaron also
The Baseball Datasets 3
Bonds (762)
Aaron (755)
Ruth (714)
600
career home runs
400
200
20 25 30 35 40
age
FIGURE 1.1
Career home runs by age for the top three home run hitters in baseball history.
declined in the final years of his career, hitting 20, 12, and 10 home runs from
1974 to 1976.
Barry Bonds had a relatively late major league debut as he did not come to
an agreement with the team that first drafted him, and was not in the career
home run race until after his 35th birthday. Towards the end of his career,
Bonds put together an impressive season home run counts of 49, 73, 46, 45,
and 45 home runs, closing in on Babe’s 714 mark. Then, after missing most
of the 2005 season because of injuries, he completed the chase to the record
with two solid seasons (26 and 28 homers) when he was 42 and 43 years old.
To compare sluggers, a researcher needs season-to-season batting data in-
cluding age and home run count for Bonds, Aaron, and Ruth. One needs this
data for a wide range of seasons, as Ruth’s career began in 1914 and Bonds’
career ended in 2007.
For many years database journalist and author Sean Lahman has been
making available at his website1 a database (Lahman, 2012) containing pitch-
ing, hitting, and fielding statistics for the entire history of professional baseball
1 www.seanlahman.com/
from 1871 to the current season. The data are available in several formats,
including a set of comma-separated-value (csv) tables, that will be used in
this book.
1.2.2 Obtaining the database

Download a copy of Lahman’s database using the following steps.
1. Access www.seanlahman.com/baseball-archive/statistics/.
2. Below the Limited Use License section, there is the section for downloading
the most recent version of the database. At the time of this writing, the
section is named Download 2012 Version. Click on the 2012 Version -
comma-delimited version.
3. Save the file to a directory of your choice.
4. Extract the content of the downloaded zipped file.
After the compressed file has been download and extracted, one has a to-
tal of 24 files with a csv extension listed in Table 1.1. In addition, a “readme
2012.txt” text file is included which gives a thorough description of the Lah-
man database. One is encouraged to read the documentation provided in the
“readme 2012.txt” file to learn about the contents of these files. Here we give
a general description of the variables in the data tables most relevant for the
studies described in this book.
1.2.3 The Master table

The Master table (Master.csv file) is a registry of baseball people. It contains
bibliographical information on every player and manager who has appeared
at the Major League Baseball level and of everyone who has been inducted
in the Baseball Hall of Fame.2 Each line of the Master.csv file constitutes a
short biography of a person, reporting on dates and places of birth and death,
height and weight, throwing hand and batting side, and the dates of the first
and last game played.
Players are identified throughout the pitching, batting, and fielding tables
in the Lahman’s database by an id code, and the Master table is useful for
retrieving the name of the player associated with a particular identifier. The
table also reports player identification codes of other databases, in particular
the ones used by Retrosheet, so one can link players from the Lahman and
Retrosheet databases.
For illustration purposes, we display below the header and first line of the
2 Examples of people who never played Major League Baseball but have been inducted
in the Hall of Fame (therefore having an entry in the Master table) are baseball pioneer
Henry Chadwick and career Negro Leaguer Josh Gibson.
TABLE 1.1
Files in the Lahman’s database.
File Description
AllStarFull.csv Players’ appearances in All-Star Games
Appearances.csv Seasonal players’ appearances by position
AwardsManagers.csv Recipients of the Manager of the Year Award
AwardsPlayers.csv Players recipients of the various Awards
AwardsShareManagers.csv Voting results for the Manager of the Year
Award
AwardsSharePlayers.csv Voting results for the various Awards for play-
ers
Batting.csv Seasonal batting statistics
BattingPost.csv Seasonal batting statistics for post-season
Fielding.csv Seasonal fielding statistics
FieldingOF.csv Seasonal appearances at the three outfield po-
sitions
FieldingPost.csv Seasonal fielding data for post-season
HallOfFame.csv Voting results for the Hall of Fame
Managers.csv Seasonal data for managers
ManagersHalf.csv Seasonal split data for managers
Master.csv Biographical information for individuals ap-
pearing in the database
Pitching.csv Seasonal pitching statistics
PitchingPost.csv Seasonal pitching statistics for post-season
Salaries.csv Seasonal salaries for players
Schools.csv List of college teams
SchoolsPlayers.csv Information on schools attended by players
SeriesPost.csv Outcomes of post-season series
Teams.csv Seasonal stats for teams
TeamsFranchises.csv Timelines of Franchises
TeamsHalf.csv Seasonal split stats for teams
Master.csv file which gives information about the first player in the database
Hank Aaron. For clarity, we place Aaron’s information in a table format in
Table 1.2.
lahmanID,playerID,managerID,hofID,birthYear,birthMonth,birthDay,
birthCountry,birthState,birthCity,deathYear,deathMonth,deathDay,
deathCountry,deathState,deathCity,nameFirst,nameLast,nameNote,
nameGiven,nameNick,weight,height,bats,throws,debut,finalGame,
college,lahman40ID,lahman45ID,retroID,holtzID,bbrefID
1,aaronha01,,aaronha01h,1934,2,5,USA,AL,Mobile,,,,,,,Hank,Aaron,
,Henry Louis,"Hammer,Hammerin’ Hank,Bad Henry",180,72,R,R,
4/13/1954,10/3/1976,,aaronha01,aaronha01,aaroh101,aaronha01,
aaronha01
From this information, we learn some details about Aaron’s life. Hank
Aaron was born February 5, 1934, in Mobile, Alabama, his full name is Henry
Louis Aaron, and his nicknames were Hammer, Hammerin’ Hank, and Bad
Henry. Aaron weighed 180 pounds and was 72 inches tall, he threw and batted
right-handed, and he played in the big leagues from 4/13/1954 to 10/3/1976.
There is a series of blank columns (those consecutive commas between “Mo-
bile” and “Hank”) corresponding to death information, which is obviously
unavailable for a living person. Finally there are the various identifying codes
for the player. The value of playerID, aaronha01, is the identifying code for
Hank Aaron in every table in the Lahman’s database. The value of the vari-
able retroID, aaroh101, is the player id specific to the Retrosheet files to be
described in Section 1.3.
1.2.4 The Batting table

The csv file Batting.csv file contains all players’ batting statistics by season
and team from 1871 to the present season. Players in this file are identified
with their playerID; for example, the season batting statistics of Hank Aaron
appear in these table with the identification playerID = aaronha01. Each
line of the file contains the statistics compiled by a player, during a single
season (variable yearID), for a particular team (variable teamID).
Players who changed teams during a particular season have multiple lines
for the season. The stint variable indicates the order in which the player
moved between teams. For example, Lou Brock, who moved during the 1964
season from the Chicago Cubs to the St. Louis Cardinals, has the following
batting lines for the 1964 season.
playerID yearID stint teamID lgID, ...
9567 brocklo01 1964 1 CHN NL, ...
9568 brocklo01 1964 2 SLN NL, ...
Batting statistics variables are identified by their traditional abbreviations
such as AB, R, H, 2B, etc., so the column names of the batting tables should be
TABLE 1.2
First line of the Master.csv file.
field name value

lahmanID 1
playerID aaronha01
managerID
hofID aaronha01h
birthYear 1934
birthMonth 2
birthDay 5
birthCountry USA
birthState AL
birthCity Mobile
deathYear
deathMonth
deathDay
deathCountry
deathState
deathCity
nameFirst Hank
nameLast Aaron
nameNote
nameGiven Henry Louis
nameNick Hammer,Hammerin’ Hank,Bad Henry
weight 180
height 72
bats R
throws R
debut 4/13/1954
finalGame 10/3/1976
college
lahman40ID aaronha01
lahman45ID aaronha01
retroID aaroh101
holtzID aaronha01
bbrefID aaronha01
easily understood by those familiar with baseball box scores. If one has doubts
about the meaning of one particular column name, the “readme 2012.txt” file
provided with the database gives the variable descriptions.
An excerpt of the Batting.csv file for Babe Ruth is conveniently format-
ted in Table 1.3. This table shows his batting statistics for his early seasons
as a Boston Red Sox pitcher, his years for the Yankees when he became a
great home run slugger, and his seasons at the twilight of his career with the
Boston Braves.
Only count statistics such as the count of at-bats and count of hits are
reported in the batting table. Derived statistics such as a batting average
need to be computed from these count statistics. For example, a researcher
who wants to know Ruth’s batting average for the 1919 season has to calculate
it following paragraph 10.21(b) of the Official Baseball Rules (Triumph Books,
2012) that instructs to “divide the number of safe hits by the total times at
bat.” The relevant columns are H and AB, and the desired result is 139 / 432 =
.322. Some statistics are not visible for Babe Ruth as they were not recorded
in the 1920s. For example, the counts of intentional walks (IBB) are blank for
Ruth’s seasons, indicating that intentional walks were not recorded for Ruth’s
seasons.
1.2.5 The Pitching table

The Pitching.csv file contains season-by-season pitching data for players.
This file contains the traditional count data for pitching such as W (number
of wins), L (number of losses), G (games played), BB (number of walks), and
SO (number of strikeouts). In addition, this dataset contains several derived
statistics such as ERA (earned run average) and BAOpp (opponent’s batting
average).
Babe Ruth also provides a good illustration of the pitching statistics tables
of the Lahman’s database since he had a great pitching record before becoming
one of the greatest home run hitters in history. Table 1.4 displays statistics
from the data file Pitching.csv for the seasons where Ruth was a pitcher.
We see from the table that Ruth pitched in more than 40 games in 1916 and
1917 (by viewing column G), mostly as a starter (see GS), then appeared on
the mound for half that many in the final two seasons for the Red Sox. When
he moved to New York he was only an occasional pitcher. Note that Ruth
always was a winning pitcher, as his wins (W) outnumbered his losses (L) for
all pitching seasons, even when he returned to the pitching mound at the end
of his career. He pitched one game both in 1930 and in 1933 (over ten years
after he was a dominant pitcher for the Red Sox) and went the full nine innings
(see variable CG) in each occasion.
TABLE 1.3
Batting statistics for Babe Ruth, taken from the Batting.csv file.
playerID yearID stint teamID lgID G G batting AB R H 2B 3B HR RBI SB CS BB SO IBB HBP SH SF GIDP G old
1 ruthba01 1914 1 BOS AL 5 5 10 1 2 1 0 0 2 0 0 4 0 0 5
2 ruthba01 1915 1 BOS AL 42 42 92 16 29 10 1 4 21 0 9 23 0 2 42
The Baseball Datasets
3 ruthba01 1916 1 BOS AL 67 67 136 18 37 5 3 3 15 0 10 23 0 4 67

4 ruthba01 1917 1 BOS AL 52 52 123 14 40 6 3 2 12 0 12 18 0 7 52
5 ruthba01 1918 1 BOS AL 95 95 317 50 95 26 11 11 66 6 58 58 2 3 95
6 ruthba01 1919 1 BOS AL 130 130 432 103 139 34 12 29 114 7 101 58 6 3 130
7 ruthba01 1920 1 NYA AL 142 142 457 158 172 36 9 54 137 14 14 150 80 3 5 142
8 ruthba01 1921 1 NYA AL 152 152 540 177 204 44 16 59 171 17 13 145 81 4 4 152
9 ruthba01 1922 1 NYA AL 110 110 406 94 128 24 8 35 99 2 5 84 80 1 4 110
10 ruthba01 1923 1 NYA AL 152 152 522 151 205 45 13 41 131 17 21 170 93 4 3 152
11 ruthba01 1924 1 NYA AL 153 153 529 143 200 39 7 46 121 9 13 142 81 4 6 153
12 ruthba01 1925 1 NYA AL 98 98 359 61 104 12 2 25 66 2 4 59 68 2 6 98
13 ruthba01 1926 1 NYA AL 152 152 495 139 184 30 5 47 150 11 9 144 76 3 10 152
14 ruthba01 1927 1 NYA AL 151 151 540 158 192 29 8 60 164 7 6 137 89 0 14 151
15 ruthba01 1928 1 NYA AL 154 154 536 163 173 29 8 54 142 4 5 137 87 3 8 154
16 ruthba01 1929 1 NYA AL 135 135 499 121 172 26 6 46 154 5 3 72 60 3 13 135
17 ruthba01 1930 1 NYA AL 145 145 518 150 186 28 9 49 153 10 10 136 61 1 21 145
18 ruthba01 1931 1 NYA AL 145 145 534 149 199 31 3 46 163 5 4 128 51 1 0 145
19 ruthba01 1932 1 NYA AL 133 133 457 120 156 13 5 41 137 2 2 130 62 2 0 133
20 ruthba01 1933 1 NYA AL 137 137 459 97 138 21 3 34 103 4 5 114 90 2 0 137
21 ruthba01 1934 1 NYA AL 125 125 365 78 105 17 4 22 84 1 3 104 63 2 0 125
22 ruthba01 1935 1 BSN NL 28 28 72 13 13 0 0 6 12 0 20 24 0 0 2 28
9
Analyzing Baseball Data with R
TABLE 1.4
Pitching statistics for Babe Ruth, taken from the Pitching.csv file. A few extra columns, not reported here, are avialable
in the Pitching.csv file.
playerID yearID stint teamID lgID W L G GS CG SHO SV IPouts H ER HR BB SO BAOpp ERA IBB WP HBP BK BFP GF R
1 ruthba01 1914 1 BOS AL 2 1 4 3 1 0 0 69 21 10 1 7 3 0.23 3.91 0 0 0 100 0 12
2 ruthba01 1915 1 BOS AL 18 8 32 28 16 1 0 653 166 59 3 85 112 0.21 2.44 9 6 1 895 3 80
3 ruthba01 1916 1 BOS AL 23 12 44 41 23 9 1 971 230 63 0 118 170 0.20 1.75 3 8 1 1300 3 83
4 ruthba01 1917 1 BOS AL 24 13 41 38 35 6 2 979 244 73 2 108 128 0.21 2.01 5 11 0 1313 2 91
5 ruthba01 1918 1 BOS AL 13 7 20 19 18 1 0 499 125 41 1 49 40 0.21 2.22 3 2 1 660 0 51
6 ruthba01 1919 1 BOS AL 9 5 17 15 12 0 1 400 148 44 2 58 30 0.29 2.97 5 2 1 591 2 59
7 ruthba01 1920 1 NYA AL 1 0 1 1 0 0 0 12 3 2 0 2 0 0.20 4.50 0 0 0 17 0 4
8 ruthba01 1921 1 NYA AL 2 0 2 1 0 0 0 27 14 9 1 9 2 0.35 9.00 0 0 0 49 1 10
9 ruthba01 1930 1 NYA AL 1 0 1 1 1 0 0 27 11 3 0 2 3 0.30 3.00 0 0 0 39 0 3
10 ruthba01 1933 1 NYA AL 1 0 1 1 1 0 0 27 12 5 0 3 0 0.30 5.00 0 0 0 42 0 5
10
1.2.6 The Fielding table

The Fielding.csv file contains season-to-season fielding statistics for all play-
ers in major league history. For a given player, there will be a separate line
for each fielding position. Outfielders positions are grouped together and la-
beled as OF for the older seasons, whereas for the more recent ones they are
conveniently distinguished as LF, CF, RF, for left fielders, center fielders, and
right fielders, respectively. For a player in a position, the data files give the
count of games played (G), the count of games started (GS), the time played
in the field expressed in terms of outs (InnOuts), the count of putouts (PO),
assists (A), and errors (E).
To illustrate fielding data, Table 1.5 displays Babe Ruth’s fielding statistics
for his career. Only one line appears for the each of the seasons between
1914 and 1917, as The Babe was exclusively employed as a pitcher. Later,
as the Boston Red Sox took advantage of his powerful bat, there are three
lines for 1918, one for each defensive position played by Ruth during this
season. Suppose one focuses on Ruth’s fielding as an outfielder. One raw way
of measuring his fielding range, proposed by Bill James in 1977 in his first
Baseball Abstract, is to sum his putouts (variable PO) and assists (variable A)
and divide the sum by the games played (G). The values of this range statistic
for the seasons 1918 through 1935 were
2.19 2.13 1.99 2.40 2.18 2.69 2.36 2.27 2.14 2.26 2.03 1.84 1.92
1.70 1.71 1.70 1.80 1.54
Clearly, Ruth’s range as an outfielder deteriorated towards the end of his

career.
1.2.7 The Teams table

The Teams.csv file contains seasonal data at the team level going back to
1871. A single line in this file includes the team’s abbreviation (teamID), its
final position in the standings (rank), its number of wins and losses (W and
L), and whether the team won the World Series (WSWin), the League (LgWin),
the Division (DivWin), or reached the post-season via the Wild Card (WCWin).
In addition, this file includes cumulative team offensive statistics such as
counts of runs scored (R), hits (H), doubles (2B), walks (BB), strikeouts (SO),
stolen bases (SB), and sacrifice flies (SF). Team defensive statistics include
opponents runs scored (RA), earned runs allowed (ER), complete games (CG),
shutouts (SHO), saves (SV), hits allowed (HA), home runs allowed (HRA), strike-
outs by pitchers (SOA), and walks by pitchers (BBA). Team fielding statis-
tics are included such as the counts of errors (E), double plays (DP), and the
fielding percentage (FP). Last, this file includes the total home attendance
(attendance) and the three-year park factors3 for batters (BPF) and pitchers
(PPF). Teams are identified, in this and other tables in the database, by a
3 See Chapter 11 for an introduction to park factors.
TABLE 1.5
Fielding statistics for Babe Ruth, taken from the Fielding.csv file. Columns
featuring statistics relevant only to catchers are not reported.
playerID yearID stint teamID lgID POS G GS InnOuts PO A E DP
1 ruthba01 1914 1 BOS AL P 4 0 7 0 0
2 ruthba01 1915 1 BOS AL P 32 17 63 2 3
3 ruthba01 1916 1 BOS AL P 44 24 83 3 6
4 ruthba01 1917 1 BOS AL P 41 19 101 2 4
5 ruthba01 1918 1 BOS AL 1B 13 130 6 5 8
6 ruthba01 1918 1 BOS AL OF 59 121 8 7 3
7 ruthba01 1918 1 BOS AL P 20 19 58 6 5
8 ruthba01 1919 1 BOS AL 1B 5 35 4 1 4
9 ruthba01 1919 1 BOS AL OF 111 222 14 1 6
10 ruthba01 1919 1 BOS AL P 17 13 35 2 1
11 ruthba01 1920 1 NYA AL 1B 2 10 0 1 1
12 ruthba01 1920 1 NYA AL OF 141 259 21 19 3
13 ruthba01 1920 1 NYA AL P 1 1 0 0 0
14 ruthba01 1921 1 NYA AL 1B 2 8 0 0 0
15 ruthba01 1921 1 NYA AL OF 152 348 17 13 6
16 ruthba01 1921 1 NYA AL P 2 1 2 0 0
17 ruthba01 1922 1 NYA AL 1B 1 0 0 0 0
18 ruthba01 1922 1 NYA AL OF 110 226 14 9 3
19 ruthba01 1923 1 NYA AL 1B 4 41 1 1 2
20 ruthba01 1923 1 NYA AL OF 148 378 20 11 2
21 ruthba01 1924 1 NYA AL OF 152 340 18 14 4
22 ruthba01 1925 1 NYA AL OF 98 207 15 6 3
23 ruthba01 1926 1 NYA AL 1B 2 10 0 0 2
24 ruthba01 1926 1 NYA AL OF 149 308 11 7 5
25 ruthba01 1927 1 NYA AL OF 151 328 14 13 4
26 ruthba01 1928 1 NYA AL OF 154 304 9 8 0
27 ruthba01 1929 1 NYA AL OF 133 240 5 4 2
28 ruthba01 1930 1 NYA AL OF 144 266 10 10 0
29 ruthba01 1930 1 NYA AL P 1 0 4 0 2
30 ruthba01 1931 1 NYA AL 1B 1 5 0 0 0
31 ruthba01 1931 1 NYA AL OF 142 237 5 7 2
32 ruthba01 1932 1 NYA AL 1B 1 3 0 0 0
33 ruthba01 1932 1 NYA AL OF 128 209 10 9 1
34 ruthba01 1933 1 NYA AL 1B 1 6 0 1 0
35 ruthba01 1933 1 NYA AL OF 132 215 9 7 4
36 ruthba01 1933 1 NYA AL P 1 1 1 0 0
37 ruthba01 1934 1 NYA AL OF 111 197 3 8 0
38 ruthba01 1935 1 BSN NL OF 26 39 1 2 0
three-character code (teamID). The column name in the Teams.csv file helps
in recognizing clubs by their full name.
To illustrate the teams dataset, we extract the data for one of the greatest
teams in baseball history, the 1927 New York Yankees.
yearID lgID teamID franchID divID Rank G Ghome W L DivWin WCWin
1927 AL NYA NYY <NA> 1 155 77 110 44 <NA> <NA>
LgWin WSWin R AB H X2B X3B HR BB SO SB CS HBP SF RA ER ERA
Y Y 975 5347 1644 291 103 158 635 605 90 64 NA NA 599 494 3.2
CG SHO SV IPouts HA HRA BBA SOA E DP FP name
82 11 20 4167 1403 42 409 431 196 123 0.96 New York Yankees
park attendance BPF PPF teamIDBR teamIDlahman45
820 Yankee Stadium I 1164015 98 94 NYY NYA
We see the 1927 Yankees finished the season with a 110-44 record and won
the World Series. The “Bronx Bombers” hit 158 home runs, stole 90 bases,
and had a total home attendance of 1,164,015.
1.2.8 Baseball questions

The following questions can be answered with the Lahman’s database.
Q What is the average number of home runs per game recorded
in each decade? Does the rate of strikeouts show any correlation
with the rate of home runs?
A The number of home runs per game soared from 0.3 in baseball’s first two
decades to 0.8 in the 1920s. After the 1920s, the home run rate showed a
steady increase up to 2.2 per games at the turn of the millennium. The
first years of the current decade seems to reflect a decline in home run
hitting as the rate has decreased to 1.9 HR per game. Strikeouts have
steadily increased over the history of baseball – the number of strikeouts
per games was 1 in the 1870s to 5.6 in the 1920s to 14.2 of the 2010s.
Relevant data to obtain this answer is found in the Teams table.
Q What effect has the introduction of the Designated Hitter (DH)
in the American League had in the difference in run scoring
between the American and National Leagues?
A The DH rule was instituted in 1973 only for the American League. Twice
in the previous three years the National League teams had scored half a
run more per game than the American League teams. From 1973 till the
end of the decade runs scoring was roughly equal. Since then, the American
League has maintained an edge of about half a run per game.
Relevant data to obtain this answer is found in the Teams table.
Q How does the percentage of games completed by the starting
pitcher from 2000 to 2010 compare to the percentage of games
100 years before?
A From 1900 to 1909 pitchers completed 79% of the games they started, from
2000 to 2010 it had dropped to 3.5%.
Data for this answer can be found in the Pitching table.
1.3 Retrosheet Game-by-Game Data

1.3.1 The 1998 McGwire and Sosa home run race
Another sacred Babe Ruth record was the 60 home runs recorded in the 1927
season. This record was eventually broken in 1961 by Roger Maris, after a
thrilling race with his teammate Mickey Mantle: the “M&M Brothers,” as they
were often dubbed, ended the season with 61 and 54 home runs, respectively.
The new home run record lasted another 37 years. In 1998 two other players,
Mark McGwire of the St. Louis Cardinals and Sammy Sosa of the Chicago
Cubs, gave life to a new home run race, that is displayed in Figure 1.2. This
graph shows the cumulative home run count of each player as a function of
the day of the 1998 season.
From the figure, we see that for much of the season, McGwire was the only
man in the chase. Then Sosa caught fire and the two were very close in home
runs starting from mid-August. “Big Mac” first broke the record, hitting his
62nd home run on September 8. Then, on September 25, the two were tied at
66 apiece. Finally, McGwire managed to hit four more in the final days of the
season, while “Slammin’ Sammy” remained at 66.
To produce the graph in Figure 1.2 and relive the 1998 season, one needs
data at a game-by-game level.
1.3.2 Retrosheet
Retrosheet is a volunteer organization, founded in 1989 by University of
Delaware professor David Smith, that aims to collect play-by-play accounts of
every game played in Major League Baseball history. Through the labor of love
of many volunteers who have unearthed old newspaper accounts, scanned mi-
crofilms, and manually entered data into computers, the Retrosheet website4
contains game-by-game summaries going back to the dawn of Major League
Baseball in the nineteenth century. The Retrosheet site also has play-by-play
data of most of the games played since the 1945 season and continues to add
games for previous seasons. This data is introduced in Section 1.4.
4 www.retrosheet.org
70
62
60
home runs in the season
50
40
30
20
McGwire (70)
10 Sosa (66)
0
Apr May Jun Jul Aug Sep Oct
date
FIGURE 1.2
Seasonal home runs for Mark McGwire and Sammy Sosa during the 1998 race.
1.3.3 Game logs

Retrosheet provides individual game data going back to 1871. A game log has
details regarding when the game was played, how many spectators attended,
the teams and the ballpark, and the score (both the final score and the inning
by inning runs scored). In addition, the game log file include teams offensive
and defensive statistics, starting players, managers, and umpire crews. There
are missing observations for some game log variables for earlier baseball sea-
sons.
Retrosheet provides a comprehensive Guide to Retrosheet Game Logs 5 doc-
ument which gives details of all 161 fields compiled for each game. Readers
are encouraged to peruse the guide to fully understand the contents of the
files. Details on the relevant data fields will be described when they are used
in later chapters.
5 www.retrosheet.org/gamelogs/glfields.txt
1.3.4 Obtaining the game logs from Retrosheet

Game log files can be found at www.retrosheet.org/gamelogs/index.html.
A zip file is provided for each season, starting from 1871, and can be down-
loaded in a folder of choice by clicking on the relevant year. When one extracts
the zip file, one obtains a plain text file (.txt extension) where fields are sep-
arated by commas. Appendix A provides an R script file for downloading and
parsing all the game log files.
1.3.5 Game log example

On September 9, 1995, Cal Ripken Jr. of the Baltimore Orioles surpassed
the seemingly unbeatable consecutive games record of 2130 belonging to the
late Lou Gehrig. One can learn more about this historic game by explor-
ing the game log files for the 1995 season. Table 1.6 contains a subset of
the copious information available for this particular game between Baltimore
and California. This data is taken from a single line in the gl1995.txt file
available at www.retrosheet.org/gamelogs/index.html. This table displays
team statistics6 as well as the players’ identities and fielding positions for the
home team; similar statistics and player information are available for the vis-
itor team.
What does one learn from this game log information displayed in Table 1.6?
This game took place on a Wednesday night in front of 46,272 people in Bal-
timore (the hometeam = BAL indicates the Orioles were the home team). The
game lasted over three and a half hours (duration = 215 minutes), thanks
in part to the standing ovation Ripken got at the end of the fifth inning, when
the game became official. (The standing ovation information is not available in
this file.) Baltimore defeated California 4-2; since we observe homehr = 4, we
observe that all of Baltimore’s runs this game were due to four home runs with
the bases empty. The Orioles infield in this game included Rafael Palmeiro at
first base, Chris Hoiles at second base, Ripken at shortstop, and Jeff Huson
at third base.

Here are some typical questions one can answer with the Retrosheet game logs
files.
Q In which months are home runs more likely to occur? What

about ballparks?
A Since 1980, July has been the month with the most home runs per game
(1.97), while September has had the lowest frequency (1.84). In the same
6 Some other team statistics, such as Stolen Bases and Caught Stealings, omitted in Table
1.6, are reported in Game log files.

TABLE 1.6
Excerpt of information available in the Game Logs. Sample from the Cal Rip-
ken’s Iron Man game (Sept. 9, 1995).
date 19950906
dayofweek Wed
visitorteam CAL
hometeam BAL
visitorrunsscored 2
homerunsscore 4
daynight N
parkid BAL12
attendence 46272
duration 215
visitorlinescore 100000010
homelinescore 10020010x
homeab 34
homeh 9
homehr 4
homerbi 4
homebb 1
homek 8
homegdp 0
homelob 7
homepo 27
homea 8
homee 0
umpirehname Larry Barnett
umpire1bname Greg Kosc
umpire2bname Dan Morrison
umpire3bname Al Clark
visitormanagername Marcel Lachemann
homemanagername Phil Regan
homestartingpitchername Mike Mussina
homebatting1name Brady Anderson
homebatting1position 8
homebatting2name Manny Alexander
homebatting3name Rafael Palmeiro
homebatting4name Bobby Bonilla
homebatting5name Cal Ripken
homebatting6name Harold Baines
homebatting7name Chris Hoiles
homebatting8name Jeff Huson
homebatting9name Mark Smith
time frame, 2.71 home runs per game have been hit in Coors Field (home
of the Colorado Rockies), and 1.14 in the Astrodome (the former home of
the Houston Astros).
Q Do runs happen more frequently when some umpires are be-
hind the plate? What is the difference between the most pitcher-
friendly and the most hitter-friendly umpires?
A Among umpires with more than 400 games called since 1980, teams scored
the highest number of runs (10.0 per game combined) when Chuck Meri-
wether was behind the plate and the lowest (7.8) when Doug Harvey was
in charge.
Q How many extra people attend ballgames during the weekend?
What’s the average attendance by day of the week?
A Close to 33,000 people attend games played on Saturdays (data from 1980
to 2011) and 31,000 on Sundays. The average goes down to 29,000 on
Fridays, 25,000 on Thursdays and Mondays, and 24,000 on Tuesdays and
Wednesdays.
1.4 Retrosheet Play-by-Play Data

1.4.1 Event files
Retrosheet has collected data to an even finer detail for most games played
since 1945. For those seasons, play-by-play accounts are available at www.
retrosheet.org/game.htm. These “event files” (as these play-by-play files
are named) contain information on every single event happening on the field
during a game. For each play, information is reported on the situation (inning,
team batting, number of outs, presence of runners on base), the players on
the field, the sequence of pitches thrown, and details on the play itself. For
example, the file indicates whether a hit occurred and if a ball in play is a
ground ball, the file gives the defender that fielded the ball.
Event files come in a format expressly devised for them. The process of
rendering the files in a format suitable for use in R (or other statistical pro-
grams) is not straightforward, thus we will provide a friendly version of the
event files for the seasons used in the book. Retrosheet gives detailed instruc-
tion on how to use the files7 and a step-by-step guide,8 plus the software to
parse the files.9 In Appendix A, R code is presented that implements the full
process of downloading, extracting, and parsing data.
7 How to use Our Event Files: www.retrosheet.org/datause.txt
8 Step-by-Step Example: www.retrosheet.org/stepex.txt
9 Software Tools: www.retrosheet.org/tools.htm
1.4.2 Event example

As a historical game was used for the purpose of showing the contents of Ret-
rosheet game logs, a famous fielding play is used to illustrate the Retrosheet
event files. This play is represented as a single line in an event file shown in
Table 1.7.
The play took place in a game played in Oakland on October 13, 2001,
as can be inferred from the value of the GAME ID variable. This game
was Game 3 of the American League Division Series featuring the home-
town Athletics against the New York Yankees (AWAY TEAM ID = NYA). The
play occurred in the seventh inning with the home team batting (variables
INN CT and BAT ID ID). There were two outs (variable OUTS CT) and the
A’s were leading 1-0 (variables AWAY SCORE CT and HOME SCORE CT). Right-
handed Mike Mussina (variables PIT ID and PIT HAND CD) was on the mound
for the Yankees, facing left-handed batter Terrence Long (variables BAT ID
and BAT HAND CD) with Jeremy Giambi standing on first base (variable
BASE1 RUN ID). The BAT FLD CD = 7 and BAT LINEUP ID = 7 fields inform
us that Giambi’s defensive position was left field (position 7 corresponding to
left field) and he was batting 7th in the lineup. The variables POS2 FLD ID
through POS9 FLD ID report the full defensive lineup for the Yankees.
The seemingly unintelligible characters appearing in the PITCH SEQ TX and
EVENT TX variables depict what happened during that particular at bat. From
looking at the pitch sequence variable PITCH SEQ TX, one sees that Mussina
quickly went ahead in the count as Long let a strike go by and swung and
missed another pitch (CS). Then Mussina followed with consecutive balls (BB)
and Long battled with a foul ball (F) before putting the ball in play (X).
The variable EVENT TX gives the results of the play. Long’s hit resulted in a
double, collected by the Yankees right fielder (D9 in the Event Text) in short
right (9S). The runner on first was thrown out on his way to home (1XH) by
a throw from right fielder Shane Spencer, relayed by shortstop Derek Jeter to
catcher Jorge Posada (962).
Once the event files are properly processed, many more fields are available
than the ones presented in Table 1.7. However these additional fields are, for
the most part, derived from what is in the table. For example, one additional
field indicates whether the at-bat resulted in a base hit, one field will identify
the fielder who collected the ball, and four fields will indicate where each
runner (and the batter) stood at the end of the play – all of this can be
inferred by the EVENT TX field.
This play-by-play information is available for most games going back to
1945, thus it is possible to recreate what happened on the field in the past
half-century. For this particular play, the Retrosheet event files cannot tell
us all of the interesting details. Derek Jeter came out of nowhere to cut off
TABLE 1.7
Excerpt of information available in Retrosheet Event files. Sample from Jeter’s
“Flip Play” (Oct. 13, 2001)
1. GAME ID OAK200110130
2. YEAR ID 2001
3. AWAY TEAM ID NYA
4. INN CT 7
5. BAT HOME ID 1
6. OUTS CT 2
7. BALLS CT 2
8. STRIKES CT 2
9. PITCH SEQ TX CSBBFX
10. AWAY SCORE CT 1
11. HOME SCORE CT 0
12. BAT ID longt002
13. BAT HAND CD L
14. PIT ID mussm001
15. PIT HAND CD R
16. POS2 FLD ID posaj001
17. POS3 FLD ID martt002
18. POS4 FLD ID soria001
19. POS5 FLD ID bross001
20. POS6 FLD ID jeted001
21. POS7 FLD ID knobc001
22. POS8 FLD ID willb002
23. POS9 FLD ID spens001
24. BASE1 RUN ID giamj002
25. BASE2 RUN ID NA
26. BASE3 RUN ID NA
27. EVENT TX D9/9S.1XH(962)
28. BAT FLD CD 7
29. BAT LINEUP ID 7
Spencer’s throw and flipped it backhand to Posada in time to nail Giambi at

home, on what has become known as “The Flip Play.”10

Below are some questions that can be explored with the Retrosheet event files.
These specific questions are about how batters perform in particular situations
in the pitch count and with runners on base.
Q During the McGwire/Sosa home run race, which player was
more successful at hitting homers with men on base?
10 In 2002, Baseball Weekly recognized “The Flip Play” as one of the ten most amazing
fielding plays of all time.

A Mark McGwire hit 37 home runs in 313 plate appearances with runners
on base, Sammy Sosa 29 in 367. Once walks (both intentional and unin-
tentional) and hit by pitches are removed, the number of opportunities
become 223 for McGwire and 317 for Sosa.
Q How many intentional walks in unusual situations (e.g., empty
bases or bases loaded) was Barry Bonds issued in his 73 HR
campaign?
A During his record 2001 season, Barry Bonds was passed intentionally only
35 times. Of those free passes one came with a runner on first and two
with runners on first and second. When he was awarded 120 intentional
walks in 2004, 19 came with nobody on, 11 with a runner on first, and 3
with runners on first and third. He was once walked intentionally with the
bases loaded in 1998.
Q What is the Major league batting average when the ball/strike
count is 0-2? What about on 2-0?
A In 2011, hitters compiled a .253 batting average on plate appearances
where they fell behind 0-2. Conversely they hit .479 after going ahead 2-0.
1.5 Pitch-by-Pitch Data

1.5.1 MLBAM Gameday and PITCHf/x
After having hit a combined 59 home runs in four seasons, Blue Jays right-
fielder Jose Bautista emerged in 2010 as an elite slugger by blasting a league
leading 54 balls out of the park. Figure 1.3 shows the location and the type
of the 54 pitches Jose sent into the stands.
Since 2005 baseball fans have had the opportunity to follow, pitch-by-pitch,
the games played by their favorite team on the Web thanks to the MLBAM11
Gameday application featured on the MLB.com website. For a couple of years
fans would only know the outcome of each pitch (whether it was a ball, a
called strike, a swinging strike, and so on). Starting from an October 2006
game played at the Metrodome in Minneapolis, a wealth of detail began to
appear for each pitch tracked on Gameday. Data on the release point, the
pitch speed, and its full trajectory, have been available for about one-third of
the games played in 2007. Starting from the 2008 season, nearly every MLB
pitch flight has been recorded by the PITCHf/x system.
PITCHf/x is a product by Sportvision, a company which produces broad-
cast effects for sports, such as the first-down virtual line for football and the
11 Major League Baseball Advanced Media
4
(ft. from ground)
vertical location
breaking ●
● fastball
3 ●
rulebook
●
● ● ● strike zone
●
● ●
●
2
−1 0 1
horizontal location
(ft. from center of plate
catcher's view)
FIGURE 1.3
Pitch type and location for Jose Bautista’s 54 home runs of the 2010 season.
FoxTrax hockey puck. Two cameras installed in each MLB park record the
flight of the baseball between the pitcher’s mound and home plate, and ad-
vanced software calculates the position, the velocity and the acceleration of
the ball, giving sufficient information to estimate the full trajectory of the ball
in its mound-to-plate trip. Raw PITCHf/x data can be accessed by anyone
from the MLB.com website, however its format (XML) might not be easy to
manage for the average reader. The data used for the examples in this book
are available in a format suitable for quick use inside R. In Appendix B we will
show how the XML package allows to manage data in XML format in R and
point to some free online resources that can be used to download PITCHf/x
data.
1.5.2 PITCHf/x Example

On April 21, 2012, Phil Humber became the 21st pitcher in Major League
Baseball history to throw a perfect game by retiring all the 27 batters he
faced. PITCHf/x captured his final pitch (like it has done for nearly every
other pitch thrown in MLB ballparks since 2008), providing the data shown
in Table 1.8. The outcome of the pitch (variable des) is recorded by a stringer,
while most of the remaining information is either captured by the Sportvision
system or calculated from the captured data.
Each pitch is assigned an identifier (sv id), that is actually a time stamp:
Humber’s final pitch was recorded on April 21, 2012, at 15:25:37. The key
information Sportvision obtains through its camera system is recorded in lines
11 through 19 of the table. Those nine parameters give the position (variables
x0, y0, z0), velocity (variables vx0, vy0, vz0), and acceleration (variables ax,
ay, az) components of the pitch at release point. With these nine parameters
the full trajectory of the pitch from release to home plate can be estimated. (In
fact, Sportvision actually estimates the parameters somewhere in the middle
of the ball’s flight, then derives the parameters at release point.)
While the nine parameters just mentioned are sufficient for learning about
the trajectory of the pitch, they are difficult to understand by casual fans who
follow the game on MLBAM Gameday.12 Other more descriptive quantitities
are calculated starting from those nine parameters. The one measure familiar
to baseball fans is the pitch speed at release, which for Humber’s final pitch
is calculated at 85.3 mph (variable start speed). PITCHf/x also provides
the speed of the ball as it crosses the plate, 79.1 mph in this case (variable
end speed). Another two important values are the variables px and pz; they
represent the horizontal and vertical location of the pitches, respectively, and
can be combined with the batter’s strike zone upper and lower limits (sz top
and sz bot) to infer whether the pitch crossed the strike zone.
Let’s focus on the location of this particular pitch. The horizontal reference
point is the middle of the plate, with positive values indicating pitches passing
on the right side of it from the umpire’s viewpoint. In this case the ball crossed
the plate 2.21 feet on the right of its midpoint. Since the plate is 17 inches
wide, it was way out of the strike zone. The pitch was also too low to be a
strike, as the vertical point at which crossed the plate is listed at 1.17 feet,
while the hitter’s lower limit of the strike zone is 1.74.13 Luckily for Humber,
since otherwise a walk would have ruined the perfect game, the home plate
umpire controversially declared that Brendan Ryan had swung the bat for
strike three.
Other interesting quantities about a pitch are available with PITCHf/x,
including the horizontal and vertical movement (variables pfx x and pfx z)
of the pitch trajectory, the spin direction, and its rate (variables spin dir
and spin rate).14 MLBAM has devised a complex algorithm which processes
12 Philip Humber’s perfect game can be relived, pitch by pitch, at mlb.mlb.com/mlb/
gameday/index.jsp?gid=2012_04_21_chamlb_seamlb_1&mode=gameday.
13 The batter’s strike zone boundaries are recorded by the human stringer at the beginning
of the at-bat, and thus are less precise than the pitch location coordinates recorded by the
advanced system.
14 Detailed explanations for the PITCHf/x fields have been provided by Mike Fast at
fastballs.wordpress.com/2007/08/02/glossary-of-the-gameday-pitch-fields/. Prof.
Alan Nathan provides a collection of PITCHf/x references at webusers.npl.illinois.edu/
~a-nathan/pob/pitchtracker.html.
TABLE 1.8
Excerpt of information available from PITCHf/x. Sample from the final pitch
of Phil Humber’s perfect game (Apr. 21, 2012)
1. des Swinging Strike (Blocked)
2. sv id 120421 152537
3. start speed 85.3
4. end speed 79.1
5. sz top 3.73
6. sz bot 1.74
7. pfx x 0.31
8. pfx z 1.81
9. px 2.211
10. pz 1.17
11. x0 -1.58
12. y0 50.0
13. z0 5.746
14. vx0 9.228
15. vy0 -124.71
16. vz0 -5.311
17. ax 0.483
18. ay 25.576
19. az -29.254
20. break y 23.8
21. break angle -4.1
22. break length 7.8
22. pitch type SL
23. spin dir 170.609
24. spin rate 344.307
the information captured by Sportvision and marks the pitch with a label
familiar to baseball fans. In this case the algorithm recognizes the pitch as a
slider (variable pitch type).

Below are questions you can answer with PITCHf/x data. The data can be
used to address specific questions about pitch type, speed of the pitch, and
play outcomes on specific pitches.
Q Who are the hitters who see the lowest and the highest percent-
age of fastballs?
A From 2008 to 2011 pitchers have thrown fastballs 35% of the time when
Ryan Howard was at the plate, 56% of the time when facing David Eck-
stein.
Q Who is the fastest pitcher in baseball currently?
A Nine of the fastest ten pitches recorded by PITCHf/x from 2008 to 2011
have been thrown by Aroldis Chapman, the highest figure being a 105.1
mph pitch thrown on September 24, 2010, in San Diego. Neftali Feliz is
the other pitcher making the top ten list with a 103.4 fastball delivered in
Kansas City on September 1, 2010.
Q What are the chances of a successful steal when the pitcher
throws a fastball compared to when a curve is delivered?
A From 2008 to 2010 baserunners were successful 73% of the times at stealing
second base on a fastball. The success rate increases to 85% when the pitch
is a curveball.
1.6 Summary
When choosing among the three main sources of baseball data (Lahman, Ret-
rosheet, and PITCHf/x), one always has to consider the trade-off between the
level of detail and the seasons covered by the source. With Lahman’s database,
for example, one can explore the evolution of the game since its beginnings
back into the nineteenth century. However only the basic season count statis-
tics are available from this source. For example, simple information such as
Babe Ruth’s batting splits by pitcher’s handedness cannot be retrieved from
Lahman’s files.
Retrosheet is steadily adding past seasons to its play-by-play database,
allowing researchers to perform studies to validate or reject common beliefs
about players of the past decades. During the years, for example, analysis
of play-by-play data has confirmed the huge defensive value of players like
Brooks Robinson and Mark Belanger, and has substantiated the greatness of
Roberto Clemente’s throwing arm.
PITCHf/x has been available only since 2008 and, contrary to Retrosheet,
there is no way to compile data for games of the past. This means we will
never be able to compare the velocity of Aroldis Chapman’s fastball to that
of Nolan Ryan or Bob Feller. However, studies performed since its inception
have provided an enhanced understanding of the game, enabling researchers
to explore issues like pitch sequencing, batter discipline, pitcher fatigue, and
the catcher’s ability to block bad pitches.
1.7 Further Reading

Schwarz (2005) provides a detailed history of baseball statistics. Adler (2006)
explains how to obtain baseball data from several sources, including Lahman’s
database, Retrosheet, and MLBAM Gameday and how to analyze them using
diverse tools, from Microsoft Excel to R, MySQL and PERL. Fast (2009)
introduces the PITCHf/x system to the uninitiated. The PITCHf/x, HITf/x,
FIELDf/x section of The Physics of Baseball website (Nathan, 2012) features
material on the subject of pitch tracking data.
1.8 Exercises
1. (Which Datafile?)
This chapter has given an overview of the Lahman database, the Ret-
rosheet game logs, the Retrosheet play-by-play files, and the PITCHf/x
database. Describe the relevant data among these four databases that can
be used to answer the following baseball questions.
(a) How has the rate of walks (per team for nine innings) changed over
the history of baseball?
(b) What fraction of baseball games in 1968 were shutouts? Compare this
fraction with the fraction of shutouts in the 2012 baseball season.
(c) What percentage of first pitches are strikes? If the count is 2-0, what
fraction of the pitches are strikes?
(d) Is it easier to steal second base or third base? (Compare the fraction
of successful steals of second base with the fraction of successful steals
of third base.)
2. (Lahman Pitching Data)

From the pitching data file from the Lahman database, the following in-
formation is collected about Bob Gibson’s famous 1968 season.
playerID yearID stint teamID lgID W L G GS CG SHO SV IPouts
gibsobo01 1968 1 SLN NL 22 9 34 34 28 13 0 914
H ER HR BB SO BAOpp ERA IBB WP HBP BK BFP GF R SH SF GIDP
198 38 11 62 268 0.18 1.12 6 4 7 0 1161 0 49 NA NA NA
(a) Gibson started 34 games for the Cardinals in 1968. What fraction of
these games were completed by Gibson?
(b) What was Gibson’s ratio of strikeouts to walks this season?
(c) One can compute Gibson’s innings pitched by dividing IPouts by

three. How many innings did Gibson pitch this season?
(d) A modern measure of pitching effectiveness is WHIP, the average
number of hits and walks allowed per inning. What was Gibson’s
WHIP for the 1968 season?
3. (Retrosheet Game Log)

Jim Bunning pitched a perfect game on Father’s Day on June 21, 1964.
Some details about this particular game can be found from the Retrosheet
game logs.
date day visitorteam visitorleague visitorgame hometeam
19640621 Sun PHI NL 60 NYN
homeleague homegame visitorscore homescore lengthgame daynight
NL 67 6 0 54 D
completioninfo forfeitinfo protestinfo parkid attendance timegame
NA NA NA NYC17 0 139
visitorlinescore homelinescore visitorab visitorh visitor2b
110004000 000000000 32 8 2
visitor3b visitorhr visitorrbi visitorsh visitorsf visitorhbp
0 1 6 2 0 0
visitorbb visitoribb visitork visitorsb visitorcs visitorgdp
4 0 6 0 1 0
visitorci visitorlob
0 5
(a) What was the time in hours and minutes of this particular game?
(b) Why is the attendance value in this record equal to zero?
(c) How many extra base hits did the Phillies have in this game? (We
know that the Mets had no extra base hits this game.)
(d) What was the Phillies’ on-base percentage in this game?
4. (Retrosheet Play-by-Play Record)

One of the famous plays in Philadelphia Phillies baseball history is second-
baseman Mickey Morandini’s unassisted triple play against the Pirates
on September 20, 1992.15 The following records from the Retrosheet
play-by-play database describe this half-inning. The variables indicate
the half-inning (variables INN CT and HOME ID), the current score (vari-
ables AWAY SCORE CT and HOME SCORE CT), the identities of the pitcher
and batter (variables BAT ID and PIT ID), the pitch sequence (variable
PITCH SEQ), the play event description (variable EVENT TEX), and the run-
ners on base (variables BASE1 RUN and BASE2 ID).
15 This play is described in detail at phillysportshistory.com/2011/09/20/.
INN_CT BAT_HOME_ID AWAY_SCORE_CT HOME_SCORE_CT BAT_ID PIT_ID

6 1 1 1 vansa001 schic002
6 1 1 1 bondb001 schic002
6 1 1 1 kingj001 schic002
PITCH_SEQ_TX EVENT_CD EVENT_TX BASE1_RUN_ID BASE2_RUN_ID
CBBBX 20 S9/L9M
C1BX 20 S/G56.1-2 vansa001
BLLBBX 2 4(B)4(2)4(1)/LTP/L4M bondb001 vansa001
Based on the records, write a short paragraph that describes the play-by-
play events of this particular inning.
5. (PITCHf/x Record of Several Pitches)
R. A. Dickey is one of the current pitchers who predominantly throws
a knuckleball. The following gives some PITCHf/x variables for the first
knuckleball and the first fastball that Dickey threw for a game against the
Kansas City Royals on April 13, 2013.
start_speed end_speed pfx_x pfx_z px pz sz_bot sz_top
73 66.3 -0.64 -7.58 -0.047 2.475 1.5 3.35
start_speed end_speed pfx_x pfx_z px pz sz_bot sz_top

81.2 75.4 -4.99 -7.67 -1.99 2.963 1.5 3.43
Describe the differences between the knuckleball and the fastball in terms
of pitch speed, movement (horizontal and vertical directions), and location
in the strike zone. Based on this data, why is a knuckleball so difficult for
a batter to make contact?
2
Introduction to R
CONTENTS
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Installing R and RStudio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.1 Career of Warren Spahn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.2 Vectors: defining and calculations . . . . . . . . . . . . . . . . . . . . . . . 32
2.3.3 Vector functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.4 Vector index and logical variables . . . . . . . . . . . . . . . . . . . . . . . 35
2.4 Objects and Containers in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4.1 Character data and matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4.2 Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.3 Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.5 Collection of R Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5.1 R scripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5.2 R functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.6 Reading and Writing Data in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.6.1 Importing data from a file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.6.2 Saving datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7 Data Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7.2 Manipulations with data frames . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.7.3 Merging and selecting from data frames . . . . . . . . . . . . . . . . 49
2.8 Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.9 Splitting, Applying, and Combining Data . . . . . . . . . . . . . . . . . . . . . . . 50
2.9.1 Using sapply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.9.2 Using ddply in the plyr package . . . . . . . . . . . . . . . . . . . . . . . 52
2.10 Getting Help . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.11 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.12 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
29
2.1 Introduction
In this chapter, we provide a general introduction to the R statistical system.
We describe the process of installing R and the program RStudio that provides
an attractive interface to the R system. Pitching data from the legend Warren
Spahn is used to motivate manipulations with vectors, a basic data struc-
ture. We describe different data types such as characters, factors, and lists,
and different “containers” for holding these different data types. The process
of executing collections of R commands by means of scripts and functions is
discussed, and methods for importing and exporting datasets from R are de-
scribed. A fundamental data structure in R is a data frame and we introduce
defining a data frame, performing manipulations, merging data frames, and
performing operations on a data frame split by values of a variable. We con-
clude the chapter by describing how to install and load R packages and how
one gets help using resources from the R system and the RStudio interface.
2.2 Installing R and RStudio

The R system (R Development Core Team (2013)) is available for
download from The Comprehensive R Archive Network (CRAN) at
www.r-project.org. R is available for Linux, Windows, and Macintosh sys-
tems; all of the commands described in this book will work in any of these
environments.
One can use R through the standard graphical user interface by launching
the R application. Recently, several new developmental environments have
been created for R, and we will demonstrate the RStudio environment (RStu-
dio (2013)) available from rstudio.org. In installation, one first installs R
and then the RStudio application, and then R is opened by launching the
RStudio application.
The RStudio opening screen is displayed in Figure 2.1. The screen is di-
vided into four windows. One can type commands directly and see output
in the lower-left Console window. Moving clockwise, the top-left window is a
blank file where one can write and execute R scripts or groups of instructions.
The top-right window shows names of objects such as vectors and data frames
created in an R session. By clicking on the History tab, one can see a record
of all commands entered during the current R session. Last, any plots are
displayed in the lower-right window. By clicking on the Files tab, one can see
a list of files stored in the current working directory. (This is the file directory
where R will expect to read files, and where any output, such as data files and
graphs, will be stored.) The Packages tab lists all of the R packages currently
Introduction to R 31
installed in the system and the Help tab will display documentation for R
functions and datasets.
FIGURE 2.1
The opening screen of the RStudio interface to R.
2.3 Vectors
2.3.1 Career of Warren Spahn
One of the authors collected the 1965 Warren Spahn baseball card pictured
in Figure 2.2. The back of the card, shown in Figure 2.3, displays many of
the standard pitching statistics for the seasons preceding Spahn’s final 1965
season. We use data from Spahn’s season statistics to illustrate some basic
components of the R system.
FIGURE 2.2
Front of the 1965 Warren Spahn Topps card.
2.3.2 Vectors: defining and calculations

A basic structure in R is a vector, a sequence of values of a given type, such
as numeric or character. A basic way of creating a vector is by means of the c
(combine) function. To illustrate, suppose we are interested in exploring the
games won and lost by Spahn for the seasons after the war when he played for
the Boston Braves. We create two vectors by use of the c function; the games
won are stored in the vector W and the games lost are stored in the vector L.
The symbol “<-” is the assignment character in R; the “=” symbol can also
be used for assignment. These lines can be directly typed into the Console
window. (By the way, R is case sensitive, so R will distinguish the vector L
from the vector l.)
W <- c(8, 21, 15, 21, 21, 22, 14)
L <- c(5, 10, 12, 14, 17, 14, 19)
One attractive feature of R is its ability to do element-by-element calculations

with vectors. Suppose we wish to compute the winning percentage for Spahn
for these seven seasons. We want to compute the fraction of winning games
and multiply this fraction by 100 to convert it to a percentage. We create a
new vector named Win.Pct by use of the basic multiplication (*) and division
(/) operators:
FIGURE 2.3
Back of the 1965 Warren Spahn Topps card.
Win.Pct <- 100 * W / (W + L)
We can display these winning percentages by simply typing the variable name:
Win.Pct
[1] 61.53846 67.74194 55.55556 60.00000 55.26316 61.11111 42.42424
A convenient way of creating patterned data is by use of the function seq.

We use this function to generate the season years from 1946 to 1952 and store
the output to the variable Year.1
Year <- seq(1946, 1952)
Year
[1] 1946 1947 1948 1949 1950 1951 1952
For a sequence of integer values, the colon notation will also work:
Year <- 1946 : 1952
Suppose we wish to calculate Spahn’s age for these seasons. Spahn was
born in April 1921 and we can compute his age by subtracting 1921 from each
season value – the resulting vector is stored in the variable Age.
Age <- Year - 1921
We construct a simple scatterplot of Spahn winning percentages (vertical)

against his age (horizontal) by use of the plot function (see Figure 2.4).
1 The function seq(a, b, s) will generate a vector of values from a to b in steps of s.
plot(Age, Win.Pct)
We see that Spahn was pretty successful for most of his Boston seasons – his
winning percentage exceeded 55% for six of his seven seasons.
●
65
●
●
60
●
Win.Pct
●
55
●
50
45
25 26 27 28 29 30 31
Age
FIGURE 2.4
Scatterplot of the winning percentage against age for Warren Spahn’s seasons
playing for the Boston Braves.
2.3.3 Vector functions

There are many built-in R functions for vectors including mean (arithmetic av-
erage), sd (standard deviation), length (number of vector entries), sum (sum
of values), max (maximum value), sort, and order. To illustrate a function,
one can use the mean function to find the average winning percentage of Spahn
during this seven-season period.
mean(Win.Pct)
[1] 57.66207
It is actually more common to compute a pitcher’s career winning percentage
by dividing his win total by the number of wins and losses. One can compute
this career winning percentage by means of the following R expression.
100 * sum(W) / (sum(W) + sum(L))

[1] 57.277
One can sort the win numbers from low to high by the sort function:
sort(W)
[1] 8 14 15 21 21 21 22
The cumsum function is useful for displaying cumulative totals of a vector

cumsum(W)
[1] 8 29 44 65 86 108 122
We see from the output that Spahn won 8 games in the first season, 29 games
in the first two seasons, and so on. The summary function applied on the
winning percentages displays several summary statistics of the vector values
such as the extremes (low and high values), the quartiles (first and third), the
median, and the mean.
summary(Win.Pct)
Min. 1st Qu. Median Mean 3rd Qu. Max.
42.42 55.41 60.00 57.66 61.32 67.74
This output tells us that his median winning percentage was 60, his mean
percentage was 57.66, and the entire group of winning percentages ranged
from 42.42 to 67.74.
2.3.4 Vector index and logical variables

To extract portions of vectors, the square is used. For example, the expression
W[c(1, 2, 5)]
[1] 8 21 21
will extract the first, second, and fifth entries of the vector W. The first four
values of the vector can be extracted by typing
W[1 : 4]
[1] 8 21 15 21
By use of a minus index, we remove entries from a vector. For example, if we

wish to remove the first and sixth entries of W, we would type
W[-c(1, 6)]
[1] 21 15 21 21 14
A logical variable is created in R by the use of a vector together with

the operations >, <, == (logical equal), and != (logical not equals). For
example, suppose we are interested in the values in the winning percentage
vector Win.Pct that exceed 60%.
Win.Pct > 60
[1] TRUE TRUE FALSE FALSE FALSE TRUE FALSE
The result of this calculation is a logical vector – the output indicates that
Spahn had a winning percentage exceeding 60% for the first, second, and sixth
seasons (TRUE), and not exceeding 60% for the remaining seasons (FALSE).
Were there any seasons where Spahn won more than 20 games and his winning
percentage exceeded 60%? We use the logical & (AND) operator to find the
years where W > 20 and Win.Pct > 60.
(W > 20) & (Win.Pct > 60)
[1] FALSE TRUE FALSE FALSE FALSE TRUE FALSE
The output indicates that both conditions were true for the second and sixth
seasons.
By using logical variables and the square bracket notion, we can find sub-
sets of vectors satisfying different conditions. During this period, when did
Spahn have his highest winning percentage? We use
Win.Pct == max(Win.Pct)
[1] FALSE TRUE FALSE FALSE FALSE FALSE FALSE
to create a logical vector which is true when this condition is satisfied. (Note
the use of the double equals sign notion to indicate logical equality.) Then we
select the corresponding year by indexing Year by this logical vector.
Year[Win.Pct == max(Win.Pct)]
[1] 1947
We see that the highest winning percentage occurred in 1947 during this
period.
What seasons did the number of decisions (wins plus losses) exceed 30?
We first create a logical vector based on W + L > 30, and then choose the
seasons using this logical vector.
Year[W + L > 30]
[1] 1947 1949 1950 1951 1952
We see that the number of decisions exceeded 30 for the five seasons 1947,
1949, 1950, 1951, and 1952.
2.4 Objects and Containers in R

The individual data components such as the years 1947, 1949, and 1950 are
called objects. These objects can be of different types such as numeric, logical,
character, and integer. We have already worked with objects of types numeric
and logical in the previous section. We store a number of objects into a con-
tainer. A vector is a simple type of container where we place a number of
objects of the same type, say objects that are all numeric or all logical. Here
we illustrate some of the different object types and containers that we find
useful in working with baseball data.
2.4.1 Character data and matrices

String variables such as the names of teams and players are stored as charac-
ters which are represented by letters and numbers enclosed by double quotes.
As a simple example, suppose we wish to explore information about the World
Series in the years 2003 through 2012. We create three character vectors NL,
AL, and Winner containing abbreviations for the National League winner, the
American League winner, and the league of the team that won the World Se-
ries. Note that we represent each character value by a string of letters enclosed
by double quotes. We also define two numeric vectors: N.Games contains the
number of games of each series, and Year gives the corresponding seasons.
NL <- c("FLA", "STL", "HOU", "STL", "COL",
"PHI", "PHI", "SFG", "STL", "SFG")
AL <- c("NYY", "BOS", "CHW", "DET", "BOS",
"TBR", "NYY", "TEX", "TEX", "DET")
Winner <- c("NL", "AL", "AL", "NL", "NL",
"NL", "AL", "NL", "NL", "NL")
N.Games <- c(6, 4, 4, 5, 4, 5, 6, 5, 7, 4)
Year <- 2003 : 2012
There are other ways to store objects besides vectors. For example, suppose
we wish to display the World Series contestants in a tabular format. A matrix
is a rectangular grid of objects of the same type. A matrix can be created by
the matrix function; the arguments are the objects to be put in the matrix
and the number of rows and the number of columns. By default the objects are
placed in the matrix by columns. Suppose we want to create a matrix with 10
rows and 2 columns with the National League contestants in the first column
and the American League contestants in the second column. We combine the
two team vectors into one vector by the c function and apply the matrix
function, storing the result in variable results.
results <- matrix(c(NL, AL), 10, 2)
results
[,1] [,2]
[1,] "FLA" "NYY"
[2,] "STL" "BOS"
[3,] "HOU" "CHW"
[4,] "STL" "DET"
[5,] "COL" "BOS"
[6,] "PHI" "TBR"
[7,] "PHI" "NYY"
[8,] "SFG" "TEX"

[9,] "STL" "TEX"
[10,] "SFG" "DET"
One can use the dimnames function to add descriptive labels to the rows and
columns of the matrix. We label the rows by the seasons using the first index
of dimnames(results) (note the use of the double square brackets) and label
the columns with "NL Team" and "AL Team". The matrix is now in a more
readable format.
dimnames(results)[[1]] <- Year
dimnames(results)[[2]] <- c("NL Team", "AL Team")
results
NL Team AL Team
2003 "FLA" "NYY"
2004 "STL" "BOS"
2005 "HOU" "CHW"
2006 "STL" "DET"
2007 "COL" "BOS"
2008 "PHI" "TBR"
2009 "PHI" "NYY"
2010 "SFG" "TEX"
2011 "STL" "TEX"
2012 "SFG" "DET"
There are a number of R functions available for exploring character data.
The table function will construct a frequency table for a vector of character
data. For example, to learn about the number of wins by each league in the
10 World Series, we use table with the variable Winner
table(Winner)
Winner
AL NL
3 7
It is interesting that the National League won 7 of these 10 World Series. If
we barplot the result from table
barplot(table(Winner))
we obtain a bar graph of the frequencies (see Figure 2.5).
2.4.2 Factors
A factor is a special way of representing character data. To motivate the con-
sideration of factors, suppose we construct a frequency table of the National
League representatives to the World Series in the character vector NL.
table(NL)
NL
COL FLA HOU PHI SFG STL
1 1 1 2 2 3
7
6
5
4
3
2
1
0
AL NL
FIGURE 2.5
Bar graph of the number of wins of the American League and National League
teams in the World Series between 2003 and 2012.
Note that R will organize the teams alphabetically (from COL to STL) in the
frequency table. It may be preferable to organize the teams by the division
(East, Central, and West). We can change the organization of the team labels
by converting this character type to a factor.
We make this conversion by means of the factor function. The basic
arguments are the vector of data to be converted and a vector levels that
gives the ordered values of the variable. Here we list the values ordered by the
East, Central, and West divisions. The result is stored in the factor variable
NL2.
NL2 <- factor(NL, levels=c("FLA", "PHI", "HOU", "STL", "COL", "SFG"))
One can understand how factor variables are stored by using the str function
to examine the structure of the variable NL2.
str(NL2)
Factor w/ 6 levels "FLA","PHI","HOU",..: 1 4 3 4 5 2 2 6 4 6
We see that a factor variable is actually encoded by integers (1, 4, 3, ...) where
the levels are the team names. If we table this new factor variable
table(NL2)
NL2
FLA PHI HOU STL COL SFG
1 2 1 3 1 2
we obtain the same frequencies as before, but the teams are now listed in the
order specified in the factor function.
Generally, we will see that R will automatically convert character-type
data to factors. Many R functions require the use of factors, and the use of
factors gives one finer control on how character labels are displayed in output
and graphs.
2.4.3 Lists
All of the containers we have described such as vectors and matrices require
that data values have the same type. For example, vectors contain all numeric
data or all character data; one cannot mix numeric and character data in a
single vector. A list is a convenient way of storing data of different types. To
illustrate, suppose we wish to collect the league that won the World Series (a
character type), the number of games played (a numeric type), and a short
description (a character type) into a single variable. Using the list function,
we create a new list World.Series with components Winner, Number.Games,
and Seasons.
World.Series <- list(Winner=Winner, Number.Games=N.Games,
Seasons="2003 to 2012")
Once a list such as World.Series is defined, there are different ways of

accessing the different components. If we wish to display the number of games
played Number.Games, we can use the list variable name together with with
the $ symbol and the component name.
World.Series$Number.Games
[1] 6 4 4 5 4 5 6 5 7 4
Or we can use the double square brackets to display the second component of
the list.
World.Series[[2]]
[1] 6 4 4 5 4 5 6 5 7 4
As an alternative, we can use the single square brackets with the name of the
component in quotes.
World.Series["Number.Games"]
$Number.Games
[1] 6 4 4 5 4 5 6 5 7 4
Note that the first two options return vectors and the third option returns a
list with a single component Number.Games.
Many R functions return a list of data of different types, so it is important

how to access the list components. Also we will see that lists provide a con-
venient way of collecting information of different types (character, numeric,
logical, factors) about teams and players.
2.5 Collection of R Commands

2.5.1 R scripts
The R expressions described in the previous sections can be typed directly in
the Console window and any output will be directly displayed in that window.
Alternatively, the R expressions can be stored in a text file, called an R script,
and executed as a group of commands.
Continuing our previous example, suppose we wish to run the following R
commands. We are entering the length and the winners of ten World Series,
tabulating and graphing the winner, and comparing the lengths of the series
won by the National and American Leagues. The by function is used to find
summaries of the numeric variable N.Games for each level of a categorical
variable Winner.
N.Games <- c(6, 4, 4, 5, 4, 5, 6, 5, 7, 4)
Winner <- c("NL", "AL", "AL", "NL", "NL",
"NL", "AL", "NL", "NL", "NL")
table(Winner)
barplot(table(Winner))
by(N.Games, Winner, summary)
A convenient way to run R scripts is through the text window in the upper-
left window of the RStudio environment. The R commands are typed in this
window and the script is executed by selecting these lines and pressing Control-
Enter (in a Windows operating system) or Command-Enter (in a Macintosh
operating system). The screenshot in Figure 2.6 shows the result of executing
this R script. The R output is displayed in the lower-left Command window.
In the Workspace window (upper-right), we see that two vectors N.Games and
Winner have been created. In the Plots window (lower-right), we see the bar
graph as a result of the barplot function.
Another way of running an R script is by saving the commands in a file,
and then “sourcing” this file within R. Suppose that a file with the above
commands has been saved in the file “World.Series.R” in the current working
directory. (Section 2.6.1 explains how one changes the working directory.)
Then one can execute this file by typing in the Console window:
source("World.Series.R", echo=TRUE)
The echo=TRUE argument is used so that the R output is displayed in the
Console window.
FIGURE 2.6
Snapshot of the RStudio interface after executing commands from an R script.
2.5.2 R functions
We have illustrated the use of a number of R built-in packages. One attractive
features of R is the capability to create one’s own function to implement
specific computations and graphs of interest.
As a simple example, suppose you are interested in writing a function to
compute a player’s home run rates for a collection of seasons. One inputs a
vector age of player ages, a vector hr of home run counts, and a vector ab of
at-bats. You want the function to compute the player’s home run rates (as a
percentage, rounded to the nearest tenth), and output the ages and rates in
a form amiable for graphing.
The following function hr.rates will perform the desired calculations. All
functions start with the syntax Name.of.function <- function(arguments),
where arguments is a list of input variables. All of the work in the function
goes inside the curly brackets that follow. The result of the last line of the
function is returned as the output. In our example, the name of the function
is hr.rates and there are three vector inputs age, hr, and ab. The round
function is used to compute the home run rates.2 The output of this function
2 The function round(x, n) rounds x to n decimal places.
is a list with two components: x is the vector of ages, and y is the vector of
home run rates.
hr.rates <- function(age, hr, ab){
rates <- round(100 * hr / ab, 1)
list(x=age, y=rates)
}
To use this function, first it needs to be read into R. This can be done by
entering it directly into the Console window, or by saving the function in a
file, say hr.rates.R, and reading it into R by the source function.
source("hr.rates.R")
We illustrate using this function on some home run data for Mickey Mantle for
the seasons 1951 to 1961. We enter Mantle’s home run counts in the vector HR,
the corresponding at-bats in AB, and the ages in Age. We apply the function
hr.rates with inputs Age, HR, AB, and the output is a list with Mantle’s
ages and corresponding home run rates.
HR <- c(13, 23, 21, 27, 37, 52, 34, 42, 31, 40, 54)
AB <- c(341, 549, 461, 543, 517, 533, 474, 519, 541, 527, 514)
Age <- 19 : 29
hr.rates(Age, HR, AB)
$x
[1] 19 20 21 22 23 24 25 26 27 28 29
$y
[1] 3.8 4.2 4.6 5.0 7.2 9.8 7.2 8.1 5.7 7.6 10.5
One can easily construct a scatterplot (not shown here) of Mantle’s rates
against age by the plot function on the output of the function.
plot(hr.rates(Age, HR, AB))
Note that Mantle’s home run rates rose steadily in the first six seasons of his
career.
2.6 Reading and Writing Data in R

2.6.1 Importing data from a file
Generally it is tedious to input data manually into R. For the large data files
that we will be working with in this book, it will be necessary to import these
files directly into R. We illustrate this importing process using the complete
pitching profile of Spahn.
We created the file “spahn.csv” containing Spahn’s pitching statistics and
placed the file in the current working directory. One can check the location of
the current working directory in R by means of typing in the Console window:
getwd()
On RStudio, one can change the working directory by selecting the “Change
Working Directory” option on the Tools menu or by use of the setwd func-
tion. One can easily import this dataset in RStudio by pressing the “Import
Dataset” button in the top right window. You select the “From Text File”
option and find the dataset of interest. After you select the file, Figure 2.7
shows a snapshot of the Import Dataset window. One sees the input file and
also the format of the data that will be saved into R. It is important to check
the button that the file contains a heading which means the first line of the
input file contains the variable names.
FIGURE 2.7
Snapshot of the Import Dataset window in the RStudio interface.
If the input file is in the current working directory, an alternative method of

importing data from a file uses the read.csv function. This function assumes
the file is stored in a “comma separated value” format, where different values
on a single row are separed by commas. For our example, the following R
expression reads the comma separated value file “spahn.csv” and saves the
data into a data frame with name spahn.
spahn <- read.csv("spahn.csv")
2.6.2 Saving datasets

We have seen that it is straightforward to read comma-delimited data files
(csv format) into R by use of the read.csv function. Similarly, we can use
the write.csv function to save datasets in R in the csv format.
As an example, we return to the Mickey Mantle example where we have
vectors of home run counts, at-bats, and ages and we use the user-defined
function hr.rates to compute home run rates. We create a matrix Mantle
attaching vectors Age, HR, AB, and the y component of the list HR.Rates
using the cbind function.3
HR <- c(13, 23, 21, 27, 37, 52, 34, 42, 31, 40, 54)
AB <- c(341, 549, 461, 543, 517, 533, 474, 519, 541, 527, 514)
Age <- 19 : 29
HR.Rates <- hr.rates(Age, HR, AB)
Mantle <- cbind(Age, HR, AB, Rates=HR.Rates$y)
We use the write.csv function to save the data to the current working
directory. This function has two arguments: the R object Mantle that we
wish to save, and the output file name “mantle.csv”. By using the argument
row.names = FALSE option, row names will be omitted in the file that is
saved.
write.csv(Mantle, "mantle.csv", row.names=FALSE)
It is good to confirm that a new file “mantle.csv” exists in the current working
directory.
2.7 Data Frames

2.7.1 Introduction
The variable spahn described in the previous section is an example of a data
frame, a fundamental variable type in R, very similar to the table of pitching
statistics that we saw on the back of Warren Spahn’s baseball card in Fig-
ure 2.3. A data frame is a rectangular table of data, where rows of the table
3 If one has three vectors a, b, c of equal lengths, the function cbind(a, b, c) combines
the vectors into a matrix where the vectors are columns of the matrix.
correspond to different individuals or seasons, and columns of the table cor-

respond to different variables collected on the individuals. Data variables can
be numerical (like a batting average or a winning percentage), integer (like
the count of home runs or number of wins), a factor (a categorical variable
such as the player’s team), or other types.
We can display portions of a data frame using the square bracket notation.
For example, if we wish to display the first five rows and the first four variables
(columns) of a data frame x, we type x[1 : 5, 1 : 4]. Below we display the
first three rows and columns 1 though 10 of the spahn data frame.
spahn[1 : 3, 1 : 10]
Year Age Tm Lg W L W.L ERA G GS
1 1942 21 BSN NL 0 0 NA 5.74 4 2
2 1946 25 BSN NL 8 5 0.615 2.94 24 16
3 1947 26 BSN NL 21 10 0.677 2.33 40 35
The header labels Year, Age, Tm, W, L, W.L, ERA, G, GS are some variable
names of the data frame; the numbers 1, 2, 3 displayed on the left give the row
numbers. We can display all variables for the first row by leaving the second
argument blank.
spahn[1, ]
Year Age Tm Lg W L W.L ERA G GS GF CG SHO SV IP H R ER HR BB
1 1942 21 BSN NL 0 0 NA 5.74 4 2 0 1 0 0 15.2 25 15 10 0 11
IBB SO HBP BK WP BF ERA. WHIP H.9 HR.9 BB.9 SO.9 SO.BB Awards
1 NA 7 0 0 0 79 59 2.298 14.4 0 6.3 4 0.64
A subset of the variables can be displayed by use of bracket notation. For

example, if we wish to display the variables Age, W, L, ERA for the first 10
seasons, we would write the code
spahn[1 : 10, c("Age", "W", "L", "ERA")]
Age W L ERA
1 21 0 0 5.74
2 25 8 5 2.94
3 26 21 10 2.33
4 27 15 12 3.71
5 28 21 14 3.07
6 29 21 17 3.16
7 30 22 14 2.98
8 31 14 19 2.98
9 32 23 7 2.10
10 33 21 12 3.14
Individual variables of a data frame can be accessed by means of the $

notation. For example spahn$ERA would contain the earned run averages for
Spahn’s seasons. We can obtain some descriptive statistics for this measure
by means of the summary function.
summary(spahn$ERA)
2.100 2.940 3.040 3.256 3.260 5.740
From this display, we see that 50% of Spahn’s season eras fell between the
lower quartile 2.940 and the upper quartile 3.260. Using logical operators, the
age when Spahn had his lowest ERA can be found by use of the following
expression.
spahn$Age[spahn$ERA == min(spahn$ERA)]
[1] 32
Using the ERA measure, Spahn had his best pitching season at the age of 32.
2.7.2 Manipulations with data frames

The pitching variables in the spahn data frame are the traditional or standard
pitching statistics. One can add new “sabermetric” variables to the data frame
by means of vector operations. Suppose that one wishes to measure pitching
by the FIP (fielding independent pitching) statistic4 defined by
13HR + 3BB − 2K
F IP = .
IP
We add a new variable to a current data frame using the $ convention. In the
following R code, the with function indicates that the variables HR, BB, SO,
and IP are understood in the environment of the spahn data frame.
spahn$FIP <- with(spahn, (13 * HR + 3 * BB - 2 * SO) / IP)
Suppose we are interested in finding the seasons where Spahn performed

the best using the FIP measure. The order function is used to give the po-
sitions of the measure where “1” corresponds to the smallest value; these
positions are stored in the vector pos. By using the bracket notation, we sort
the data frame using these positions. The head function is used to display
only the first few rows of the data frame.
pos <- order(spahn$FIP)
head(spahn[pos, c("Year", "Age", "W", "L", "ERA", "FIP")])
Year Age W L ERA FIP
8 1952 31 14 19 2.98 0.3448276
9 1953 32 23 7 2.10 0.3619910
2 1946 25 8 5 2.94 0.4153355
15 1959 38 21 15 2.96 0.6746575
3 1947 26 21 10 2.33 0.6950207
12 1956 35 20 11 2.78 0.8004269
4 FIP is a measure of pitching performance dependent only on plays that do not involve
fielders.
It is interesting that Spahn’s best FIP seasons occurred during the middle of
his career. Also, note that Spahn had a smaller (better) FIP in 1952 compared
to 1953, although his ERA was significantly larger in 1952.
Since Spahn pitched primarily for two cities, Boston and Milwaukee, sup-
pose we are interested in comparing his pitching for the two cities. We first
create a new data frame spahn1 containing only the statistics for the two
teams. This is done using the subset function with two arguments – the first
argument is the data frame to subset, and the second argument is the logical
condition defining the new data frame. (We introduce the logical OR operator
|.)
spahn1 <- subset(spahn, Tm == "BSN" | Tm == "MLN")
The current factor variable Tm has three possible values, “BSN,” “MLN,”
and “TOT” (for the total statistics for the 1965 season when Spahn played
for two teams). We redefine Tm using the factor function so that there are
only two possible values.
spahn1$Tm <- factor(spahn1$Tm, levels=c("BSN", "MLN"))
To compare various pitching statistics for the two teams, we use the by
function. The three arguments to by are the data frame to be summarized
(here the subset of the main data frame consisting of the variables W.L, ERA,
WHIP, and FIP), the grouping variable Tm, and the function (here summary)
that will be applied to each group. The output gives the summary statistics
for the Boston seasons and the Milwaukee seasons.
by(spahn1[, c("W.L", "ERA", "WHIP", "FIP")], spahn1$Tm, summary)
spahn1$Tm: BSN
W.L ERA WHIP FIP
Min. :0.4240 Min. :2.330 Min. :1.136 Min. :0.3448
1st Qu.:0.5545 1st Qu.:2.970 1st Qu.:1.154 1st Qu.:0.6251
Median :0.6000 Median :3.025 Median :1.222 Median :0.8219
Mean :0.5766 Mean :3.364 Mean :1.331 Mean :0.7922
3rd Qu.:0.6130 3rd Qu.:3.297 3rd Qu.:1.230 3rd Qu.:0.9836
Max. :0.6770 Max. :5.740 Max. :2.298 Max. :1.2500
NA’s :1
------------------------------------------------------------
spahn1$Tm: MLN
W.L ERA WHIP FIP
Min. :0.3160 Min. :2.100 Min. :1.058 Min. :0.3620
1st Qu.:0.5780 1st Qu.:2.757 1st Qu.:1.123 1st Qu.:0.8345
Median :0.6405 Median :3.030 Median :1.163 Median :0.9944
Mean :0.6202 Mean :3.121 Mean :1.187 Mean :0.9839
3rd Qu.:0.6695 3rd Qu.:3.170 3rd Qu.:1.226 3rd Qu.:1.0764
Max. :0.7670 Max. :5.290 Max. :1.474 Max. :1.7263
It is interesting that Spahn’s ERAs were higher in Boston (the middle 50%
between 2.970 and 3.297 in Boston, compared to the middle 50% between
2.757 and 3.170 in Milwaukee), but Spahn’s FIPs were lower in Boston. This
indicates that Spahn may have had a weaker defense or unlucky with hits in
ball in play in Boston.
2.7.3 Merging and selecting from data frames

In baseball research, it is common to have several data frames containing
batting and pitching data for teams. Here we describe several ways of merg-
ing data frames and extract a portion of a data frame that satisfies a given
condition.
Suppose we read into R data frames NLbatting and ALbatting containing
batting statistics for all National League and American League teams in the
2011 season. Suppose we wish to merge or combine these data frames into a
new data frame batting. To join two data frames vertically, we can use the
rbind (for row combine) function.
NLbatting <- read.csv("NLbatting.csv")
ALbatting <- read.csv("ALbatting.csv")
batting <- rbind(NLbatting, ALbatting)
This command assumes that the two data frames NLbatting and ALbatting
have the same variables; otherwise an error message will be displayed.
Suppose instead that we have read in the batting data NLbatting and the
pitching data NLpitching for the NL teams in the 2011 season and we wish to
merge these data frames horizontally. In this case, a row of the merged data
frame would contain the batting and pitching statistics for a particular team.
In this case, we use the function merge where we specify the two data frames
and the by argument indicates the common variable (Tm) to merge by.
NLpitching <- read.csv("NLpitching.csv")
NL <- merge(NLbatting, NLpitching, by="Tm")
The new data frame NL contains 16 (the number of NL teams) rows and all
of the variables from both the NLbatting and ALbatting data frames.
A third useful operation is choosing a subset of a data frame that satisfies
a particular condition. Suppose one has the data frame NLbatting and one
wishes to focus on the batting statistics for only the teams who hit over 150
home runs this season. We use the subset function – the arguments are the
original data frame and the logical condition that describes how teams are
selected.
NL.150 <- subset(NLbatting, HR > 150)
The new data frame NL.150 contains the batting statistics for the eight teams
who hit over 150 home runs.
2.8 Packages
This book will focus on functions available on the base R system that is
installed. But one attractive feature of R is the availability of collections of
functions and datasets in R packages. Currently, there are over 4000 packages
contributed by R users available on the R website cran.r-project.org/ and
these packages expand the capabilities of the R system. In our book, we focus
on a few contributed packages that we find useful in our baseball work.
To illustrate installing and loading an R package, we recently found a pack-
age Lahman, that contains the data files from the Lahman database described
in Section 1.2. Assuming one is connected to the Internet, one can install the
current version of this package into R by means of the command
install.packages("Lahman")
Alternately, one can install packages by use of the Install Packages button on
the Package tab in RStudio.
After a package has been installed, then one needs to load the package into
R to have access to the functions and datasets. For example, to load the new
package Lahman, one types
library(Lahman)
To confirm that the package has been loaded correctly, we use the help func-
tion to learn about the dataset Batting in the Lahman package. (A general
discussion of the help function is given in Section 2.10.)
?Batting
When one launches R, one needs to load the packages that are not automati-
cally loaded in the system.
2.9 Splitting, Applying, and Combining Data

In many situations, one is interested in splitting a data frame into parts,
applying some operation on each part, and then combining the results in a
new data frame. This type of “split, apply, combine” operation is facilitated
using the function sapply and special functions in the plyr package. Here we
illustrate this process on the Lahman batting database. In this work, we review
some other handy data frame manipulation functions previously discussed.
2.9.1 Using sapply

Suppose we are interested in looking at the great home run hitters in baseball
history. Specifically, we want to answer the question “Who hit the most home
runs in the 1960s?”
We begin by reading in Lahman’s batting data file “Batting.csv” contain-
ing the season batting statistics for all players in baseball history – the data
frame is stored in the variable Batting.
Batting <- read.csv("Batting.csv")
Since we are focusing on the 1960s, the subset function is used to select
batting data only for the seasons between 1960 and 1969, creating the new
data frame Batting.60.
Batting.60 <- subset(Batting, yearID >= 1960 & yearID <= 1969)
The sapply function is useful for repeating a particular operation over a

set of values in a vector. In this example, we would like to compute the total
number of home runs for each player in the data frame Batting.60. First,
a function compute.hr is written that computes the total home runs for a
player with playerID equal to pid.
compute.hr <- function(pid){
d <- subset(Batting.60, playerID == pid)
sum(d$HR)
}
By use of the unique function, a vector of the ids for all of the players
in the 1960s is created. The home runs for all players is accomplished by the
sapply function – the arguments are the vector players and the function
compute.hr that will be applied to each element in the vector. The output is
a vector S containing the total home run count for all players in the vector.
players <- unique(Batting.60$playerID)
S <- sapply(players, compute.hr)
A new data frame R is created using the data.frame function. The syntax
indicates there are two variables in the new data frame – Player correspond
to the player ids contained in the vector players and HR corresponds to the
home run counts contained in the vector S. Using the order function, we sort
this data frame so that the best home run hitters are on the top, and display
the first lines of this data frame.
R <- data.frame(Player=players, HR=S)
R <- R[order(R$HR, decreasing=TRUE), ]
head(R)
players S
857 killeha01 393
1 aaronha01 375
1045 mayswi01 350

1373 robinfr02 316
1058 mccovwi01 300
752 howarfr01 288
The best home run hitters in the 1960s were Harmon Killebrew, Hank Aaron,
Willie Mays, and Frank Robinson.
2.9.2 Using ddply in the plyr package

The plyr package provides a more extensive collection of tools for this “split,
apply, combine” operation. Using the same Batting data frame of season
batting statistics, suppose we are interested in collecting the career at-bats,
career home runs, and career strikeouts for all players in baseball history with
at least 5000 career at-bats. Both home runs and strikeouts are of interest
since we suspect there may be some association between a player’s strikeout
rate (defined by SO/AB) and his home run rate HR/AB.
A new data frame consisting of hitting data of batters with at least 5000
career at-bats will be created. This operation is done in three steps. First, a
new data frame is created consisting of the career AB for all batters. Second,
this new data frame is merged with the original data frame Batting, creating
a new variable Career.AB. Last, by use of the subset function, the batting
seasons are selected from the data frame for the players with 5000 AB.
The function ddply in the plyr package is useful for the first operation.
We want to compute the sum of AB over the seasons of a player’s career.
There are three arguments to the ddply function. Batting is the data frame
we wish to split, .(player.ID) indicates we wish to split the data frame by
the player id variable, and summarize, Career.AB=sum(AB, na.rm=TRUE)
indicates we wish to summarize each data frame “part” by computing the
sum of the AB. (Some of the AB values will be missing and coded as “NA”,
and the na.rm=TRUE will remove these missing values before taking the sum.)
The new data frame dataframe.AB contains the career AB for all players.
library(plyr)
dataframe.AB <- ddply(Batting, .(playerID), summarize,
Career.AB=sum(AB, na.rm=TRUE))
We want to add a new variable Career.AB to the original data frame.

This is done by use of the merge function, merging data frames Batting and
dataframe.AB, matching by the common variable playerID.
Batting <- merge(Batting, dataframe.AB, by="playerID")
Now that we have this new variable Career.AB, one can now use the subset
function to choose only the season batting statistics for the players with 5000
AB.
Batting.5000 <- subset(Batting, Career.AB >= 5000)
For each player in the data frame Batting.5000, we want to compute the
career AB, career HR, and career SO. This is another example of the “split,
apply, combine” operation done conveniently by the ddply function. We first
write a small function ab.hr.so that works on the season batting statistics
for a single player. The input is the data frame d containing the statistics for
one player and the output is a data frame with the career AB, HR, and SO.
ab.hr.so <- function(d){
c.AB <- sum(d$AB, na.rm=TRUE)
c.HR <- sum(d$HR, na.rm=TRUE)
c.SO <- sum(d$SO, na.rm=TRUE)
data.frame(AB=c.AB, HR=c.HR, SO=c.SO)
}
To illustrate the use of ab.hr.so, we extract Hank Aaron’s batting statistics
and apply this function on Aaron’s data frame aaron.
aaron <- subset(Batting.5000, playerID == "aaronha01")
ab.hr.so(aaron)
AB HR SO
1 12364 755 1383
This confirms that Aaron had 755 career home runs and 1383 career strikeouts.
To apply this function to each batter and collect the results, we again use
the function ddply. The arguments are the data frame Batting.5000 to split,
the splitting variable player.ID, and the function ab.hr.so to apply on each
part.
d.5000 <- ddply(Batting.5000, .(playerID), ab.hr.so)
The resulting data frame d.5000 contains the career AB, HR, and SO for all
batters with at least 5000 career AB. To confirm, the first six lines of the data
frame are displayed by the head function.
head(d.5000)
playerID AB HR SO
1 aaronha01 12364 755 1383
2 abreubo01 8128 284 1763
3 adamssp01 5557 9 223
4 adcocjo01 6606 336 1059
5 alfoned01 5385 146 617
6 allendi01 6332 351 1556
Is there an association between a player’s home run rate and his strike-
out rate? Using the plot function, we construct a scatterplot of HR/AB and
SO/AB. Using the lines function we add a smoothing curve (using the function
lowess) to the scatterplot. (See Figure 2.8.)
with(d.5000, plot(HR/AB, SO/AB))
with(d.5000, lines(lowess(HR/AB, SO/AB)))
It is clear from the graph that batters with higher home run rates tend to
have higher strikeout rates.
●
0.30
●
● ●
● ●●
●
● ● ●
0.25
● ● ● ●
●● ●● ●
● ● ●●
● ● ● ●
● ● ●●
● ● ●● ●
● ●
● ● ● ● ● ● ●
● ●● ● ● ●●
● ●●
●
0.20
● ● ● ●●●
● ● ● ● ● ● ●● ●
●● ● ●● ● ● ●
●●● ● ● ●
● ● ● ●
● ●
● ● ● ● ●● ● ● ● ● ●●
SO/AB
●● ●● ●
●
●●● ● ●● ● ● ● ●●
●
●
●●
● ● ●
●
●● ● ● ● ●●● ● ● ● ● ● ● ●●
● ●
● ●
● ● ● ●● ●● ●● ● ● ●●
●● ● ● ●● ●● ● ●●●●● ●● ● ●●●
0.15
● ● ●
●● ●● ● ●●● ● ● ● ●● ● ●● ● ● ●
● ● ●●● ● ● ● ● ●●
● ● ● ● ● ● ●●● ● ● ● ● ● ●● ●● ● ●● ● ●
● ●
●● ● ●● ● ● ● ● ● ●
● ● ●● ● ● ● ● ● ●● ●● ●● ●
● ●
● ● ●●
● ●
●
● ●●● ●●●● ● ● ● ● ●
●
● ●
● ● ● ● ●● ●
● ●●● ●● ●● ●●● ● ● ●
● ●● ● ● ● ● ● ● ●
●● ●●● ● ●● ●●●●●●
● ● ● ●
●●● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ●
0.10
● ●●● ●● ●
●● ● ●● ● ● ●●● ●
● ● ●● ●● ● ●● ●● ●●●● ● ●● ● ● ● ●● ●
● ●●● ●● ●
● ● ● ● ● ●
●● ● ● ●● ● ●●●● ●
●
●● ●● ● ● ●●●
● ●
●
● ●
●● ● ●● ● ● ●
●● ●
● ●
● ● ●● ●
●● ●●●●● ●● ●●●●● ●● ● ● ●● ●● ●
● ●● ● ●●● ●●
● ●
● ● ●●
●
●
●●●● ●●
●●●● ●●●●●● ● ●
●
● ●● ●●●●● ● ●● ● ●
●● ● ●● ● ● ●
●●● ●
●
0.05
●●● ● ● ●●●● ●●● ● ●

●● ● ● ●● ●● ● ●
● ●
● ●●● ●● ● ●
● ●●
●● ●●
● ●●●●●●●
● ●●●
●●● ● ●
●● ●●
●
●● ●● ●●●
● ●● ●
●
●
●
●●
●
●●
●●●●●● ●●●
●●● ●●
●●●●●●● ●● ●●
●● ●● ●●
●
●● ●●● ●
0.00
●● ●● ● ● ●●●●
● ● ●●●● ●
0.00 0.02 0.04 0.06 0.08
HR/AB
FIGURE 2.8
Scatterplot of the homerun rates and strikeout rates of all players with at
least 5000 career at-bats. A smoothing curve is added to the plot to show that
home run rates and strikeout rates have a positive association.
2.10 Getting Help

The Help menu in RStudio provides general documentation about the R sys-
tem (see the R Help option). From the Help menu, we also find general infor-
mation about the RStudio system such as keyboard shortcuts. In addition, R
contains an online help system providing documentation on R functions and
datasets. For example, suppose you wish to learn about the dotchart func-
tion which constructs a dot chart, a new type of statistical graphical display
discussed in Chapter 3. By typing in the Console window a question mark
followed by the function name,
?dotchart
you see a long description of this function including all of the possible function
arguments. To find out about related functions, one can preface “dotchart”
by two question marks to find all objects that contain this character string:
??dotchart
On my system, I learn about the function dotchart2 in the Hmisc package

that constructs an enhanced dot chart display.
RStudio provides an additional online help system which is especially help-
ful when one does not know the exact spelling of an R function. For example,
suppose I want to construct a dot chart, but all I know is that the function
contains the string “dot”. In the Console window, I type “dot” followed with
a Tab. RStudio will complete the code, forming “dotplot” and show an ab-
breviated description of the function. In the case where the character string
does not uniquely define the function, RStudio will display all of the functions
with that string.
2.11 Further Reading

R is an increasingly popular system for performing data analysis and graphics
and a large number of books are available which introduce the system. The
manual “An Introduction to R” (Venables et al., 2011) available on the R and
RStudio systems provides a broad overview of the R language and the manual
“R Data Import/Export” provides an extended description of R capabilities
to import and export datasets. Kabacoff (2011) and the accompanying web-
site www.statmethods.net provide helpful advice on specific R functions on
data input, data management, and graphics. Albert and Rizzo (2012) provide
an example-based introduction to R, where different chapters are devoted to
specific statistics topics such as exploratory fitting, modeling, graphics, and
simulation.
2.12 Exercises
1. (Top Base Stealers in the Hall of Fame)
The following table gives the number of stolen bases (SB), the number of
times caught stealing (CS), and the number of games played (G) for nine
players currently inducted in the Hall of Fame.
Player SB CS G
Rickey Henderson 1406 335 3081
Lou Brock 938 307 2616
Ty Cobb 897 212 3034
Eddie Collins 741 195 2826
Max Carey 738 109 2476
Joe Morgan 689 162 2649
Luis Aparicio 506 136 2599
Paul Molitor 504 131 2683
Roberto Alomar 474 114 2379
(a) In R, place the stolen base, caught stealing, and game counts in the
vectors SB, CS, and G.
(b) For all players, compute the number of stolen base attempts SB + CS
and store in the vector SB.Attempt.
(c) For all players, compute the success rate Success.Rate = SB /
SB.Attempt.
(d) Compute the number of stolen bases per game SB.Game = SB /
Game.
(e) Construct a scatterplot of the stolen bases per game against the suc-
cess rates. Are there particular players with unusually high or low
stolen base success rates? Which player had the greatest number of
stolen bases per game?
2. (Character, Factor, and Logical Variables in R)

Suppose one records the outcomes of a batter in ten plate appearances:
Single, Out, Out, Single, Out, Double, Out, Walk, Out, Single
(a) Use the c function to collect these outcomes in a character vector

outcomes.
(b) Use the table function to construct a frequency table of outcomes.
(c) In tabulating these results, suppose one prefers the results to be
ordered from least-successful to most-successful. Use the following
code to convert the character vector outcomes to a factor variable
f.outcomes.
f.outcomes <- factor(outcomes,
levels=c("Out", "Walk", "Single", "Double"))
Use the table function to tabulate the values in f.outcomes. How
does the output differ from what you saw in part (b)?
(d) Suppose you want to focus only on the walks in the plate appearances.
Describe what is done in each of the following statements.
outcomes == "Walk"
sum(outcomes == "Walk")
3. (Pitchers in the 350 Wins Club)

The following table lists nine pitchers who have won at least 350 career
wins.
Player W L SO BB
Pete Alexander 373 208 2198 951
Roger Clemens 354 184 4672 1580
Pud Galvin 364 310 1806 745
Walter Johnson 417 279 3509 1363
Greg Maddux 355 227 3371 999
Christy Mathewson 373 188 2502 844
Kid Nichols 361 208 1868 1268
Warren Spahn 363 245 2583 1434
Cy Young 511 316 2803 1217
(a) In R, place the wins and losses in the vectors W and L, respectively.
Also, create a character vector Name containing the last names of
these pitchers.
(b) Compute the winning percentage for all pitchers defined by 100 ×
W/(W +L) and put these winning percentages in the vector Win.PCT.
(c) By use of the command
Wins.350 <- data.frame(Name, W, L, Win.PCT)
create a data frame Wins.350 containing the names, wins, losses, and
winning percentages.
(d) By use of the order function, sort the data frame Wins.350 by win-
ning percentage. Among these pitchers, who had the largest and
smallest winning percentages?
4. (Pitchers in the 350 Wins Club, Continued)
(a) In R, place the strikeout and walk totals from the 350 win pitchers
in the vectors SO and BB, respectively. Also, create a character vector
Name containing the last names of these pitchers.
(b) Compute the strikeout-walk ratio by SO/BB and put these ratios in
the vector SO.BB.Ratio.
(c) By use of the command
SO.BB <- data.frame(Name, SO, BB, SO.BB.Ratio)
create a data frame SO.BB containing the names, strikeouts, walks,
and strikeout-walk ratios.
(d) By use of the subset function, find the pitchers who had a strikeout-
walk ratio exceeding 2.8.
(e) By use of the order function, sort the data frame by the number of
walks. Did the pitcher with the largest number of walks have a high
or low strikeout-walk ratio?
5. (Pitcher Strikeout/Walk Ratios)
(a) Read the Lahman “pitching.csv” data file into R into a data frame
Pitching.
(b) The following function computes the cumulative strikeouts, cumula-
tive walks, mid career year, and the total innings pitched (measured
in terms of outs) for a pitcher whose season statistics are stored in
the data frame d.
stats <- function(d){
c.SO <- sum(d$SO, na.rm=TRUE)
c.BB <- sum(d$BB, na.rm=TRUE)
c.IPouts <- sum(d$IPouts, na.rm=TRUE)
c.midYear <- median(d$yearID, na.rm=TRUE)
data.frame(SO=c.SO, BB=c.BB, IPouts=c.IPouts,
midYear=c.midYear)
}
Using the function ddply (plyr package) together with the function
stats, find the career statistics for all pitchers in the pitching dataset.
Call this new data frame career.pitching.
(c) Use the merge function to merge the Pitching and career.pitching
data frames.
(d) Use the subset function to construct a new data frame career.10000
consisting of data for only those pitchers with at least 10,000 career
IPouts.
(e) For the pitchers with at least 10,000 career IPouts, construct a scat-
terplot of mid career year and ratio of strikeouts to walks. Comment
on the general pattern in this scatterplot.
3
Traditional Graphics
CONTENTS
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2 Factor Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.1 A bar graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.2 Add axes labels and a title . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.2.3 Other graphs of a factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3 Saving Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.4 Dot plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.5 Numeric Variable: Stripchart and Histogram . . . . . . . . . . . . . . . . . . . . 65
3.6 Two Numeric Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.6.1 Scatterplot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.6.2 Building a graph, step-by-step . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.7 A Numeric Variable and a Factor Variable . . . . . . . . . . . . . . . . . . . . . . 73
3.7.1 Parallel stripcharts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.7.2 Parallel boxplots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.8 Comparing Ruth, Aaron, Bonds, and A-Rod . . . . . . . . . . . . . . . . . . . . 76
3.8.1 Getting the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.8.2 Creating the player data frames . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.8.3 Constructing the graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.9 The 1998 Home Run Race . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.9.1 Getting the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.9.2 Extracting the variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.9.3 Constructing the graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.10 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.1 Introduction
To illustrate basic methods for creating graphs in R in the graphics package,
consider all the career batting statistics for the current members of the Hall of
Fame. If we remove the pitchers’ batting statistics from the dataset, then we
have statistics for 147 non-pitchers. The data file “hofbatting.csv” contains
59
the career batting statistics for this group. We read this data file into R by
the read.csv function; the statistics are stored in a data frame named hof.
hof <- read.csv("hofbatting.csv")
The type of graph we use depends on the measurement scale of the variable.
There are two fundamental data types – measurement and categorical – which
are represented in R as numeric and factor variables. We initially describe
graphs for a single factor variable and a single numeric variable, and then
describe graphical displays helpful for understanding relationships between
the variables. Using the traditional graphics package of R, it is easy to modify
the attributes of a graph by adding labels and changing the style of plotting
symbols and lines. After describing the graphical methods, we describe the
process of creating graphs for two home run stories. The first graph compares
the home run career progress of four great sluggers in baseball history and
the second graph illustrates the famous home run race of Mark McGwire and
Sammy Sosa during the 1998 season.
3.2 Factor Variable

3.2.1 A bar graph
The Hall-of-Famers played during different eras of baseball; one common clas-
sification of eras is “19th Century” (up to the 1900 season), “Dead Ball” (1901
through 1919), “Lively Ball” (1920 though 1941), “Integration” (1942 through
1960), “Expansion” (1961 through 1976), “Free Agency” (1977 through 1993),
and “Long Ball” (after 1993). We want to create a new factor variable Era
giving the era for each player. First a player’s mid career (variable MidCareer)
is defined as the average of his first and last seasons in baseball. The cut func-
tion creates the new factor variable Era – the arguments of the function are
the numeric variable to be converted, the vector of cut points, and the vector
of labels for the categories of the factor variable.
hof$MidCareer <- with(hof, (From + To) / 2)
hof$Era <- cut(hof$MidCareer,
breaks=c(1800, 1900, 1919, 1941, 1960, 1976, 1993, 2050),
labels=c("19th Century", "Lively Ball", "Dead Ball",
"Integration", "Expansion", "Free Agency",
"Long Ball"))
A bar graph of a factor variable is constructed using the barplot func-
tion. We first construct a frequency table of the variable Era using the table
function and store the output into the variable T.Era. The output of table is
the input for barplot. Figure 3.1 (a) shows the resulting graph. We see that
a large number of these Hall of Fame players played during the Lively Ball
era.
Traditional Graphics 61
T.Era <- table(hof$Era)

T.Era
19th Century Dead Ball Lively Ball Integration Expansion

17 19 46 24 21
Free Agency Long Ball
18 2
barplot(T.Era)
Era of the Nonpitching Hall of Famers

40
40
30
30
Frequency
20
20
10
10
0
19th Century Lively Ball Expansion Long Ball 19th Century Lively Ball Expansion Long Ball
Era
(a) (b)
FIGURE 3.1
Bar graphs of era of the Hall of Fame non-pitchers. The right graph adds axes
labels and a title to the basic plot.
3.2.2 Add axes labels and a title

As good practice, graphs should have descriptive axes labels and a title for
describing the main message of the display. In the traditional graphics displays
in R, the arguments xlab and ylab add horizontal and vertical axis labels and
the main argument adds a title. In the following barplot function, we add
the labels “Era” and “Frequency” and add the title “Era of the Nonpitching
Hall of Famers.” The enhanced plot is shown in Figure 3.1(b).
barplot(table(hof$Era), xlab="Era", ylab="Frequency",
main="Era of the Nonpitching Hall of Famers")
3.2.3 Other graphs of a factor

There are alternative graphical displays for a table of frequencies of a factor
variable. For the table of era frequencies, the function
plot(table(hof$Era))
draws a vertical line graph of the frequencies and the function

pie(table(hof$Era))
constructs a pie chart. Figure 3.2 shows these alternative displays A line graph
Dead Ball
40
Lively Ball
19th Century
30
T
Long Ball
20
Free Agency
10
Integration
Expansion
0
19th Century Lively Ball Expansion Long Ball
(a) (b)
FIGURE 3.2
Line graph and pie chart of the frequencies of era of the Hall of Fame non-
pitchers.
is helpful when there are a large number of categories of the factor. Although
pie charts are popular in displaying frequencies in the media, we prefer a bar
graph since it can be more difficult for a reader to visually compare the relative
sizes of slices of a pie chart than lengths of bars in a bar chart.
3.3 Saving Graphs

After a graph is produced in R, it is straightforward to export the graph to one
of the the usual graphics formats so that it can be used in a document, blog,
or website. We outline the steps for saving graphs in the RStudio interface.
If a graph appears in the Plots window of RStudio, then the Export menu
allows one to “Save Plot as Image,” “Save Plot as PDF,” or “Copy Plot to the
Clipboard.” If one chooses the “Save Plot as Image” option, then by choosing
an option from a drop-down menu, one can save the graph in PNG, JPEG,
TIFF, BMP, metafile, clipboard, SVG, or EPS formats. The PNG format is
convenient for uploading to a web page and the EPS and PDF formats are
well-suited for use in a LaTex document. The metafile and clipboard options
are useful for insertion of the graph into a Microsoft Word document.
Alternately, plots can be saved by use of R functions typed in the Console
window. For example, suppose we wish to save the bar graph shown in Figure
3.1(b) in a graphics file of PNG format. We use the special png function where
the argument is the name of the saved graphics file. We follow this with the
R commands to produce the graph, and conclude with the dev.off function.
(Note that no graph will be displayed in this operation.)
png("bargraph.png")
barplot(table(hof$Era), xlab="Era", ylab="Frequency",
main="Era of the Nonpitching Hall of Famers")
dev.off()
RStudioGD
2
If we look at the current directory, we will see a new file “bargraph.png”

containing the image in PNG format. If one types the help command
?png
one will see command instructions for saving graphs in other graphics formats.
This method of saving graphs is especially useful if one wishes to save a number
of graphs in a single file. For example, if one types
pdf("graphs.pdf")
barplot(table(hof$Era))
plot(table(hof$Era))
dev.off()
RStudioGD
2
then the bar graph and the line graph will be saved together in the PDF file
“graphs.pdf.”
3.4 Dot plots

A modern method of displaying labeled data is a dot plot, introduced by
Cleveland (1994). This is a useful alternative to a bar graph, and is nice for
displaying statistics for a group of baseball players.
We revisit the example where we are exploring the era for the Hall of Fame
non-pitchers. The table of era frequencies is stored in the variable T.Era. We

first convert T.Era to a numeric variable by use of the as.numeric function.
To construct a dot plot, we use the dotchart function – the first argument is
the vector of numeric values to display and the label argument is a vector of
the corresponding labels.
T.Era <- table(hof$Era)
dotchart(as.numeric(T.Era), labels=names(T.Era), xlab="Frequency")
The dotplot, shown in Figure 3.3 simply represents each frequency by an open
circle on a scale against the corresponding label. It is easy to compare the era
frequencies from this display.
Long Ball ●
Free Agency ●
Expansion ●
Integration ●
Lively Ball ●
Dead Ball ●
19th Century ●
10 20 30 40
Frequency
FIGURE 3.3
Dot plot of era of the Hall of Fame non-pitchers.
Dot plots can be used to display any collection of labeled numeric data.
Suppose we are interested in exploring the career OPS values for the Hall of
Fame players with at least 500 career home runs. We first use the subset
function to create a new data frame hof.500 consisting of the statistics of
the players with at least 500 career home runs. By use of the order function,
we order the rows of this matrix by the career OPS. We use the dotchart
function to construct a dot plot of the OPS values where the labels are the
names of the players (variable X in the data frame hof.500).
hof.500 <- subset(hof, HR >= 500)
hof.500 <- hof.500[order(hof.500$OPS), ]
dotchart(hof.500$OPS, labels=hof.500$X, xlab="OPS")
In this display (Figure 3.4), we see that Babe Ruth, Ted Williams, and Jimmie
Foxx stand out as the top OPS players in this 500-home run group.
Babe Ruth HOF ●
Ted Williams HOF ●
Jimmie Foxx HOF ●
Mickey Mantle HOF ●
Mel Ott HOF ●
Willie Mays HOF ●
Hank Aaron HOF ●
Frank Robinson HOF ●
Mike Schmidt HOF ●
Willie McCovey HOF ●
Eddie Mathews HOF ●
Harmon Killebrew HOF ●
Reggie Jackson HOF ●
Eddie Murray HOF ●
Ernie Banks HOF ●
0.85 0.90 0.95 1.00 1.05 1.10 1.15
OPS
FIGURE 3.4
Dot plot of career OPS values for Hall of Famers with at least 500 career home
runs.
3.5 Numeric Variable: Stripchart and Histogram

When one collects a numeric variable such as a batting average, an on-base
percentage, or an OPS from a group of players, one typically wants to learn
about its distribution. For example, if we examine OPS values for the non-
pitcher HOF inductees, we are interested in learning about the general shape
of the OPS values. For example, is the distribution of OPS values symmetric,
or is the distribution right or left skewed? Also we are interested in learn-
ing about the typical or representative Hall-of-Fame OPS value, and how the
OPS values are spread out. Graphical displays provide a quick visual way of
studying distributions of collections of baseball statistics.
For a single numeric variable, two useful displays for visualizing a distribu-
tion are the stripchart or one-dimensional scatterplot, and the histogram. A
stripchart is basically a number line graph, where the values of the statistics
are plotted over a number line ranging over all possible values of the variable.
This graph is constructed in R by the stripchart function. To begin, we use
the windows function to open a new graphics window 7 inches wide and 3.5
inches tall.1 This is done since we don’t want to use the default 7 in. by 7 in.
format. The only required argument in the stripchart function is the vector
of data to be graphed. The optional argument method="jitter" indicates
that the points are randomly placed in a band over their values; this “jitter-
ing” method of plotting is helpful when you have multiple plotting points with
the same value. The pch=1 argument indicates that the plotting symbol 1 (an
open circle) is to be used,2 and the xlab argument indicates that the x axis
is labeled by “Mid Career”.
windows(width=7, height=3.5)
stripchart(hof$MidCareer, method="jitter", pch=1,
xlab="Mid Career")
The resulting graph is shown in Figure 3.5. One interesting observation from
this graph is the presence of a gap between the seasons 1910 and 1920 with
no Hall of Famers represented.
A second graphical display for a numeric variable is a histogram where
the values are grouped into bins of equal width and the bin frequencies are
displayed as non-overlapping bars over the bins. A histogram is constructed
in R using the function hist. The only required input to hist is the vector
of mid careers hof$MidCareer. The xlab adds a label to the x axis and the
main="" argument removes the default title that is produced with hist.
hist(hof$MidCareer, xlab="Mid Career", main="")
The histogram of mid career values in Figure 3.6, as expected, resembles the
bar graph of the variable Era described in the previous section. One issue
in constructing a histogram is the choice of bins and the function hist will
typically make reasonable choices for the bins to produce a good display of the
data distribution. One can select one’s own bins in the function hist by use of
the argument breaks. For example, if one wanted to choose the alternative bin
endpoints 1880, 1900, 1920, 1940, 1960, 1980, 2000, then one could construct
the histogram by the following code (the figure is not displayed):
1 On a Macintosh computer, the quartz function is used to open up a new graphics
window.
2 See www.statmethods.net/advgraphs/parameters.html for a display of all the possible
plotting symbols possible with the pch argument.

●● ●
● ●● ● ●
●● ● ● ● ● ●●
● ●●●●●
●●
●● ●
● ● ●●● ●●●●
● ●●●●
● ● ●● ●●●
●
●● ● ●
●
● ● ●●
●
● ● ●●●● ● ● ●●● ● ●●●●● ● ● ●●●
●
●● ●
●● ●●● ● ● ● ● ●● ●●●●
●● ● ●
●● ● ● ●●● ● ● ● ● ● ● ● ● ●
●●
● ● ● ●● ● ● ●
●
● ● ● ● ●● ● ● ● ●
1880 1900 1920 1940 1960 1980 2000
Mid Career
FIGURE 3.5
One-dimensional scatterplot (stripchart) of mid career values of HOF non-
pitchers.
hist(hof$MidCareer, xlab="Mid Career", main="",

breaks=seq(1880, 2000, by=20))
3.6 Two Numeric Variables

3.6.1 Scatterplot
When one collects two numeric variables for many players, one is interested in
exploring their relationship. A scatterplot is a standard method for graphing
two numeric variables and the workhorse function for constructing a scatter-
plot in R is plot.
A good measure of batting performance is the OPS statistic, the sum of
the on-base percentage (OBP) and the slugging percentage (SLG). Is there
any relationship between a player’s OPS and the baseball era? Were there
particular seasons where the HOF OPS values were unusually high or low?
We can answer these questions by constructing a scatterplot of the vari-
ables MidCareer and OPS. The plot function has two arguments, the variable
to be plotted on the horizontal scale and the variable on the vertical scale.
As it can be difficult to visually detect scatterplot patterns, it is helpful to
add a smoothing curve to show the general association pattern. The popu-
lar loess smoothing method (Cleveland, 1979) is implemented by the lowess
function with three arguments: the two variables and a constant f between
0 and 1 indicating the degree of smoothness in the curve. (Larger values of
f result in smoother curves with fewer wrinkles.) By using the lines func-
tion, we add this smoothing curve to the current scatterplot. In viewing the
20
15
Frequency
10
5
0
1880 1900 1920 1940 1960 1980 2000
Mid Career
FIGURE 3.6
Histogram of mid career values of HOF nonpitchers.
scatterplot (see Figure 3.7), we notice four unusual career OPS values, three
large values and one very small value, and we’d like to identify the players
with these extreme values. Identification of specific points is accomplished by
the identify function; the inputs are the two variables in the scatterplot, the
vector giving labels (here, names) for the points, and the number of points to
identify. When identify is executed, a hairline will appear when the mouse
is placed on the graph, and the user clicks the mouse near the extreme points.
Figure 3.7 shows the scatterplot with points identified.
with(hof, plot(MidCareer, OPS))
with(hof, lines(lowess(MidCareer, OPS, f=0.3)))
with(hof, identify(MidCareer, OPS, X, n=4))
What do we learn from Figure 3.7? The typical OPS of a Hall of Famer
has stayed pretty constant through the years. But there was an increase in the
OPS during the 1930s when Babe Ruth and Lou Gehrig were in their primes.
There has been a steady decline in the average OPS (among these Hall of
Famers) over the least 30 years. It is interesting to note that the variability of
the OPS values among these players seems small in recent seasons.
Babe Ruth HOF ●
● Ted Williams HOF

1.1
Lou Gehrig HOF ●
●
●
1.0
● ● ●
●
● ● ● ● ● ●
● ● ● ●●
● ●● ●
0.9
●
●●
●
● ●● ● ● ●● ● ● ●●
●
● ● ●
● ● ●●● ● ● ● ● ●● ●● ●
● ● ● ● ● ●
● ● ● ● ● ● ● ●●
● ●● ● ●
OPS
● ● ● ●●● ● ● ●●
● ● ●
0.8
● ● ● ● ●● ● ●
● ● ● ●● ● ● ● ● ●●
● ●
● ● ● ●
●
● ● ●
● ● ●
● ● ●● ●
● ● ●
●
0.7
● ●
● ●
● ● ●
● ● ● ●
0.6
0.5
Willard Brown HOF ●
1880 1900 1920 1940 1960 1980 2000
MidCareer
FIGURE 3.7
Scatterplot of mid career against season for HOF nonpitchers.
3.6.2 Building a graph, step-by-step

Generally, constructing a graph is an iterative process. One begins by choosing
variables of interest and a particular graphical method (such as a scatterplot).
By inspecting the resulting display, one will typically find ways for the graph
to be improved. By using several of the optional arguments, one can make
changes to the graph that result in a clearer and more informative display.
We illustrate this graph construction process in the situation where one is
exploring the relationship between two variables.
There are two dimensions of hitting, the ability to get on base, measured
by the on-base percentage OBP, and the ability to advance runners already on
base, measured by the slugging percentage SLG. One can better understand
the hitting performances of players by constructing a scatterplot of these two
measures. We use the plot function to construct a scatterplot of OBP and
SLG. (See Figure 3.8.)
with(hof, plot(OBP, SLG))
Looking at this figure, we see several problems with this display. First,
due to the one outlier in the bottom-left section of the graph, most of the
0.7
●
● ●
●
0.6
● ●
● ●●
● ●
● ●●
● ●
● ● ●
● ● ●
● ●
● ●
● ● ●
0.5
● ●
● ●● ● ● ● ●
●● ●
SLG
● ●● ●
● ● ●
●● ● ●
● ●● ● ● ●●
●● ●
● ● ●
● ● ● ●
● ● ●●
●● ● ●● ●● ●●
● ● ●● ●●● ●
●
● ●●
●● ●
●● ●
● ●● ● ●
●
● ● ●● ●
●
●●
0.4
●
●
● ●●
● ●● ●
● ●● ●
● ● ●
● ●●
●
●●●
●
●
●
0.3
0.20 0.25 0.30 0.35 0.40 0.45
OBP
FIGURE 3.8
Scatterplot of OBP against SLG for HOF members.
points fall in a relatively small region of the plotting region. Second, it may
be preferable to use an alternative plotting symbol such as a filled circle that
is more distinctive than the default open circle symbol. Last, the graph would
be easier to read if more descriptive labels were used for the two axes. A new
figure is plotted to incorporate these new ideas. By use of the xlim and ylim
arguments, we change the limits of the horizontal and vertical axes. By the
new choices of the limits (0.28, 0.50) for the horizontal and (0.28, 0.75) for
the vertical, we remove the outlier and allow for more space in the upper-left
section of the graph for labels. By use of the pch=19 argument, we change the
plotting symbol to a sold black circle. We use the xlab and ylab arguments to
replace OBP and SLG respectively with “On-Base Percentage” and “Slugging
Percentage.” The updated display is shown in Figure 3.9.
with(hof, plot(OBP, SLG, xlim=c(0.25, 0.50),
ylim=c(0.28, 0.75), pch=19,
xlab="On-Base Percentage",
ylab="Slugging Percentage"))
A good measure of batting performance is the OPS statistic defined by

0.7
0.6
Slugging Percentage
0.5
0.4
0.3
0.25 0.30 0.35 0.40 0.45 0.50

On−Base Percentage
FIGURE 3.9
Scatterplot of OBP against SLG, changing axes limits and axes labels.
OPS = OBP + SLG. To evaluate hitters in our graph on the basis of OPS, it
would be helpful to draw constant values of OPS on the graph. If we represent
OBP and SLG by x and y, suppose we wish to draw a line where OPS = 0.7 or
where x+y = 0.7. Equivalently, we want to draw the function y = 0.7−x on the
graph; this is accomplished in R by the curve function where the argument
of the function is represented by the variable x. The add=TRUE arguments
indicate that this function is to be drawn on the current graph. Similarly, we
apply the curve function three more times to draw lines on the graph where
OPS takes on the values 0.8, 0.9, and 1.0. The resulting display is shown in
Figure 3.10.
curve(.7 - x, add=TRUE)
curve(.8 - x, add =TRUE)
curve(.9 - x, add=TRUE)
curve(1.0 - x, add=TRUE)
In our final iteration, we add labels to the lines showing the constant values
of OPS, and we label the points corresponding to players having a lifetime
OPS exceeding one. Each of the line labels is accomplished using the text
OPS = 1.0
0.7
OPS = 0.9
0.6
Slugging Percentage
OPS = 0.8
0.5
OPS = 0.7
0.4
0.3
0.25 0.30 0.35 0.40 0.45 0.50

FIGURE 3.10
Scatterplot of OBP against SLG, adding lines of constant values of OPS =
OBP + SLG.
function – the three arguments are the x location and y location where the
text is to be drawn, and the string of text to be displayed.
text(.27, .42, "OPS = 0.7")
text(.27, .52, "OPS = 0.8")
text(.27, .62, "OPS = 0.9")
text(.27, .72, "OPS = 1.0")
To label the points for the best hitters by use of mouse clicks, the identify
function is used. The inputs are the x and y plotting variables, the vector of
point labels, and the number of points to label. The final graph is displayed
in Figure 3.11.
with(hof, identify(OBP, SLG, X, n=6))
This final graph is very informative about the batting performance of these
Hall of Famers. We see that a large group of these batters have career OPS
values between 0.8 and 0.9, and only six players (Hank Greenberg, Roger
Hornsby, Jimmie Foxx, Ted Williams, Lou Gehrig, and Babe Ruth) had ca-
reer OPS values exceeding 1.0. Points to the right of the major point cloud
OPS = 1.0
0.7
Babe Ruth HOF
Lou Gehrig HOF

Ted Williams HOF
OPS = 0.9
Hank Greenberg HOF Jimmie Foxx HOF
0.6
Rogers Hornsby HOF

Slugging Percentage
OPS = 0.8
0.5
OPS = 0.7
0.4
0.3
0.25 0.30 0.35 0.40 0.45 0.50

FIGURE 3.11
Scatterplot of OBP against SLG, adding text.
correspond to players with strong skills in getting on-base, but relatively weak
in advancing runners home. In contrast the two points to the left of the major
point cloud correspond to hitters who are better in slugging than in reaching
base.
3.7 A Numeric Variable and a Factor Variable

When one collects a numeric variable such as OPS and a factor such as era, one
is typically interested in comparing the distributions of the numeric variable
across different values of the factor. In R, the stripchart function can be
used to construct parallel stripcharts or number line graphs for values of the
factor, and the boxplot constructs parallel boxplots (graphs of summaries of
the numeric variable) across the factor.
Home run hitting has gone through dramatic changes in the history of
baseball and suppose we are interested in exploring these changes over baseball
eras. Suppose one focuses on the home run rate defined by HR / AB for our
Hall of Fame players. We add a new variable HR.AB to the data frame hof:
hof$HR.Rate <- with(hof, HR / AB)
3.7.1 Parallel stripcharts

One constructs parallel stripcharts of HR.Rate by Era by using the
stripchart function; the argument of the function has the form HR.Rate
∼ Era and we indicate by the data argument that these variables are part of
the hof data frame.
stripchart(HR.Rate ~ Era, data=hof)
There are some problems with the basic display (not shown here). First, by
default, the era labels are drawn parallel to the y axis and all of the labels
are not displayed. Second, the points are drawn on top of each other and it is
difficult to see all of the points. One can have the labels drawn perpendicular
to the axes by using the argument option las=2. Unfortunately, when we do
this, the labels don’t fit in the space to the left of the plotting box. We can
create more space by changing the location of the plot region. The location
of the plot region is given by the graphical parameter plt which by default
is the vector (0.09048276 0.95365517 0.17516995 0.85917710) which represent
fractions (xlo, xhi, ylo, yhi) of the figure region. By assigning the xlo value
the larger value of 0.2, we add more space to the left of the plot region. The
par function is used to set graphical parameters; here it is used to change
the plot region by setting the value of plt. The method="jitter" and pch=1
arguments in the stripchart function are used to jitter the points and make
the points open circles.
par(plt=c(.2, .94, .145, .883))
stripchart(HR.Rate ~ Era, data=hof,
method="jitter", pch=1, las=2)
This display in Figure 3.12 shows how the rate of hitting home runs has
changed over eras. Home runs were rare in the 19th Century and Dead Ball
eras. In the Lively Ball era, home run hitting was still relatively low, but there
were some unusually good home run hitters such as Babe Ruth. The home
run rates in the Integration, Expansion, and Free Agency eras were pretty
similar.
3.7.2 Parallel boxplots

An alternative display for comparing distributions is the boxplot function.
This function has the same basic argument HR.Rate ∼ Era. As before, we
use the argument las=2 to indicate the labels are drawn perpendicular to the
axes, horizontal=TRUE will display the boxplots horizontally and we add the
label “HR Rate” to the x axis.
●
Long Ball ●
● ●● ●● ●● ● ●
Free Agency ●● ● ● ● ●● ● ●
● ● ●● ●
Expansion ●
●
● ●
● ●● ●
● ● ●
●
● ●
●
●
● ● ●● ● ●
Integration ●
●
●● ●
● ● ● ● ● ●●● ● ●● ●
● ●● ●
Lively Ball ●●●●
●●●●● ● ●● ●●
●● ●●● ●
● ● ● ● ●●
●● ● ● ●
● ● ●● ● ●
● ●
●● ●
●●
●●● ●●●●
Dead Ball ●● ● ●●● ● ●
● ●●● ●
● ●
19th Century ●
● ●
●●
● ●●
●
0.00
0.02
0.04
0.06
HR.Rate 0.08
FIGURE 3.12
Stripcharts of home run rates of HOFers for each era.
par(plt=c(.2, .94, .145, .883))

boxplot(HR.Rate ~ Era, data=hof, las=2,
horizontal=TRUE, xlab="HR Rate")
The parallel boxplot display is shown in Figure 3.13. Each rectangle in the
display shows the location of the lower quartile, the median, and the upper
quartile, and lines are drawn to the extreme values. Unusual points (outliers)
that fall far from the rest of the distribution are indicated by open circle points.
This graph confirms the observations we made when we viewed the stripchart
display. Home run hitting was low in the first two eras and started to increase
in the Lively Ball era. It is interesting that the only “outlier” among these
Hall of Famers was Babe Ruth’s career home run rate of 0.085.
Long Ball
Free Agency
Expansion
Integration
Lively Ball ●
Dead Ball
19th Century
0.00
0.02
0.04
0.06
HR Rate 0.08
FIGURE 3.13
Boxplots of home run rates of HOFers for each era.
3.8 Comparing Ruth, Aaron, Bonds, and A-Rod

In Chapter 1, we constructed a graph comparing the career home run trajec-
tories of four great sluggers in baseball history. In this section, we describe
how we use R to create this graph. First, we need to read in the relevant data
files into R. Next, we need to construct data frames containing the home run
and age data for the sluggers. Last, we use R functions to construct the graph.
3.8.1 Getting the data

To obtain the graph, we need to collect the number of home runs, the at-
bats, and the age for each season of each slugger’s career. From the Lahman
database, the relevant data files are Master.csv and batting.csv. From the
master data file Master.csv, we obtain the player ids and birth years for the
four players. The batting data file batting.csv is used to extract the home
run and at-bats information.
We begin by reading in the Lahman master file, storing the file in the data
frame master.
master <- read.csv("Master.csv")
From the Master table, we wish to extract the player id and the birth year
for a particular player. Since we will be doing this operation for four players,
it is convenient to write a new function getinfo to get this information for an
arbitrary player of interest. The inputs to this function are the first and last
names of the player and the output will be a list (a special data structure in
R) giving the player’s id and birth year. Some comments can be made about
the R code in the function.
• The subset function is used to extract the row in the master data frame
matching the player’s first and last names; this row of this data frame is
stored in the variable playerline.
• From the player information in the data frame playerline, we ex-

tract the player’s birth year, birth month, birthday, and player id; these
items are stored in the variables birthyear, birthmonth, birthday, and
name.code.
• In Major League Baseball, a player’s age for a season is defined to be his
age on June 30. So we make a slight adjustment to a player’s birth year
depending if his birthday falls in the first six months or not. The adjusted
birth year is stored in the variable byear. (The ifelse function is useful
for assignments based on a condition; if birthmonth <= 6 is TRUE, then
byear <- birthyear, otherwise byear <- birthyear + 1.)
getinfo <- function(firstname, lastname){

playerline <- subset(master,
nameFirst==firstname & nameLast==lastname)
name.code <- as.character(playerline$playerID)
birthyear <- playerline$birthYear
birthmonth <- playerline$birthMonth
birthday <- playerline$birthDay
byear <- ifelse(birthmonth <= 6, birthyear, birthyear + 1)
list(name.code=name.code, byear=byear)}
We use the function getinfo to get the information for the sluggers Babe
Ruth, Hank Aaron, Barry Bonds, and Alex Rodriguez and store the infor-
mation in variables. By displaying ruth.info, we see the player id and birth
year for Babe Ruth.
ruth.info <- getinfo("Babe", "Ruth")
aaron.info <- getinfo("Hank", "Aaron")
bonds.info <- getinfo("Barry", "Bonds")
arod.info <- getinfo("Alex", "Rodriguez")
ruth.info
$name.code
[1] "ruthba01"
$byear
[1] 1895
Ruth’s id code is “ruthba01” and he was born in 1895.
3.8.2 Creating the player data frames

Now that we have the player id codes and birth years, we use this information
together with the Lahman batting file to create data frames for each player.
We read the Lahman batting data file “Batting.csv” into R, storing the
file in the data frame batting.
batting <- read.csv("Batting.csv")
One of the variables in the batting data frame is playerID. To get the
batting data for Babe Ruth, we use the subset function to extract the rows
of the batting data from where playerID is equal to “ruthba01”. We create a
new variable Age defined to be the season year minus the player’s birth year.
(Recall that we made a slight modification to the byear variable so that one
obtains a player’s correct age for a season.)
ruth.data <- subset(batting, playerID == ruth.info$name.code)
ruth.data$Age <- ruth.data$yearID - ruth.info$byear
We perform similar commands to get batting data frames for the sluggers
Hank Aaron, Barry Bonds, and Alex Rodriguez.
aaron.data <- subset(batting, playerID == aaron.info$name.code)
aaron.data$Age <- aaron.data$yearID - aaron.info$byear
bonds.data <- subset(batting, playerID == bonds.info$name.code)
bonds.data$Age <- bonds.data$yearID - bonds.info$byear
arod.data <- subset(batting, playerID == arod.info$name.code)
arod.data$Age <- arod.data$yearID - arod.info$byear
From the data frames ruth.data, aaron.data, bonds.data, and arod.data,

it is straightforward to use R data analysis operations to compare the batting
performance of these four sluggers.
3.8.3 Constructing the graph

We want to plot the cumulative home run counts for each of the four players
against age. In each player data frame, the relevant variables are HR and Age.
The function cumsum computes the cumulative sums of a vector. For example,
below we illustrate the use of cumsum to compute the cumulative sums of the
vector {1, 2, 3, 4}.
cumsum(c(1, 2, 3, 4))
[1] 1 3 6 10
We use the plot function to graph Ruth’s cumulative home run count
against his age. The arguments to plot indicate that a line graph will be drawn
(type="l") using a dotted line type (lty=3) of double-thickness (lwd=2). The
xlab and ylab arguments label the horizontal and vertical axes and the xlim
and ylim arguments give the limits of the two axes.
with(ruth.data, plot(Age, cumsum(HR), type="l", lty=3, lwd=2,
xlab="Age", ylab="Career Home Runs",
xlim=c(18, 45), ylim=c(0, 800)))
Using three applications of the lines function, three lines are added to the
current graph corresponding to the cumulative home runs of Aaron, Bonds,
and Rodriguez. Different line styles are applied by use of the lty argument so
we can distinguish the four lines of the graph. Using the legend function, a
legend is added to the graph connecting the line styles with the players. The
argument to legend are the x and y coordinates of the location, a vector of
character strings to display, the corresponding vector of line styles (lty), and
the line width (lwd). Figure 3.14 displays the completed graph.
with(aaron.data, lines(Age, cumsum(HR), lty=2, lwd=2))
with(bonds.data, lines(Age, cumsum(HR), lty=1, lwd=2))
with(arod.data, lines(Age, cumsum(HR), lty=4, lwd=2))
legend(20, 700, legend=c("Bonds", "Aaron", "Ruth", "ARod"),
lty=1 : 4, lwd=2)
3.9 The 1998 Home Run Race

The Retrosheet play-by-play files are helpful for learning about patterns of
player performance during a particular baseball season. We illustrate the use
of R to read in the files for the 1998 season and graphically view the famous
home run duel between Mark McGwire and Sammy Sosa.
3.9.1 Getting the data

There are three important data files to read into R. One file denoted by
“all1998.csv” contains all the play-by-play data for the 1998 season. The file
“fields.csv” contains the names of all variables in the play-by-play file, and
the data file “retrosheetIDS.csv” contains the player id codes used to extract
data for particular players.
We begin by reading in the 1998 play-by-play data and storing it in the
data frame data1998. The first line of this data file does not contain the
800
Bonds
Aaron
600
Ruth
ARod
Career Home Runs
400
200
0
20 25 30 35 40 45
Age
FIGURE 3.14
Career home runs by age for four great home run hitters in baseball history.
variable names, so we use the header=FALSE argument option. The header

information with the variable names is stored in the file “fields.csv”. We read
this file, storing the names in the vector fields, and the names function is
used to add this header information to the data frame.
data1998 <- read.csv("all1998.csv", header=FALSE)
fields <- read.csv("fields.csv")
names(data1998) <- fields[, "Header"]
In the play-by-play database, the variable BAT ID gives the identification

code for the player who is batting. To extract the batting data for McGwire
and Sosa, we need to find the codes for these two players. The data file “ret-
rosheetIDs.csv” is useful for this purpose. We read this data file into R and
store the information in the data frame retro.ids. There are three variables
in this data frame – LAST, FIRST, and ID. By use of the subset function, we
find the id code where FIRST-"Sammy" and LAST="Sosa". Likewise, we find
the id code corresponding to Mark McGwire; these codes are stored in the
variables sosa.id and mac.id.
retro.ids <- read.csv("retrosheetIDs.csv")
sosa.id <- as.character(subset(retro.ids,

FIRST=="Sammy" & LAST=="Sosa")$ID)
mac.id <- as.character(subset(retro.ids,
FIRST=="Mark" & LAST=="McGwire")$ID)
Now that the player id codes are obtained, McGwire’s and Sosa’s plate ap-
pearance data are extracted from the play-by-play data frame data1998. The
two player data frames are stored in the variables sosa.data and mac.data.
sosa.data <- subset(data1998, BAT_ID == sosa.id)
mac.data <- subset(data1998, BAT_ID == mac.id)
3.9.2 Extracting the variables

For each player, we are interested in collecting the current number of home
runs hit for each plate appearance and graphing the date against the home
run count. For each player data frame, the two important variables are the
date and the home run count. We write a new function createdata that will
extract these two variables given a player’s play-by-play batting data.
In the play-by-play data frame, the variable GAME ID identifies the game
location and date. For example, the value GAME ID of “ARI199805110” indi-
cates that this particular play occurred at the game played in Arizona on
May 11, 1998. (The variable is displayed in the “location, year, month, day”
format.) Using the substr function, we select the 4th through 11th characters
of this string variable and assign this date to the variable Date. (The as.Date
converts the date to the more readable “year-month-day” format.) Using the
order function, we sort the play-by-play data from the beginning to the end of
the season. The variable EVENT CD contains the outcome of the batting play; a
value EVENT CD of 23 indicates that a home run has been hit. A new variable
HR is defined to be either 1 or 0 depending if a home run occurred or not,
and the new variable cumHR computes the cumulative number of home runs
hit in the season. The output of the function is a new data frame containing
the date and the cumulative number of home runs for all plate appearances
during the season.
createdata <- function(d){
d$Date <- as.Date(substr(d$GAME_ID, 4, 11),
format="%Y%m%d")
d <- d[order(d$Date), ]
d$HR <- ifelse(d$EVENT_CD == 23, 1, 0)
d$cumHR <- cumsum(d$HR)
d[, c("Date", "cumHR")]
}
We use the function createdata twice, once on Sosa’s batting data and
once on McGwire’s batting data, obtaining the new data frames mac.hr and
sosa.hr. We display the first few lines (using the head function) of sosa.hr
to show the format of these new data frames.
mac.hr <- createdata(mac.data)

sosa.hr <- createdata(sosa.data)
head(sosa.hr)
Date cumHR
71539 1998-03-31 0
71561 1998-03-31 0
71581 1998-03-31 0
71600 1998-03-31 0
71621 1998-03-31 0
71628 1998-04-01 0
3.9.3 Constructing the graph

Once these new data frames are created, it is straightforward to produce the
graph of interest. The plot function constructs a graph of the cumulative
home run count against the date. McGwire’s data is plotted using a thick
(lwd=2) line and Sosa’s data is overlaid using the lines function and using a
grey color (col="grey"). The abline function is used to add a horizontal line
at the home run value of 62 and the text function is applied to place the text
string “62” above this plotted line. We conclude using the legend function to
identify McGwire and Sosa’s home run trajectories. (See Figure 3.15.)
plot(mac.hr, type="l", lwd=2, ylab="Home Runs in the Season")
lines(sosa.hr, lwd=2, col="grey")
abline(h=62, lty=3)
text(10440, 65, "62")
legend(10440, 20, legend=c("McGwire (70)", "Sosa (66)"),
lwd=2, col=c("black", "grey"))

A good reference to the traditional graphics system in R is Murrell (2011).
Kabacoff (2011) together with the Quick-R website at www.statmethods.net
provide a useful reference for specific graphics functions. Chapter 4 of Albert
and Rizzo (2012) provides a number of examples of modifying traditional
graphics in R such as changing the plot type and symbol, using color, and
overlying curves and mathematical expressions.
70
62
60
50
Home Runs in the Season
40
30
20
McGwire (70)
Sosa (66)
10
0
Apr May Jun Jul Aug Sep Oct
Date
FIGURE 3.15
Seasonal home runs for Mark McGwire and Sammy Sosa during the 1998 race.
3.11 Exercises
1. (Hall of Fame Pitching Dataset)
The data file “hofpitching.csv” contains the career pitching statistics for
all of the pitchers inducted in the Hall of Fame. This data file can be read
into R by means of the read.csv function.
hofpitching <- read.csv("hofpitching.csv")
The variable BF is the number of batters faced by a pitcher in his ca-

reer. Suppose we group the pitchers by this variable using the intervals
(0, 10,000), (10,000, 15,000), (15,000, 20,000), (20,000, 30,000). One can
reexpress the variable BF to the grouped variable BF.group by use of the
cut function.
hofpitching$BF.group <- with(hofpitching,
cut(BF, c(0, 10000, 15000, 20000, 30000),
labels=c("Less than 10000", "(10000, 15000)",

"(15000, 20000)", "more than 20000")))
(a) Construct a frequency table of BF.group using the table function.

(b) Construct a bar graph of the output from table. How many HOF
pitchers faced more than 20,000 pitchers in their career?
(c) Construct a pie graph of the BF.group variable. Compare the effec-
tiveness of the bar graph and pie graph in comparing the frequencies
in the four intervals.
2. (Hall of Fame Pitching Dataset (Continued))

The variable WAR is the total wins above replacement of the pitcher during
his career.
(a) Using the hist function, construct a histogram of WAR for the pitchers
in the Hall of Fame dataset.
(b) There are two pitchers who stand out among all of the Hall of Famers
on the total WAR variable. Identify these two pitchers.
To understand a pitcher’s season contribution, suppose we define the new
variable WAR.Season defined by
hofpitching$WAR.Season <- with(hofpitching, WAR / Yrs)
(a) Use the stripchart function to construct parallel stripcharts of

WAR.Season for the different levels of BP.group.
(b) Use the boxplot function to construct parallel boxplots of
WAR.Season across BP.group.
(c) Based on your graphs, how does the wins above replacement per
season depend on the number of batters faced?

Suppose we limit our exploration to pitchers whose mid-career was 1960
or later. We first define the MidYear variable and then use the subset
function to construct a data frame consisting of only these 1960+ pitchers.
hofpitching$MidYear <- with(hofpitching, (From + To) / 2)
hofpitching.recent <- subset(hofpitching, MidYear >= 1960)
(a) By use of the order function, order the rows of the data frame by
the value of WAR.Season.
(b) Construct a dot plot of the values of WAR.Season where the labels
are the pitcher names.
(c) Which two 1960+ pitchers stand out with respect to wins above re-
placement per season?

The variables MidYear and WAR.Season are defined in the previous exer-
cises.
(a) Construct a scatterplot of MidYear (horizontal) against WAR.Season
(vertical).
(b) Is there a general pattern in this scatterplot? Explain.
(c) There are two pitchers whose mid careers were in the 1800s who had
relatively low WAR.Season values. Use the identify function with
the scatterplot to find the names of these two pitchers.
6. (Working with the Lahman Batting Dataset)
(a) Read the Lahman “Master.csv” and “batting.csv” data files into R.
(b) Use the getinfo to obtain three data frames for the season batting
statistics for the great hitters Ty Cobb, Ted Williams, and Pete Rose.
(c) Add the variable Age to each data frame corresponding to the ages
of the three players.
(d) Using the plot function, construct a line graph of the cumulative hit
totals against age for Pete Rose.
(e) Using the lines function, overlay the cumulative hit totals for Cobb
and Williams.
(f) Write a short paragraph summarizing what you have learned about
the hitting pattern of these three players.
7. (Working with the Retrosheet Play-by-Play Dataset)
In Section 3.9, we used the Retrosheet play-by-play data to explore the
home run race between Mark McGwire and Sammy Sosa in the 1998 sea-
son. Another way to compare the patterns of home run hitting of the
two players is to compute the spacings, the number of plate appearances
between home runs.
(a) Following the work in Section 3.9, create the two data frames
mac.data and sosa.data containing the batting data for the two
players.
(b) Use the following R commands to restrict the two data frames to
the plays where a batting event occurred. (The relevant variable
BAT EVENT FL is either TRUE or FALSE.)
mac.data <- subset(mac.data, BAT_EVENT_FL == TRUE)
sosa.data <- subset(sosa.data, BAT_EVENT_FL == TRUE)
(c) For each data frame, create a new variable PA that numbers the plate
appearances 1, 2, ... (The function nrow gives the number of rows of
a data frame.)
mac.data$PA <- 1:nrow(mac.data)
sosa.data$PA <- 1:nrow(sosa.data)
(d) The following commands will return the numbers of the plate appear-
ances when the players hit home runs.
mac.HR.PA <- mac.data$PA[mac.data$EVENT_CD==23]
sosa.HR.PA <- sosa.data$PA[sosa.data$EVENT_CD==23]
(e) Using the R function diff, the following commands compute the
spacings between the occurrences of home runs.
mac.spacings <- diff(c(0, mac.HR.PA))
sosa.spacings <- diff(c(0, sosa.HR.PA))
(f) By use of the summary and hist functions on the vectors
mac.spacings and sosa.spacings, compare the home run spacings
of the two players.
4
The Relation Between Runs and Wins
CONTENTS
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2 The Teams Table in Lahman’s Database . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3 Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 The Pythagorean Formula for Winning Percentage . . . . . . . . . . . . . 93
4.5 The Exponent in the Pythagorean Formula . . . . . . . . . . . . . . . . . . . . . 95
4.6 Good and Bad Predictions by the Pythagorean Formula . . . . . . . 96
4.7 How Many Runs for a Win? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.8 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.1 Introduction
The goal of a baseball team is, just like in any other sport, winning games.
Similarly, the goal of the baseball analyst is being able to measure what hap-
pens on the field in term of wins. Answering a question such as “Who is the
better player between Brett Gardner and Prince Fielder?” becomes an easier
task if one succeeds in estimating how much Gardner’s speed and slick field-
ing contribute to his team’s victories and how many wins can be attributed
to Prince’s powerful bat.
Victories are obtained by outscoring opponents, thus the percentage of
wins obtained by a team over the course of a season is strongly related with
the number of runs it scores and allows. This chapter explores the relation-
ship between runs and wins. Understanding this relationship is a critical step
towards answering questions on players’ value. In fact, while it’s impossible to
directly quantify the impact of players in terms of wins, it will be seen in the
following chapters that it is possible to measure their contributions in term of
runs.
87
4.2 The Teams Table in Lahman’s Database

The teams.csv file from Lahman’s database contains seasonal stats for major
league teams going back to the first professional season in 1871. We begin by
loading this database file into R and exploring its contents by looking at the
final lines of this dataset, using the tail function.
teams <- read.csv("teams.csv")
tail(teams)
yearID lgID teamID franchID divID Rank G Ghome W L

2683 2011 NL LAN LAD W 3 161 NA 82 79
2684 2011 NL COL COL W 4 162 NA 73 89
2685 2011 NL SDN SDP W 5 162 NA 71 91
DivWin WCWin LgWin WSWin R AB H X2B X3B HR BB SO
2683 N N N N 644 5436 1395 237 28 117 498 1087
2684 N N N N 735 5544 1429 274 40 163 555 1201
2685 N N N N 593 5417 1284 247 42 91 501 1320
SB CS HBP SF RA ER ERA CG SHO SV IPouts HA HRA BBA
2683 126 40 45 43 612 563 3.56 7 17 40 4296 1287 132 507
2684 118 42 57 44 774 713 4.44 5 7 41 4343 1471 176 522
2685 170 44 48 47 611 551 3.43 0 10 44 4348 1324 125 521
SOA E DP FP name park
2683 1265 85 121 0.986 Los Angeles Dodgers Dodger Stadium
2684 1118 98 156 0.984 Colorado Rockies Coors Field
2685 1139 94 138 0.985 San Diego Padres Petco Park
attendance BPF PPF teamIDBR teamIDlahman45 teamIDretro
2683 2935139 98 98 LAD LAN LAN
2684 2909777 116 116 COL COL COL
2685 2143018 92 92 SDP SDN SDN
The description of every column is provided in the readme file accompanying
the Lahman’s database.
Suppose that one is interested in relating the proportion of wins with the
runs scored and runs allowed for all of the teams. Towards this goal, the
relevant fields of interest in this table are the number of games played G, the
number of team wins W, the number of losses L, the total number of runs
scored R, and the total number of runs allowed RA. A new data frame myteams
is created containing only the above five columns plus information on the team
(teamID), the season (yearID), and the league (lgID). We are interested in
studying the relationship between wins and runs for recent seasons, so by use
of the subset function, we focus our exploration to seasons since 2001.
myteams <- subset(teams, yearID > 2000)[ , c("teamID", "yearID",
"lgID", "G", "W", "L", "R", "RA")]
The Relation Between Runs and Wins 89
tail(myteams)
teamID yearID lgID G W L R RA

2680 FLO 2011 NL 162 72 90 625 702
2681 ARI 2011 NL 162 94 68 731 662
2682 SFN 2011 NL 162 86 76 570 578
2683 LAN 2011 NL 161 82 79 644 612
2684 COL 2011 NL 162 73 89 735 774
2685 SDN 2011 NL 162 71 91 593 611
The run differential is defined as the difference between the runs scored and
the runs allowed by a team. The winning proportion is the fraction of games
won by a team. In baseball (and generally in sports) winning percentage is
commonly used instead of the more appropriate winning proportion. In the
remainder of this chapter we have chosen to adopt the most widely used term.
Two new variables RD (run differential) and Wpct (winning percentage) are
calculated with the following lines of code.
myteams$RD <- with(myteams, R - RA)
myteams$Wpct <- with(myteams, W / (W + L))
A scatterplot of the run differential and the winning percentage using
the plot function gives a first indication on the association between the two
variables.
plot(myteams$RD, myteams$Wpct,
xlab="run differential",
ylab="winning percentage")
As expected, Figure 4.1 shows a strong positive relationship – teams with
large run differentials are more likely to be winning.
4.3 Linear Regression

One simple way to predict a team’s winning percentage using runs scored and
allowed is with linear regression. A simple linear model is
W pct = a + b × RD + ǫ,
where a and b are unknown constants and ǫ is the error term which
captures all other factors influencing the dependent variable (Wpct).
This is a special case of a linear model fit using the lm function
from the stats package (which is installed and loaded in R by de-
fault). The most basic call to the function requires a formula, specified
as response ~ predictor1 + predictor2 + ..., data=dataset, in which
●
0.7
●
●● ●
● ● ●
● ● ●●
● ●
0.6
● ●
● ●
●
winning percentage
● ●●● ● ●
●● ●●●● ● ●● ●
● ●● ●●●●● ●● ●●●
●
● ● ● ● ● ●● ●
●● ● ● ●●● ●
●● ● ●●●● ●
● ●●●● ●●●● ●
● ●
● ●●●●● ● ●
● ●●
●● ●●
● ●●●● ●
● ● ● ●●● ● ●
●●● ● ●● ●●●●● ●●
● ● ●● ●● ●
●●●● ●●
●
● ● ● ●
0.5
● ●●● ●
● ● ●●●
●●
●●● ● ●●● ●
●
● ●● ●● ● ● ● ●
●●● ●● ●● ●● ●
● ● ●● ● ● ●
● ●● ● ●
● ● ●●●●●
●● ●
● ●●●● ● ● ●
● ● ● ●● ● ●
●
●● ●●● ● ●
● ● ● ●●●● ● ●●
● ● ●●● ●●● ●
●●
● ● ●
● ● ●● ●
●● ●
● ● ● ●● ●●
● ● ● ●●
0.4
●●●● ● ●
●●● ●●
● ●● ●
●●●
● ●●●
●● ● ●
● ● ●
● ●
●
●
● ●● ●●
●
0.3
−300 −200 −100 0 100 200 300

run differential
FIGURE 4.1
Scatterplot of team run differential against team winning percentage for major
league teams from 2001 to 2011. A best-fitting line is overlaid on top of the
scatterplot.
the variable to be predicted (the dependent variable) is indicated on the left

side of the tilde character and the the variables used to predict the response
are specified on the right side. In the following illustration of the lm function,
the data argument in lm is used to specify which data frame to use.
linfit <- lm(Wpct ~ RD, data=myteams)
linfit
Call:
lm(formula = Wpct ~ RD, data=myteams)
Coefficients:
(Intercept) RD
0.499992 0.000623
The fitted line in the plot of Figure 4.1 is obtained with the abline func-
tion1 and the coef function for extracting the model (linfit) coefficients.
abline(a=coef(linfit)[1], b=coef(linfit)[2], lwd=2)
From the above output, a team’s winning percentage can be estimated

from its run differential RD by the equation:
W pct = 0.499992 + 0.000623 × RD
This formula tells us that a team with a run differential of zero (RD = 0)
will win half of its games (estimated intercept ≈ .500) which is reasonable. In
addition, a one-unit increase in run differential corresponds to an increase of
0.000623 in winning percentage. To give further insight into this relationship,
a team scoring 750 runs and allowing 750 runs is predicted to win half of its
games corresponding to 81 games in a typical MLB season of 162 games. In
contrast, a team scoring 760 runs and allowing 750 has a run differential of +10
and is predicted to have a winning percentage of 0.500 + 10 · 0.000623 ≈ 0.506.
A winning percentage of 0.506 in a 162-game schedule corresponds to 82 wins.
Thus an increase of 10 runs in the run differential of a team corresponds,
according to the straight-line model, to an additional win in the standings.
One concern is that predictions from this fitted line can assume values
outside the range [0, 1]. For example, a hypothetical team that outscores its
opponent by a total of 805 runs would be predicted to win more than 100
percent of its games which is impossible. However, since over 99 percent of
teams throughout major league baseball history have run differentials between
-350 and +350, the straight-line model is a reasonable approximation.
Once one has a fitted model, the function predict can be used to calculate
the predicted values from the model, while the function residuals computes
the difference between the response values and the fitted values (i.e., between
the actual and the estimated winning percentages).
myteams$linWpct <- predict(linfit)
myteams$linResiduals <- residuals(linfit)
Figure 4.2 displays a plot of the residuals against the run differential using
the following code. The abline and text functions have been introduced in
Chapter 3, while points is used to draw data points at specified coordinates
(in particular, the points and text functions are used to mark and label a
few anomalous data points).
plot(myteams$RD, myteams$linResiduals,
xlab="run differential",
ylab="residual")
1 In this case, rather than a single argument h or v for plotting an horizontal or a vertical
line, abline is supplied with two arguments a and b indicating the intercept and the slope
of the line to be drawn.
abline(h=0, lty=3)
points(c(68, 88), c(.0749, -.0733), pch=19)
text(68, .0749, "LAA ’08", pos=4, cex=.8)
text(88, -.0733, "CLE ’06", pos=4, cex=.8)
● ● LAA '08
● ● ●
●
0.05
●●
● ●
● ●
●
● ●● ● ●
● ●
● ● ● ●● ●
● ●
●● ●●
● ● ● ● ●
● ● ●● ●●● ●● ●●● ●●
●
●
● ● ● ● ●● ● ● ● ● ●
● ●●
●●● ●●●● ● ●
● ● ●
● ● ●
● ● ● ● ● ● ●● ● ● ●● ●● ● ●●
residual
● ● ●
● ● ● ●
●
● ● ●● ●● ● ●● ● ●●●● ●●●● ●
●
0.00
● ● ● ● ● ●
● ●
●● ● ● ● ● ●● ●● ●●●● ●● ●
● ● ● ● ●●
● ● ●●● ●● ●●●●● ●●
● ● ● ● ● ● ● ●●●● ●
● ● ● ● ● ● ●● ● ●●
●● ●● ● ● ● ●
● ● ●● ●● ●●
●
●● ● ● ●● ●●● ● ●●
● ● ● ● ●
● ● ●●●●●● ● ● ●●●● ●● ●● ●
● ●● ● ● ● ● ● ● ●●
● ●
● ● ● ●
● ● ● ● ● ● ● ●
●●●● ●● ●
●
● ●
● ● ●● ● ● ● ●
● ● ●
● ●● ●● ● ●
−0.05
● ● ● ● ●
● ●
● ●
● ●
●
● CLE '06
−300 −200 −100 0 100 200 300

run differential
FIGURE 4.2
Residuals versus run differential for the fitted linear model. Two large resid-
uals are labeled corresponding to the 2008 Los Angeles Angels and the 2006
Cleveland Indians.
Residuals can be interpreted as the error of the linear model in predicting

the actual winning percentage. Thus the points in Figure 4.2 farthest from
the zero line correspond to the teams where the linear model fared worst in
predicting the winning percentage.
One of the extreme values at the top of the residual graph in Figure 4.2
corresponds to the 2008 Los Angeles Angels: given their +68 run differential,
they were supposed, according to the linear equation (4.3), to have a 0.542
winning percentage; they ended the season at 0.617. The residual value for
this team is 0.617 − 0.542 = 0.075. At the other end of the spectrum, the 2006
Cleveland Indians, with a +88 run differential, are seen as a 0.555 team by
the linear model, but they actually finished at a mere 0.481, corresponding to
the residual 0.481 − 0.555 = −0.073.
The average value of the residuals for this model is equal to zero, which
means that the model predictions are equally likely to overestimate than to un-
derestimate the winning percentage, or that the method for fitting the model
is unbiased. In order to estimate the average magnitude of the errors, one
first squares the residuals so that each error has a positive value, calculates
the mean of the squared residuals, and takes the square root of such mean
value to get back to the original scale. The value so calculated is the root
mean square error, abbreviated as RMSE. (The square root function sqrt is
introduced.)
mean(myteams$linResiduals)
[1] -2.952603e-19
linRMSE <- sqrt(mean(myteams$linResiduals ^ 2))
linRMSE
[1] 0.02507176
Approximately two thirds of the residuals fall between −RM SE and
+RM SE, while 95% of the residuals are between −2·RM SE and 2·RM SE.2
These statements can be confirmed with the following lines of code. (The func-
tion abs computes the absolute value.)
nrow(subset(myteams, abs(linResiduals) < linRMSE)) /
nrow(myteams)
[1] 0.6757576
nrow(subset(myteams, abs(linResiduals) < 2 * linRMSE)) /

nrow(myteams)
[1] 0.9545455
Above the function nrow is used to obtain the number of rows of a data frame.
In the numerators of the expressions, we obtain the number of residuals sizes
(computed using the abs function) that are smaller than one and two RM SE.
The computed fractions are close to the 68% and 95% values stated above.
4.4 The Pythagorean Formula for Winning Percentage

Bill James, regarded as the father of Sabermetrics, empirically derived
the following non-linear formula to estimate winning percentage, called the
Pythagorean expectation.
R2
W pct =
R2 + RA2
2 Equivalently, it can be stated that, over a 162-game season, the number of wins predicted
by the linear model comes within four wins of the actual number of wins in two-thirds of
the cases, while for 19 out of 20 teams the difference is not higher than 8 wins.
One can use this formula to predict winning percentages by use of the following
R code.
myteams$pytWpct <- with(myteams, R ^ 2 / (R ^ 2 + RA ^ 2))
Here the residuals need to be calculated explicitly, but that’s not a hard task.
A new variable pytResiduals is defined that is the difference between the
actual and predicted winning percentages. The RMSE is computed for these
new predictions.
myteams$pytResiduals <- myteams$Wpct - myteams$pytWpct
sqrt(mean(myteams$pytResiduals ^ 2))
[1] 0.02545247
The RMSE calculated on the Pythagorean predictions is similar in value to
the one calculated with the linear predictions (it’s actually higher for the 2000-
2011 data we have been using here). Thus it does not seem justifiable using
a more complex model. However, the Pythagorean expectation has several
desirable properties missing in the linear model. Both of these advantages can
be illustrated with several examples.
Suppose there exists a powerhouse team that scores an average of ten runs
per game, while allowing a close to average five runs per game. In a 162-
game schedule, this team would score 1620 runs, while allowing 810, for a run
differential of 810. Replacing RD with 810 in the linear equation, one obtains a
winning percentage of over 1, which is impossible. On the other hand, replacing
R and RA with 1620 and 810 respectively in the Pythagorean expectation, the
resulting winning percentage is equal to 0.8, a more reasonable prediction. A
second hypothetical team has pitchers who never allow runs, while the hitters
always manage to score the only run they need. Such a team will score 162 runs
in a season and win all of its games, but the linear equation would predict it to
be merely a .601 team. The Pythagorean formula, instead, correctly predicts
this team to win all of its games.
While neither of the above examples is ever going to materialize, there are
some extreme situations in modern baseball history. For example, the 2001
Seattle Mariners had 116 wins and 46 losses for a +300 run differential and
the 2003 Detroit Tigers had a 43-119 recored with a -337 run differential.
In these unlikely scenarios, the Pythagorean formula will give more sensible
winning percentage estimates.
Recall our statement at the end of the introductory section that the runs-
to-wins relationship is crucial in assessing the contribution of players to their
team’s wins. Once we estimate the number of runs players contribute to their
teams (as it will be shown in the following chapters), runs-to-wins formulas
can be used to convert these run values to wins. One can now answer questions
like “Home many wins would a lineup of nine Albert Pujols’ accumulate in
a season?” For these kind of investigations, the scenarios in which the linear
formula break down are more likely to occur, thus highlighting the need for
a formula such as the Pythagorean expectation that gives reasonable predic-
tions.
4.5 The Exponent in the Pythagorean Formula

Subsequent refinements to the Pythagorean formula by Bill James and other
analysts have aimed at finding an exponent which would give a better fit
relative to the originally proposed exponent value of 2. In this section, we
describe how one finds the value of the Pythagorean exponent leading to
predictions closest to the actual winning percentages.
Replacing the value 2 in the exponents in the Pythogorean expectation
with an unknown variable k, the formula is written as
Rk
W% =
Rk + RAk
With some algebra, the equation can be rewritten as follows:
W Rk
=
L RAk
Taking the logarithm on both sides of the equation (using the function log),
one obtains the linear relationship

W R
log = k · log
L RA
The value of k can now be estimated using linear regression, where the re-
sponse variable is log(W/L) and the predictor is log(R/RA). In the following
R code, we compute the logarithm of the ratio of wins to losses, the logarithm
of the ratio of runs to runs allowed, and fit a simple linear model with these
transformed variables. (In the call of the lm function, a model with a zero
intercept is indicated by a zero term on the right side of the formula.)
myteams$logWratio <- log(myteams$W / myteams$L)
myteams$logRratio <- log(myteams$R / myteams$RA)
pytFit <- lm(logWratio ~ 0 + logRratio, data=myteams)
pytFit
Call:
lm(formula = logWratio ~ 0 + logRratio, data=myteams)
Coefficients:
logRratio
1.903
The R output suggests a Pythagorean exponent of 1.903 which is significantly
smaller than the value 2.
4.6 Good and Bad Predictions by the Pythagorean For-

mula
The 2011 Boston Red Sox scored 875 runs, while allowing 737. According
to the Pythagorean formula with exponent 2, they were expected to win 95
games – we obtain this number by plugging 875 and 737 into the Pythagorean
formula and multiplying by the number of games in a season:
8752
162 × ≈ 95
8752+ 7372
The Red Sox actually won 90 games. The five games difference was quite
costly to the Red Sox, as they missed clinching the Wild Card (that went to
the Rays) in the final game (actually in the final minute) of the season. The
Pythagorean formula is more on target with the Tampa Bay Rays of the same
season, as the prediction of 92 (coming from their 707 runs scored versus 614
runs allowed) is just a bit higher than the actual 91.
Why does the Pythagorean formula miss so poorly on the Red Sox? In
other words, why did they win five less games than expected from their run
differential? Let’s have a look at their season game by game.
The gl2011.txt (a game log file downloaded from Retrosheet, see Section
1.3.3) contains detailed information on every game played in the 2011 season.
The following commands load the file into R, select the lines pertaining to the
Red Sox games, and keep only the runs related columns.
gl2011 <- read.table("gl2011.txt", sep=",")
glheaders <- read.csv("retrosheet/game_log_header.csv")
names(gl2011) <- names(glheaders)
BOS2011 <- subset(gl2011, HomeTeam=="BOS" | VisitingTeam=="BOS")[
, c("VisitingTeam", "HomeTeam", "VisitorRunsScored",
"HomeRunsScore")]
head(BOS2011)
VisitingTeam HomeTeam VisitorRunsScored HomeRunsScore

16 BOS TEX 5 9
31 BOS TEX 5 12
45 BOS TEX 1 5
61 BOS CLE 1 3
76 BOS CLE 4 8
88 BOS CLE 0 1
Using the results of every game featuring the Boston team, run differentials
(ScoreDiff) are calculated both for games won and lost and a column W is
added indicating whether the Red Sox won the game.
BOS2011$ScoreDiff <- with(BOS2011, ifelse(HomeTeam == "BOS",

HomeRunsScore - VisitorRunsScored,
VisitorRunsScored - HomeRunsScore))
BOS2011$W <- BOS2011$ScoreDiff > 0
Summary statistics are computed on the run differentials for games won
and for games lost using the aggregate function. For this function, the first
argument is the variable for which the summary statistics are to be calculated
(the absolute value of the score differential), the second argument is a list of
grouping factors (in this case only one factor is specified–whether the game
resulted in a win for Boston), and the final argument is the summarizing
function we want to apply (summary which was introduced in Chapter 2).
aggregate(abs(BOS2011$ScoreDiff), list(W=BOS2011$W), summary)
W x.Min. x.1st Qu. x.Median x.Mean x.3rd Qu. x.Max.

1 FALSE 1.000 1.000 3.000 3.458 4.000 11.000
2 TRUE 1.000 2.000 4.000 4.300 6.000 14.000
The 2011 Red Sox had their victories decided by a larger margin than their
losses (4.3 vs 3.5 runs on average), leading to their underperformance of the
Pythagorean prediction by five games. A team overperforming (or underper-
forming) its Pythagorean winning percentage is often seen, in sabermetrics
circles, as being lucky (or unlucky), and consequently is expected to get closer
to its expected line as the season progresses.
A team can overperform its Pythagorean winning percentage by winning a
disproportionate number of close games. This claim can be confirmed by a brief
data exploration. With the following lines of code, a data frame (results) is
created from the previously loaded 2011 game logs, containing the names of
the teams and the runs scored. Two new columns are created, the variable
winner contains the abbreviation of the winning team and a second variable
diff contains the margin of victory.
results <- gl2011[,c("VisitingTeam", "HomeTeam",

"VisitorRunsScored", "HomeRunsScore")]
results$winner <- ifelse(results$HomeRunsScore >
results$VisitorRunsScored, as.character(results$HomeTeam),
as.character(results$VisitingTeam))
results$diff <- abs(results$VisitorRunsScored -
results$HomeRunsScore)
Suppose we focus on the games won by only one run. The data frame
onerungames is created containing only the games decided by one run, and
the table function is used to count the number of wins in such contests for
each team. The as.data.frame function converts this table to a data frame
and the names function adds names to the columns.
onerungames <- subset(results, diff == 1)

onerunwins <- as.data.frame(table(onerungames$winner))
names(onerunwins) <- c("teamID", "onerunW")
Using the myteams data frame previously created, we look at the relation
between the Pythagorean residuals and the number of one-run victories. Note
that the team abbreviation for the Angels needs to be changed because it is
coded as “LAA” in the Lahman’s database and as “ANA” in the Retrosheet
game logs.
teams2011 <- subset(myteams, yearID == 2011)
teams2011[teams2011$teamID == "LAA", "teamID"] <- "ANA"
teams2011 <- merge(teams2011, onerunwins)
plot(teams2011$onerunW, teams2011$pytResiduals,
xlab="one run wins",
ylab="Pythagorean residuals")
The final line of code produces the plot in Figure 4.3 which shows a positive
relationship between the number of one-run games won and the Pythagorean
residuals. The identify function is used to identify teams on the plot.
identify(teams2011$onerunW, teams2011$pytResiduals,
labels=teams2011$teamID)
Figure 4.3 shows that San Francisco is a team with a large number of one-run
victories and a large positive Pythogorean residual. In contrast, San Diego
had few one-run victories and a negative residual.
Winning a disproportionate number of close games is sometimes attributed
to plain luck. However teams with certain attributes may be more likely to
systematically win contests decided by a narrow margin. For example, teams
with top quality closers will tend to preserve small leads, and will be able to
overperform their Pythogorean expected winning percentage. To check this
conjecture, we look at the data.
The pitching.csv file in the Lahman’s database contains individual sea-
sonal pitching stats. This file is read into R and the subset function is used
select the pitchers/seasons where more than 50 games were finished with an
ERA lower than 2.50. The data frame top closers contains only the columns
identifying the pitcher, the season and the team.
pit <- read.csv("pitching.csv")
top_closers <- subset(pit, GF > 50 & ERA < 2.5)[ ,c("playerID",
"yearID", "teamID")]
The top closers data frame is merged with our myteams dataset, creating
the new teams top closers data frame – this contains the teams featuring
a top closer. Summary statistics on the Pythagorean residuals are obtained
using the summary function.
0.04 ●
SFN ●
●
●
●
0.02
● ●
●
Pythagorean residuals
● ●
● ●
●
●
0.00
● ●
−0.02
● ●
●
● ●
●
●
●
−0.04
●
●
SDN
●
20 22 24 26 28 30 32
one run wins
FIGURE 4.3
Scatterplot of number of one-run games won and Pythagorean residuals for
major league teams in 2011.
teams_top_closers <- merge(myteams, top_closers)

summary(teams_top_closers$pytResiduals)
-0.048690 -0.011660 0.003359 0.005189 0.022990 0.071400
The mean of the residuals is only slightly above zero (0.005189), but when one
multiplies it by the number of games in a season (162), one finds that teams
with a top closer win, on average, 0.8 games more than would be predicted
by the Pythagorean formula.
4.7 How Many Runs for a Win?

Readers familiar with websites like www.insidethebook.com,
www.hardballtimes.com, and www.baseballprospectus.com/ are surely fa-
miliar with the ten-runs-equal-a-win rule of thumb. Over the course of a
season, a team scoring ten more runs is likely to have one more win in the
standings. The number comes directly from the Pythagorean formula with
an exponent of two. Suppose a team scores an average of five runs per game,
while allowing the same number of runs. In a 162-game season, the team would
score (and allow) 810 runs. Inserting 810 in the Pythagorean formula one gets
(as expected) a perfect .500 expected winning percentage with 81 wins. If one
substitutes 810 with 820 for the number of runs scored in the formula, one
obtains a .506 winning percentage that translates to 82 wins in 162 games.
The same result is obtained for a team scoring 810 runs and allowing 800.
Ralph Caola has derived the number of extra runs needed to get an extra
win in a more rigorous way using calculus. He starts from the equivalent
representation of the Pythagorean formula.
R2
W =G·
R2 + RA2
If one takes a partial derivative of the right side of the above equation with
respect to R, holds RA constant, the result is the incremental number of wins
per run scored. Taking the reciprocal of this result, one can derive the number
of runs needed for an extra win.
R is capable of calculating partial derivatives, thus we can retrace Ralph’s
steps in R by using the functions D and expression to take the partial deriva-
tive of R2 /(R2 + RA2 ) with respect to R.
D(expression(G * R ^ 2 / (R ^ 2 + RA ^ 2)), "R")
G * (2 * R)/(R^2 + RA^2) G * R^2 * (2 * R)/(R^2 + RA^2)^2
Unfortunately R does not do the simplifying. The reader has the choice of
either doing the tedious work himself or believing the final equation for incre-
mental runs per win (IR/W) is the following3 :
2
R2 + RA2
IR/W =
2 · G · R · RA2
If R and RA are expressed in runs per game, the G is removed from the above
formula.
Using this formula, one can compute the incremental runs needed per
one win for various runs scored/runs allowed scenarios. As a first step, a
function IR is created to calculate the incremental runs, according to Caola’s
formula; this function takes runs scored per game and runs allowed per game
as arguments.
IR <- function(RS=5, RA=5){
round((RS ^ 2 + RA ^ 2)^2 / (2 * RS * RA ^ 2), 1)
}
3 The formula is the result of algebraic simplification and taking the reciprocal.
This function is used to create a table for various runs scored/runs al-
lowed combinations. We perform this step by using the functions seq and
expand.grid. The seq function is used create a vector containing a regular
sequence specifying, as arguments, the start value, the end value, and the in-
crement value. Here seq creates a vector of values from 3 to 6 in increments
of 0.5. Then the expand.grid function is used to obtain a data frame con-
taining all the combinations of the elements of the supplied vectors. In the
following code the first and the final few lines of the new data frame IRtable
are displayed.
IRtable <- expand.grid(RS=seq(3, 6, .5), RA=seq(3, 6, .5))
rbind(head(IRtable), tail(IRtable))
RS RA
1 3.0 3
2 3.5 3
3 4.0 3
4 4.5 3
5 5.0 3
6 5.5 3
44 3.5 6
45 4.0 6
46 4.5 6
47 5.0 6
48 5.5 6
49 6.0 6
Finally, the incremental runs are calculated for the various scenarios. The
xtabs function in the second line of the following code is used to show the
results in a tabular form. The formula specified as the first argument has the
variable which populates the cells on the left side of the tilde character, while
on the right side the cross-classifying variables are separated by a + sign.
IRtable$IRW <- IR(IRtable$RS, IRtable$RA)
xtabs(IRW ~ RS + RA, data=IRtable)
RA
RS 3 3.5 4 4.5 5 5.5 6
3 6.0 6.1 6.5 7.0 7.7 8.5 9.4
3.5 7.2 7.0 7.1 7.5 7.9 8.5 9.2
4 8.7 8.1 8.0 8.1 8.4 8.8 9.4
4.5 10.6 9.6 9.1 9.0 9.1 9.4 9.8
5 12.8 11.3 10.5 10.1 10.0 10.1 10.3
5.5 15.6 13.4 12.2 11.4 11.1 11.0 11.1
6 18.8 15.8 14.1 13.0 12.4 12.1 12.0
Looking at the results we notice that the rule of ten is appropriate in typical
run scoring environments (4 to 5 runs per game). However, in very low scoring
environments (the upper-left corner of the table), a lower number of runs is
needed to gain an extra win; on the other hand, in high scoring environments
(lower-right corner), one needs a larger number of runs for an added win.
4.8 Further Reading

Bill James first mentioned his Pythagorean formula in James (1980) which,
like other early works by James, was self-published and it is currently hard
to find. Reference to the formula is present in James (1982), the first edition
published by Ballantine Books. Davenport and Woolner (1999) and Heipp
(2003) revisited Bill James’ formula, deriving exponents that vary according
to the total runs scored per game. Caola (2003) algebraically derived the
relation between run scored and allowed and winning percentage. Star (2011)
recounts the final moments of the 2011 regular season, when in the turn of a
few minutes the Rays and the Red Sox fates dramatically turned; the page also
features a twelve-minute video chronicling the events of the wild September
28, 2011 night.
4.9 Exercises
1. (Relationship Between Winning Percentage and Run Differential
Across Decades)
Section 4.3 used a simple linear model to predict a team’s winning per-
centage based on its run differential. This model was fit using team data
since the 2001 season.
(a) Refit this linear model using data from the seasons 1961-1970, the
seasons 1971-1980, the seasons 1981-1990, and the seasons 1991-2000.
(b) Compare across the five decades the predicted winning percentage
for a team with a run differential of 10 runs.
2. (Pythagorean Residuals for Poor and Great Teams in the 19th
Century)
As baseball was evolving into its ultimate form, nineteenth century leagues
often featured abysmal teams that did not even succeed in finishing their
season, as well as some dominant clubs.
(a) Fit a Pythagorean formula model to the run-differential, win-loss data
for teams who played in the 19th century.
(b) By inspecting the residual plot of your fitted model from (a), did the
great and poor teams in the 19th century do better or worse than
one would expect on the basis of their run differentials?
3. (Exploring the Manager Effect in Baseball)

Retrosheet game logs report, for every game played, the managers of both
teams.
(a) Select a period of your choice (encompassing at least ten years) and
fit the Pythagorean formula model to the run-differential, win-loss
data.
(b) On the basis of your fit in part (a) and the list of managers, compile
a list of the managers who most overperformed their Pythagorean
winning percentage and the managers who most underperformed it.
4. (Pythagorean Relationship for Other Sports)
Bill James’ Pythagorean formula has been used for predicting winning per-
centage in other sports. Since the pattern of scoring is very different among
sports (compare for example points in basketball and goals in soccer), the
formula needs to be adapted to the scoring environment. Find the neces-
sary data for a sport of your choice and compute the optimal exponent
to the Pythagorean formula. (The website www.opensourcesports.com
provides databases for NBA and WNBA basketball and for NHL hockey
in a format similar to Lahman’s baseball database.)
5
Value of Plays Using Run Expectancy
CONTENTS
5.1 The Runs Expectancy Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2 Runs Scored in the Remainder of the Inning . . . . . . . . . . . . . . . . . . . . 106
5.3 Creating the Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.4 Measuring Success of a Batting Play . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.5 Albert Pujols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.6 Opportunity and Success for All Hitters . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.7 Position in the Batting Lineup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.8 Run Values of Different Base Hits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.8.1 Value of a home run . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.8.2 Value of a single . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.9 Value of Base Stealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.10 Further Reading and Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.1 The Runs Expectancy Matrix

An important matrix in sabermetrics research is the runs expectancy matrix.
As each base (first, second, and third) can be occupied by a runner or empty,
there are 2×2×2 = 8 possible arrangements of runners on the three bases. The
number of outs can be 0, 1, or 2 (three possibilities), and so there are a total of
8 × 3 = 24 possible arrangements of runners and outs. For each combination of
runners on base and outs, one is interested in computing the average number
of runs scored in the remainder of the inning. When these average runs are
arranged as a table classified by runners and outs, the display is often called
the runs expectancy matrix. We illustrate using R to compute this matrix
using play-by-play data for the 2011 season. This matrix is used to define
the average run value (or run value) of a batter’s plate appearance. Then the
distribution of average runs values is explored for all batters in the 2011 season.
The runs values for Albert Pujols are used to help understand the pattern of
run values. We continue by exploring how players in different positions in the
batting lineup perform with respect to this criterion. The notion of runs value
is helpful for understanding the relative benefit of different batting plays and
105
we explore the value of a home run and a single. This chapter is concluded by
using the runs expectancy matrix and runs values to understand the benefit
of stealing a base and the cost of being caught stealing.
5.2 Runs Scored in the Remainder of the Inning

We begin by reading into R the play-by-play database for the 2011 season
and storing this data into the R data frame data2011. The fields.csv file
contains the names of the variables of this database and the names function
is used to attach these variable names to the data frame.
At a given plate appearance, there is a potential to score runs. Clearly,

this potential is greater with runners on base, specifically runners in scoring
position (second or third base), and when there are few outs. This runs po-
tential is measured by computing the average number of runs scored in the
remainder of the inning for each combination of runners on base and number
of outs. Certainly, the average runs scored is dependent on many variables
such as home versus away, the current score, the pitching and the defense.
But this runs potential represents the opportunity to create runs in a typical
situation during an inning and is useful for measuring contributions of players
in an average scenario.
To compute the number of runs scored in the remainder of the inning, we
need to know the total runs scored by both teams at the plate appearance and
also the total runs scored by the teams at the end of the specific half-inning.
The runs scored in the remainder of the inning, denoted by RUNS.ROI, is the
difference
RUNS.ROI = Total Runs Scored in Inning − Current Runs Scored.
To begin, a variable RUNS is created that is equal to the sum of the visitor’s
score (AWAY SCORE CT) and the home team’s score (HOME SCORE CT) at each
plate appearance.
data2011$RUNS <- with(data2011, AWAY_SCORE_CT + HOME_SCORE_CT)
A new variable HALF.INNING is also created, using the paste function, com-
bining the game id, the inning, and the team at bat.
data2011$HALF.INNING <- with(data2011,
paste(GAME_ID, INN_CT, BAT_HOME_ID))
Value of Plays Using Run Expectancy 107
The variable HALF.INNING creates a unique identification for each half-inning

of every game during the season.
We wish to compute the maximum total score for each half-inning, combin-
ing home and visitor scores. To accomplish this, a new variable RUNS.SCORED
is created that gives the number of runs scored for each play. (The variables
BAT DEST ID, RUN1 DEST ID, RUN2 DEST ID, and RUN3 DEST ID give the des-
tination bases for the batter and each runner, and runs are scored for each
destination base that exceeds 3.) By use of the aggregate function with the
sum function, we compute the total runs scored in each half-inning and store
in the data frame RUNS.SCORED.INNING.
data2011$RUNS.SCORED <- with(data2011, (BAT_DEST_ID > 3) +
(RUN1_DEST_ID > 3) + (RUN2_DEST_ID > 3) + (RUN3_DEST_ID > 3))
RUNS.SCORED.INNING <- aggregate(data2011$RUNS.SCORED,
list(HALF.INNING=data2011$HALF.INNING), sum)
By another application of the aggregate function with the “[” function, we

find the total game runs at the beginning of each half-inning.
RUNS.SCORED.START <- aggregate(data2011$RUNS,
list(HALF.INNING=data2011$HALF.INNING), "[", 1)
The maximum total score in a half-inning is the sum of the initial total runs
and the runs scored. A new data frame MAX is created, defining a new variable
x equal to the maximum runs scored. The merge function is used to merge this
information with the data frame data2011 and create the new maximum total
score variable MAX.RUNS. (The function ncol gives the number of columns of
a data frame.)
MAX <- data.frame(HALF.INNING=RUNS.SCORED.START$HALF.INNING)
MAX$x <- RUNS.SCORED.INNING$x + RUNS.SCORED.START$x
data2011 <- merge(data2011, MAX)
N <- ncol(data2011)
names(data2011)[N] <- "MAX.RUNS"
Now the runs scored in the remainder of the inning (new variable RUNS.ROI)
can be computed by taking the difference of MAX.RUNS and RUNS.
data2011$RUNS.ROI <- with(data2011, MAX.RUNS - RUNS)
5.3 Creating the Matrix

Now that the runs scored in the remainder of the inning variable have been
computed for each plate appearance, it is straightforward to compute the runs
expectancy matrix.
Currently, there are three variables BASE1 RUN ID, BASE2 RUN ID, and
BASE2 RUN ID containing the player codes of the baserunners (if any) who
are respectively on first, second, or third base. Three new binary variables
RUNNER1, RUNNER2, and RUNNER3 are created that are either 1 or 0 if the corre-
sponding base is respectively occupied or empty. (The as.character function
converts a factor variable to a character variable.)
RUNNER1 <- ifelse(as.character(data2011[ ,"BASE1_RUN_ID"]) == "", 0, 1)
A short function get.state is written to create a state variable, combining

(using the function paste) the runner indicators and the number of outs. This
function is used to create a current state variable STATE.
get.state <- function(runner1, runner2, runner3, outs){
runners <- paste(runner1, runner2, runner3, sep="")
paste(runners, outs)
}
data2011$STATE <- get.state(RUNNER1, RUNNER2, RUNNER3, data2011$OUTS_CT)
One particular state value would be “011 2” which indicates that there are
currently runners on second and third base with two outs. A second state
value “100 0” indicates there is a runner at first with no outs.
We want to only consider plays in our data frame where there is a change
in the runners on base, number of outs, or the runs scored. Three new variables
NRUNNER1, NRUNNER2, NRUNNER3 are created which indicate, respectively, if first
base, second base, and third base are occupied after the play. (The function
as.numeric converts a logical variable to a numeric variable.) The variable
NOUTS is the number of outs after the play, and RUNS.SCORED is the number of
runs scored on the play. Again the get.state function is used to create the
variable NEW.STATE giving the runners on each base and the number of outs
after the play.
NRUNNER1 <- with(data2011, as.numeric(RUN1_DEST_ID == 1 |
BAT_DEST_ID == 1))
RUN2_DEST_ID == 2 | BAT_DEST_ID==2))
RUN2_DEST_ID == 3 | RUN3_DEST_ID == 3 | BAT_DEST_ID == 3))
NOUTS <- with(data2011, OUTS_CT + EVENT_OUTS_CT)
data2011$NEW.STATE <- get.state(NRUNNER1, NRUNNER2, NRUNNER3, NOUTS)
By use of the subset function, attention is restricted to plays where either

there is a change between STATE and NEW.STATE (indicated by the not equal
logical operator “!=”) or there are runs scored on the play.
data2011 <- subset(data2011, (STATE != NEW.STATE) | (RUNS.SCORED > 0))
Before the runs expectancies are computed, one final adjustment is nec-
essary. The play-by-play database includes scoring information for all half-
innings during the 2011 season, including partial half-innings at the end of
the game where the winning run is scored with less than three outs. In our
work, we want to work only with complete half-innings where three outs are
recorded. The ddply function in the plyr package is applied to compute the
number of outs for each half-inning, and the merge function is used to add
a new variable Outs.Inning to the data frame. The subset function is used
to extract the data from the half-innings in data2011 with exactly three outs
– the new data frame is named data2011C. (By removing the noncomplete
innings, one is introducing a small bias since these innings are not complete
due to the scoring of at least one run.)
library(plyr)
data.outs <- ddply(data2011, .(HALF.INNING), summarize,
Outs.Inning=sum(EVENT_OUTS_CT))
data2011 <- merge(data2011, data.outs)
data2011C <- subset(data2011, Outs.Inning == 3)
The expected number of runs scored in the remainder of the inning (the
runs expectancy) is computed for each of the 24 bases/outs situations by use
of the aggregate function, grouping by STATE with the mean function.
RUNS <- with(data2011C, aggregate(RUNS.ROI, list(STATE), mean))
To display these run values as an 8 × 3 matrix, a new variable Outs is

defined, and the RUNS data frame is sorted by the number of outs.
RUNS$Outs <- substr(RUNS$Group, 5, 5)
RUNS <- RUNS[order(RUNS$Outs), ]
The matrix function is used to create a matrix RUNS.out. For readability,

each value is rounded to two decimal places and labels are assigned to the
rows and columns by two applications of the dimnames function.
RUNS.out <-matrix(round(RUNS$x, 2), 8, 3)
dimnames(RUNS.out)[[2]] <- c("0 outs", "1 out", "2 outs")
dimnames(RUNS.out)[[1]] <- c("000", "001", "010", "011", "100", "101",
"110", "111")
To see how the run expectancy values have changed over time, the 2002
season values as reported in Albert and Bennett (2003) are collected in the
vector RUNS.2002. The 2011 and 2002 expectancies are displayed side-by-side
for comparison purposes.
RUNS.2002 <- matrix(c(.51, 1.40, 1.14, 1.96, .90, 1.84, 1.51, 2.33,
.27, .94, .68, 1.36, .54, 1.18, .94, 1.51,
.10, .36, .32, .63, .23, .52, .45, .78), 8, 3)
dimnames(RUNS.2002) <- dimnames(RUNS.out)
cbind(RUNS.out, RUNS.2002)
0 outs 1 out 2 outs 0 outs 1 out 2 outs
000 0.47 0.25 0.10 0.51 0.27 0.10
001 1.45 0.94 0.32 1.40 0.94 0.36
010 1.06 0.65 0.31 1.14 0.68 0.32
011 1.93 1.34 0.54 1.96 1.36 0.63
100 0.84 0.50 0.22 0.90 0.54 0.23
101 1.75 1.15 0.49 1.84 1.18 0.52
110 1.41 0.87 0.42 1.51 0.94 0.45
111 2.17 1.47 0.76 2.33 1.51 0.78
It is somewhat remarkable that these run expectancy values have not changed
over the recent history of baseball. That indicates that there have been little
changes in the average run scoring tendencies of MLB teams between 2002
and 2011.
5.4 Measuring Success of a Batting Play

When a player comes to bat with a particular runners and outs situation,
the run expectancy matrix tells us the average number of runs a team will
score in the remainder of the half-inning. Based on the outcome of the plate
appearance, the state (runners on base and outs) will change and there will be
a updated runs expectancy value. The value of the plate appearance, called
the runs value, is measured by the difference in run expectancies of the old
and new states plus the number of runs scored on the particular play.
RUNS VALUE = RUNSN ew State − RUNSOld State + RUNSScored on P lay
The run values for all plays in the original data frame data2011 are com-
puted using the following R script. First a 32 × 1 matrix RUNS.POTENTIAL is
defined that contains the run expectancies for the 32 situations including 3
outs. The run expectancy of a situation with 3 outs is obviously zero, so we
append some zeros to the previously calculated run expectancy values. The
new variable RUNS.STATE is defined to be the runs expectancy of the current
state and the variable RUNS.NEW.STATE is defined to be the runs expectancy
of the new state. The new variable RUNS.VALUE is set equal to the difference
in RUNS.NEW.STATE and RUNS.STATE plus the RUNS.SCORED variable.
RUNS.POTENTIAL <- matrix(c(RUNS$x, rep(0, 8)), 32, 1)
dimnames(RUNS.POTENTIAL)[[1]] <- c(RUNS$Group, "000 3", "001 3",
"010 3", "011 3", "100 3", "101 3", "110 3", "111 3")
data2011$RUNS.STATE <- RUNS.POTENTIAL[data2011$STATE, ]
data2011$RUNS.NEW.STATE <- RUNS.POTENTIAL[data2011$NEW.STATE, ]
data2011$RUNS.VALUE <- data2011$RUNS.NEW.STATE - data2011$RUNS.STATE +
data2011$RUNS.SCORED
5.5 Albert Pujols

To better understand runs values, let’s focus on the plate appearances for the
great hitter Albert Pujols for the 2011 season. To find Pujols’ player id, we
read in the roster2011.csv data file and use the subset function to extract
the Player.ID variable. The as.character function converts Player.ID from
a factor to a character variable called albert.id.
Roster <- read.csv("roster2011.csv")
albert.id <- subset(Roster, First.Name == "Albert" &
Last.Name == "Pujols")$Player.ID
albert.id <- as.character(albert.id)
Using the subset function, a data frame of Pujols plate appearances is

found, where the batter id (variable BAT ID) is equal to albert.id.
albert <- subset(data2011, BAT_ID == albert.id)
We wish to consider only the batting plays where Albert was the hitter, so
the subset function is used to select the rows where the batting flag (variable
BAT EVENT FL) is true.1
albert <- subset(albert, BAT_EVENT_FL == TRUE)
How did Albert do on his first two plate appearances this season? To
answer this, we display the first two rows of the data frame albert, showing
the original state, new state, and run value variables:
albert[1:2, c("STATE", "NEW.STATE", "RUNS.VALUE")]
STATE NEW.STATE RUNS.VALUE
6556 100 1 000 3 -0.4960492
6574 001 2 001 3 -0.3173913
On his first plate appearance, there was a runner on first with one out. The
outcome of this plate appearance was three outs, indicating that Albert hit
into a double-play, and the runs value for this play was −0.496. On his second
plate appearance, there was a runner on third with two outs. Evidently Albert
got out (the final state had three outs) and the run value was −0.317. Based
on the runs values of these first plate appearances, Albert didn’t have a very
good start to the 2011 season.
When one evaluates the run values for any player, there are two primary
questions. First, we need to understand the player’s opportunities for produc-
ing runs. What were the runner/outs situations for the player’s plate appear-
ances? Second, what did the batter do with these opportunities to score runs?
1 The variable BAT EVENT FL distinguishes batting events from non-batting events such as
steals and wild pitches.

The batter’s success or lack of success on these opportunities can be measured

by use of the runs values.
Let’s focus on the runners states to understand Albert’s opportunities.
Since a few of the counts of the runners/outs states over the 32 outcomes are
close to zero, we focus on runners on base and use the substr function to
create a new variable RUNNERS containing the runners state of the three bases.
The table function is applied to tabulate the runners state for all of Albert’s
plate appearances.
albert$RUNNERS <- substr(albert$STATE, 1, 3)
table(albert$RUNNERS)
000 001 010 011 100 101 110 111

354 27 60 7 135 19 36 13
We see that Albert generally was batting with the bases empty (000) or with
only a runner on first (100). Most of the time, Albert was batting with no
runners in scoring position.
How did Albert perform with these opportunities? Using the following R
code, we construct a stripchart (using the stripchart function) that shows
the runs values for all plate appearances organized by the runners state. (See
Figure 5.1.) A horizontal line at the value zero is added to the graph – points
above the line (below the line) correspond to positive (negative) contributions.
with(albert, stripchart(RUNS.VALUE ~ RUNNERS, vertical=TRUE, jitter=0.2,
xlab="RUNNERS", method="jitter", pch=1, cex=0.8))
abline(h=0)
There are many duplicate runs values, so we jitter the points (using the
method="jitter" argument) to better show the density of runs values. When
the bases were empty (000), the range of possible runs values was relatively
small. For this state, the large cluster of points at a negative runs value corre-
sponds to the many occurrences when Albert got an out with the bases empty.
The cluster of points at (000) at the value 1 corresponds to Albert’s home
runs with the bases empty. (A home run with runners empty will not change
the bases/outs state and the value of this play is exactly one run.) For other
situations, say the bases-loaded situation (111), there is much variation in the
runs values. For one plate appearance, the state moved from 111 1 to 000
3, indicating that Albert hit into a inning-ending double play with the bases
loaded with a runs value of −1.41. In contrast, Albert did hit a home run with
the bases loaded with no outs and the runs value of this outcome was 2.38.
(The runs value of a grand slam is not 4 since the run potential of the end
state of bases empty is much smaller than the run potential of a bases-loaded
state.)
To understand Albert’s total run production for the 2011 season, the
aggregate function together with the sum and length functions can be used
to compute the number of opportunities and sum of runs values for each of
the runners situations.
●
2
●
● ●●●
●
● ●● ● ●
● ●
●
● ● ●
●
●
●●● ●
1
●
●●●●
● ●
●●
●●● ●●
●●● ● ●●
●●●
RUNS.VALUE
● ● ● ●●●● ●
●
● ●● ●
●
●●●
● ●● ●●●●●
●● ●
● ●
● ●
●
●●● ● ● ●●●
● ●●●●●
●● ●
●
●●
●
●● ●
●
●●
●●
● ●● ●●
● ●
●
●●
● ●
●●
●●
●●●●
●
●● ●●●
●
●●
●●●● ●●● ●●
●● ●● ●●
●
●
●●
●●
●●
●
●●
●
●●
●●● ● ●●
● ● ● ●
●
●●●●●
●
0
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●● ● ●● ●
●
●
●●
●
●●●
●●
●
●●
●
●●
●
●●
●
●●
●
●● ● ●●
● ●
● ●
●
●
●●
●
●●
●●
●●
●
●●
●●
●●
●●
●
●
●● ●
●●
●●
●●
●●
●●●
●●●
●● ●
●●●
●
●●
●●●
●●●● ●
●
●●●● ● ●
●
●
●●
●●
● ●
●
●●
● ●●
●● ●● ●●●
● ● ●
●●●●●●
●●●
●
●●
●
●●●●
●
●●● ● ●●●
●●
● ● ●
●●
●●● ● ●
● ●
●● ● ●
● ●●●
● ●
●
−1
●●●●●
● ●
000 001 010 011 100 101 110 111
RUNNERS
FIGURE 5.1
Dotplot of runs values of Albert Pujols for all 2011 plate appearances as a
function of the runners state. The points have been jittered since there are
many plate appearances with identical runs values.
A.runs <- aggregate(albert$RUNS.VALUE, list(albert$RUNNERS), sum)

names(A.runs)[2] <- "RUNS"
A.PA <- aggregate(albert$RUNS.VALUE, list(albert$RUNNERS), length)
names(A.PA)[2] <- "PA"
A <- merge(A.PA, A.runs)
A
Group.1 PA RUNS
1 000 354 10.7568515
2 001 27 -0.9417766
3 010 60 9.6399002
4 011 7 -0.5680439
5 100 135 14.2174870
6 101 19 -1.9853612
7 110 36 -4.5593651
8 111 13 0.6971003
We see, for example, that Albert came to bat with the runners empty 354
times, and his total runs value contribution to these 354 PAs was 10.76. Albert
didn’t do particularly well with runners in scoring position. For example, there
were 36 PAs where he came to bat with runners on first and second, and
his net contribution in runs for this situation was −4.56. Albert’s total runs
contribution for the 2011 season can be computed by summing the last column
of this matrix.
sum(A$RUNS)
[1] 27.25679
It is not surprising that Albert has a positive total contribution in his PAs in
2011, but it is difficult to understand the size of 27 runs unless this value is
compared with the contribution of other players. In the next section, we will
see how Albert compares to all hitters in the 2011 season.
5.6 Opportunity and Success for All Hitters

The runs value criterion can be used to compare the batting effectiveness
of players. We focus on batting plays, so a new data frame data2011b is
constructed that is the subset of the main data frame data2011 where the
BAT EVENT FL variable is equal to TRUE:
data2011b <- subset(data2011, BAT_EVENT_FL == TRUE)
It is difficult to compare the total run values of two players on face value,
since they have different opportunities to create runs for their teams. One
player in the middle of the batting order may come to bat many times when
there are runners in scoring position and good opportunities to create runs.
Other players towards the bottom of the batting order may not get the same
opportunities to bat with runners on base. One can measure a player’s op-
portunity to create runs by the sum of the runs potential state (variable
RUNS.STATE) over all of his plate appearances. We can summarize a player’s
batting performance in a season by the total number of plate appearances,
the sum of the runs potentials, and the sum of the runs values.
The R function aggregate is helpful in obtaining these summaries. In the
following R code, the data frame runs.pa contains the number of plate appear-
ances for all batters in the 2011 season, the data frame runs.sums contains
the total runs value for these players, and the data frame runs.start contains
the total starting runs potential for the players. Using two applications of the
merge function, we merge the two matrices, creating the data frame runs. A
row of this data frame will contain the number of plate appearances, the total
run value, and the total run potential for a particular player.
runs.sums <- aggregate(data2011b$RUNS.VALUE, list(data2011b$BAT_ID), sum)
runs.pa <- aggregate(data2011b$RUNS.VALUE, list(data2011b$BAT_ID), length)
runs.start <- aggregate(data2011b$RUNS.STATE, list(data2011b$BAT_ID), sum)
names(runs.sums) <- c("Batter", "Runs")

names(runs.pa) <- c("Batter", "PA")
names(runs.start) <- c("Batter", "Runs.Start")
runs <- merge(runs.sums, runs.pa)
runs <- merge(runs, runs.start)
The data frame runs contains batting data for both pitchers and nonpitch-
ers. It seems reasonable to restrict attention to nonpitchers, since pitchers and
nonpitchers have very different batting abilities. Also we limit our focus on the
players who are primarily starters on their teams. One can remove pitchers
and nonstarters by focusing on batters with at least 400 plate appearances. A
new data frame runs400 is created by an application of the subset function.
There are 203 players in this data frame; we display the first few rows by use
of the head function.
runs400 <- subset(runs, PA >= 400)
head(runs400)
Batter Runs PA Runs.Start
1 abreb001 9.695694 585 251.6928
15 andir001 -8.511571 511 246.0340
16 andre001 4.847767 665 323.6004
18 ankir001 -4.326331 415 183.3090
19 arenj001 6.358087 486 226.9625
24 avila001 32.125231 551 266.5247
Is there a relationship between batters’ opportunities and their success in

converting these opportunities to runs? To answer this question, we construct
a scatterplot of run opportunity (Runs.Start) against runs value (Runs) for
these hitters with at least 400 at bats (see Figure 5.2). To help see the pattern
in this scatterplot, the lowess function is used to smooth this scatterplot and
the smoothing curve is placed on top of the scatterplot by use of the lines
function. To interpret this graph, it is helpful to add a horizontal line (using
the abline function) at Runs=0; points above this line correspond to hitters
who had a total positive runs value contribution in the 2011 season.
with(runs400, plot(Runs.Start, Runs))
with(runs400, lines(lowess(Runs.Start, Runs)))
abline(h=0)
From viewing Figure 5.2, we see that batters with larger values of Runs.Start
tend to have larger runs contributions. But there is a wide spread in the
runs values for these players. In the group of players who have Runs.Start
values between 300 and 350, four of these players actually have negative runs
contributions and other players created over 60 runs in the 2011 season.
From the graph, we see that only a limited number of players created
more than 40 runs for their teams. Who are these players? In the R code, we
create a new data frame runs400.top containing the runs statistics for only
the players who created more than 40 runs. For labeling purposes, we would
like to obtain the last names of the players available on the “roster2011.csv”
data file. We read in this file using the read.csv function and use the merge
function to merge the roster information with the data frame runs400.top.
By use of the text function, point labels are added to the previous scatterplot
for these outstanding hitters. (See Figure 5.2.)
runs400.top <- subset(runs400, Runs >= 40)
roster2011 <- read.csv("roster2011.csv")
runs400.top <- merge(runs400.top,
roster2011, by.x="Batter", by.y="Player.ID")
with(runs400.top, text(Runs.Start, Runs, Last.Name, pos=1))
From this figure, we learn that the best hitters using the runs criterion are
Miguel Cabrera (71.11), Jose Bautista (67.37), Prince Fielder (60.82), Joey
Votto (60.65), and Matt Kemp (59.64). There is an interesting outlier in this
figure – Mike Napoli created over 40 runs for his team despite only having a
Runs.Start value close to 200. Napoli was a very productive batter for his
team given his opportunities to produce runs.
5.7 Position in the Batting Lineup

Managers like to put their best hitters in the middle of the batting lineup.
Traditionally, a team’s “best hitter” bats third and the clean-up hitter in the
fourth position is the best batter for advancing runners on base. What are the
batting positions of the hitters in our sample? Specifically, are the best hitters
using the runs value criterion the ones who bat in the middle of the lineup?
A player may bat in several positions in the lineup during the season. We
define a player’s batting position as the position that he bats most frequently.
A function get.batting.pos is defined that will find a player’s batting po-
sition. This function will compute a table of a player’s batting position and
identify the position which has the highest frequency. By use of the sapply
function together with the function get.batting.pos, we find the batting
position for all players in the runs400 data frame.
get.batting.pos <- function(batter){
TB <- table(subset(data2011, BAT_ID == batter)$BAT_LINEUP_ID)
names(TB)[TB == max(TB)][1]}
position <- sapply(as.character(runs400$Batter), get.batting.pos)
In the following R code, the players’ run opportunities are plotted against
their runs values. By use of the type="n" option, the axes are drawn but points
are not plotted. Instead, we use the text function with labels contained in
the vector position to display the batting positions as plotting points. (See
Figure 5.3.)
Cabrera
Bautista Fielder
Votto
60
Kemp
Braun Ellsbury
Berkman Gonzalez
Napoli MartinezCano
40
Granderson
Runs
20
0
−20
200 250 300 350

Runs.Start
FIGURE 5.2
Scatterplot of total run value against the runs potential for all players in the
2011 season with at least 400 plate appearances. A smoothing curve is added
to the scatterplot – this shows that players who had more run potential tend
to have large run values. The players with a total runs value at least 40 are
labeled.
with(runs400, plot(Runs.Start, Runs, type="n"))

with(runs400, lines(lowess(Runs.Start, Runs)))
abline(h=0)
with(runs400, text(Runs.Start, Runs, position))
From this figure, we better understand the relationship between batting posi-
tion, run opportunities, and run values. The best hitters, the ones who create
a large number of runs, generally bat third, fourth, and fifth in the batting
order. The number of runs created by the leadoff (first) and second batters in
the lineup are much smaller than the runs created by the best hitters in the
middle (third and fourth positions) of the lineup. There are some surprises
from this general pattern of batting positions. Mike Napoli, the unusual hitter
who created over 40 runs with only 200 run opportunities, bats only sixth in
the lineup. Also, there are many cleanup hitters displayed who have mediocre
values of runs created.
4
3
3 44
60
3 1
5 3
5 5
6
2
40
14 3
34 33 53
3 8 34 4 2
4 3 4 2
3 3
3 4 4
1
4
●
3 3
2
3
11 4 6 4 4 2
Runs
20
3
6 2 1
9 8 1 7 11
3
7 17 1 5 26
3 47 8 4 35 3
5 54
3 42 7 3 7 4 1
311 3 31 5 3 7 2 444 6
2
3 5 28 7 9 31 5 1
5 5 8 16 5 2 5 6 1 4 6 44 2
7 9 2 2 1
79 6 1
0
2 25 6
2 6 6 58 783 88 1 24 1
6 9 816 632 1 14
2 81 61 4 9 58 26
4 51 2 1 2
3 2
7 5
1 64 9 1 1
7
−20
3 2 2
8 6 6 9
5
7
6
200 250 300 350
Runs.Start
FIGURE 5.3
Scatterplot of total run value against the runs potential for all players in the
2011 season with at least 400 plate appearances. The points are labeled by
the position in the batting lineup and the large point corresponds to Albert
Pujols.
How does Albert Pujols and his total runs value of 27.2 compare among the
group of hitters with at least 400 plate appearances? By use of the subset
function, we find Pujols’ data from the runs400 matrix. Using the points
function, we display Albert’s (Runs.Start, Runs) value by a large solid dot.
In this particular season (2011), Albert was one of the better hitters in terms
of creating runs for his team.
AP <- subset(runs400, Batter == albert.id)
points(AP$Runs.Start, AP$Runs, pch=19, cex=3)
5.8 Run Values of Different Base Hits

There are many applications of runs values in studying baseball. Here we look
at the values of a home run and a single from the perspective of creating runs.
One criticism of a batting average is that it gives equal value to the four
possible base hits (single, double, triple, and home run). One way of distin-
guishing the values of the base hits is to assign the number of bases reached –
1 for a single, 2 for a double, 3 for a triple, and 4 for a home run. The slugging
percentage is the total number of bases divided by the number of at-bats. But
it is not clear that the values 1, 2, 3, and 4 represent a reasonable measure of
the value of the four possible base hits. We can get a better measure of the
importance of these base hits by the use of runs values.
5.8.1 Value of a home run

Let’s focus on the value of a home run from a runs perspective. The home run
plays are extracted from the data frame runs2011 using the EVENT CD play
event variable. A value of EVENT CD of 23 corresponds to a home run. Using
the subset function with the EVENT CD == 23 condition, a new data frame
d.homerun is created with the home run plays.
d.homerun <- subset(data2011, EVENT_CD == 23)
When are the runners/outs states for the home runs hit during the 2011
season? We answer this question by use of the table function.
table(d.homerun$STATE)
000 0 000 1 000 2 001 0 001 1 001 2 010 0 010 1 010 2 011 0 011 1
1226 812 630 15 50 54 57 94 144 17 28
011 2 100 0 100 1 100 2 101 0 101 1 101 2 110 0 110 1 110 2 111 0
32 262 301 286 32 50 61 60 115 128 15
111 1 111 2
42 41
By use of the prop.table function, the relative frequencies are computed and
the round function is used to round the values to three decimal spaces.
round(prop.table(table(d.homerun$STATE)), 3)
000 0 000 1 000 2 001 0 001 1 001 2 010 0 010 1 010 2 011 0 011 1
0.269 0.178 0.138 0.003 0.011 0.012 0.013 0.021 0.032 0.004 0.006
011 2 100 0 100 1 100 2 101 0 101 1 101 2 110 0 110 1 110 2 111 0
0.007 0.058 0.066 0.063 0.007 0.011 0.013 0.013 0.025 0.028 0.003
111 1 111 2
0.009 0.009
We see from this table that the fraction of home runs hit with the bases empty
is 0.269 + 0.178 + 0.138 = 0.585. So over half of the home runs are hit with
no runners on base.
What are the runs values of these home runs? We already observed in the
analysis of Pujols’ data that the runs value of a home run with the bases
empty is one. A histogram of the run values for all home runs is constructed
using the truehist function in the MASS package.2 (See Figure 5.4.)
library(MASS)
truehist(d.homerun$RUNS.VALUE)
It is obvious from this graph that most home runs (the ones with the bases
empty) have a runs value of one. But there is a cluster of home runs with
values between 1.5 and 2.0, and there is a small group of home runs with
runs value exceeding three. Which runners/outs situation leads to the most
6
Mean Runs Value

5
4
3
2
1
0
1.0 1.5 2.0 2.5 3.0
d.homerun$RUNS.VALUE
FIGURE 5.4
Histogram of the runs values of the home runs hit during the 2011 season.
The vertical line shows the location of the mean runs value of a home run.
2 We prefer truehist to the hist function since by default density values, instead of
counts, are graphed and it has a nice appearance.

valuable home runs? Using the subset function, the row of the data frame is
extracted corresponding to the largest runs value.
subset(d.homerun, RUNS.VALUE == max(RUNS.VALUE))[1,
c("STATE", "NEW.STATE", "RUNS.VALUE")]
6748 111 2 000 2 3.336175
As one might expect, the most valuable home run occurs when there are bases
loaded with two outs. From our earlier work, it was seen that this type of home
run occurred 41 times during the season, and the runs value of this home run
is 3.34.
Overall, what is the runs value of a home run? This question is answered by
computing the runs value of all the home runs in the data frame d.homerun.
mean.HR <- mean(d.homerun$RUNS.VALUE)
mean.HR
[1] 1.392393
A vertical line is drawn on the graph showing the mean runs value and a label
is added to this line. (See Figure 5.4.)
abline(v=mean.HR, lwd=3)
text(1.5, 5, "Mean Runs Value", pos=4)
This average runs value is pretty small, but this value partially reflects the
fact that most home runs are hit with the bases empty.
5.8.2 Value of a single

Runs values can also be used to evaluate the benefit of a single. Unlike a home
run, the runs value of a single will depend both on the initial state (runners
and outs) and on the final state. The final state of a home run will always
have the bases empty; in contrast, the final state of a single will depend on
the movement of any runners on base.
In R, the subset function is used to select the plays where EVENT CD ==
20 (corresponding to a single); the new data frame is called d.single. A
histogram is constructed of the run values for all of the singles in the 2011
season. (See Figure 5.5.)
d.single <- subset(data2011, EVENT_CD == 20)
library(MASS)
truehist(d.single$RUNS.VALUE)
Looking at the histogram of run values of the single, there are three large
spikes between 0 and 0.5. These large spikes can be explained by constructing
a frequency table of the beginning state.
table(d.single$STATE)
Mean Runs Value

5
4
3
2
1
0
−0.5 0.0 0.5 1.0 1.5 2.0 2.5
d.single$RUNS.VALUE
FIGURE 5.5
Histogram of the runs values of the singles hit during the 2011 season. The
vertical line shows the location of the mean runs value of a single.
000 0 000 1 000 2 001 0 001 1 001 2 010 0 010 1 010 2 011 0 011 1
7078 5067 3814 85 315 356 502 844 959 87 225
011 2 100 0 100 1 100 2 101 0 101 1 101 2 110 0 110 1 110 2 111 0
216 1593 2063 1826 154 344 385 346 728 770 100
111 1 111 2
284 277
We see that most of the singles occur with the bases empty, and the three
spikes in the histogram, as one moves from left to right in Figure 5.5, corre-
spond to singles with no runners on and two outs, one out, and no outs. The
small cluster of runs values in the interval 0.5 to 2.0 correspond to singles hit
with runners on base.
What is the most valuable single from the runs value perspective? We use
the subset function to find the beginning and end states for the single that
resulted in the largest runs value.
subset(d.single, d.single$RUNS.VALUE ==
max(d.single$RUNS.VALUE))[ , c("STATE", "NEW.STATE", "RUNS.VALUE")]

105749 111 2 001 2 2.556382
In this particular play, the hitter came to bat with the bases loaded and two
outs, and the final state was a runner on third with two outs. How could have
this happened with a single? The data frame does contain a brief description
of the play. But from the data frame we identify the play happening during the
bottom of the 7th inning of a game between the Brewers and Twins on July
3, 2011. We check with www.espn.com to find the following play description.
“D Valencia singled to deep left, J Mauer and M Cuddyer scored, J Thome to
second, J Thome scored, D Valencia to third on error by left fielder M Kotsay.”
So evidently, the left fielder made an error on the fielding of the single that
allowed all three runners to score and the batter to reach third base.
At the other extreme, by use of the subset function, two plays are iden-
tified which achieved the smallest runs value.
subset(d.single, d.single$RUNS.VALUE == min(d.single$RUNS.VALUE))[
, c("STATE", "NEW.STATE", "RUNS.VALUE")]
69618 010 0 100 1 -0.5622312
138351 010 0 100 1 -0.5622312
How could the run value of a single be negative one-half a run? With further
investigation, we find that in each case, there was a runner on second who was
thrown out at the plate as a result of the single.
As in the case of the home run, it is straightforward to compute the mean
runs value of a single. We display this mean value on the histogram in Figure
5.5.
mean.single <- mean(d.single$RUNS.VALUE)
mean.single
[1] 0.4424186
abline(v=mean.single, lwd=3)
text(.5, 5, "Mean Runs Value", pos=4)
In this case, we see that the mean value of a single is approximately equal to
the runs value when a single is hit with the bases empty with no outs. It is
interesting that the runs value of a single can be large (in the 1 to 2 range).
These large runs values reflect the fact that the benefit of the single depends
on the advancement of the runners.
5.9 Value of Base Stealing

The runs expectancy matrix is also useful in understanding the benefits of
stealing bases. When a runner attempts to steal a base, there are two likely
outcomes – either the runner will be successful in stealing the base or the
runner will be caught stealing. Overall, is there a net benefit to attempting to
steal a base?
The variable EVENT CD gives the code of the play and codes of 4 and 6
correspond respectively to a stolen base (SB) or caught stealing (CS). Using
the subset function, a new data frame stealing is created that consists of
only the plays where a stolen base is attempted.
stealing <- subset(data2011, EVENT_CD == 6 | EVENT_CD == 4)
By use of the table function, we find the frequencies of the SB and CS

outcomes.
table(stealing$EVENT_CD)
4 6
2863 864
Among all stolen base attempts, the proportion of stolen bases is 2863 /(2863
+ 864) = 0.768.
What are common runners/outs situations for attempting a stolen base?
This is answered by constructing a frequency table for the STATE variable.
table(stealing$STATE)
001 1 001 2 010 0 010 1 010 2 011 1 100 0 100 1 100 2 101 0 101 1
10 1 19 114 101 1 758 1018 1160 35 111
101 2 110 0 110 1 110 2 111 1
180 27 102 89 1
We see that stolen base attempts typically happen with a runner only on first
(state “100”). But there is a wide variety of situations where runners attempt
to steal.
Every stolen base attempt has a corresponding runs value that is stored in
the variable RUNS.VALUE. This runs value reflects the success of the attempt
(either SB or CS) and the situation (runners and outs) where this attempt
occurs. Using the truehist function, a histogram is constructed of all of the
runs created for all the stolen base attempts.
library(MASS)
truehist(stealing$RUNS.VALUE)
Generally, all of the successful SBs have positive runs value, although most of
the values fall in the interval from 0 to 0.3. In contrast, the unsuccessful CSs
(as expected) have negative runs values. In further exploration, one can show
the three spikes for negative runs values correspond to CS when there is only
a runner on first with 0, 1, and 2 outs.
Let’s focus on the benefits of stolen base attempts in a particular situation.
We create a new data frame which gives the attempted stealing data when
there is a runner on first base with one out (state “100 1”).
5
4
3
2
1
0
−1.0 −0.5 0.0 0.5 1.0
stealing$RUNS.VALUE
FIGURE 5.6
Histogram of the runs values of all steal attempts during the 2011 season.
stealing.1001 <- subset(stealing, STATE == "100 1")
By tabulating the EVENT CD variable, we see the runner successfully stole 753
times out of 753 + 265 attempts for a success rate of 74.0%.
table(stealing.1001$EVENT_CD)
4 6
753 265
Another way to look at the outcome is to look at the frequencies of the

NEW STATE variable.
with(stealing.1001, table(NEW.STATE))
NEW.STATE
000 2 001 1 010 1
262 52 704
This provides more information than simply recording a stolen base. On 704
occurrences, the runner successfully advanced to second base. On an additional
52 occurrences, the runner advanced to third. Perhaps this extra base was due
to a bad throw from the catcher or a misplay by the infielder; more can be
learned about the details of these plays by further examination of the other
variables.
We are most interested in the value of attempting stolen bases in this
situation – we address this by computing the mean run values of all of the
attempts with a runner on first with one out.
mean(stealing.1001$RUNS.VALUE)
[1] 0.02649315
Stolen base attempts are worthwhile, although the value overall is about 0.02
runs per attempt. Of course, the actual benefit of the attempt depends on the
success or failure and on the situation (runners and outs) where the stolen
base is attempted.
5.10 Further Reading and Software

Lindsey (1963) was the first researcher to analyze play-by-play data in the
manner described in this chapter. Using data collected by his father for the
1959-60 season, Lindsey obtained the runs expectancy table that gives the
average number of runs in the remainder of the inning for each of the run-
ners/outs situations. Chapters 7 and 9 of Albert and Bennett (2003) illustrate
the use of the runs expectancy table to measure the value of different base hits
and to assess the benefits of stealing and sacrifice hits. Dolphin et al. (2007),
in their Toolshed chapter, describe the runs expectancy table as one of the
fundamental tools used throughout their book. Also, runs expectancy plays a
major role in the essays in Keri et al. (2007).
Ben Baumer and Gregory Matthews have developed an R package openWAR
providing a convenient way of reading play-by-play data from the web and
computing run expectancies. This package uses these run expectancies in the
computation of WAR (wins above replacement) measures for players. The
website www.fangraphs.com/library/misc/war/ introduces WAR, a useful
way of summarizing a player’s total contribution to his team.
5.11 Exercises
1. (Runs Values of Hits)
In Section 5.8, we found the average runs value of a home run and a single.
(a) Use similar R code as described in Section 5.8 for the 2011 season
data to find the mean run values for a double, and for a triple.
(b) Albert and Bennett (2001) use a regression approach to obtain the
weights 0.46, 0.80, 1.02, and 1.40 for a single, double, triple, and home
run, respectively. Compare the results from Section 5.8 and part (a)
with the weights of Albert and Bennett.
2. (Value of Different Ways of Reaching First Base)
There are three different ways for a runner to get on base, a single, walk
(BB), or hit-by-pitch (HBP). But these three outcomes have different runs
values due to the different advancement of the runners on base. Use runs
values based on data from the 2011 season to compare the benefit of a
walk, a hit-by-pitch, and a single when there is a single runner on first
base.
3. (Comparing Two Players with Similar OBPs.)
Rickie Weeks (batter id “weekr001”) and Michael Bourne (batter id
“bourm001”) both were leadoff hitters during the 2011 season. They had
similar on-base percentages – .350 for Weeks and .349 for Bourne. By ex-
ploring the runs values of these two payers, investigate which player was
really more valuable to his team. Can you explain the difference in runs
values in terms of traditional batting statistics such as AVG, SLG, or
OBP?
4. (Create Probability of Scoring a Run Matrix)
In Section 5.3, the construction of the runs expectancy matrix from 2011
season data was illustrated. Suppose instead that one was interested in
computing the proportion of times when at least one run was scored for
each of the 24 possible bases/outs situations. Use R to construct this
probability of scoring matrix.
5. (Runner Advancement with a Single)
Suppose one is interested in studying how runners move with a single.
(a) Using the subset function, select the plays when a single was hit.
(The value of EVENT CD for a single is 20.) Call the new data frame
d.single.
(b) Use the table function with the data frame d.single to construct
a table of frequencies of the variables STATE (the beginning run-
ners/outs state) and NEW.STATE (the final runners/outs state).
(c) Suppose there is a single runner on first base. Using the table from
part (b), explore where runners move with a single. Is it more likely
for the lead runner to move to second, or to third base?
(d) Suppose instead there are runners on first and second. Explore where
runners move with a single. Estimate the probability a run is scored
on the play.
6. (Hitting Evaluation of Players by Runs Values)

Choose several players who were good hitters in the 2011 season. For
each player, find the runs values and the runners on base for all plate
appearances. As in Figure 5.1, construct a graph of the runs values against
the runners on base. Was this particular batter successful when there were
runners in scoring position?
6
Advanced Graphics
CONTENTS
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.2 The lattice Package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2.2 The verlander dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2.3 Basic plotting with lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.2.4 Multipanel conditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.2.5 Superposing group elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.2.6 Scatterplots and dot plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.2.7 The panel function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.2.8 Building a graph, step-by-step . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.3 The ggplot2 Package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.3.2 The cabrera dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.3.3 The first layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.3.4 Grouping factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.3.5 Multipanel conditioning (faceting) . . . . . . . . . . . . . . . . . . . . . . 149
6.3.6 Adding elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.3.7 Combining information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.3.8 Adding a smooth line with error bands . . . . . . . . . . . . . . . . . 151
6.3.9 Dealing with cluttered charts . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.3.10 Adding a background image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.1 Introduction
Chapter 3 introduced graphics in R, illustrating a variety of displays with
functions provided by the graphics package. While this traditional package
has sufficient flexibility for many purposes, the functions are more difficult
to use in graphing complicated data structures such as associations among
three variables. In this chapter, the lattice and the ggplot2 packages are
introduced that are well-suited for constructing more sophisticated graphical
129
displays. In this chapter we focus on the use of these packages for baseball
data.
The datasets used in this chapter are contained in an R workspace, a file
with extension .Rdata which can be loaded with the following command line:
load("data/balls_strikes_count.Rdata")
6.2 The lattice Package

6.2.1 Introduction
The lattice graphics system, written by Deepayan Sarkar, provides an R
implementation of Trellis displays, a regular gridlike framework for visualizing
multivariate data, popularized by William Cleveland. The lattice package is
currently included with the R base distribution and it becomes available for
use by loading with the library function.
library(lattice)
This package produces color figures by default, so the output one obtains
by typing the code in this chapter differs from the black and white figures
printed in these pages. (One who wishes to exactly reproduce the black and
white graphs should change the lattice default theme with the following
command: trellis.par.set(canonical.theme(color=FALSE)).
6.2.2 The verlander dataset

The lattice package is illustrated using data from the verlander data frame
containing five years of PITCHf/x data for 2011 Cy Young Award and MVP
recipient Justin Verlander. To give a glimpse of the contents of verlander,
a random sample of 20 rows is selected from the data frame in the following
R code. The function sample selects a random sample of size 20 from a set
of integers 1 to nrow(verlander), where the function nrow gives the number
of rows of a data frame. The bracket notation is used to choose the rows
of the data frame verlander corresponding to the random integer values in
sampleRows.1
sampleRows <- sample(1:nrow(verlander), 20)
verlander[sampleRows,]
The sample rows are reported in tabular form in Table 6.1.
1 Further arguments in sample that can be specified are replace, which indicates whether
sampling should be performed with replacement and is set as FALSE by default, and a vector
probs if one wants to give different selection probabilities for the set of integers.
TABLE 6.1
A random sample of twenty rows of the verlander dataset.
(row#) season gamedate pitch type balls strikes pitches speed px pz batter hand
Advanced Graphics
7924 2011 2011-04-11 FF 1 0 68 95.80 0.35 3.32 L

11850 2012 2012-09-14 CU 1 2 39 81.20 -1.05 1.11 L
3028 2009 2009-08-30 FT 1 2 94 95.90 0.14 2.43 L
10363 2011 2011-07-31 FF 1 1 64 98.50 -0.45 2.99 R
2890 2009 2009-08-24 FF 2 1 65 99.00 0.10 2.85 R
4370 2010 2010-04-27 FT 0 2 58 98.40 -0.05 4.47 L
8117 2011 2011-04-22 CH 0 1 27 89.00 0.77 1.21 L
11774 2012 2012-09-29 CH 1 1 2 85.20 -1.18 2.04 L
13308 2012 2012-07-04 CU 0 2 86 79.70 -1.63 1.79 L
4147 2010 2010-04-17 FT 0 2 61 97.30 -0.53 4.16 L
12223 2012 2012-09-08 SL 2 0 8 85.60 1.43 1.65 R
10609 2011 2011-08-11 CU 1 1 86 79.30 -0.44 2.99 L
9500 2011 2011-06-25 FF 0 2 49 97.00 0.65 3.88 R
326 2009 2009-04-22 FF 0 0 42 93.80 -0.42 3.63 R
10065 2011 2011-07-21 FF 1 1 17 97.70 0.67 4.11 L
8611 2011 2011-05-13 CU 0 0 80 78.70 -1.09 2.44 R
1795 2009 2009-07-01 FT 0 0 96 97.10 -0.21 2.28 L
13397 2012 2012-06-29 SL 1 1 116 84.80 0.52 3.75 R
7228 2010 2010-09-12 SL 1 0 82 87.00 0.61 2.24 R
882 2009 2009-05-20 CU 2 2 30 82.60 -0.48 1.51 L
131
The first two columns of the verlander data frame, season and gamedate,
contain the season and the date in which the game was played. The
pitch type column indicates the type of pitch thrown by the pitcher as a
two-character abbreviation. The five pitches in Verlander’s repertoire are the
four-seam fastball (FF), the two-seam fastball (FT), the curveball (CU), the
change-up (CH), and the slider (SL.) The balls and strikes columns re-
port the ball-strike count on the batter when the pitch was delivered and the
pitches column contains the number of pitches already thrown by the pitcher
in that particular game. Reading the first row in Table 6.1, we see that Verlan-
der threw a four-seam fastball, delivered as the 69th pitch in the game, with
a 1-0 count on the batter. The next three columns include PITCHf/x data:
speed is the recorded speed at the release of the pitch, px and pz define the
location of the pitch when it crosses the front of the plate. For example, the
first pitch in the table crossed the plate 0.35 feet to the right of the middle of
the plate, 3.32 feet from the ground. The final column indicates whether the
opposing player was batting from the left or the right side of the plate.
6.2.3 Basic plotting with lattice

The lattice package can be used to construct the basic statistical graphs
described in Chapter 3. For example, suppose we are interested in exploring
the speeds of Verlander’s pitches. The histogram function in the lattice
package constructs a histogram analogous to the hist function in the base R
package. For functions in the lattice package, the first argument is a formula
usually requiring at least one variable on the left side of the tilde character
(to be plotted on the y-axis) and one variable on the right side of the tilde (to
be plotted on the x-axis). The data argument gives the data frame containing
the variables. For a one-variable graph such as a histogram only one variable
(such as speed) is needed in the formula.
histogram(~ speed, data=verlander)

The graphical display is shown in Figure 6.1(a).
The densityplot function constructs a density plot (Figure 6.1(b)), an
alternative method of displaying the distribution of a single variable. The
code to produce a density plot is similar to that of the histogram; the option
plot.points=FALSE prevents the display of the data points at the base of the
curve.
densityplot(~ speed, data=verlander, plot.points=FALSE)
The histogram and density plot give similar messages about the speeds of
Verlander’s pitches. There are two noticeable humps in the distribution; the
smaller hump in the mid-80s corresponds to the speeds of the off-speed pitches
and the larger hump in the mid-90s corresponds to the fastballs.
Advanced Graphics 133
0.10
30
25 0.08
20
Percent of Total
0.06
Density
15
0.04
10
0.02
5
0 0.00
60 70 80 90 100 60 70 80 90 100
speed speed
(a) (b)
FIGURE 6.1
Histogram and density plot of Verlander’s pitch speeds using functions in
lattice package.
6.2.4 Multipanel conditioning

One of the main enhancements of the lattice package over the traditional
graphics package is that it allows multipanel conditioning. In other words,
given a grouping factor, it is straightforward to produce a set of as many
displays corresponding to the groups identified by the factor.
As an example, suppose one is interested in constructing separate density
plots for the speed of Verlander’s pitches where each plot is based on the speeds
of a particular pitch type. In this case, the grouping factor is pitch type and,
by use of the lattice package, one can display these “conditional” density
plots on a series of panels using the same scales for the horizontal and vertical
variables.
The multipanel conditioning is obtained in the R graphing function by
adding to the formula the pipeline (|) symbol, followed by the conditioning
variable. For example, the set of conditional density plots is constructed by
means of the following command.2
densityplot(~ speed | pitch_type, data=verlander,

layout=c(1, 5), plot.points=FALSE)
The argument option plot.points=FALSE suppresses the plotting of the data
on the horizontal axes. In Figure 6.2 density plots are plotted for each of the
five different pitch types composing Justin Verlander’s arsenal. This figure
2 The layout argument is used to indicate how the panels are to be arranged. A two-
element vector is specified, indicating the number of columns and rows, respectively.
clearly shows that Verlander’s fastballs are thrown in the mid-90s, his sliders
and changeups in the mid-80s, and his curveballs about 80 mph.
SL
0.20
0.15
0.10
0.05
0.00
FT
0.20
0.15
0.10
0.05
0.00
FF
0.20
Density
0.15
0.10
0.05
0.00
CU
0.20
0.15
0.10
0.05
0.00
CH
0.20
0.15
0.10
0.05
0.00
60 70 80 90 100
speed
FIGURE 6.2
Panels of density plots of Verlander’s pitch speed by pitch type.
6.2.5 Superposing group elements

An alternative way to graphically compare the pitch speeds by pitch type is to
plot the lines of the density plots on a single panel, where one uses different line
types or colors to distinguish the groups. To obtain this superposed display,
one uses the densityplot function with the group argument and the grouping
variable pitch type. The resulting display is shown in Figure 6.3.
densityplot(~ speed, data=verlander, groups=pitch_type,
plot.points=FALSE, auto.key=TRUE)
For a reader to identify the pitch type for each line of the plot, a legend is
needed. The auto.key=TRUE line automatically creates a legend on top of the
plot region. This superposed display gives a similar message as the multipanel
display – Verlander throws pitches at three different speeds.
CH
CU
FF
FT
SL
0.20
0.15
Density
0.10
0.05
0.00
60 70 80 90 100
speed
FIGURE 6.3
Density plot of Verlander’s pitch speed by pitch type using superposed lines.
6.2.6 Scatterplots and dot plots

Multipanel conditioning plots and superposed plots can be constructed using
a variety of graphical methods in the lattice package. The application of
the histogram and densityplot functions have already been illustrated and
other methods are illustrated in this section.
Scatterplots similar to those produced with the plot function in Chapter
3 can be drawn using the xyplot function of the lattice package. Sup-
pose we want to display the trend of the speed of Verlander’s four-seam
fastballs throughout the season, and contrast this trend across the four sea-
sons. As a first step, the subset function is used to create a new data frame
F4verl containing only Verlander’s four-seam fastballs. Using the format
and as.integer functions, the gamedate variable is converted to a variable
gameDay containing the day of the year as an integer from 1 to 365. Using
the aggregate function, the average speeds by day of the year and season are
computed and stored in the variable dailySpeed.
F4verl <- subset(verlander, pitch_type == "FF")
F4verl$gameDay <- as.integer(format(F4verl$gamedate, format="%j"))
dailySpeed <- aggregate(speed ~ gameDay + season, data=F4verl,

FUN=mean)
The xyplot function is used to construct a multipanel conditioning scat-
terplot of speed against gameDay. The syntax speed ∼ gameDay indicates
the scatterplot, and the vertical line (|) followed by the grouping variable
factor(season) indicates that separate scatterplots are to be constructed
for each season. The xlab and ylab arguments control the labels appearing
on the axes.
xyplot(speed ~ gameDay | factor(season),
data=dailySpeed,
xlab="day of the year",
ylab="pitch speed (mph)")
There are interesting patterns in these scatterplots. For example, Verlander’s
fastballs seem to get faster later in the 2012 season, and some of Verlander’s
slowest pitches in the 2010 season occurred towards the end of the season.
To illustrate a different graph, suppose we are interested in comparing Ver-
lander’s fastball and change-up speeds in this four-year period. The verlander
data frame is restricted to only include the pitches labeled as either four-seam
fastballs or change-ups. Then the aggregate function is used to calculate the
average speeds by season and pitch type. (The droplevels function is applied
to get rid of the unused factor levels.)
speedFC <- subset(verlander, pitch_type %in% c("FF", "CH"))
avgspeedFC <- aggregate(speed ~ pitch_type + season,
data=speedFC, FUN=mean)
avgspeedFC <- droplevels(avgspeedFC)
avgspeedFC
pitch_type season speed

1 CH 2009 85.06900
2 FF 2009 96.46576
3 CH 2010 86.72249
4 FF 2010 96.23772
5 CH 2011 87.56312
6 FF 2011 95.83171
7 CH 2012 87.38355
8 FF 2012 95.46240
The dotplot function in the lattice package is used to construct a dot
plot to compare the fastball/change-up speed differential across seasons. By
use of the pch argument, the pitch type first letter is used as the plotting
symbol and cex is used for controlling the symbol size.
dotplot(factor(season) ~ speed, groups=pitch_type,
data=avgspeedFC,
pch=c("C", "F"), cex=2)
100 150 200 250
2011 2012
99
●
98
● ●
●
● ● 97
● ● ● ●
● ● ● ● ● ● ●
● ●
● ● ●
● ● ●
●● ● ● ● ● ●●● 96
● ●
● ● ● ● ●
● ● ● ● ●
●●
● ●
● ●
● ● ●●
● ● ● ● 95
pitch speed (mph)
● ● ● ●
● ●●
94
2009 2010
●
99
● ●
98 ● ● ● ●
●
● ● ● ●
● ● ●
97 ●
● ● ●
●
●
●● ● ● ● ●
●● ● ● ●
● ● ● ●
96 ● ●● ● ● ●● ●● ● ● ●
● ● ● ● ● ● ●
● ●
● ● ●● ●
95 ●
●
● ●●
94
●
100 150 200 250
day of the year
FIGURE 6.4
Scatterplots of the speeds of Verlander’s four-seam fastball against the day of
the year for the four seasons 2009 through 2012.
The graph of Figure 6.5 shows that Verlander’s fastball/change-up speed dif-
ferential has slightly diminished through the years, as Justin lost a very small
amount of speed on his fastball, while simultaneously throwing his change-up
progressively harder from 2009 to 2011.
6.2.7 The panel function

Is Verlander able to maintain the speed of his fastball throughout a game?
The xyplot function in the lattice package is used to construct a scatterplot
of pitch speed against the pitch count. This example illustrates the use of the
panel function argument in lattice to overlay a vertical reference line at
the 100 pitches mark and a horizontal reference line at the four-year average
speed, and add some text and arrows to the graph.
R code to display the basic version of the function xyplot constructs
a scatterplot of speed (pitch speed) against pitches (pitch count) is shown
where the dataset (avgSpeed) is obtained by calculating the average of fastball
2012 C F
2011 C F
2010 C F
2009 C F
86 88 90 92 94 96
speed
FIGURE 6.5
Dot plot of Verlander fastball and change-up speeds for the 2009 through 2012
seasons.
speeds at every pitch count for the four seasons considered. The overall average
fastball speed is stored in the avgSpeedComb object.
avgSpeed <- aggregate(speed ~ pitches + season, data=F4verl,
FUN=mean)
xyplot(speed ~ pitches | factor(season),
data=avgSpeed)
avgSpeedComb <- mean(F4verl$speed)
To add the reference lines and text, one adds a panel argument which is
a function describing the different components graphed. The particular panel
function used for our example is displayed below. The panel.xyplot function
draws the scatterplot, panel.abline functions add the vertical line at the 100
pitch count and the horizontal one at the average speed value, the panel.text
function adds the text, and the panel.arrows function draws the arrows.
panel=function(...){
panel.xyplot(...)
panel.abline(v=100, lty="dotted")
panel.abline(h=avgSpeedComb)
panel.text(25, 100, "avg. speed")
panel.arrows(25, 99.5, 0, avgSpeedComb,
length .1)
}
The panel.abline and panel.text are the lattice equivalent functions to
the abline and text functions in the base graphics described in Chapter
3. Similarly the lattice function panel.arrows corresponds to the arrows
function in base graphics.3
If the panel function argument is added to the basic xyplot function, a
scatterplot is obtained with the reference lines and text added (see Figure
6.6).
xyplot(speed ~ pitches | factor(season),
data=avgSpeed,
panel.xyplot(...)
panel.abline(v=100, lty="dotted")
panel.abline(h=avgSpeedComb)
panel.text(25, 100, "avg. speed")
panel.arrows(25, 99.5, 0, avgSpeedComb,
length=.1)
}
)
This figure clearly shows that the speed of Verlander’s fastball steadily in-
creases during a game, even past the 100-pitch count, a commonly used limit
for starting pitchers.
6.2.8 Building a graph, step-by-step

In this section a graph is built, step-by-step using the lattice package, using
the material introduced in the previous sections of this chapter. For Justin
Verlander’s second no-hitter of his career (Tigers - Blue Jays game of May 7,
2011), we are interested in plotting the location of all of his pitches, taking
in account the batter handedness and the pitch type. The data frame for the
desired game NoHit is obtained using the subset function and the original
data frame verlander.
NoHit <- subset(verlander, gamedate == "2011-05-07")
The variables px and pz give the horizontal and vertical pitch locations,
the variable batter hand contains the batter handedness and the variable
3 The reader will find useful referring to the help entries for the base functions (for
example, by typing ?arrows in the R console) for figuring out the arguments that can be
specified to the various panel functions.
0 50 100
2011 2012
●
●
avg. speed avg. speed 100
● ●
●
●●
● ●● ●
●●
● ●
● ● ● 98
●● ● ●
● ●●
● ●●
● ●● ●● ●●●●● ● ●●●
● ●
● ●●
● ●●●● ●
●● ●
●
●
●●●● ●●
● ●
● ●● ● ● ● ●
● ● ●●● ● ●
● ●● ● ● ●● ● ●
●●● ● ● ● ●
●●● ●
●● ● ●●● ●● ●●
● ●●●● ● ●● ● ● ●● ● ●●● ● ● ●● 96
●●●● ● ● ●●●●● ●
●● ● ● ●
● ● ●● ● ●
●● ●● ●● ● ●● ●● ● ● ●
●● ●
●●●● ● ● ●● ● ●● ●● ● ● ● ●
●●
●
●
●●
● ● ●●●●●● ●●
●
● ● ●
● ●●●
●●● ●●
●
●●●● ● ●●
●● ●●
● ●
●● ●●
●●● ● ● ●●● ●
● ●●●
●
●● ● 94
● ●●
● ●
●
speed
2009 2010
●
● ●
100 avg. speed avg. speed
● ● ●
●
●● ●● ●
98 ●● ●
●●
●●
● ● ●●●
● ● ● ●● ●●
●● ● ● ●● ●● ● ●● ●● ●
●● ● ● ●●●●●●●● ● ●●●●●●●● ●●●●● ●● ● ●●
●
● ●
● ●●●● ● ●●● ●●
●● ● ●●● ●● ● ●● ●●● ●● ●● ● ● ● ● ● ●● ●●●● ●● ● ●●
●● ● ● ●● ●
●● ●
● ● ●●● ● ●● ● ● ●●● ● ● ●●●
●●●●●●●●● ● ● ● ●●● ●● ● ●● ●●● ●
● ●●●● ●●● ●
96 ●
●● ● ●
●● ●
●●●●●●●● ● ●●● ●● ● ●● ●● ●● ●
●● ● ● ●
● ●●● ●● ●
● ●●● ●
●
●●●● ● ● ● ● ●● ● ● ●
● ●
●
● ●
● ●
94
0 50 100
pitches
FIGURE 6.6
Verlander’s four-seam fastball speed through the game - Scatterplot with ver-
tical reference line at 100 pitches and horizontal reference line drawn at the
value corresponding overall fastball speed (as indicated by the text labels and
the arrows).
pitch type contains the pitch type. A scatterplot of the pitch locations is
constructed using the function xyplot; by use of the “| batter hand” option,
separate panels are created for left-handed and right-handed batters, and by
use of the groups=pitch type argument, each of the pitch types is represented
by a different symbol. The auto.key=TRUE argument creates a legend for the
plotting symbols. This first graph is displayed in Figure 6.7.
xyplot(pz ~ px | batter_hand, data=NoHit, groups=pitch_type,
auto.key=TRUE)
Since the axes are both measured in feet, it is desirable that the units on
the x-axis are expressed in the same scale of the units on the y-axis. This
can be ensured by choosing isometric scales, which is accomplished by the
aspect="iso" argument option. With this change, one obtains the graph
shown in Figure 6.8.
CH ●
CU
FF
FT
SL
−2 −1 0 1 2
L R
●
3
●
● ●
●
pz
2 ●
●
●
●
●
1
−2 −1 0 1 2
px
FIGURE 6.7
Location of Verlander’s pitches in his second career no-hitter - base graph.

auto.key=TRUE,
aspect="iso")
Next, we want to limit the plotting area from two feet to the left (of the
center of the plate) to two feet to the right, and from the ground to five feet
above it. The limits of the x and y axes are controlled by the arguments xlim
and ylim. The axes can be labeled with meaningful text using the xlab and
ylab arguments.4 The third graph in our sequence is displayed in Figure 6.9.
auto.key=TRUE,
aspect="iso",
xlim=c(-2.2, 2.2),
ylim=c(0, 5),
xlab="Horizontal Location\n(ft. from middle of plate)",
ylab="Vertical Location\n(ft. from ground)")
4 The \n character sequence is used to split the text over multiple lines.
CH ●
CU
FF
FT
SL
●
3
●
●
●
●
pz
3
●
2 ●
● ●
●
●
1 ●
−2 −1 0 1 2
px
FIGURE 6.8
Location of Verlander’s pitches in his second career no-hitter - graph with
change in aspect ratio.
Further improvements can be made to the display in Figure 6.9. This graph
is missing the rectangle of the strike zone and the legend uses abbreviations
of the pitch type and has neither the title (“pitch type”) nor the border.
The first change is to improve the legend. Other than a logical argument,
the auto.key argument can accept a list of parameters to fine-tune the legend.
The following code prepares this list, after a vector of labels for the pitch
type is created.
pitchnames <- c("change-up", "curveball", "4S-fastball"

, "2S-fastball", "slider")
myKey <- list(space="right",
border=TRUE,
cex.title=.8,
title="pitch type",
text=pitchnames,
padding.text=4)
CH ●
CU
FF
FT
SL
−2 −1 0 1 2
L R
4
Vertical Location
(ft. from ground)
●
3
●
● ●
●
2 ● ●
●
●
1 ●
−2 −1 0 1 2
Horizontal Location
(ft. from middle of plate)
FIGURE 6.9
changes in axes limits and labels.
The meaning of most of the parameters should be clear. Of the others,

padding.text indicates the spacing between the lines of the legend and
cex.title is the font size for the title text.
To plot the strike zone on the figure, it is necessary to consult the baseball
rule book on the size of this region. According to the rule book, the home plate
defines the width of the strike zone. The front side of home plate measures
17 inches. Since the PITCHf/x coordinates system assumes the middle of the
plate is located at the origin of the x-axis, the front side of the plate is located
from −17/12/2 ≃ −0.71 to 17/12/2 ≃ 0.71 feet. However, since just a piece of
the baseball needs to cross home plate in order for the pitch to be considered
a strike, the diameter of the ball (about 2.9 inches) has to be added to both
sides, making the range [−0.95; 0.95]. The height of the strike zone varies from
batter to batter and we chose to use average values for the top and the bottom
of the zones (3.5 and 1.6 feet, respectively.) Based on this discussion, the limits
for the strike zone rectangle can be defined in R by the assignments to the
variables topKzone, botKzone, inKzone, and outKzone.
topKzone <- 3.5

botKzone <- 1.6
inKzone <- -.95
outKzone <- 0.95
The following R code produces the final graph of the locations of Ver-
lander’s pitches displayed in Figure 6.10. The auto.key argument has been
changed from auto.key=TRUE to auto.key=myKey where myKey is the list
of our parameters for the legend. In addition, the panel function argu-
ment is new. The panel.xyplot function produces the scatterplot and the
panel.rect function5 draws the strike zone.
auto.key=myKey,
aspect="iso",
xlim=c(-2.2, 2.2),
ylim=c(0, 5),
xlab="horizontal location\n(ft. from middle of plate)",
ylab="vertical location\n(ft. from ground)",
panel.xyplot(...)
panel.rect(inKzone, botKzone, outKzone, topKzone,
border="black", lty=3)
}
)
This final graph is very informative about the location of Verlander’s pitches
during this no-hit game. For example, we see that Verlander was successful
in throwing his slider down and away to right-handed hitters. We also see
Verlander’s tendency to throw his change-up down and away to left-handed
hitters. His fastballs generally were thrown within the strike zone.
6.3 The ggplot2 Package

6.3.1 Introduction
The ggplot2 graphics system developed by Hadley Wickham is an implemen-
tation in R of Leland Wilkinson’s framework described in Wilkinson (2005).
Wilkinson describes a statistical graphic as a combination of independent com-
ponents. Variables in a dataset are assigned particular roles or aesthetics; the
aesthetic for one variable might be the plotting position along the horizontal
axis and the aesthetic for a second variable might be the color or shape of the
plotting point. Once the aesthetics for a set of variables are defined, then one
5 Equivalent to the rect function for base graphics.
−2 −1 0 1 2
L R
pitch type
4 change−up
(ft. from ground)
●
vertical location
3 ● curveball
●
● ● ● 4S−fastball
2 ●
●
●
●
2S−fastball
●
1 ● slider
●
−2 −1 0 1 2
horizontal location
(ft. from middle of plate)
FIGURE 6.10
changes in legend and addition of strike zone box.
uses a geometric object or geom to construct a graph. Examples of geoms are

points, lines, bars, histograms, and boxplots.
Graphics in ggplot2 are constructed by progressively adding layers, start-
ing with the raw data, then adding strata of statistical summaries and an-
notations. This section illustrates the layer-by-layer building of a ggplot2
graphic.
6.3.2 The cabrera dataset

To illustrate ggplot2 graphics, we consider the cabrera dataset, featuring
PITCHf/x batting data for five seasons of Miguel Cabrera’s career including
the historical 2012 season in which he was the first winner of the batting Triple
Crown since Carl Yastrzemski’s accomplishment in 1967. A random sample of

twenty rows of the dataset is displayed in Table 6.2 using the following code.
sampleRows <- sample(1:nrow(cabrera), 20)
cabrera[sampleRows,]
This dataset introduces four variables not featured in the verlander data
frame: swung indicates whether the batter attempted a swing on the pitch
(coded 1 in case of an attempt and 0 otherwise), and hitx and hity are the
coordinates of the batted ball in feet (for pitches resulting in a ball hit in-play)
where home plate is the origin. Last, the variable hit outcome is an indicator
of the result of the at-bat (again reported only on balls put into play), and
this variable can be either O (out), H (base hit), or E (batter reaching on an
error).
6.3.3 The first layer

One general objective in pitch analysis is the construction of a graph of the lo-
cations of the balls hit into play by Miguel Cabrera – this type of plot is known
among baseball enthusiasts as a spray chart. The first step in constructing this
graph in the ggplot2 package is to find the relevant data frame and assign
roles to the scatterplot variables. The ggplot2 package is first loaded by the
library function. In the ggplot function, the data argument indicates the
use of the cabrera data frame and the aes argument indicates that hitx (the
horizontal coordinate of the batted ball) is assigned the position on the x-axis
and hity (the vertical coordinate of the batted ball) is assigned the position
on the y-axis.
library(ggplot2)
p0 <- ggplot(data=cabrera, aes(x=hitx, y=hity))
The ggplot object is assigned to the variable p0, allowing for the addition of
the ensuing layers. Typing p0 in the R console at this point does not produce
a plot as we have just set up the defaults for the plot and no plotting layer
has been yet created.6
To produce a graph, one needs to add at least one layer to the ggplot
object. For example, the geom point function adds a points layer and creates
a scatterplot. One sees the graph shown in Figure 6.11 by typing p1 in the R
console.
p1 <- p0 + geom_point()
p1
No argument is passed to the geom point function, indicating the defaults set
in the previous ggplot call are to be maintained. One can override the set
defaults for one or more layers, as demonstrated in Sections 6.3.6 and 6.3.8.
6 In fact R would return the following error message: Error: No layers in plot.
TABLE 6.2
Twenty rows sample of the cabrera dataset.
(row#) season gamedate pitch type balls strikes speed px pz swung hitx hity hit outcome
Advanced Graphics
1512 2009 2009-10-01 SL 0 1 80.50 -0.02 2.94 1

1866 2010 2010-05-09 CU 1 1 77.90 0.47 2.66 1 78.60 169.60 O
2776 2010 2010-08-31 FF 0 0 90.80 -2.46 1.82 0
2645 2010 2010-08-18 FC 0 1 85.40 0.77 2.80 0
5361 2012 2012-07-04 SL 0 0 85.50 -1.25 2.33 0
5227 2012 2012-07-15 SL 0 2 83.60 2.14 1.92 0
4285 2011 2011-09-02 CH 1 1 83.30 0.49 3.05 1 94.80 56.19 O
1759 2010 2010-04-25 FF 0 1 87.90 -0.40 2.73 1 -32.15 123.71 O
5048 2012 2012-08-03 SL 0 0 83.90 2.69 1.69 0
5186 2012 2012-07-13 SL 3 2 85.40 -0.43 1.93 1 32.68 339.77 O
5091 2012 2012-08-06 FF 2 0 95.10 0.24 2.02 0
5655 2012 2012-06-07 SI 3 0 89.00 0.46 1.99 0
2510 2010 2010-08-03 FT 3 1 86.00 1.01 3.33 1
3939 2011 2011-07-28 SI 0 0 93.90 0.01 2.58 0
2185 2010 2010-06-23 KN 0 0 78.70 -0.41 2.05 1 -48.34 77.79 O
2410 2010 2010-07-21 FF 0 0 92.40 -1.56 1.72 0
5427 2012 2012-06-22 FF 1 2 95.90 -0.20 1.84 1 13.77 37.28 O
2897 2010 2010-09-18 FF 0 1 96.50 -0.24 3.76 1
5187 2012 2012-07-13 FF 2 2 94.50 -1.72 3.36 0
5477 2012 2012-06-12 SL 3 2 83.10 -0.46 1.98 1 -61.84 77.79 O
147
● ●
●
500 ● ● ●●
● ●
● ● ● ● ●●
● ● ● ●
● ● ●● ●●
● ● ● ● ● ●
● ● ● ● ● ● ●
● ● ●● ●● ● ● ● ●● ●
●●
● ● ● ●● ●● ● ● ●
● ● ● ●
●
● ● ●● ●
●● ●●
● ●
● ● ● ● ●● ● ● ● ●
●
● ●
●● ●● ● ●● ● ● ● ●
● ● ● ● ●● ● ●
●
●
● ● ●● ● ●● ●
● ● ●●
●
●
●● ● ●●● ●● ● ● ● ●●
● ● ●● ● ●● ● ● ● ●
● ● ● ● ● ●● ●
400 ●● ● ●● ● ● ● ●
●● ●●●
● ● ● ●● ● ●● ● ●
● ● ●
●●● ● ● ● ●● ● ●●● ● ●
● ●● ● ●● ● ●●
● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ●● ● ●●● ●
● ● ●● ● ● ● ● ●● ●● ●● ●
●● ● ● ● ● ●
●●● ● ● ● ● ● ●
●●●● ●
● ● ● ●
● ● ●●●
●● ● ● ● ● ● ● ● ●
●
● ●●
●● ●●
●●●●●
● ● ●
● ●
●
● ● ● ●● ● ● ●
●● ● ●● ● ●
● ● ● ●● ● ● ● ●● ●●● ● ●
● ● ● ● ● ●
●
● ● ●● ●● ●
●● ●●
● ●●●
●● ● ● ●
●●● ● ●● ● ● ●
● ● ●● ● ●
● ● ● ● ●●● ● ● ● ●● ●
● ● ● ● ● ●
● ●●● ● ● ● ●● ●●● ●● ●
●● ●●●
● ●● ● ●●
● ● ● ●●
●
● ● ●●● ●●● ●●●● ● ●●
●● ● ● ●● ● ● ●
●
●
●
●
●
●● ●●● ● ● ●●
● ●●
● ●● ● ●
●
●● ●● ● ● ●
●●
●
●●● ●
● ●● ●
● ● ●● ● ● ●
●
● ● ● ● ●●● ● ● ●● ● ●●●●
● ●
●●●●●● ●●●● ● ● ●●● ●
● ●● ● ● ●
● ●
● ● ● ●
● ● ●●●
● ●●●●● ● ● ●● ●● ● ● ● ● ● ● ● ●
● ● ● ●●●
● ● ● ●●●
300 ● ●
●
●
● ●
●●
●
●●●
● ●●●
●●●
●●
●
●
●
●●
●
●
●
●● ●
● ●
●●●
●●● ● ●
● ●
●● ● ● ●
●
●●● ●
● ● ●●●●●● ●●●
●●
●●●● ● ● ● ●● ●●
● ● ●● ●● ●
● ● ● ●●● ●●●
● ●●● ●●●
● ● ●● ● ●●●
●● ●
●●● ●
●
●● ● ● ●● ●● ●
●
● ● ● ● ●●●
● ●● ●● ●●●● ● ● ● ●●● ●● ●
● ● ● ● ●
●
●●●● ● ● ● ●
● ●
● ● ●●●●● ● ● ● ●●● ●
●●●
●●●
●●●● ● ●
●● ●● ●
● ● ●● ● ● ●●● ●● ●●● ●● ● ● ● ●
● ● ● ● ● ●● ●● ●
●
●
●
●
●
●
● ●●●●
●●●
● ●●●●
●
●
●●●● ●●● ●● ● ● ●
●●●●
●●
●●
●● ●
●●
●
●●
●
●
● ●
● ●●●
●●●●● ● ●● ● ● ●● ● ● ● ●● ● ● ●● ●●
●● ● ●●
● ●●●●●● ●● ● ● ● ●●●
●
●●●●● ●● ● ●●●
hity
●● ●●●● ● ● ● ● ●●●
●
● ● ●●
● ●●● ● ●●
●● ● ●● ● ●●●●●● ●● ● ●●
● ●●
● ●●● ●
●● ●●● ● ● ●●
● ● ● ●●● ●
● ●●● ●●● ●●●●
●●●●
● ● ● ● ●● ● ●
● ●● ●● ● ● ● ●
● ● ● ●
●● ● ●●● ● ● ● ●● ●●
●●● ● ● ●● ● ●●
● ● ●● ● ●● ● ●●●●●
●
● ●●●
● ●●●● ● ● ● ● ●
●
● ●
●●● ●
● ●● ●● ●●
●● ● ● ● ● ● ● ●●● ●
● ● ●
200 ● ● ●● ●● ● ●
●●●● ●
●
●●
●
● ●
●
●
●
●● ● ●● ● ● ● ●●
●●● ● ● ●● ●
●●● ●●●● ●●
● ● ●
● ● ● ● ●● ●
● ● ●● ●● ● ●
● ● ● ● ● ● ● ● ● ●●
● ●● ● ●● ●● ●
● ● ● ● ● ●●
● ●●● ●●●● ● ●● ●● ● ●●
● ● ●●
●●
●● ●● ●●
●●●
●●
●
●●● ● ●
●●●●
●●
●●●
●
●●●●
● ●
●● ●●●●
●●
●●
●● ●
●●
● ● ● ●
●●
●●●
●●
●
●
●●●
●●
●
●
●●
●●
●● ●● ●●●●●
●●●●●
●
●● ●
● ●
●●
●●
●● ●
●●●
● ●● ●●●
●●●●●
● ●●● ● ●
●●●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ● ●●●●
●
●
●
●●
●
●
●
●
●●●●
●● ●●
●● ●●● ●●
●
●●
●●
●●●
●
●● ● ●
● ●● ● ●● ●
● ●●●
●●
● ● ●●●● ●●●● ● ● ●
100 ●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●
●
●
●●●●●●
●
●
●
●●
● ●
● ●●● ● ●
●●●
● ●● ● ●
●
●●
● ●●
●
●●
●
●●
●
●
●●
●●●●● ●●
●● ●●●
● ●●
●●
●
●●●
●
●
●●
●
●●
●
●
●●
● ● ●●●●
●● ●
●● ●
●●●
●
●●
●●
●
●●
●●
●
●●●
●● ● ●●●●● ●●●● ●
●●●
●
●
●
● ●●●
●●● ● ●● ●
● ●●
●●● ●●● ● ●
●● ●● ●●●
●●
●●
● ●● ●
● ●●●● ●●
●●
●
● ●● ●
●
● ●●● ● ●●
● ●
●●●●●● ●● ●●●
●●● ●
●● ●● ● ●●
● ●
● ● ● ●
● ●
●
0 ●
●
●
−300 −200 −100 0 100 200 300

hitx
FIGURE 6.11
Scatterplot of Miguel Cabrera’s batted balls (2009 - 2012).
6.3.4 Grouping factors

The hit outcome variable in the cabrera data frame indicates whether a
batted ball resulted in a base hit (H), an out (O), or an error (E). By assigning
the variable hit outcome to the color aesthetic in the ggplot function, the
three different outcomes are represented with different colors in the plot. The
coord equal function ensures the units are equally scaled on the x-axis and
on the y-axis (as they are measured in feet in both cases); it is the equivalent
of passing the aspect="iso" argument to a lattice plot. The plot (Figure
6.12) is viewed after typing p2 in the R console.
p0 <- ggplot(data=cabrera, aes(hitx, hity))

p1 <- p0 + geom_point(aes(color=hit_outcome))
p2 <- p1 + coord_equal()
p2
● ●
●
500 ●● ● ● ●● ●
●
●
●●
● ● ●● ● ●●
●●
●●
● ●● ● ● ●
●
●
● ●●
●
● ●● ●
● ●●●● ●● ●
●●
● ● ● ●●
●● ● ● ● ● ● ● ●
● ● ● ●●● ●● ● ●● ●●● ● ●● ● ●
● ●●● ● ●● ● ●● ●●● ● ● ●
●
● ●
●●●● ● ● ●● ● ●●
● ● ● ● ● ●● ● ● ● ●● ●
●● ● ● ● ● ●● ● ● ● ● ● ● ● ●●
● ● ●● ●●● ●● ● ●
400 ● ●● ● ● ●●● ●
●●●●●●
● ●
● ●
● ●● ●
● ●● ● ●● ●●●●
● ● ●
●● ● ● ● ● ●● ●● ● ● ● ●●
● ● ● ●● ● ●●● ●●●● ● ●●●● ● ● ●● ●
● ● ● ●●● ● ●
● ●● ● ● ● ●● ●●●●●●● ● ●● ● ●● ● ● ● ● ● ●●
●● ●●●● ● ● ● ● ● ● ● ●●● ● ●● ●●● ● ●
●● ●●●● ●● ● ● ●
●● ● ● ● ● ● ● ● ● ●●● ●● ●
● ●● ● ● ●
● ● ● ● ● ●
● ●● ●
●●
●● ●
● ● ●●● ●
●●● ●●● ●●●
●●●● ●
●●●●● ●●
●● ● ●●
●
● ●● ● ● ●
● ●●
● ● ● ●● ● ● ● ●● ●● ● ● ● ●
● ●●●●● ●● ●● ●●●
●● ●● ● ● ● ●
●
●
●● ●●● ●●●
●● ● ● ● ●●●●
●●
●
● ●
●● ● ●●●
●●●●
●● ●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●
● ●
● ●●●●●● ●● ●●
●●●●● ●
● ● ● ●
●
●●●●●
● ●
●●●
●
●
●
●
●
●●
● ●●
● ● ●
●●● ● ●●●
hit_outcome
● ●● ● ● ●● ●●●● ●●
●●● ●●
● ●● ● ● ●●●● ● ● ●
● ●
300 ● ●
● ●
●
●
● ● ●
●
●
●● ●●
●
●● ●● ● ●
●●●●●
●
●
●
●
●
●
●
●●●●
● ●●● ● ●
● ●
●
●
● ●
●●
●
●●
●
●
●●●● ● ●
●
●
●●● ●
● ● ● ●●
● ●● ●●●●●
●●● ● ●●● ● ●● ●● ●●●●● ●● ● ● ● ●● ● ● ● ●● ●● ● ●
● ●
●● ● ●● ● ● ●●
● ●●
●
●
●
●●
●
●
●●
●
●
●
●
● ●● ●●● ●●
●● ●●
●● ● ●
●
●
●●
●●
●●
● ●●●●●
●●● ● ●●●● ●●● ●
●● ●●●
●
●●
●●
●
●●
●●
●
●●
●●
●● ●
●●
●
●
●
●● ●● ●●● ●● ● E
● ●
● ●● ● ●
●●●●●●●●●● ●●●●●●●●●●●●●●●● ●● ● ● ● ●● ●●●●●● ●●●●
●●●
●●
hity
● ●●
●●
●●
●● ●
● ●●● ● ● ● ● ● ● ● ●● ●
●●● ●
●● ●●
● ●
● ● ●● ●
●●● ●● ● ●● ● ●● ● ●●●●●●●● ●● ●● ●●●
● ●● ●● ●●●●●●● ●● ● ● ● ● ● ●
● ● ● ●●
●● ●●
●● ● ●● ● ●●
● ●●● ●
●● ●● ● ●● ●● ●●●● ●●●
●●
●●● ● ● ●●
●● ● ●
●●●●●●● ●
●
● ●●
● ●●
●●● ●
●● ●● ●
●
●●
●●
● ●● ●●
●● ●
●
●●●
●
●
●●●
●
●
●● ●●
● ● ● ●
● H
200 ● ● ●●●●
●
● ●● ●● ●●
● ●●●
● ●
●●●●
●●
●●
● ● ● ●●●
●
●●
● ● ● ●● ● ●
● ●●● ●●●● ● ●● ●
● ●
● ● ●●● ●●●
●●
●●● ●●● ●
●
●● ●● ●●
● ●●● ● O
●● ● ●
●●● ●● ● ●● ●
●●
●●● ●
●●●●● ● ●●●●●● ●
●
● ● ●●●
● ●● ●●●
●●●● ● ●●
● ●●
●●●
●●
●●●
●
●
●●●
●
●●
●●
● ● ● ●●●●
●● ●
●●
●
●
●● ●●● ● ●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
● ●● ● ●
●
●● ●
●●
●
●
●●
●●●●
●●● ●
●●●●●●●●
●
● ●
●●
●●
●
●●
●●
●●●
● ● ● ●●
●●
●
●
●
●
●●
●
●
●● ●
● ● ●
100 ●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●●●
● ● ●
●●●●● ●●●●●
●●
●
●●●● ●●●
●
●●●
● ●●●
● ●
●●
●●
●●●● ●
● ●
●● ●●
●●
●●
●
●
●●
●●●●
●
●
●
●●
●●●● ●●● ●
●●●●●●●●
●●● ● ●
●●
● ●●
●
●● ● ●●● ●
●●● ●●● ● ●
●●
●●● ● ●● ●
● ●● ●●●●
● ●● ●
● ●● ●
●●●●● ●●
●●
●● ●
●● ●●
● ●● ●●
●
● ● ● ●● ● ●
●
0 ●
●
●
●
−300 −200 −100 0 100 200 300

hitx
FIGURE 6.12
Outcomes of Miguel Cabrera’s batted balls (2009 - 2012).
6.3.5 Multipanel conditioning (faceting)

Similar to lattice graphics, multipanel conditioning plots are straightforward
to construct in ggplot2 graphics. The different panels in which one subsets
data are described in ggplot2 as facets. A multipanel conditioning plot is con-
structed with season as the conditioning variable by adding the facet wrap
function with the ∼ season argument to the current plot p2. (See Figure
6.13.)
p3 <- p2 + facet_wrap(~ season)
p3
This figure illustrates how the locations of Cabrera’s batted balls have changed
over the four seasons.
2009 2010
● ●
500 ●●●
●●
●●
● ● ●● ●
●
●
● ● ● ●●●● ● ● ● ● ●●●
●● ● ● ●
●●
●
●●● ●● ● ● ●● ● ●●
● ●● ● ●●●
● ●● ●●● ● ●● ●● ●●
● ● ●● ●● ●●●● ●
400 ●●
●● ●● ●
●●
● ●● ● ●● ●●
●●● ● ●● ● ●●
●●●●●●● ●●●● ● ● ● ● ● ●● ●●●● ● ●●
● ● ●
● ● ● ●●● ●● ● ● ●● ●● ● ● ●
●●●● ● ●●● ● ● ● ● ●●
●●
●●●● ●●● ●●●● ● ● ●● ●●● ●●
● ● ●● ●●● ●
●
● ● ●● ●●
● ● ● ● ●●● ●●● ●
●●● ●
●●●●●●● ● ● ●
● ●● ●●●● ●●● ●● ●● ● ●● ●●● ●●●●● ● ● ● ● ●●● ●●
300 ●● ●● ●●●●
●
●●
●●
●●
● ● ●● ●●● ● ● ●●
●
●
●●
●
●●●●
● ●
● ●
●● ●
●● ●●● ● ● ●● ● ● ●●● ●
● ● ●
●●
●
● ●
●● ●
●● ●●● ●● ●
●●●●
●
●●● ●
●●●●
● ● ●●●● ● ●
●
●●●●● ●●●
● ●●
● ●● ●● ●
●●●
● ● ● ● ● ●● ●●
● ●● ● ● ●●●● ●● ● ●● ●●●
● ●●
●●
●●●●
●●
● ●● ● ● ●●●
●●
●● ● ●● ●● ●
●●●
●● ● ●
●● ●
●●
●●●●● ●
●●● ● ●● ●
●● ● ●
● ●● ● ● ●
200 ●
● ● ●
●
● ●● ● ●
● ●●
●
●
●●●●● ●
●●● ●
●●●
●● ●●● ●
● ● ● ● ●●●● ●
● ●● ●●
●
●
● ● ● ● ● ● ●
●●●●● ●●
●
●
●●
●
●●
●
●
●●
●
● ●
●●
●
●●
●
●
● ●●●
●
●●
●●
●
●
●
●
●● ● ●●
●
●
●●●
●
● ● ●
● ●● ● ● ● ● ● ● ●● ●
100 ●● ●
●
●
●
●
●
●
●
●
●
●
●
●●● ●●●●
●
●
●●
●●
●
●
●
●●
● ●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●● ●●
●●●● ●
●
●
●
● ●●
●●
●●
●●●
●● ●●
●
●●
● ●●●
●● ● ●●● ●● ●● ●
●● ●●● hit_outcome
● ●
0 ● ●
●
● ● E
hity
2011 2012 ● H
500 ● ●● ●●
● ●
● ● ● ● ●● ●
● ●●
●
● ● ●● ●● ● ● ●
●
●●
●● ●
●● ●
● ●
● ●● ● ●●● ●●● ● ●
● ● O
●● ●● ●●
● ●● ● ●●●●●● ●
400 ●
●●●●
● ●●
● ●● ● ● ● ●● ● ●
●
●●●●● ● ● ●●
●● ● ● ●
● ● ●● ●●
●●● ●
●● ●
● ●● ● ● ●●● ●● ●
●● ● ●●● ●● ● ●
●●
● ●●●● ●
● ●●● ● ●
●● ●●●●●●
● ●
●
●●
●● ● ● ● ●
●● ● ●● ● ● ● ●● ●●●●
●
●● ●
●●
●● ● ●
● ●●● ● ● ●
●●●● ●● ● ●● ●
●●●
● ● ●●●●
● ● ● ● ●● ● ●● ●●
●●
●
●●●●●● ● ● ●●● ●● ●●
● ● ●● ● ●● ● ●●●●●●
●● ●● ● ●
●●●
300 ●
● ●
●
● ● ●●●
●
●
●
●
●●
● ●
●
●●
●
●
●
●
●
●●
● ●
●●●
●
● ●
●●●
● ●
● ●
●
●
●
● ●● ●●
●
● ●
● ●
●
●●
●● ● ●●
●
●
●
●●●
●
●● ● ● ●
●
●●●
●
●
●●
● ●●●
● ●
●●●● ●● ●●●●● ●● ●●● ● ●● ●● ●● ●●●●● ●
●● ●
● ●●●● ●● ● ●●●●●
● ●●
●
●
●●
●●
●● ●
● ●
●● ● ●●
●
● ●● ● ●●
●
● ●● ● ● ●●●
●● ●
●
● ●●
●● ●
●● ● ● ●●●● ●●
●
● ● ●
●●
●●
●
●●
●●●
●●
● ● ●● ●●●
●● ● ● ● ●● ●●● ● ● ●
●
●●
●●●
●●●●
● ●
●● ●
200 ●●
● ●● ●●● ● ● ●●
● ● ● ●●● ●●● ●●
●●● ●
●●●
● ●● ● ● ●
● ● ●● ●
● ●● ●●● ● ●
● ●●● ●
● ●● ●●●●● ●● ● ● ●
●
●
●●
●
●●
●
● ●
●
● ●
●●
●
● ●● ● ●●●●
●●●
●●●● ●
●
●●
●●●
● ● ●●●●● ●
● ●
●
●
● ●●
100 ●●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●●●
●●●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●● ●●●●
●
●●
●●
● ●
●● ● ●● ● ● ●●
● ●● ●
●
●● ● ● ●● ● ●
● ● ●● ●
●● ●●● ●
● ●● ●
0 ●
−300−200−100 0 100 200 300−300−200−100 0 100 200 300

hitx
FIGURE 6.13
Outcomes of Miguel Cabrera’s batted balls by season.
6.3.6 Adding elements

One can add reference lines in ggplot2 as in other graphics systems. For
example, one may want to display the base paths in the previous plots. Since
home plate is set to be at the origin and feet are the unit of measurement
in the [hitx, hity] coordinates system, the positions of the remaining three
bases are obtained by a few applications of the Pythagorean theorem.7 A new
data frame bases is defined that gives the location of the bases in (x, y)
coordinates.
bases <- data.frame(x=c(0, 90/sqrt(2), 0, -90/sqrt(2), 0),
y=c(0, 90/sqrt(2), 2 * 90/sqrt(2), 90/sqrt(2), 0)
)
7 The distance between bases as defined by the baseball rule book is 90 feet.
Note that, since we are going to draw a path, the coordinates of home plate
are contained both at the beginning and at the end of the values of x and y.
The displaying of the base paths lines (Figure 6.14) is obtained by adding a
layer to the previous plot.
The new plot with base paths added is constructed by adding the
geom path function component to the current plot p3. In this case, data and
the mapping of aesthetics (aes) are specified as arguments in the geom path
function, overriding the defaults set in the original ggplot function. When
calling the p4 object, two additional layers are added for the display of the
foul lines using the geom segment functions.
p4 <- p3 + geom_path(aes(x=x, y=y), data=bases)
p4 +
geom_segment(x=0, xend=300, y=0, yend=300) +
geom_segment(x=0, xend=-300, y=0, yend=300)
6.3.7 Combining information

Using the ggplot2 graphics system it is easy to combine multiple information
on a single plot. For example, the graph in Figure 6.15 is an enhanced display
of Cabrera’s regular season batted ball data from September and October
2012. This plot simultaneously portrays the batted ball outcome (mapped to
the point shape), the pitch speed (mapped to the point size), and the pitch
type (mapped to the point color). This graph is constructed by use of the
ggplot function with the variable hit outcome assigned the shape aesthetic,
the variable pitch type assigned the color aesthetic, and speed assigned the
size aesthetic. The guides function allows control on how the legend is placed
– in this case, the col=guide legend(ncol=2) argument indicates the color
key is distributed over two columns.
cabreraStretch <- subset(cabrera, gamedate > "2012-08-31")
p0 <- ggplot(data=cabreraStretch, aes(hitx, hity))
p1 <- p0 + geom_point(aes(shape=hit_outcome, colour=pitch_type,
size=speed))
p2 <- p1 + coord_equal()
p3 <- p2 + geom_path(aes(x=x, y=y), data=bases)
p4 <- p3 + guides(col=guide_legend(ncol=2))
p4 +
geom_segment(x=0, xend=300, y=0, yend=300) +
geom_segment(x=0, xend=-300, y=0, yend=300)
6.3.8 Adding a smooth line with error bands

We return to the earlier scatterplot of the pitch speed against pitch count
of Justin Verlander discussed in Section 6.2.7. Using ggplot2, one can cre-
ate a figure analogous to Figure 6.6, with the added bonus of also provid-
2009 2010
● ●
500 ●●●
●●
●●
● ● ●● ●
●
●
● ● ● ●●●● ● ● ● ● ●●●
●● ● ● ●
●●
●
●●● ●● ● ● ●● ● ●●
● ●● ● ●●●
● ●● ●●● ● ●● ●● ●●
● ● ●● ●● ●●●● ●
400 ●●
●● ●● ●
●●
● ●● ● ●● ●●
●●● ● ●● ● ●●
●●●●●●● ●●●● ● ● ● ● ● ●● ●●●● ● ●●
● ● ●
● ● ● ●●● ●● ● ● ●● ●● ● ● ●
●●●● ● ●●● ● ● ● ● ●●
●●
●●●● ●●● ●●●● ● ● ●● ●●● ●●
● ● ●● ●●● ●
●
● ● ●● ●●
● ● ● ● ●●● ●●● ●
●●● ●
●●●●●●● ● ● ●
● ●● ●●●● ●●● ●● ●● ● ●● ●●● ●●●●● ● ● ● ● ●●● ●●
300 ●● ●● ●●●●
●
●●
●●
●●
● ● ●● ●●● ● ● ●●
●
●
●●
●
●●●●
● ●
● ●
●● ●
●● ●●● ● ● ●● ● ● ●●● ●
● ● ●
●●
●
● ●
●● ●
●● ●●● ●● ●
●●●●
●
●●● ●
●●●●
● ● ●●●● ● ●
●
●●●●● ●●●
● ●●
● ●● ●● ●
●●●
● ● ● ● ● ●● ●●
● ●● ● ● ●●●● ●● ● ●● ●●●
● ●●
●●
●●●●
●●
● ●● ● ● ●●●
●●
●● ● ●● ●● ●
●●●
●● ● ●
●● ●
●●
●●●●● ●
●●● ● ●● ●
●● ● ●
● ●● ● ● ●
200 ●
● ● ●
●
● ●● ● ●
● ●●
●
●
●●●●● ●
●●● ●
●●●
●● ●●● ●
● ● ● ● ●●●● ●
● ●● ●●
●
●
● ● ● ● ● ● ●
●●●●● ●●
●
●
●●
●
●●
●
●
●●
●
● ●
●●
●
●●
●
●
● ●●●
●
●●
●●
●
●
●
●
●● ● ●●
●
●
●●●
●
● ● ●
● ●● ● ● ● ● ● ● ●● ●
100 ●● ●
●
●
●
●
●
●
●
●
●
●
●
●●● ●●●●
●
●
●●
●●
●
●
●
●●
● ●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●● ●●
●●●● ●
●
●
●
● ●●
●●
●●
●●●
●● ●●
●
●●
● ●●●
●● ● ●●● ●● ●● ●
●● ●●● hit_outcome
● ●
0 ● ●
●
● ● E
hity
2011 2012 ● H
500 ● ●● ●●
● ●
● ● ● ● ●● ●
● ●●
●
● ● ●● ●● ● ● ●
●
●●
●● ●
●● ●
● ●
● ●● ● ●●● ●●● ● ●
● ● O
●● ●● ●●
● ●● ● ●●●●●● ●
400 ●
●●●●
● ●●
● ●● ● ● ● ●● ● ●
●
●●●●● ● ● ●●
●● ● ● ●
● ● ●● ●●
●●● ●
●● ●
● ●● ● ● ●●● ●● ●
●● ● ●●● ●● ● ●
●●
● ●●●● ●
● ●●● ● ●
●● ●●●●●●
● ●
●
●●
●● ● ● ● ●
●● ● ●● ● ● ● ●● ●●●●
●
●● ●
●●
●● ● ●
● ●●● ● ● ●
●●●● ●● ● ●● ●
●●●
● ● ●●●●
● ● ● ● ●● ● ●● ●●
●●
●
●●●●●● ● ● ●●● ●● ●●
● ● ●● ● ●● ● ●●●●●●
●● ●● ● ●
●●●
300 ●
● ●
●
● ● ●●●
●
●
●
●
●●
● ●
●
●●
●
●
●
●
●
●●
● ●
●●●
●
● ●
●●●
● ●
● ●
●
●
●
● ●● ●●
●
● ●
● ●
●
●●
●● ● ●●
●
●
●
●●●
●
●● ● ● ●
●
●●●
●
●
●●
● ●●●
● ●
●●●● ●● ●●●●● ●● ●●● ● ●● ●● ●● ●●●●● ●
●● ●
● ●●●● ●● ● ●●●●●
● ●●
●
●
●●
●●
●● ●
● ●
●● ● ●●
●
● ●● ● ●●
●
● ●● ● ● ●●●
●● ●
●
● ●●
●● ●
●● ● ● ●●●● ●●
●
● ● ●
●●
●●
●
●●
●●●
●●
● ● ●● ●●●
●● ● ● ● ●● ●●● ● ● ●
●
●●
●●●
●●●●
● ●
●● ●
200 ●●
● ●● ●●● ● ● ●●
● ● ● ●●● ●●● ●●
●●● ●
●●●
● ●● ● ● ●
● ● ●● ●
● ●● ●●● ● ●
● ●●● ●
● ●● ●●●●● ●● ● ● ●
●
●
●●
●
●●
●
● ●
●
● ●
●●
●
● ●● ● ●●●●
●●●
●●●● ●
●
●●
●●●
● ● ●●●●● ●
● ●
●
●
● ●●
100 ●●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●●●
●●●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●● ●●●●
●
●●
●●
● ●
●● ● ●● ● ● ●●
● ●● ●
●
●● ● ● ●● ● ●
● ● ●● ●
●● ●●● ●
● ●● ●
0 ●
−300−200−100 0 100 200 300−300−200−100 0 100 200 300

hitx
FIGURE 6.14
Outcomes of Miguel Cabrera’s batted balls by season (with base lines super-
posed).
ing smoothed lines with shading reflecting the error bands (Figure 6.16).8
This graph is constructed in ggplot2 in five layers. The geom point function
gives the plotted points, the facet wrap produces separate panels by year,
the geom smooth function provides the smoothed line with error bands, the
geom vline function gives the vertical line at 100 pitches, and the geom line
function draws the horizontal line.9
ggplot(F4verl, aes(pitches, speed)) +
facet_wrap(~ season) +
geom_line(stat="hline", yintercept="mean", lty=3) +
geom_point(aes(pitches, speed),
8 Adding smoothed lines to the lattice version of the plot would also be possible, but
would require the use of much more complicated code.

9 Note that in this instance we chose to display in each panel the average speed for the
season displayed in the panel.

hit_outcome
● H
500 O
●
● ● speed
●
400 ● ● ● ●
●
● ● 70
● ●
●
● ● ● 80
300 ● ●
●
●
●
●
● ● ● ● ● ● ● 90
hity
● ● ●●
200 ● ● ● 100
● ●
pitch_type
100 ●● ● CH ● FT
● CU ● SC
0 ● FC ● SI
−300 −200 −100 0 100 200 ● FF ● SL

hitx
● FS
FIGURE 6.15
Miguel Cabrera’s batted balls in the final month of his Triple Crown season
(September/October 2012).
data=F4verl[sample(1:nrow(F4verl), 1000),]) +
geom_smooth(col="black") +
geom_vline(aes(xintercept=100), col="black", lty=2)
To avoid an over-cluttering of data points in the graph, the aesthetics mapping
is respecified in the geom point layer, where a random sample of 1000 points
is used from the F4verl data frame. The smoothing has been calculated on
the original F4verl data frame, because no explicit aesthetics mapping has
been specified in the geom smooth layer.
6.3.9 Dealing with cluttered charts

In our example, plots have been built in ggplot2 by adding geom layers. How-
ever, this graphics system also allows the addition of stat layers, consisting
2009 2010
102.5 ●
●
● ●
● ● ●
● ●
100.0 ●
●
● ●● ●● ● ●
●● ●
●
●● ●●●● ● ●● ●
●
●●
●● ● ● ● ● ● ● ● ● ● ●●
●● ●● ● ●● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●
● ● ●● ●● ●● ●● ●●
● ●● ● ● ●● ● ●●● ●● ●
● ● ● ● ●
● ●● ● ● ● ●●
● ●●
● ● ●● ●● ● ● ●
● ● ● ● ● ●● ● ● ●
● ● ● ●
●
97.5 ●●
● ●
● ●●● ●● ● ● ●
●●
●
●●
●
●
●● ● ●●
● ● ● ●
● ●
●
● ●●
●
●●●
● ●
● ● ● ● ●
● ●
●●●●
●● ●●●●● ●● ●●●●●● ● ●● ●● ● ● ● ●
● ● ● ●● ● ●● ● ●● ●
● ●
● ●● ●● ● ● ● ●●●●●
●●●●
●●●●●● ●●●●●●●● ●● ● ● ● ●●● ● ● ●● ● ● ● ● ●●
● ● ●
● ●●● ● ●● ● ● ●● ● ● ● ● ● ● ●●● ●●●●● ● ●
●● ● ● ● ● ● ● ● ● ● ● ●
●● ● ● ●● ● ●● ●●●
● ●● ● ●
● ● ●
● ● ● ●●
●
●● ● ●● ●
● ●●●
● ● ● ●●
● ●● ●● ● ●
●●● ●
● ●
●● ● ● ●● ●● ●● ●
●●● ●
● ● ●●
95.0 ●
●●●
●
●●● ● ●● ●● ●
● ● ● ●● ● ●
● ● ● ●●●● ● ●
● ●● ●●
●●●● ●● ● ● ●●● ●●
●
●●
●● ● ●
●●● ●● ●
●
●● ● ● ● ● ●
●
● ● ● ● ● ● ●
● ●● ● ● ●
● ● ● ● ● ● ●●
●● ● ●● ●
● ●
● ● ●
●● ● ●
●● ● ●● ● ● ●● ●
●
● ● ● ● ● ● ● ● ●
92.5 ● ●
●
● ● ●
●
● ● ●
● ●
● ●
speed
2011 2012
102.5 ●
● ●
● ● ●●
● ● ● ●●
● ●
100.0 ● ● ●
● ● ● ● ● ●
● ● ● ● ●
● ● ● ● ●
●● ● ●● ● ● ● ● ●●● ●
●● ● ● ●
● ● ● ● ●● ●● ● ● ● ●●●● ● ● ●
●● ● ● ●● ●● ● ●● ● ●● ●
97.5 ● ●●●● ● ● ● ●● ●● ● ●
● ● ● ● ●●
●●
●● ● ● ● ● ● ● ● ● ● ●● ●
●
● ●●●● ●
●● ●● ●●● ● ●● ● ●● ● ●● ●●
● ● ● ●● ● ●
● ● ● ●●● ●● ● ●●
●● ● ● ● ● ● ●● ●
● ●● ● ●● ● ● ●
● ● ●● ● ●●● ● ●
● ●●
●● ●● ● ● ●
● ● ●●● ● ●
●
●● ●● ●● ● ●
●● ● ● ● ●● ● ● ●● ● ● ●
● ● ● ●●
●
● ●● ●
●● ● ● ●
● ● ● ●● ●
95.0 ● ● ●●● ● ●● ● ● ●
●● ●
● ● ● ●● ●●●● ●
● ●●● ●
●●● ● ● ● ● ●
●
● ●● ● ● ● ●
● ●● ● ● ● ● ● ● ● ●● ● ●
● ●● ● ● ● ● ● ●● ● ● ●●● ● ●
●●
●● ● ●● ●
●●
●
● ●● ● ● ● ● ●● ● ● ● ● ● ●
● ● ●● ● ●
● ●● ● ●● ● ● ●
● ●●● ●● ●●
●
●
●
●● ● ● ●●● ● ●●●● ●● ● ● ●● ● ●
● ● ●
● ● ● ●● ● ●● ●
●
● ●
●● ● ●
92.5 ●
● ●● ●
● ● ●
●
● ●● ● ●●
● ●● ● ● ●●
● ●
●● ●
● ●
● ●
● ●
0 50 100 0 50 100
pitches
FIGURE 6.16
Verlander’s four-seam fastball speed through the game - Scatterplot with ref-
erence line at 100 pitches and a smooth line with shading for errors.
of statistical transformations of the data. For example, one can add summary
statistics such as means and standard deviations to the current graph.
These stat layers are useful for dealing with cluttered data. As an ex-
ample, if we are to use a scatterplot to graph the locations of Verlander’s
four-seam fastballs from 2009 to 2012 using the following code, we see an
indistinguishable cloud of black points and it is difficult to see any patterns
(Figure 6.17).
kZone <- data.frame(
x=c(inKzone, inKzone, outKzone, outKzone, inKzone),
y=c(botKzone, topKzone, topKzone, botKzone, botKzone)
)
ggplot(F4verl, aes(px, pz)) +
geom_point() +
facet_wrap(~ batter_hand) +
coord_equal() +
geom_path(aes(x, y), data=kZone, lwd=2, col="white")
One way of handling the cluttering of data points is tiling the [x; y] plane
L R
●
6
● ●
● ● ●
● ● ● ●● ●
●● ● ●●●● ● ● ● ●
● ● ●● ● ● ● ●● ●● ● ● ●●●
●●● ●● ● ●
● ●●●● ●●● ●● ●●●●● ●●
● ●●●
●●● ●● ● ●
● ● ● ●
● ● ●● ●● ●
●●
● ●● ● ●●
● ●
●●
● ●
●
●●
● ●● ● ●● ●● ● ● ● ●
●●● ●● ●●●● ●
●● ● ●●●● ●
●● ●●● ●●● ●● ●
●●●● ●●● ●● ●● ● ●●
●●●●● ●●●
●
●
● ●●●● ● ● ●●●●● ●
●
●●●● ● ●●
●● ●
●● ● ● ●● ● ●● ●● ●
●● ●●●●● ●●
● ●●●●● ● ● ●●
●●● ● ●● ●●
● ● ● ●● ●
● ● ●● ● ● ●● ●●
● ● ●● ●●●
●●●●●●
● ●●●●● ●●●●
●●
●
● ●
●●
●●
●
● ● ●
●●
● ●●
● ●●
●
●●●
●●
●●● ●●
● ● ●●●●● ●● ●●
●●●● ●● ● ●● ●●
● ●●●●●●● ●●● ●● ● ●
●● ●●● ●
●●●
●●●●●
●●●●●● ●●●
●●
●●●● ●●
● ●● ● ● ●● ●●
● ●●
●● ●
●●●● ●●●●●
● ● ● ●●● ●
4 ● ●● ●
●● ●● ● ●●● ●●●●
●●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●●
●●●
●●
●●●
● ●●
●●
●●●
●
●●
●●●
●●● ●
●
●●●
●●●
●
●
●
●
●●●●●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●●
●
● ●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
● ● ●
●
●
● ●●● ●●●●
● ● ●●●●●
●●●
● ●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●●
●●●●
●
●
●
●●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
● ●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●●
●
●●
● ●
●
●
●●●
●
●
● ● ●● ●●● ●●
●●●
●● ●●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●●●●
●●
●
●●
●●
●●
●
●●●
●
●●●●
●●
●
●
●
●●
●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●●●
●●●●●
●●
●
●
●● ● ●
●●●● ●
● ● ●● ●●●●
●●● ● ●
●●●
● ●●
●●●●●●●
●●
●●
● ●
●● ●
●●●
●●
●
●● ● ● ●
●
●● ●● ●●●● ● ●
●● ●● ●
●
●●●● ●●
●
●●●
● ● ●●●●●●
●
●
●●●● ●●●
●
●●
●
●●●●● ●
●●●●
● ●●●●
●●● ●●●●
●●●●
●
●●
●●
●
●
●
●●●
●●
●
●
●●
●●
●●●
●
●●
●
●
●●
●●
●●●●
●
●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●●
● ●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●●
●
●
●
●●
●●●
●
●
●
●
●
●●●
●● ● ● ●
●● ●
●●●●●
●●
●●
●●●
●
● ●
●
● ●●● ●
●
●
●
●
●●
●●
●●●●
●
●●● ●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●● ●
●●
●
●
●●●
●●
●●
●●● ●
●●●●●
●● ● ●
● ●●●
●●● ● ●●●●● ●●
● ●
●
● ●●●
●
●●
● ●●●
●●
●●
●● ●
●●
●●
● ●
●●
●
●
●●
●●
●
●●
●●
●
●
●●
●
●●
●
●
●
●●●
●
●
●●
●●
●●
●
●
●
●
●●
●●
●
●●
●●
●
●●
●
●●●
●
●●●
●●
●
●
●●
●●
●●●
● ●
● ● ●● ●● ● ●●●●● ●●
● ●●●●●
●● ●
●●
●
●●
●
●●
●●●
●
●●●
●●
●●
●●●●
●●●
●
●
●
●●
●
●●
●●
●
●
●●
●●
●●●
●
●●
●
●●●
●
●●
●
●
●●
● ●●●●● ●
● ●●
●
● ●
●
●
●
●●●
●
●
●●
●●
●●
●●
●●
●●
●
●●●
●●●
●●
●●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●●●
●
●●
●
●
●
●●●
●
●●●
●
●
●
●●●
●
●
●●●●
● ●
●●●
●● ●●
●●●
● ●● ●●●●
●●
●●●
●
●
●
●●
●●●●
●
●
●
●
●●●
●
●●
●●
● ● ●
●●
●
●
●
●
●
●
●●
●
●
●●●
● ●●
●●
●●●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●●
●●●
●●●
●
●●
●●● ●●
pz
● ● ●
● ● ●
●● ●
●● ● ● ● ● ●● ●
●
● ●●
●●●● ●● ● ● ● ●●●
● ● ●
●● ● ●●●●● ● ● ● ● ● ● ●
● ● ●●●●●
●●●
●
● ●
●●
●
●●
●
●
●●●
●
●
●●
●●
●●●
●●●
●●
●
●●●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●●●
●
●●
●
●●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●●●
●
●●
●●
●●● ●
●●●● ●●●
●
●
●●●● ●
●
●● ●
●
● ●
●
●●
● ●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●●
●●
●●●
●●
●
●
●●
●
●
●
●●●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●●
●
●● ●● ●●
● ●
●
●
●●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●●
● ●
●
●●
●● ●● ● ●●
●●●●
● ●
●●
●
●●
●
●
● ●
●●●
●
●
●
●
●
●
●●
● ●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
● ●●● ●
●● ● ●● ● ●
● ●●
●● ●● ●
●● ● ●● ●●●
● ●
●●●●●●●●●
●●
● ●
●●
●● ●
● ●
● ●●●● ●
● ● ●● ● ●● ● ● ● ●
●
●● ●● ●●●
●● ●● ● ●● ● ● ●
●● ●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●●
●●
●
●
●
●
●
●●●
● ●●● ● ● ●
●●●
●
●●●
● ●●●
●●●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●●
●●●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●●●
●●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●
●●●
●
●
●●
●●
●
●
●●●
● ●●
●●
●
●
●● ●
●
●●●●● ● ●● ●●
● ●●
●●
●●
●●
●●
●●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●●
●
●●●
●●
●
●●
●
●
●
●
●
●●
●
●
●●
●●●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●●●●●●
●●●●●●● ● ●●●●
● ●● ●
●● ● ●● ●
●●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●●
●
●●
●●●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●●●●●●
●
● ●
●●●●
●●●
●
●●●● ●
●●●●●
●
●●●●
●
●●
●●● ●●
●●●
●●
●●
●
●●
●●
●
●●
●●
●●
●
●●
●●
●
●●
●●●
●●
●
●●
●●
●
●
●
●●●
●
●
●
●
●●
●
●●
●●
●
●●●
●
●
●●
●
●●
●●
●
●●●●
● ●●
●●
●
● ●
●●●● ● ●●●●
●●●
●
●
●●● ●
●●●
●●
●●
●
●●
●●●
●●
●
●●●●
●●●
●●
●●
●●●
●●
●
●●
● ●
●
●
●●●
●
●
●●
●
●●
●●
●●
●
●
● ●
●
●●
●
●●
●
●●
●
●
●●
●
●●●●● ● ●
● ●●
●●●
● ●●●
●
● ● ●●● ●●
●●
●●●● ●
●
● ●
●
●
●●
●●
●●
●●
● ●
●
●●
●●●●
●
●●
●●
●●●
●
●●
●
●●
●●
●
●●
● ●
●●
●
●●
●●
●●●
●
●
●
●
●●
● ●
●
●
●●
●●
●●●
●
●
●●
●●
●●
●
●●
●
● ●
●●
●●●●
● ●● ●●●●
● ●●●
● ●●●
●● ●●●
●●●●
●●
●
●●
●●
●●
●
●●
●●
● ●●●
●●
●
●●
●
●●●●
●
●●
●●
●●
●
●●
●●●
●
●
●●
●●
●●
●●●
●●
●
●●
●●
●
●●
● ●
●
●●●
●
● ●
●●●
●● ●●●
●
● ● ●●
●●
●
●
●●●
●
●
●●
●●
●●
●
● ●
●
●●●
●
●
●
●●●
●
●●●
●
● ●
●●●
●●●
●
●
●
●●
●
●●
●
●●
●●
●
●
●
●●
●●
●
●●
●
●
●●●
●
●●
●●
●●
● ●
●
●
●●
●●
●
●●
●
● ●
●
●
●●●
●●●
●
●
●
● ●●
●● ● ● ● ●
●●
●
●●
●
●●●
●
●●
● ●●
●
● ●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●●●
●●
●●
●●
●
●●
●
●●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●●●
● ●
● ●●
●
●
●●
●
● ●● ●●
● ●
●●
●
●●●●
●
●
●●
●
●
●
●
●
●●●●
●
●
●
●
●●
●
●
●
●
●●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●● ●
●
●●●
●●
●
●
●
●
●●●●● ●●
● ●
● ●● ●
●●
●●●● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●● ●●
●
●●
●●
●
●
●●● ● ●
● ● ● ● ● ● ●● ● ● ● ● ● ●
2 ●●
●● ●
● ●●
● ●
●●
●●
●
●
●●
●
●
●
●
●
●
●●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●● ●
●●●
● ●
●●
● ●●
●
●
●
●●●●
●●
●
●●●●
●●
●●
●●●●
●
●
●●
●
●
●●●
●●●
●
● ●●
●
●●●●● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●●●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●●
●
● ●
●●●● ●
●● ● ●
●●●●●
●● ●●
●● ●
●● ●● ●
●●●
● ● ●
● ●
●
●● ●● ● ● ● ● ●● ●
● ●● ●●
●● ● ●●●
● ● ●●●
●● ●●●
●●● ●●● ●
● ●
● ●● ●
●●●● ●
●●●
● ●
●● ●●●
●●
●
●●
● ●●
●
● ●●
●●●●
●
●
●●●● ●●● ●●● ●● ● ●●● ●●●●
●●●● ●● ●●
●
●●
●●●●●
●●● ●
●●●●●●
● ●
●●
●● ●
●●
●● ●
●●●● ●●●● ●
● ●●
●●
●●
●●● ●●
●●●●
●
●●●
●
●
●
●●●●●
●
●●● ●
●●
●●● ●●●● ●●
●
●
●
● ●
●● ● ● ●● ● ● ● ● ●●
● ●
●
●
●
●●
●
●●
●● ●
● ●●●●●●
● ●●●● ●● ● ●
●● ● ● ● ●● ● ● ●● ● ● ●
● ●● ● ● ● ●
●●● ●●●● ● ● ●● ●● ●● ● ●● ●●●●●●● ●
● ●
● ●●
●
●●●● ●● ●
●
●●● ●● ●●
● ● ●● ● ● ●
● ●●
●●●●●● ●●
● ● ●● ●● ●● ●● ●● ●● ●
● ●● ● ● ● ●● ● ●● ●
● ●
● ● ●● ● ● ●
● ● ●
●●
0 ●
−3 −2 −1 0 1 2 −3 −2 −1 0 1 2
px
FIGURE 6.17
A cluttered scatterplot - Locations of Justin Verlander’s four-seam fastballs
by batter handedness (2009-2012).
with hexagonal bins and having the bins colored according to the number
of data points they contain. One can create a new graph by replacing the
geom point layer with a stat binhex layer. The new graph is displayed in
Figure 6.18.
ggplot(F4verl, aes(px, pz)) +
stat_binhex() +
facet_wrap(~ batter_hand) +
coord_equal() +
geom_path(aes(x, y), data=kZone, lwd=2, col="white", alpha=.3)
Note how the alpha argument in the geom path layer is used to adjust the
transparency of the strike zone border.
6.3.10 Adding a background image

Adding images to the plot background is usually not a good idea, as they
often do nothing more than obscure the data points. However, plotting batted
L R
6
count
40
4 30
pz
20
2
10
0
−3 −2 −1 0 1 2 −3 −2 −1 0 1 2
px
FIGURE 6.18
Hexagonal binning as a way to portray 2D density - Locations of Justin Ver-
lander’s four-seam fastballs by batter handedness (2009-2012).
ball locations might be an exception, since the figure of the baseball field may
give a better reference guide than the horizontal and vertical axes.
Packages exist for reading common format images into R. For example, if
one obtains a diagram of Detroit’s Comerica Park10 as a jpeg file, the jpeg
package can be used for reading the image into R by the readJPEG function.11
library(jpeg)
diamond <- readJPEG("Comerica.jpg")
The diamond object is a three-dimensional array of dimension x × y × 3,
where x and y correspond to the dimension of the image in pixels. The array
can be seen as a collection of three x × y matrices, containing information
on the red, green, and blue (RGB) components at every pixel of the image
respectively, expressed as values in the [0, 1] range.
10 This diagram was retrieved as a svg file from MLBAM Gameday at gd2.mlb.com/
images/gameday/fields/svg/2394.svg and then converted to a jpeg format.

11 Similar packages named png and bmp provide analogous functionality for reading other
common image formats.

The diagram can be added in ggplot2 as a layer of the plot using the
annotation raster function. The following code is used to obtain Figure
6.19.
ggplot(cabrera, aes(hitx, hity)) +
coord_equal() +
annotation_raster(diamond, -310, 305, -100, 480) +
stat_binhex(alpha=.9, binwidth=c(5, 5)) +
scale_fill_gradient(low="grey70", high="black")
The four numbers passed to the annotation raster layer indicate where
the image has to be positioned and were found by trial and error.12 In the
stat binhex layer two arguments have been specified: alpha gives the degree
of transparency to the hexagons so that their coloring does not completely
hide the diamond diagram and binwidth sets the dimensions of the hexagons.
Finally, the scale fill gradient layer sets the coloring of the hexagonal bins
as a gradient starting from the color grey70 and ending at the color black.13
6.4 Further Reading

Detailed descriptions of the lattice and ggplot2 packages are available in
Sarkar (2008) and Wickham (2009) by the respective developers.
6.5 Exercises
1. (Location of Pitches for Left- and Right-Handed Batters)
Use a density plot to display the horizontal location of Justin Verlander’s
pitchers by opponent’s handedness. Choose the conditioning and grouping
variables so that one can easily detect the differences in location by hand-
edness. Add a legend (if necessary) and vertical reference lines indicating
the borders of the strike zone.
2. (Comparing Pitch Locations for Two Pitchers)
12 Since the image is imported in R as an array of RGB values, it would be possible to
retrieve the position of two points (for example home plate and second base) and compute
the four values to be passed in the annotation raster layer. For an example on how to
identify an object inside an image with R look at is-r.tumblr.com/post/36874307174/
finding-a-bright-object.
13 A map of R colors by name is available at research.stowers-institute.org/efg/R/
Color/Chart/ColorChart.pdf.
500
400
count
15
300
10
hity
200
5
100
−300 −200 −100 0 100 200 300

hitx
FIGURE 6.19
Scatterplot of Cabrera’s batted balls (2009 - 2012) with Detroit’s Comerica
Park diagram in the background. Note: batted balls for games in other ball-
parks are included as well.
The sanchez data frame contains 2008-2012 PITCHf/x data for pitcher
Jonathan Sanchez. The structure of this data frame is similar to the
verlander data frame described in the chapter. Use a graphical display
to compare the ability of Sanchez and Verlander in maintaining their fast-
ball speed through the game. (See Sections 6.2.7 and 6.3.8.) Use either
the lattice or ggplot2 graphics package and display the data either as
a multipanel plot or a superposed lines plot.
3. (Graphical View of the Speeds of Justin Verlander’s Fastballs)
(a) The cut function is useful for recoding a continuous variable into
intervals. Use this function to categorize the pitches variable in the
verlander data frame in groups of ten pitches.
(b) Use the bwplot function from the lattice package to produce a
boxplot of Verlander’s four-seam fastball speed (use the F4verl data
frame) for each ten-pitches group. Compare the information conveyed
by the resulting chart with that of Figure 6.6.
4. (Exploring Miguel Cabrera’s Slugging Power)
(a) Create a data frame by selecting, from the cabrera data frame, the
instances where the hit outcome variable assumes the value H (for
base hit).
(b) Using the hitx and hity variables, create a new variable equal to
the distance, in feet from home plate, of Cabrera’s base hits. (This
variable is computed by simply applying the Pythagorean Theorem–
remember that home plate is at the origin.)
(c) In the newly created data frame, create a gameDay variable indicating
the day of the year (from 0 to 365) in which the game took place (see
Section 6.2.6).
(d) Build a scatterplot featuring gameDay on the x-axis, distance on the
y-axis and a smooth line with error bands. Does the resulting plot
appear to indicate changes in Cabrera’s power during the season?
7
Balls and Strikes Effects
CONTENTS
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.2 Hitter’s Counts and Pitcher’s Counts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.2.2 An example for a single pitcher . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.2.3 Pitch sequences on Retrosheet . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.2.3.1 Functions for string manipulation . . . . . . . . . . . . 165
7.2.3.2 Finding plate appearances going through a
given count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.2.4 Expected run value by count . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.2.5 The importance of the previous count . . . . . . . . . . . . . . . . . . 170
7.3 Behaviors by Count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7.3.1 Swinging tendencies by count . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7.3.1.1 Propensity to swing by location . . . . . . . . . . . . . . 173
7.3.1.2 Effect of the ball/strike count . . . . . . . . . . . . . . . . 176
7.3.2 Pitch selection by count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.3.3 Umpires’ behavior by count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
7.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.1 Introduction
In this chapter we explore the effect of the ball/strike count on the behav-
ior of players and umpires and on the final outcome of a plate appearance.
Retrosheet data from the 2011 season is used to estimate how the ball/strike
count affects the run expectancy. Also PITCHf/x data is used to explore how
one pitcher modifies his pitch selection based on the count and, similarly, how
one batter alters his swing zone, and umpires judge pitches according to the
count. Functions for string manipulation are introduced that are useful for
managing the pitch sequences from the Retrosheet files. Level plots and con-
tour plots, created with the use of the lattice package, will be used for the
explorations of batters’ swing tendencies and umpires’ strike zones.
161
7.2 Hitter’s Counts and Pitcher’s Counts

7.2.1 Introduction
When watching a broadcast of a baseball game, one often hears an announcer’s
concern for a pitcher who is constantly “falling behind” in the count, or his/her
anticipation for a particular pitch because it’s a “hitter’s count” and the batter
has a chance to do some damage. We will see if there is actual evidence that
the so-called hitter’s count really leads to more favorable outcomes for batters,
while “getting ahead” in the count (a pitcher’s count) is beneficial for pitchers.
7.2.2 An example for a single pitcher

The Baseball Reference1 website provides for every player a plethora of
splits statistics. The website provides splits by ball/strike counts for all
seasons since 1988. Mike Mussina’s split statistics are found by entering
the player’s profile page (typing “Mussina” on the search box brings one
there), clicking on the “Splits” tab in the “Standard Pitching” table, and
clicking on “Career” (or whatever season we are interested in) on the pop-
up menu that appears. One finds the “Count Balls/Strikes” table scrolling
down on the splits page. Alternatively, the table can be reached by a di-
rect link: in this case the career splits by count for Mike Mussina are cur-
rently available at www.baseball-reference.com/players/split.cgi?id=
mussimi01&year=Career&t=p#count.
The first series of lines (from “First Pitch” to “Full Count”) shows the
statistics for events happening in that particular count. Thus, for example, a
batting average (BA) of .337 on 1-0 counts indicates batters hit safely 34%
of the time when putting the ball in play on a 1-0 count against Mussina.
We are more interested in the second group of rows, those beginning with the
word “After”. In fact, in these cases the statistics are relative to every plate
appearance that goes through that count. Thus a .337 on-base percentage
(OBP) after 1-0 means that, whenever Mike Mussina started a batter with a
ball, the batter successfully got on base 34% of the time, no matter how many
pitches it took to end the plate appearance.
The last column on every table in the splits page is ”tOPS+”. It’s an index
for comparing the overall player’s OPS (the sum of on base percentage and
slugging percentage2 ) with his OPS in that particular situation. A value over
100 is interpreted as a higher OPS in the situation compared to the overall
OPS; conversely values below 100 indicate a OPS value that is lower than the
overall OPS.
1 baseball-reference.com
2 OPS is widely used as a measure of offensive production because, while being very easy
to calculate, it correlates very well, at the team level, with runs scored.
Balls and Strikes Effects 163
Figure 7.1 uses a heat map to display Mussina’s tOPS+ through the vari-
ous counts. If one focuses on a particular number of strikes, a higher number
of balls in the count makes the outcome more likely to be favorable to the
hitter (lighter shades). Conversely, if one fixes the number of balls, the bal-
ance moves towards the pitcher (darker shades) as one increases the number
of strikes. This figure emphasizes the importance from a pitcher’s perspec-
strikes
0 1 2
0 100 72 30
1 118 82 38
balls
2 157 114 64
3 207 171 122
FIGURE 7.1
Heat map of tOPS+ for Mike Mussina through each balls/strikes count. Data
from Baseball Reference website.
tive of beginning the duel with a strike. When Mussina fell behind 1-0 in his
career, batters performed 18% better than usual in the plate appearance; con-
versely, after a first pitch strike, they were limited to 72% of their potential
performance.
How is the heatmap display of Figure 7.1 created in R? First a data
frame mussina is prepared with all the possible balls/strikes counts, using
the expand.grid function as previously illustrated in Section 4.7. A new vari-
able value is added with the tOPS+ values taken from the Baseball-Reference
website.
mussina <- expand.grid(balls=0 : 3, strikes=0 : 2)
mussina$value <- c(100, 118, 157, 207, 72, 82, 114, 171, 30, 38,
64, 122)
mussina
balls strikes value

1 0 0 100
2 1 0 118
3 2 0 157
4 3 0 207
5 0 1 72
6 1 1 82
7 2 1 114
8 3 1 171
9 0 2 30
10 1 2 38
11 2 2 64
12 3 2 122
The graph is created using the new function countmap with argument
mussina, making use of the color2D.matplot function in the plotrix pack-
age.
countmap <- function(data){
require(plotrix)
data <- xtabs(value ~ ., data)
color2D.matplot(data, show.values=2, axes=FALSE
, xlab="", ylab="")
axis(side=2, at=3.5:0.5, labels=rownames(data), las=1)
axis(side=3, at=0.5:2.5, labels=colnames(data))
mtext(text="balls", side=2, line=2, cex.lab=1)
mtext(text="strikes", side=3, line=2, cex.lab=1)
}
countmap(mussina)
Let’s explore the commands inside the countmap function. The first line
loads the package plotrix. (This package needs to be installed beforehand for
the function to work.) The second line reshapes the data in a contingency table
format that is needed for creating the heatmap table. For example, passing
the mussina data frame as the data argument, the data frame looks like the
following after applying the xtabs function:
strikes
balls 0 1 2
0 100 72 30
1 118 82 38
2 157 114 64
3 207 171 122
The color2D.matplot function is called to draw the plot. By use of the argu-
ment show.values=2, we indicate that the values (tOPS+) are to be shown.3
3 The parameter show.values requires either a logical value (TRUE/FALSE), indicating
The arguments axes=FALSE, xlab="", ylab="" indicate that axes and labels
are not drawn. The axes and labels are explicitly drawn by use of the axis
and mtext functions.
The axis function controls how axes are drawn. The side of the plot where
the axis should appear is controlled by the side parameter; to place an axis
on top of the plot, as was done in Figure 7.1, one specifies side=3. The at
parameter controls the positions at which tick marks are to be drawn, while
labels indicates the text that is to be placed at the tick points. Several
other optional parameters can be passed to the axis function to control the
positioning of the axis, the appearance of the line and the tick marks and the
font for the labels (type ?axis on the R console for more details).
The mtext function allows to place text at the margins of a plot. Similarly
to what happens for axis, the side parameter specifies whether the text
should be placed to the bottom (side=1), the left (2), the top (3), or the right
(4) of the plot. Other parameters are used to indicate the distance from the
margin at which the text should be placed (line) and its appearance.
7.2.3 Pitch sequences on Retrosheet

From viewing Figure 7.1, one obtains an initial view of hitter’s counts (lighter
shades) and pitcher’s counts (darker shades) on the basis of offensive produc-
tion. However this figure is based on data for a single pitcher – does a similar
pattern emerge when using league-wide data?
Retrosheet provides pitch sequences beginning with the 1988 season. Se-
quences are stored in strings such as the example FBSX. Each character encodes
the description of a pitch. In this example, the pitch sequence is a foulball,
followed by a ball, a swinging strike, and the ball put into play. Table 7.1
provides the description for every code used in Retrosheet pitch sequences.4
7.2.3.1 Functions for string manipulation

Sequence strings from Retrosheet need some initial processing to place in
a form suitable for analysis. A quick tutorial on some R functions for the
manipulation of strings is provided in this section. Readers not interested in
string manipulation functions may skip to Section 7.2.4.
The function nchar returns the number of characters in a string. This func-
tion is helpful for obtaining the number of pitches delivered in a Retrosheet
pitch sequence. For example, the number of pitches in the sequence BBSBFFFX
is given by
nchar("BBSBFFFX")
[1] 8
whether the values should be displayed inside the cells, or an integer number, dictating the
maximum number of decimal places to be displayed.
4 Source: www.retrosheet.org/eventfile.htm.
TABLE 7.1
Pitch codes used by Retrosheet.
Symbol description
+ following pickoff throw by the catcher
* indicates the following pitch was blocked by the catcher
. marker for play not involving the batter
1 pickoff throw to first
2 pickoff throw to second
3 pickoff throw to third
> indicates a runner going on the pitch
B ball
C called strike
F foul
H hit batter
I intentional ball
K strike (unknown type)
L foul bunt
M missed bunt attempt
N no pitch (on balks and interference calls)
O foul tip on bunt
P pitchout
Q swinging on pitchout
R foul ball on pitchout
S swinging strike
T foul tip
U unknown or missed pitch
V called ball because pitcher went to his mouth
X ball put into play by batter
Y ball put into play on pitchout
However, as indicated in Table 7.1, there are some characters in the Retrosheet
strings denoting actions that are not pitches, such as pickoff attempts.
The functions grep and grepl are used to find a pattern within elements
of character vectors. The function grep returns the indices of the elements
for which a match is found, and the function grepl returns a logical vector,
indicating for each element of the vector whether a match is found. For both
functions, the first argument is the string pattern to search and the second
argument is the vector of strings where matches are sought. For example, we
apply the two functions to the vector of pitch sequences sequences in search
for pickoff attempts to first base denoted by the code 1.
sequences <- c("BBX", "C11BBC1S", "1X")
grep("1", sequences)
[1] 2 3
grepl("1", sequences)
[1] FALSE TRUE TRUE
The function grep tells us that “1” is contained in the second (2) and third
(3) components of the character vector sequences, and grepl outputs this
same information by means of a logical vector.
The pattern parameter to search does have to be a single character. For
example we may want to look for consecutive pickoff attempts to first which
is the pattern “11”. The below output shows that “11” is contained in the
second component of sequences.
grepl("11", sequences)
[1] FALSE TRUE FALSE
Also the function gsub allows for the substitution of the pattern found with
a replacement. The replacement can be an empty string, in which case the
pattern is simply removed. For example, the following code removes the pickoff
attempts to first from the pitch sequences.
gsub("1", "", sequences)
[1] "BBX" "CBBCS" "X"
7.2.3.2 Finding plate appearances going through a given count

Since we are interested only in pitch counts, it is necessary to remove the char-
acters not corresponding to actual pitches from the pitch sequences. Regular
expressions are the computing tool needed for this particular task. While it’s
beyond the scope of this book to fully explain how regular expressions work,
we will instead show a few examples on how to use them.5
We begin by loading the all2011.csv file containing Retrosheet’s play-
by-play for the 2011 season.
5 The website www.regular-expressions.info/ is a very comprehensive online resource
on regular expressions, featuring examples, tutorials, reference for syntax, and a list of
related books.
pbp2011 <- read.csv("retrosheet/all2011.csv")

headers <- read.csv("retrosheet/fields.csv")
names(pbp2011) <- headers$Header
The gsub function is used to create the new variable pseq of pitch se-
quences which removes the symbols from the Retrosheet pitch sequence vari-
able PITCH SEQ TX that don’t correspond to actual pitches.
pbp2011$pseq <- gsub("[.>123N+*]", "", pbp2011$PITCH_SEQ_TX)
The square brackets indicate, in a regular expression, the collection of charac-
ters to search. The above code removes pickoff attempts at any base (1, 2, 3)
either by the pitcher or the catcher (+), balks and interference calls (N), plays
not involving the batter (.), indicators of runners going on the pitch (>), and
of catchers blocking the pitch (*).6
Another special character is used to identify the plate appearances that
go through a 1-0 count. In a regular expression, the ˆ character means the
pattern has to be matched at the beginning of the string. Looking at Table
7.1 for the four different ways a ball can be coded, the following line script
creates the desired variable c10.
pbp2011$c10 <- grepl("^[BIPV]", pbp2011$pseq)
Similarly, plate appearances going through a 0-1 count are identified by use
of the new variable c01 variable.
pbp2011$c01 <- grepl("^[CFKLMOQRST]", pbp2011$pseq)
To check our work, the values of PITCH SEQ TX, c10, and c01 are displayed
for the first ten lines of the data frame.
pbp2011[1:10, c("PITCH_SEQ_TX", "c10", "c01")]
PITCH_SEQ_TX c10 c01

1 X FALSE FALSE
2 CBCS FALSE TRUE
3 CBBBB FALSE TRUE
4 BCSBS TRUE FALSE
5 CBB1>S FALSE TRUE
6 CBB1>S.FBFB FALSE TRUE
7 CCX FALSE TRUE
8 BX TRUE FALSE
9 CBBFX FALSE TRUE
10 BFCX TRUE FALSE
Writing regular expressions for every pitch count is a tedious task and we
will defer the reader to Appendix A for the full code. For the purpose of this
chapter a play-by-play file containing additional information on ball/strike
counts is provided.
6 Applying the nchar function to the newly create variable pseq gives the number of
pitches delivered in each at-bat.

7.2.4 Expected run value by count

The file pbp11rc.csv contains an enhanced version of the play-by-play data
for the 2011 season. Other than the typical information provided by Ret-
rosheet, this data file reports the run value for each play as calculated in
Chapter 5 and additional variables such as c00 and c10 indicating for each
possible ball/strike count whether the at-bat has gone through that particular
count. This data is stored into the data frame pbp11rc.
pbp11rc <- read.csv("pbp11rc.csv")
pbp11rc[1:5, c("GAME_ID", "EVENT_ID", "c00", "c10", "c20", "c11",
"c01", "c30", "c21", "c31", "c02", "c12", "c22", "c32",
"RUNS.VALUE")]
GAME_ID EVENT_ID c00 c10 c20 c11 c01 c30 c21 c31 c02 c12
1 ANA201104080 2 1 0 0 0 0 0 0 0 0 0
2 ANA201104080 3 1 0 0 1 1 0 0 0 0 1
3 ANA201104080 4 1 0 0 1 1 0 1 1 0 0
4 ANA201104080 5 1 1 0 1 0 0 0 0 0 1
5 ANA201104080 6 1 0 0 1 1 0 1 0 0 0
c22 c32 RUNS.VALUE
1 0 0 -0.1555661
2 0 0 -0.0953746
3 0 0 0.3571207
4 1 0 -0.3347205
5 1 0 -0.3919715
For example, the at-bat in the second line of the data frame started with a 0-1
count (value 1 in column c01), then moved to the counts 1-1 and to 1-2 and
generated a run value of -0.095. The pbp11rc data frame has all the necessary
information to calculate the run values of the various balls/strikes counts,
in the same way the value of a home run and of a single were calculated in
Chapter 5.
As an illustration, one can measure the importance of getting ahead on
the first pitch. The mean run value is calculated for at-bats starting with a
ball and for the at-bats starting with a strike.
ab10 <- subset(pbp11rc, c10 == 1)
ab01 <- subset(pbp11rc, c01 == 1)
c(mean(ab10$RUNS.VALUE), mean(ab01$RUNS.VALUE))
[1] 0.03969483 -0.03546708
The conclusion is that the difference between a first pitch strike and a first
pitch ball, as estimated with data from the 2011 season, is over 0.07 runs.
The runs value can be calculated for each possible ball/strike count. First
a data.frame named runs.by.count is prepared with the twelve possible
counts as done in Section 7.2 and a zero value is temporarily assigned to all
counts.
runs.by.count <- expand.grid(balls=0 : 3, strikes = 0 : 2)

runs.by.count$value <- 0
A function bs.count.run.value is created that calculates the mean of the

column RUNS.VALUE given the number of balls and strikes as arguments. In
the function, note that in the pbp11rc data frame, the plate appearances are
limited to the ones that have gone through the particular ball-strike count.
bs.count.run.value <- function(b, s){
column.name <- paste("c", b, s, sep="")
mean(pbp11rc[pbp11rc[, column.name] == 1, "RUNS.VALUE"])
}
The mapply function is used to apply bs.count.run.value to every row of

the runs.by.count data frame, creating a new variable value.
runs.by.count$value <- mapply(FUN=bs.count.run.value,
b=runs.by.count$balls,
s=runs.by.count$strikes
)
Finally, using the countmap function introduced in Section 7.2 with the
runs.by.count data frame, the run values are visualized for all of the possible
balls/strikes counts. (See Figure 7.2.)
countmap(runs.by.count)
By glancing at the values and shading colors in Figure 7.2, one can con-
struct reasonable definitions for the terms “hitter’s count” and “pitcher’s
count.” Ball/strike counts can be roughly divided in the following four cate-
gories7 :
• Pitcher’s counts: 0-2, 1-2, 2-2, 0-1;

• Neutral counts: 0-0, 1-1;
• Modest hitter’s counts: 3-2, 2-1, 1-0;
• Hitter’s counts: 3-0, 3-1, 2-0.
7.2.5 The importance of the previous count

In the previous section run values were calculated for any ball/strike count. In
performing this calculation we simply looked at whether a plate appearance
went through a particular count, without considering how it got there. In
other words, we considered, for example, all the at-bats going through a 2-2
7 The proposed categorization, based on an observation of Figure 7.2 reflects the one pro-
posed by analyst Tom Tango (see www.insidethebook.com/ee/index.php/site/comments/

plate_counts/).
strikes
0 1 2
0 0 −0.04 −0.09
1 0.04 −0.01 −0.07

balls
2 0.1 0.03 −0.03
3 0.18 0.12 0.05
FIGURE 7.2
Run value for plate appearances through each balls/strikes count. Values es-
timated on data from the 2011 season.
count as having the same run expectancy, no matter if the pitcher started
ahead 0-2 or fell behind 2-0. The implicit assumption in these calculations is
that the previous counts have no influence on the outcome on a particular
count. However, a pitcher getting ahead 0-2 is likely to “waste some pitches.”
That is, he would likely throw a few balls out of the strike zone with the sole
intent of making the batter (who cannot afford another strike) swing at them
and possibly miss or make poor contact. On the other hand, with a plate
appearance starting with two balls, the batter has the luxury of not swinging
at strikes in undesirable locations and wait for the pitcher to deliver a pitch
of his liking.
Given the above discussion, it would seem that the run expectancy on a
2-2 count would be higher if the plate appearance started with two balls than
if the pitcher started quickly with a 0-2 count. Let’s investigate if there is
numerical evidence to actually reflect this conjecture.
We begin by taking the subset of plays from the 2011 season that went
through a 2-2 count and calculate their mean run value.
count22 <- subset(pbp11rc, c22 == 1)
mean(count22$RUNS.VALUE)
[1] -0.03134373
A new variable after2 is created, denoting the ball/strike count after two
pitches. The mean run value is calculated for each of the three possible levels
of after2.
count22$after2 <- ifelse(count22$c20 == 1, "2-0",
ifelse(count22$c02 == 1, "0-2", "1-1"))
aggregate(RUNS.VALUE ~ after2, data=count22, FUN=mean)
after2 RUNS.VALUE
1 0-2 -0.02440420
2 1-1 -0.03277434
3 2-0 -0.03570539
The above results appear counterintuitive, as they seemingly imply that plate
appearances going through a 2-2 count after having started with two strikes
are more favorable to the hitter than those beginning with two balls.
This surprising result is actually a byproduct of a selection bias. Many
plate appearances starting with two strikes end without ever reaching the
2-2 count, in most cases with an unfavorable outcome for the batter.8 The
plate appearances that survive a 0-2 count reaching 2-2 are hardly a random
sample of all the plate appearances. Likely hard-to-strike-out batters are over-
represented in such sample, as well as pitchers who do not posses a quality
pitch to finish off opponents.
Using a similar study, comparing the paths leading to 1-1 counts gives
results more in line with common sense as this count is less susceptible to the
same selection bias.
count11 <- subset(pbp11rc, c11 == 1)
count11$after1 <- ifelse(count11$c10 == 1, "1-0", "0-1")
aggregate(RUNS.VALUE ~ after1, data=count11, FUN=mean)
after1 RUNS.VALUE
1 0-1 -0.013234572
2 1-0 -0.009274024
The numbers above suggest that after reaching a 1-1 count, the batter is
expected to perform slightly better if the first pitch was a ball than if it was
a strike.
8 In 2011, 80% of plate appearances beginning with two strikes and not reaching the 2-2
count ended with the batter making an out.

7.3 Behaviors by Count

In this section we explore how the roles of three individuals in the pitcher-
batter duel are affected by the ball/strike count. How does a batter alter his
swing according to being ahead or behind in the count? How does the pitcher
vary the mixing of pitches according to the count? Does an umpire, more
or less consciously, shrink or expand his strike zone depending on the pitch
count?
An R Workspace (a file with extension .Rdata) is available containing all
the datasets used in this section. Once the workspace is loaded into R, the
available data frames cabrera, umpires, and verlander are displayed by use
of the ls function.
load("balls_strikes_count.Rdata")
ls()
[1] "cabrera" "umpires" "verlander"
These datasets contain pitch-by-pitch data, including the location of pitches as
recorded by Sportvision’s PITCHf/x system. The cabrera data frame contains
four years of batting data for 2012 American League Triple Crown winner
Miguel Cabrera. The data frame umpires has information about every pitch
thrown in 2012 where the home plate umpire had to judge whether it crossed
the strike zone. The verlander data frame has four years of pitching data for
2011 Cy Young Award and MVP recipient Justin Verlander.
7.3.1 Swinging tendencies by count

It was seen in Section 7.2 that batters perform worse when falling behind in
the count. For example, when there are two strikes in the count, the batter
may be forced to swing at pitches he would normally let pass by to avoid
being called out on strikes. Using PITCHf/x data, we explore how a very
good batter like Miguel Cabrera alters his swinging tendencies according to
the ball/strike count.
7.3.1.1 Propensity to swing by location

The contents of the cabrera data frame was introduced to the reader in
Chapter 6. In this section we focus on the relationships between the variables
balls and strikes indicating the count on the batter, the variables px and
pz identifying the pitch location as it crosses the front of the plate, and the
swung binary variable, denoting whether or not the batter attempted a swing
on the pitch.
Using the xyplot function in the lattice package, Miguel Cabrera’s
swinging tendency is displayed by a scatterplot in Figure 7.3.9
9 The provided code produces a color figure in which swings and takes are better dis-
sampCabrera <- cabrera[sample(1:nrow(cabrera), 500),]

topKzone <- 3.5
botKzone <- 1.6
inKzone <- -.95
outKzone <- 0.95
library(lattice)
xyplot(pz ~ px, data=sampCabrera, groups=swung,
aspect="iso",
xlab="horizontal location (ft.)",
ylab="vertical location (ft.)",
auto.key=list(points=TRUE, text=c("not swung", "swung")
, space="right"),
panel.xyplot(...)
border="black")
})
The overlapping in the scatterplot is reduced by use of a random sample of

500 pitches from the data frame cabrera. From Figure 7.3, one can see that
Cabrera is less likely to swing at pitches delivered farther away from the strike
zone (the black box). However, it is difficult to determine Cabrera’s preferred
pitch location from this figure.
A contour plot is an effective alternative method to visualize batters’
swinging preferences. The plot is used to visualize three dimensional data
on a two-dimensional surface. Widely used in cartography and meteorology,
the contour plot usually features spatial coordinates as the first two variables,
while the third variable (which can be, for example, elevation in cartography
or barometric pressure in meteorology) is plotted as a contour line, also called
an isopleth. The contour line is a curve joining points sharing equal values of
the third variable.
As a first step in producing a contour plot, a polynomial surface is fit using
the loess function using the horizontal and vertical locations px and pz as
predictors and swung as the outcome variable. The output of this fit is stored
in the variable miggy.loess.
miggy.loess <- loess(swung ~ px + pz, data=cabrera,
control=loess.control(surface="direct"))
After the surface on the dataset has been fit, we are interested in predicting
the likelihood of a swing by Cabrera at various pitch locations. Using the
expand.grid function, a data frame is built consisting of combinations of
horizontal locations from −2 (two feet to the left of the middle of home plate)
to +2 (two feet to the right of the middle of the plate) and vertical locations
cerned than in Figure 7.3 If one wants to obtain the black and white version displayed
in this book, the following line of code has to be inserted before the calling of xyplot:
trellis.par.set(canonical.theme(color=FALSE))
●
5
●
●
● ● ●
●
● ● ● ●
●
4 ●
●
●●
● ●
● ●
● ● ●
● ●
●●● ● ● ● ●● ●
●● ● ● ● ● ●● ●
●
● ●● ●
● ● ●● ● ●
● ● ● ●
vertical location (ft.)
●● ●
● ●● ●
3 ● ● ●
●
●●
●● ● ●
●
●● ● ●●
● ●● ●
● ● ●
● ● ● ●
● ●● ●
● ● ●● ● ● ●●
● ● ●●
●●
● ● ●● ● ● ●
●
●
●
● ● ●● ● ●
●
not swung ●
● ●● ●
2 ●
●
●● ●
●● ● ●● ● ●
● ● ●
●
●
swung
● ● ● ● ●● ● ●● ●
● ● ● ●
● ● ● ●● ●
● ● ● ● ●
● ● ● ●
● ●
●● ● ● ● ●
● ● ●
● ● ● ● ●●
● ● ●
1 ● ● ●
● ●
●
● ●
● ● ● ●
● ● ●
● ●
● ●
● ●
0
●
●
−1
−2 −1 0 1 2
horizontal location (ft.)
FIGURE 7.3
Scatterplot of Miguel Cabrera’s swinging tendency by location. Sample of 500
pitches. View from the catcher’s perspective.
from the ground (value of zero) to six feet of height, using subintervals of
0.1 feet. By using the predict function,10 the likelihood of Miguel’s swing is
obtained at every location in the data frame.
pred.area <- expand.grid(px=seq(-2, 2, 0.1), pz=seq(0, 6, 0.1))
pred.area$fit <- c(predict(miggy.loess, pred.area))
From the data frame pred.area the likelihood that Miguel will swing is
estimated for three different locations – a pitch down the middle and two and
a half feet from the ground (“down Broadway”), a ball that hits the ground in
the middle of the plate (“ball in the dirt”), and another one delivered at mid-
height (2.5 feet from the ground) but way outside (two feet from the middle
of the plate). In each case, the subset function is used to take a subset of the
prediction data frame pred.area with specific values of the horizontal and
vertical locations px and pz.
10 The c function used in the second assignment converts the matrix resulting from the
predict call into a vector.

subset(pred.area, px == 0 & pz == 2.5) #down Broadway

px pz fit
1046 0 2.5 0.8202528
subset(pred.area, px == 0 & pz == 0) #ball in the dirt
px pz fit
21 0 0 0.1729128
subset(pred.area, px == 2 & pz == 2.5) #way outside
px pz fit
1066 2 2.5 0.06238047
The results are quite consistent with what one would expect: the pitch right
in the heart of the strike zone generates Cabrera’s swing more than 80 percent
of the time, while the ball in the dirt and the ball outside generates a swing
at 17 percent and six percent rates, respectively.
A contour plot of the likelihood of the swing as a function of the horizon-
tal and vertical locations of the pitch is constructed using the contourplot
function in the lattice package. The meaning of most arguments in the R
code has been explained in Chapter 6; the at parameter is used to indicate
the levels at which we want the contour lines to be drawn. Figure 7.4 shows
the resulting contour plot.
contourplot(fit ~ px * pz, data=pred.area,
at=c(.2, .4, .6, .8),
aspect="iso",
xlim=c(-2, 2), ylim=c(0, 5),
panel.contourplot(...)
border="black", lty="dotted")
})
As expected, the likelihood of a swing decreases the further the ball is delivered
from the middle of the strike zone. The plot also shows that Cabrera has a
tendency to swing at pitches on the inside part of the plate.
7.3.1.2 Effect of the ball/strike count

Figure 7.4 reports Miguel’s swinging tendency over all pitch counts. Can we
visualize how Cabrera varies his approach according to the ball/strike count?
Specifically, does Cabrera become more selective when he is ahead and can
afford to wait for a pitch of his liking and, conversely, does he “expand his
zone” when there are two strikes and he cannot allow to let another called
strike go by? The process of calculating the swing propensity by location has
been described in Section 7.3.1.1. Here the same process is performed on a
subset of data, specifically on pitches delivered on a 0-0 count.
In the following R code, a new variable bscount is defined giving the
0.2
4
0.6
3 0.8
1
0.4
−1 0 1
FIGURE 7.4
Contour plot of Miguel Cabrera’s swinging tendency by location, where the
view from the catcher’s perspective. The contour lines are labeled by the
probability of swinging at the pitch.
ball-strike count. Using the subset function, a new data frame miggy00 is
constructed consisting of pitch data only when the ball-strike out is 0-0. The
loess function is used to find the likelihood of swing surface, and the predict
function computes the likelihood for every horizontal and vertical location in
the data frame pred.area.
cabrera$bscount <- paste(cabrera$balls, cabrera$strikes, sep="-")
miggy00 <- subset(cabrera, bscount == "0-0")
miggy00loess <- loess(swung ~ px + pz, data=miggy00, control=
loess.control(surface="direct"))
pred.area$fit00 <- c(predict(miggy00loess, pred.area))
The same procedure was repeated for the 0-2 and 2-0 counts. The code
is not shown here, but the reader can write the R code by slight modifica-
tions of the code for the 0-0 count. Once the additional variables fit02 and
fit20 were obtained for the pred.area data frame, contour plots on separate
panels can be constructed to compare Cabrera’s swinging tendencies by pitch
count (Figure 7.5.) In the contourplot function, the separate panels display
is produced by means of the formula fit00 + fit02 + fit20 ∼ px * pz.
contourplot(fit00 + fit02 + fit20 ~ px * pz, data=pred.area,
at=c(.2, .4, .6),
aspect="iso",
xlim=c(-2, 2), ylim=c(0, 5),
panel.contourplot(...)
border="black", lty="dotted")
})
As expected, Cabrera expands his swing zone when behind 0-2 (his 40% con-
tour line on 0-2 counts has an area comparable to his 20% contour line on 0-0
counts). The third panel on Figure 7.5 does not suggest a shrinkage of Cabr-
era’s swinging zone on 2-0 counts. Miguel seems to be increasingly looking for
pitches up-and-in when ahead in the count.
−1 0 1
0.2
fit00 fit02 fit20
0.2 0.2
0.4
0.2 0.4
4 0.6
0.6
3
2
0.6
0.4
1 0.
2
−1 0 1 −1 0 1
FIGURE 7.5
Miguel Cabrera’s 50/50 swing zone in different balls/strikes counts. View from
the catcher’s perspective.
7.3.2 Pitch selection by count

We now move to the other side of the pitcher/batter duel in our learning about
the effect of the pitch count. Pitchers generally possess arsenals of two to five
different pitch types. Nearly all pitchers have a fastball at their disposal, which
is generally a pitch that is easy to locate at the desired location. So-called
complimentary pitches, such as curve balls or sliders, while often effective
(especially when hitters are not expecting them), are harder to control and
rarely used by pitchers behind in the count. In this section we look at one
pitcher (arguably one of the best in MLB at the time of this writing) and
explore how he chooses from his pitch repertoire according to the ball/strike
count.
The verlander data frame, consisting of over 15 thousand observations,
consists of pitch data for Justin Verlander for five seasons. Using the table
command, a frequency table of the types of pitches Verlander has thrown since
2009 is obtained.
table(verlander$pitch_type)
CH CU FF FT SL
2550 2716 6756 2021 1264
As is the case with most major league pitchers, Verlander most frequently
uses the fastball. He uses two variations of a fastball, a four-seamer (FF) and
a two-seamer (FT). He complements his fastballs with a curve ball (CU) ,a
change-up (CH), and a slider (SL).
The prop.table function is used to obtain the pitch type proportions
rather than their frequencies. The input to prop.table is the table of fre-
quencies. To obtain percentages, the result is multiplied by 100 and rounded
(using the round function) to the nearest integer.
round(100 * prop.table(table(verlander$pitch_type)))
CH CU FF FT SL
17 18 44 13 8
It can be seen from the table that 44% of Verlander’s pitches during this
five-season period were four-seamers.
Before moving to exploring pitch selection by ball/strike count, a frequency
table is used to explore the pitch selection by batter handedness. One con-
structs frequencies of pitch type for each batter hand by specifying two pa-
rameters in the table function. To compute proportions of pitch type for
each batter hand, the prob.table function is used with argument margin=2.
(The function computes proportions by row if margin=1 and by column if
margin=2.)
type_verlander_hand <- with(verlander, table(pitch_type,
batter_hand))
round(100 * prop.table(type_verlander_hand, margin=2))
batter_hand
pitch_type L R
CH 23 8
CU 17 18
FF 43 45
FT 15 11
SL 2 17
Note that the pitch selection is quite different depending on the handedness
of the opposing batter. In particular, the right-handed Verlander uses his
change-up nearly a quarter of the time against left-handed hitters, but only
eight percent of the time against right-handed hitters. Conversely the slider is
nearly absent from his repertoire when he faces lefties, while he uses it close
to one out of five times against righties.
Batter-hand differences in pitch selection are common among major league
pitchers and they exist because the effectiveness of a given pitch depends
on the handedness of the pitcher and the batter. The slider and change-up
comparison is a typical example, a slider is very effective against batters of
the same handedness and a change-up can be successful when facing opposite-
handed batters.
Justin’s pitch selection can be explored by pitch count. First, a new
bscount variable is created by merging the values in the columns balls and
strikes. Since pitch selection depends on batter handedness, pitch mixing is
examined separately by the opponent handedness. In the following code, the
subset function is used to select Verlander’s pitches delivered to right-handed
batters. The table function constructs a table of frequencies by count and
pitch type and the prop.table function with the margin=1 argument com-
putes row proportions.
verlander$bscount <- paste(verlander$balls, verlander$strikes,
sep="-")
verl_RHB <- subset(verlander, batter_hand == "R")
verl_type_cnt_R <- table(verl_RHB$bscount, verl_RHB$pitch_type)
round(100 * prop.table(verl_type_cnt_R, margin=1))
CH CU FF FT SL
0-0 7 11 53 16 13
0-1 6 24 40 10 19
0-2 16 28 27 6 22
1-0 5 11 52 11 21
1-1 8 24 40 10 19
1-2 14 33 28 6 19
2-0 2 2 70 15 12
2-1 6 8 51 14 20
2-2 10 29 36 9 16
3-0 8 0 81 10 0
3-1 2 0 78 12 8
3-2 4 4 69 12 11
The effect of the ball/strike count on the choice of pitches is apparent when
comparing pitcher’s counts and hitter’s counts. When behind 2-0, Verlander
uses his four-seamer seven times out of ten; the percentage goes up to 78%
when trailing 3-1 and 81% on 3-0 counts. Conversely, when Justin has the
chance to strike the batter out, the use of the four-seamer diminishes. In fact
he throws it less than 30 percent of the time both on 0-2 and 1-2 counts. On
a full count, Verlander’s propensity to go with the fastball is similar to the
one in hitters’ counts – this is consistent with the numbers in Figure 7.2 that
indicate the 3-2 count being slightly favorable to the hitter. Similarly, one can
explore Verlander’s choices by count when facing a left-handed hitter.
7.3.3 Umpires’ behavior by count

Hardball Times author John Walsh wrote a 2010 article The Compassionate
Umpire,11 in which he showed that home plate umpires tend to modify their
ball/strike calling behavior by slightly favoring the player who is behind in
the count. In other words, umpires tend to enlarge their strike zone in hitter’s
counts and to shrink it when pitchers are ahead. In this section we visually
explore John’s finding, by plotting contour lines for three different ball/strike
counts.
The umpires data frames are similar to the verlander and cabrera
ones. A sample of its contents, obtained by typing the command
umpires[sample(1:nrow(umpires), 20),] in the R console is shown in Ta-
ble 7.2. The data consist of every pitch of the 2012 season for which the
home plate umpire had to judge whether it crossed the strike zone. Additional
columns not present in either the verlander or the cabrera data frames iden-
tify the name of the umpire (variable umpire) and whether the pitch was called
for a strike (variable called strike).
We proceed similarly to the analysis of Section 7.3.1.2, using the loess
function to estimate the umpires’ likelihood of calling a strike, based on the
location of the pitch. Following is the R code for the 0-0 count. Note that the
analysis is limited to plate appearances featuring right-handed batters, as it
has been shown that umpires tend to call pitches slightly differently depending
on the handedness of the batter. In addition, the loess smoother is applied
on a subset of 3000 randomly selected pitches, to facilitate the computation
time of the script.
umpiresRHB <- subset(umpires, batter_hand == "R")
ump00 <- subset(umpiresRHB, balls == 0 & strikes == 0)
ump00smp <- ump00[sample(1:nrow(ump00), 3000),]
ump00.loess <- loess(called_strike ~ px + pz, data=ump00smp,
control=loess.control(surface="direct"))
By slightly modifying the code above, the reader can easily repeat the process
for other counts. In this section we compare the 0-0 count to the most extreme
batter and pitcher counts, 3-0 and 0-2 counts, respectively.
In this instance the contour lines of balls/strikes calling in different counts
will be plotted in a single panel, rather than in separate panels as was done
for swinging tendencies in Figure 7.5. The contourLines function12 is helpful
for this task, as it calculates coordinates to plot contour lines. The following
code calculates, for the 0-0 count, the 50% strike calling contour line and
11 www.hardballtimes.com/main/article/the-compassionate-umpire/.
12 The function comes with the grDevices package, which is loaded by default in R.
Analyzing Baseball Data with R
TABLE 7.2
A twenty rows sample of the umpires dataset.
(row #) season umpire batter hand pitch type balls strikes px pz called strike
80357 2012 Doug Eddings R SL 0 0 -0.17 2.88 1
33199 2012 Alfonso Marquez L FF 0 0 0.03 2.53 1
187437 2012 Angel Hernandez L SL 1 1 -1.48 2.35 0
260385 2012 Ed Rapuano R FT 1 1 -1.44 2.05 0
169709 2012 Sam Holbrook L FF 2 1 0.80 2.42 0
195437 2012 Dan Bellino L FT 0 0 1.07 1.59 0
166333 2012 Eric Cooper R FF 0 0 -0.77 2.35 1
264734 2012 Gerry Davis L FF 0 0 0.55 2.08 1
191453 2012 D.J. Reyburn R FC 1 1 -1.13 2.69 1
10811 2012 Kerwin Danley R SL 1 1 0.89 2.89 0
12474 2012 Mark Ripperger R FF 0 0 -1.11 1.80 0
134379 2012 Laz Diaz L FF 0 0 -1.84 3.21 0
55770 2012 Jeff Nelson R CU 2 2 1.70 2.02 0
203183 2012 Tim Timmons R FF 1 0 1.18 2.97 0
202977 2012 Sam Holbrook L SL 2 1 -0.23 1.01 0
58915 2012 Mark Wegner L CH 0 1 -1.65 2.42 0
300411 2012 Tim Timmons L FF 0 0 -1.21 2.40 0
84853 2012 Paul Nauert R CU 3 2 -0.23 1.04 0
312177 2012 Jim Joyce L SL 3 2 -0.25 1.65 0
216589 2012 Brian Runge L FT 1 1 -0.66 2.29 1
182
converts the resulting list into a data frame, to which a new variable bscount
indicating the balls/strikes count is added.
ump00contour <- contourLines(x=seq(-2, 2, 0.1),
y=seq(0, 6, 0.1),
z=predict(ump00.loess, pred.area),
levels=c(.5))
ump00df <- as.data.frame(ump00contour)
ump00df$bscount <- "0-0"
The function contourLines from the grDevices package (which is loaded

by default in R) requires two vectors (x and y) of values in ascending order,
that combined define the locations where the values are measured. The val-
ues are stored in a matrix z whose dimensions are given by length(x) ×
length(y). Then either a number of levels (nlevels) can be specified, or a
vector (levels) which indicates the levels at which to draw contour lines.
The function returns values that can then be supplied to plotting functions in
order to draw the contour lines.
The 50% strike calling contour lines for the counts of interest can be plotted
as line plots. The output produced by the following code (after ump02df and
ump30df are created similarly to ump00df) can be seen in Figure 7.6.
umpireContours <- rbind(ump00df, ump02df, ump30df)
trellis.par.set(theme=canonical.theme(color=FALSE))
myKey <- list(lines=TRUE
, points=FALSE
, space="right"
, title="balls/strikes count"
, cex.title=1
, padding=4)
xyplot(y ~ x , data=umpireContours
, groups=bscount
, type="l", aspect="iso"
, col="black"
, xlim=c(-2, 2), ylim=c(0, 5)
, xlab="horizontal location (ft.)"
, ylab="vertical location (ft.)"
, auto.key=myKey
, panel=function(...){
panel.xyplot(...)
border="grey70", lwd=2)
})
This figure shows that the umpire’s strike zone is shrunk in a 0-2 pitch count,
and slightly expanded in a 3-0 count.
4
3
balls/strikes count
0−0
0−2
2
3−0
−1 0 1
FIGURE 7.6
Umpires’ 50/50 strike calling zone in different balls/strikes counts viewed from
the catcher’s perspective.
7.4 Further Reading

Palmer (1983) is possibly one of the first examinations of the balls/strikes
count effect on the outcome of plate appearances: it is based on data from
World Series games from 1974 to 1977 and features a table resembling Figures
7.1 and 7.2. Walsh (2008) calculates the run value of a ball and of a strike at
every count and uses the results for ranking baseball’s best fastballs, sliders,
curveballs, and change-ups. Walsh (2010) shows how umpires are (perhaps
unconsciously) affected by the balls/strikes count when judging pitches. In
particular, he presents a scatterplot showing a very high correlation between
the strike zone area and the count run value (see Figure 7.2). Allen (2009a,
2009b) and Marchi (2010) illustrate so-called platoon splits (i.e. the different
effectiveness against same-handed versus opposite-handed batters) for various
pitch types.
7.5 Exercises
1. (Run Value of Individual Pitches)
(a) Calculate the run value of a ball and of a strike at any count. For
3-ball and 2-strike counts you need the value of a walk and a strike-
out respectively (you can calculate them as done for other events in
Chapter 5).
(b) Compare your values to the ones proposed by John Walsh in the
article www.hardballtimes.com/main/article/
searching-for-the-games-best-pitch/.
2. (Length of Plate Appearances)

(a) Calculate the length, in term of pitches, of the average plate appear-
ance by batting position using Retrosheet data for the 2011 season.
(b) Does the eighth batter in the National League behave differently than
his counterpart in the American League?
(c) Repeat the calculations in (a) and (b) for the 1991 and 2011 seasons
and comment on any differences between the seasons that you find.
3. (Pickoff Attempts)
Identify the baserunners who, in the 2011 season, drew the highest number
of pickoff attempts when standing at first base with second base unoccu-
pied.
4. (Umpire’s Strike Zone)
By drawing a contour plot, compare the umpire’s strike zone for left-
handed and right-handed batters. Use only the rows of the data frame
where the pitch type is a four-seam fastball.
5. (Umpire’s Strike Zone, Continued)
By drawing one or more contour plots, compare the umpire’s strike zone
by pitch type. For example, compare the 50/50 contour lines of four-seam
fastballs and curveballs when a right-handed batter is at the plate.
8
Career Trajectories
CONTENTS
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
8.2 Mickey Mantle’s Batting Trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
8.3 Comparing Trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
8.3.1 Some preliminary work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
8.3.2 Computing career statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
8.3.3 Computing similarity scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.3.4 Defining age, OBP, SLG, and OPS variables . . . . . . . . . . . 197
8.3.5 Fitting and plotting trajectories . . . . . . . . . . . . . . . . . . . . . . . . . 198
8.4 General Patterns of Peak Ages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.4.1 Computing all fitted trajectories . . . . . . . . . . . . . . . . . . . . . . . . 202
8.4.2 Patterns of peak age over time . . . . . . . . . . . . . . . . . . . . . . . . . . 203
8.4.3 Peak age and career at-bats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.5 Trajectories and Fielding Position . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
8.6 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.1 Introduction
The R system is well-suited for fitting statistical models to data. One popular
topic in sabermetrics is the rise and fall of a player’s season batting, fielding,
or pitching statistics as from his MLB debut to retirement. Generally, it is
believed that most players peak in their late 20s, although some players tend
to peak at later ages. A simple way of modeling a player’s trajectory is by
means of a quadratic or parabolic curve. By use of the lm (linear model)
function in R, it is straightforward to fit this model using the player’s age and
his OPS statistics.
We begin in Section 8.2 by considering a famous career trajectory. Mickey
Mantle made an immediate impact on the New York Yankees at age 19 and
quickly matured into one of the best hitters in baseball. But injuries took a toll
on Mantle’s performance and his hitting declined until his retirement at age
36. We use Mantle to introduce the quadratic model – using this model, one
187
can define his peak age, the maximum performance, and the rate of increase
and decline in performance.
To compare career performances of similar players, it is helpful to contrast
their trajectories and Section 8.3 illustrates the computation of many fitted
trajectories. Using Bill James’ notion of similarity scores, we write a function
that will find players who are most similar to a given hitter. Then we graphi-
cally compare the OPS trajectories of these similar players; by viewing these
graphs we gain a general understanding of the possible trajectory shapes.
A general problem focuses on a player’s peak age. In Section 8.4, we look
at the fitted trajectories of all hitters with at least 2000 career at-bats. The
pattern of peak ages across eras and as a function of the number of career
at-bats is explored. Also, since it is common to compare players who play the
same position, in Section 8.5 we focus on the period 1985-1995 and contrast
the peak ages for players who play different fielding positions.
8.2 Mickey Mantle’s Batting Trajectory

To start looking at career trajectories, we consider batting data from the great
slugger Mickey Mantle. To obtain his season-by-season hitting statistics, the
Lahman data files Master.csv and Batting.csv are read into R and saved
in the data frames Master and Batting.
Batting <- read.csv("Batting.csv")
Master <- read.csv("Master.csv")
We extract Mantle’s playerID from the Master data frame. By use of the
subset function, the line in the Master data file is found where nameFirst ==
"Mickey and nameLast == "Mantle". His player id is stored in the variable
mantle.id.
mantle.info <- subset(Master,
nameFirst == "Mickey" & nameLast == "Mantle")
mantle.id <- as.character(mantle.info$playerID)
One small complication is that certain statistics such as SF and HBP were
not recorded for older seasons and are currently coded as NA. A convenient
way of recoding these missing values to 0 is by the recode function in the car
package.
library(car)
Batting$SF <- recode(Batting$SF, "NA = 0")
Batting$HBP <- recode(Batting$HBP, "NA = 0")
To compute Mantle’s age for each season, we need to know his birth year
which is available in the Master data frame. Major League Baseball defines a
Career Trajectories 189
player’s age as his age on June 30 of that particular season. To facilitate the
computation of ages, a new function get.birthyear is defined which gives
the “official” birth year of a player with id player.id, similar to what was
done in the getinfo function of Section 3.8.
get.birthyear <- function(player.id){
playerline <- subset(Master, playerID == player.id)
birthyear <- playerline$birthYear
birthmonth <- playerline$birthMonth
ifelse(birthmonth >= 7, birthyear + 1, birthyear)
}
To check this function, the MLB birth year for Mantle is found using his id
stored in mantle.id.
get.birthyear(mantle.id)
[1] 1932
Mantle’s batting statistics are obtained by means of the user-defined func-

tion get.stats. The inputs are the player id and the output is a data frame
containing the player’s hitting statistics. This function computes the player’s
age (variable Age) for all seasons, and also computes the player slugging per-
centage (SLG), on-base percentage (OBP), and OPS for all seasons. Note that
the function get.birthyear is used to find the player’s MLB birth year.
get.stats <- function(player.id){
d <- subset(Batting, playerID==player.id)
byear <- get.birthyear(player.id)
d$Age <- d$yearID - byear
d$SLG <- with(d, (H - X2B - X3B - HR +
2 * X2B + 3 * X3B + 4 * HR) / AB)
d$OBP <- with(d, (H + BB) / (H + AB + BB + SF))
d$OPS <- with(d, SLG + OBP)
d
}
After reading the function get.stats into R, we obtain Mantle’s statistics

by applying this function with input mantle.id – the resulting data frame of
hitting statistics is stored in Mantle.
Mantle <- get.stats(mantle.id)
A good measure of batting performance is OPS, the sum of a player’s

slugging percentage and his on-base percentage. How does Mantle’s OPS sea-
son values vary as a function of his age? To address this question, the plot
function is used to construct a scatterplot of OPS against age. (See Figure
8.1.)
with(Mantle, plot(Age, OPS, cex=1.5, pch=19))
1.05
1.00
0.95
0.90
OPS
0.85
0.80
0.75
20 25 30 35
Age
FIGURE 8.1
Scatterplot of OPS against age for Mickey Mantle.
Looking at this figure, it is clear that Mantle’s OPS values tend to increase
from age 19 to his late 20s, and then generally decrease until his retirement
at age 36. One can model this up-and-down relationship by use of a smooth
curve. This curve will help us understand and summarize Mantle’s career
batting trajectory and will make it easier to compare Mantle’s trajectory
with other players with similar batting performances.
A convenient choice of smooth curve is a quadratic function of the form
A + B(Age − 30) + C(Age − 30)2 ,
where the constants A, B, and C are chosen so that curve is the “best” match
to the points in the scatterplot. This quadratic curve has the following nice
properties that make it easy to use.
1. The constant A is the predicted value of OPS when the player is 30 years
old.
2. The function reaches its largest value at
B
P EAK.AGE = 30 − .
2C
This is the age where the player is estimated to have his peak batting
performance during his career.
3. The maximum value of the curve is

B2
M AX = A − .
4C
This is the estimated largest OPS of the player over his career.
4. The coefficient C, typically a negative value, tells us about the degree of

curvature in the quadratic function. If a player has a “large” value of C,
this indicates that he more rapidly reaches his peak level and more rapidly
decreases in ability until retirement. One simple interpretation is that C
represents the change in OPS from his peak age to one year later.
A new function fit.model is written to fit this quadratic curve to a player’s

batting data. The input to this function is a data frame d containing the
player’s batting statistics including the variables Age and OPS. The function
lm is used to fit the quadratic curve – the formula
2
OPS ∼ I(Age - 30) + I((Age - 30) )
indicates that OPS is the response and (Age - 30) and (Age - 30)2 are the
predictors. The estimated coefficients A, B, and C are saved in the vector b.
The peak age and maximum value are stored in the variables Age.max and
Max.
fit.model <- function(d){
fit <- lm(OPS ~ I(Age - 30) + I((Age - 30)^2), data=d)
b <- coef(fit)
Age.max <- 30 - b[2] / b[3] / 2
Max <- b[1] - b[2] ^ 2 / b[3] / 4
list(fit=fit,
Age.max=Age.max, Max=Max)
}
The function fit.model is applied to Mantle’s data frame – the output of

this function includes the object fit that stores all of the calculations of the
quadratic fit. In addition, this function outputs the peak age and maximum
value displayed in the following code.
F2 <- fit.model(Mantle)
coef(F2$fit)
(Intercept) I(Age - 30) I((Age - 30)^2)
0.955433417 -0.020289562 -0.003520738
c(F2$Age.max, F2$Max)
I(Age - 30) (Intercept)
27.118564 0.984665
The best fitting curve is given by
0.955 − 0.0202(Age − 30) − 0.00352(Age − 30)2 ,
Using this model, Mantle peaked at age 27 and his maximum OPS for the curve
is estimated to be 0.985. The estimated value of the curvature parameter is
−0.00352, thus Mantle’s decrease in OPS between his peak age and one year
older is 0.00352.
This best quadratic curve is placed on the scatterplot. The predict func-
tion is used to estimate Mantle’s OPS from the curve for the sequence of age
values and the lines function overlays these values as a line on the current
plot. Two applications of abline show the locations of the peak age and the
maximum, and the text function is used to label these values. The resulting
graph is displayed in Figure 8.2.
lines(Mantle$Age, predict(F2$fit, Age=Mantle$Age), lwd=3)
abline(v=F2$Age.max, lwd=3, lty=2, col="grey")
abline(h=F2$Max, lwd=3, lty=2, col="grey")
text(29, .72, "Peak.age" , cex=2)
text(20, 1, "Max", cex=2)
Although the focus was on the best fitting quadratic curve, more details
about the fitting procedure are stored in the output of lm that is stored in the
variable fit. We display part of the output display by finding the summary of
the fit.
summary(F2$fit)
...
Residual standard error: 0.07501 on 15 degrees of freedom
Multiple R-squared: 0.6093, Adjusted R-squared: 0.5572
F-statistic: 11.69 on 2 and 15 DF, p-value: 0.0008692
The value of R2 is 0.6093 – this means that approximately 61% of the vari-
ability in Mantle’s OPS values can be explained by the quadratic curve. The
residual standard error is equal to 0.075. Approximately 2/3 of the vertical
deviations (the “residuals”) from the curve fall between plus and minus one
residual standard error. In this case, the interpretation is that approximately
2/3 of the residuals fall between −0.075 and 0.075.
8.3 Comparing Trajectories

8.3.1 Some preliminary work
When we think about hitting trajectories of players, one relevant variable
seems to be a player’s fielding position. Hitting expectations of a catcher, an
1.05
1.00
Max
0.95
0.90
OPS
0.85
0.80
0.75
Peak.age
20 25 30 35
Age
FIGURE 8.2
Scatterplot of OPS against age for Mickey Mantle with a quadratic fit added.
The location of the peak age and the maximum OPS fit are displayed.
important defensive position, may be different from the hitting expectations

of a first baseman. To compare trajectories of players with the same position,
fielding position should be recorded in our database.
The data frame Batting has already been created. We read in the fielding
data file Fielding.csv and store it in the data frame Fielding.
Fielding <- read.csv("Fielding.csv")
Many players in the history of baseball have had short careers and in our
study of trajectories, it seems reasonable to limit our analysis to players who
have had a minimum of at-bats. We consider only players with 2000 at-bats –
this will remove hitting data of pitchers and other players with short careers.
To take this subset of the Batting data frame, we use the ddply function
in the plyr package to compute the career at-bats for all players – the new
variable is called Career.AB. By use of the merge function, we add this new
variable to the Batting data frame. Finally, using the subset function, a new
data frame Batting.2000 is created consisting only of the “minimum 2000
AB” hitters.
library(plyr)
AB.totals <- ddply(Batting, .(playerID),
summarize,
Career.AB=sum(AB, na.rm=TRUE))
Batting <- merge(Batting, AB.totals)
Batting.2000 <- subset(Batting, Career.AB >= 2000)
To add fielding information to our data frame, a new function

find.position is written that finds the primary fielding position for a given
player. In the function, the subset function is used to extract the lines of the
Fielding data frame for a player with player id equal to p. The number of
games played at each possible position are tallied, and the function returns
the position where the most games are played.1
find.position <- function(p){
positions <- c("OF", "1B", "2B", "SS", "3B", "C", "P", "DH")
d <- subset(Fielding, playerID == p)
count.games <- function(po)
sum(subset(d, POS == po)$G)
FLD <- sapply(positions, count.games)
positions[FLD == max(FLD)][1]
}
In the following code, PLAYER is a character vector of player ids for all players
with at least 2000 career at-bats. The sapply function is used to find the pri-
mary fielding position for all players and the new data frame Fielding.2000
is created using the function data.frame containing the player ids and the
fielding positions (variable POS). This new information is placed into the
Batting.2000 data frame by use of the merge function.
PLAYER <- as.character(unique(Batting.2000$playerID))
POSITIONS <- sapply(PLAYER, find.position)
Fielding.2000 <- data.frame(playerID=names(POSITIONS),
POS=POSITIONS)
Batting.2000 <- merge(Batting.2000, Fielding.2000)
8.3.2 Computing career statistics

Groups of similar hitters will be found on the basis of their career statistics.
Towards this goal, one needs to compute the career games played, at-bats,
runs, hit, etc., for each player in the Batting.2000 data frame. This is conve-
niently done using the ddply function. In the R code, the summarize argument
option is used to find the sum of different batting statistics for each hitter.
A new data frame C.totals is created with the player id variable playerID
and new career variables C.G, C.AB, C.R, and so on.
1 In the rare case where there are two or more positions with the most games played, the
function will take the first position.

library(plyr)
C.totals <- ddply(Batting.2000, .(playerID),
summarize,
C.G=sum(G, na.rm=TRUE),
C.AB=sum(AB, na.rm=TRUE),
C.R=sum(R, na.rm=TRUE),
C.H=sum(H, na.rm=TRUE),
C.2B=sum(X2B, na.rm=TRUE),
C.3B=sum(X3B, na.rm=TRUE),
C.HR=sum(HR, na.rm=TRUE),
C.RBI=sum(RBI, na.rm=TRUE),
C.BB=sum(BB, na.rm=TRUE),
C.SO=sum(SO, na.rm=TRUE),
C.SB=sum(SB, na.rm=TRUE))
In the new data frame, we compute each player’s career batting average
C.AVG and his career slugging percentage C.SLG.
C.totals$C.AVG <- with(C.totals, C.H / C.AB)
C.totals$C.SLG <- with(C.totals,
(C.H - C.2B - C.3B - C.HR + 2 * C.2B +
3 * C.3B + 4 * C.HR) / C.AB)
The career statistics data frame C.totals is merged with the fielding data
frame Fielding.2000. Each fielding position has an associated value, and a se-
ries of ifelse functions are used to define a value position variable Value.POS
from the position variable POS.
C.totals <- merge(C.totals, Fielding.2000)
C.totals$Value.POS <- with(C.totals,
ifelse(POS == "C", 240,
ifelse(POS == "SS", 168,
ifelse(POS == "2B", 132,
ifelse(POS == "3B", 84,
ifelse(POS == "OF", 48,
ifelse(POS == "1B", 12, 0)))))))
8.3.3 Computing similarity scores

Bill James introduced the concept of similarity scores to facilitate the com-
parison of players on the basis of career statistics. To compare two hitters, one
starts at 1000 points and subtracts points based on the differences in different
statistical categories. One point is subtracted for each of the following differ-
ences: (1) 20 games played, (2) 75 at-bats, (3) 10 runs scored, (4) 15 hits, (5)
5 doubles, (6) 4 triples, (7) 2 home runs, (8) 10 runs batted in, (9) 25 walks,
(10) 150 strikeouts, (11) 20 stolen bases, (12) 0.001 in batting average, and
(13) 0.002 in slugging percentage. In addition, one adds the difference between
the fielding position values of the two players.
The function similar will find the players most similar to a given player
using similarity scores on career statistics and fielding position. One inputs
the id for the particular player and the number of similar players to be found
(including the given player). The output is a data frame of player statistics,
ordered in decreasing order by similarity scores.
similar <- function(p, number=10){
P <- subset(C.totals, playerID == p)
C.totals$SS <- with(C.totals,
1000 -
floor(abs(C.G - P$C.G) / 20) -
floor(abs(C.AB - P$C.AB) / 75) -
floor(abs(C.R - P$C.R) / 10) -
floor(abs(C.H - P$C.H) / 15) -
floor(abs(C.2B - P$C.2B) / 5) -
floor(abs(C.3B - P$C.3B) / 4) -
floor(abs(C.HR - P$C.HR) / 2) -
floor(abs(C.RBI - P$C.RBI) / 10) -
floor(abs(C.BB - P$C.BB) / 25) -
floor(abs(C.SO - P$C.SO) / 150) -
floor(abs(C.SB - P$C.SB) / 20) -
floor(abs(C.AVG - P$C.AVG) / 0.001) -
floor(abs(C.SLG - P$C.SLG) / 0.002) -
abs(Value.POS - P$Value.POS))
C.totals <- C.totals[order(C.totals$SS, decreasing=TRUE), ]
C.totals[1:number, ]
}
To illustrate the use of this function, suppose one is interested in finding

the five players who are most similar to Mickey Mantle. Recall that the player
id for Mantle is stored in the variable mantle.id. We use the function similar
with inputs mantle.id and 6.
similar(mantle.id, 6)
playerID C.G C.AB C.R C.H C.2B C.3B C.HR C.RBI C.BB C.SO C.SB
1282 mantlmi01 2401 8102 1677 2415 344 72 536 1509 1733 1710 153
1308 matheed01 2391 8537 1509 2315 354 72 512 1453 1444 1487 68
1805 schmimi01 2404 8352 1506 2234 408 59 548 1595 1507 1883 174
1844 sheffga01 2576 9217 1636 2689 467 27 509 1676 1475 1171 253
2013 thomafr04 2322 8199 1494 2468 495 12 521 1704 1667 1397 32
1900 sosasa01 2354 8813 1475 2408 379 45 609 1667 929 2306 234
C.AVG C.SLG POS Value.POS SS
1282 0.2980745 0.5567761 OF 48 1000
1308 0.2711725 0.5094295 3B 84 853
1805 0.2674808 0.5272989 3B 84 848
1844 0.2917435 0.5139416 OF 48 847
2013 0.3010123 0.5549457 DH 0 844
1900 0.2732327 0.5337569 OF 48 831
From reading the player ids, we see five similar players, in terms of career
hitting statistics and position: Eddie Mathews, Mike Schmidt, Gary Sheffield,
Frank Thomas, and Sammy Sosa.
8.3.4 Defining age, OBP, SLG, and OPS variables

To fit and graph hitting trajectories for a group of similar hitters, one needs to
have age and OPS statistics for all seasons for each player. One complication
with working with the Batting Lahman data base is that separate batting
lines are used for batters who played with multiple teams during a season.
There is a variable stint that gives different values (1, 2, ...) in the case of a
player with multiple teams. A new function collapse.stint is written which
collapses the counts over the stint variable. This function also computes
the batting measures SLG, OBP, and OPS. (Recall we had earlier replaced any
missing values for HBP and SF with zeros, so there will be no missing values
in the calculation of the OBP and OPS variables.)
collapse.stint <- function(d){
G <- sum(d$G); AB <- sum(d$AB); R <- sum(d$R)
H <- sum(d$H); X2B <- sum(d$X2B); X3B <- sum(d$X3B)
HR <- sum(d$HR); RBI <- sum(d$RBI); SB <- sum(d$SB)
CS <- sum(d$CS); BB <- sum(d$BB); SH <- sum(d$SH)
SF <- sum(d$SF); HBP <- sum(d$HBP)
SLG <- (H - X2B - X3B - HR + 2 * X2B +
3 * X3B + 4 * HR) / AB
OBP <- (H + BB + HBP) / (AB + BB + HBP + SF)
OPS <- SLG + OBP
data.frame(G=G, AB=AB, R=R, H=H, X2B=X2B,
X3B=X3B, HR=HR, RBI=RBI, SB=SB,
CS=CS, BB=BB, HBP=HBP, SH=SH, SF=SF,
SLG=SLG, OBP=OBP, OPS=OPS,
Career.AB=d$Career.AB[1], POS=d$POS[1])
}
By use of the ddply function together with the collapse.stint function, we
create a new version of the Batting.2000 function where the hitting statistics
for a player for a season are recorded on a single line.
Batting.2000 <- ddply(Batting.2000,
.(playerID, yearID), collapse.stint)
The next task is to obtain the ages for all players for all seasons. The
vector player.list is a character vector of the player ids for all players in the
Batting.2000 data frame. Recall that a function get.birthyear was written
in Section 8.2 to compute the MLB birth year for a particular player. By using
the sapply function together with get.birthyear, we find the birth years for
all players – this vector of birth years is stored in the variable birthyears.
By use of the merge function, this birth year information is merged with the
batting data. Now that we have birth years for all players, we can define the
new variable Age as the difference between the season year and the birth year.
player.list <- as.character(unique(Batting.2000$playerID))

birthyears <- sapply(player.list, get.birthyear)
Batting.2000 <- merge(Batting.2000,
data.frame(playerID=player.list,
Birthyear=birthyears))
Batting.2000$Age <- with(Batting.2000, yearID - Birthyear)
A small complication is that the birth year is not recorded for a few 19th
century ballplayers, and so the age variable is missing for these variables.
The complete.cases function is used to record the age records that are not
missing, and the updated data frame Batting.2000 only contains players for
which the Age variable is available.
Batting.2000 <- Batting.2000[complete.cases(Batting.2000$Age), ]
8.3.5 Fitting and plotting trajectories

Given a group of similar players, we want to fit quadratic curves to each player
and graph the trajectories in a way that facilitates comparisons. We begin by
writing a function fit.trajectory that fits the curve for a single player with
season batting statistics in data frame d. The output is a data frame with two
variables – Age contains the player ages and Fit contains the predicted OPS
using the quadratic model.
fit.trajectory <- function(d){
fit <- lm(OPS ~ I(Age - 30) + I((Age - 30)^2), data = d)
data.frame(Age = d$Age, Fit = predict(fit, Age = d$Age))
}
The function plot.trajectories graphs the trajectories for a given player

together with a group of similar players. One inputs the first and last name
of the player, the number of players to compare (including the one of inter-
est), and the number of columns in the multipanel plot. This function first
uses the Master data frame to find the player id for the player. Using the
similar function, a vector of player ids player.list is found. The data frame
Batting.new consists of the season batting statistics for only the players in
the player list. The ddply computes the trajectories for all players (using the
fit.trajectory function) and the ages and fitted OPS values for all players
is stored in the data frame F. We use the merge function to add the players’
full names to the data frame F. The graphing is done by use of the ggplot2
package. The use of geom line constructs trajectory curves of Age and Fit
for all players. The facet wrap function with the ncol argument places these
trajectories on separate panels where the number of columns in the multipanel
display is the value specified in the argument of the function.
plot.trajectories <- function(first, last, n.similar=5, ncol){
require(plyr)
require(ggplot2)
get.name <- function(playerid){
d1 <- subset(Master, playerID == playerid)
with(d1, paste(nameFirst, nameLast))
}
player.id <- subset(Master,
nameFirst == first & nameLast == last)$playerID
player.id <- as.character(player.id)
player.list <- as.character(similar(player.id, n.similar)$playerID)
Batting.new <- subset(Batting.2000, playerID %in% player.list)
F2 <- ddply(Batting.new, .(playerID), fit.traj)

F2 <- merge(F2,
data.frame(playerID=player.list,
Name=sapply(as.character(player.list), get.name)))
print(ggplot(F2, aes(Age, Fit)) + geom_line(size=1.5) +

facet_wrap(~ Name, ncol=ncol) + theme_bw())
return(Batting.new)
}
Here are several examples of the use of plot.trajectories. Mickey Man-

tle’s trajectory is compared with the trajectories of five similar hitters.
d <- plot.trajectories("Mickey", "Mantle", 6, 2)
We compare Derek Jeter’s OPS trajectory with eight similar players.

d <- plot.trajectories("Derek", "Jeter", 9, 3)
Looking at Figure 8.3 and Figure 8.4, we see notable differences in these
trajectories.
• There are players such as Eddie Mathews, Frank Thomas, Mickey Man-
tle, and Roberto Alomar who appeared to peak early in their careers.
• In contrast, other players such as Mike Schmidt, Craig Biggio, and Julio
Franco who peaked in their 30s.
• The players also show differences in the shape of the trajectory. Johnny
Damon and Julio Franco had relatively constant trajectories, and Frankie
Frisch and Roberto Alomar had trajectories with high curvature.
One can summarize these trajectories by the peak age, the maximum value,
and the curvature. A short function summarize.trajectory is written to
compute these quantities for a particular fitted trajectory. The input is the
data frame d containing the batting statistics for a player and the output is a
data frame with three variables Age.max, Max, and Curve.
Eddie Mathews Frank Thomas

1.1
1.0
0.9
0.8
0.7
0.6
Gary Sheffield Mickey Mantle

1.1
1.0
0.9
Fit
0.8
0.7
0.6
Mike Schmidt Sammy Sosa

1.1
1.0
0.9
0.8
0.7
0.6
20 25 30 35 40 20 25 30 35 40
Age
FIGURE 8.3
Fitted trajectories of OPS against age for Mickey Mantle and five similar
players.
summarize.trajectory <- function(d){

f <- lm(OPS ~ I(Age - 30) + I((Age - 30) ^ 2), data=d)
b <- coef(f)
Age.max <- round(30 - b[2] / b[3] / 2, 1)
Max <- round(b[1] - b[2] ^ 2 / b[3] / 4, 3)
data.frame(Age.max=Age.max, Max=Max,
Curve=round(b[3], 5))
}
Recall that the output of plot.trajectories was a data frame containing the
season batting statistics for a group of players. One can use the ddply function
together with summarize.trajectory to find the summary statistics for all
players. This is illustrated for Jeter and eight similar players.
d <- plot.trajectories("Derek", "Jeter", 9, 3)
S <- ddply(d, .(playerID), summarize.trajectory)
S
playerID Age.max Max Curve
1 alomaro01 28.3 0.885 -0.00309
Barry Larkin Charlie Gehringer Craig Biggio

0.95
0.85
0.75
0.65
0.55
Derek Jeter Frankie Frisch Johnny Damon
0.95
0.85
Fit
0.75
0.65
0.55
Julio Franco Roberto Alomar Robin Yount
0.95
0.85
0.75
0.65
0.55
20 30 40 50 20 30 40 50 20 30 40 50
Age
FIGURE 8.4
Fitted trajectories of OPS against age for Derek Jeter and eight similar play-
ers.
2 biggicr01 31.8 0.862 -0.00229

3 damonjo01 30.4 0.812 -0.00100
4 francju01 33.9 0.765 -0.00048
5 friscfr01 28.2 0.866 -0.00315
6 gehrich01 30.9 0.934 -0.00285
7 jeterde01 28.7 0.877 -0.00223
8 larkiba01 31.2 0.877 -0.00227
9 yountro01 28.7 0.844 -0.00228
To help understand the differences between the nine player trajectories,

the plot function is used to construct a scatterplot of the peak ages and the
curvature statistics. The text function is used to add player labels. (The last
names of the players are found from the Master data frame.)
with(S, plot(Age.max, Curve, pch=19, cex=1.5,
xlab="Peak Age", ylab="Curvature",
xlim=c(27, 36), ylim=c(-0.0035, 0)))
S$lastNames <- as.character(subset(Master,
playerID %in% S$playerID)$nameLast)

with(S, text(Age.max, Curve, lastNames, pos=3))
0.0000
−0.0035 −0.0030 −0.0025 −0.0020 −0.0015 −0.0010 −0.0005
Franco
Damon
Curvature
Jeter Larkin
Biggio
Yount
Gehringer
Alomar
Frisch
28 30 32 34 36
Peak Age
FIGURE 8.5
Scatterplot of peak age and curvature measures for Derek Jeter and eight
similar players. The points are labeled by the player last names.
Looking at Figure 8.5, this clearly indicates that Alomar peaked at an early
age, Franco at a late age, and Alomar and Frisch exhibited the greatest cur-
vature, indicating they rapidly declined in performance after the peak.
8.4 General Patterns of Peak Ages

8.4.1 Computing all fitted trajectories
We have explored the hitting career trajectories of groups of similar players.
How have career trajectories changed over the history of baseball? We’ll focus
on a player’s peak age and explore how this has changed over time. Also the
relationship between peak age and the number of career at-bats is explored.
For each player, the variable playerID contains the seasons played. We
define a new variable Midyear defined to be the average of a player’s first and
last seasons. The function ddply is used to compute Midyear for all players
and this new variable is added to the Batting.2000 data frame using the
merge function.
library(plyr)
midcareers <- ddply(Batting.2000, .(playerID),
summarize,
Midyear=(min(yearID) + max(yearID))/2)
Batting.2000 <- merge(Batting.2000, midcareers)
Quadratic curves to all of the career trajectories are fit by another appli-
cation of the ddply function. A short function coefficients.trajectory is
defined that fits the quadratic model for the season data for a particular player
and returns the coefficients A, B, C, Midyear, and the player’s career at-bats
Career.AB. We apply ddply where playerID is the grouping variable and
coefficients.trajectory is the function to be applied on each subset. The
output Beta.coef is a data frame containing the coefficients for all players,
where a row corresponds to a particular player.
coefficients.trajectory <- function(d){
b <- coef(lm(OPS ~ I(Age - 30) + I((Age - 30) ^ 2), data=d))
data.frame(A=b[1], B=b[2], C=b[3],
Midyear=d$Midyear[1], Career.AB=d$Career.AB[1])
}
Beta.coef <- ddply(Batting.2000, .(playerID), coefficients.trajectory)
The estimated peak ages are computed for all players using the formula
P eak.age = 30 − B/(2C). The new variable Peak.age is added to the data
frame Beta.coef.
Beta.coef$Peak.age <- with(Beta.coef, 30 - B / 2 / C)
8.4.2 Patterns of peak age over time

To investigate how the peak age varies over the history of baseball, we con-
struct a scatterplot of Peak.age against Midyear by use of the plot function.
It is difficult to see the general pattern by just looking at the scatterplot. So
we use the lowess function to fit a smoothing curve and this curve is added
to the current graph by the lines function. (See Figure 8.6.) (Sometimes a
player’s peak age estimate is infinite due to a division by zero operation, and
the is.finite function removes those players before the application of the
lowess function.)
with(Beta.coef,
plot(Midyear, Peak.age, ylim=c(20, 40),
xlab="Mid Career", ylab="Peak Age"))
i <- is.finite(Beta.coef$Peak.age)
with(Beta.coef,
lines(lowess(Midyear[i], Peak.age[i]), lwd=3))
40
35
Peak Age
30
25
20
1880 1900 1920 1940 1960 1980 2000

Mid Career
FIGURE 8.6
Scatterplot of peak age against mid career for all players with at least 2000
career at-bats.
Looking at this figure, we see a gradual increase in peak age over time. The
peak age for an average player was approximately 27 in 1880 and this average
has gradually increased to 28 from 1880 to 2000.
8.4.3 Peak age and career at-bats

Is there any relationship between a player’s peak age and his career at-bats?
Using the plot function, we construct a graph of Peak.age against the loga-
rithm (base 2) of the career at-bats variable Career.AB. We plot the at-bats
on a log scale, so that the points are more evenly spread out over all possible
values. Again we overlay a loess smoothing curve to see the pattern and Figure
8.7 shows the graph.
with(Beta.coef,
plot(log2(Career.AB[i]), Peak.age[i], ylim=c(20, 40),
xlab="log2 Career AB", ylab="Peak Age"))

with(Beta.coef,
lines(lowess(log2(Career.AB[i]), Peak.age[i]), lwd=3))
40
35
Peak Age
30
25
20
11.0 11.5 12.0 12.5 13.0 13.5

log2 Career AB
FIGURE 8.7
Scatterplot of peak age against log career AB for all players with at least 2000
career at-bats.
Here we see a clear relationship. Players with relatively short careers with
2000 career at-bats tend to peak about age 27. In contrast, players with long
careers, say 9000 or more at-bats, tend to peak at ages closer to 30.
8.5 Trajectories and Fielding Position

In comparing players, we typically want to compare players at the same field-
ing position. The primary fielding position POS was already defined and we
use this variable to compare peak ages of players categorized by position.
Suppose we consider the players whose mid-career is between 1985 and
1995. Using the subset function, a new data frame Batting.2000a is created
consisting of only these players.
Batting.2000a <- subset(Batting.2000, Midyear >= 1985 & Midyear <= 1995)
A short function trajectory.peak.pos is written that fits a player’s tra-

jectory based on his (Age, OPS) data. The output is the estimated coefficients
A, B, C, his estimated peak age Peak.Age, and his fielding position Position.
By use of the ddply function, the trajectories are fit for all players in the
Batting2000a data frame, and the trajectory fits for all players is stored in
the data frame Beta.estimates.
trajectory.peak.pos <- function(d){
b <- coef(lm(OPS ~ I(Age - 30) + I((Age - 30)^2), data=d))
data.frame(A=b[1], B=b[2], C=b[3],
Peak.Age=30 - b[2] / 2 / b[3],
Position=d$POS[1])
}
Beta.estimates <- ddply(Batting.2000a, .(playerID), trajectory.peak.pos)
We focus on the primary fielding positions excluding pitcher and desig-

nated hitter. The subset function removes these other positions and the
Beta.estimates1$Position[ , drop=TRUE] statement removes the levels
from the factor Position that do not occur. The trajectory and fielding in-
formation are combined with the Master info by use of the merge function –
the combined information is stored in the data frame Beta.estimates1.
Beta.estimates1 <- subset(Beta.estimates, Position %in%
c("1B", "2B", "3B", "SS", "C", "OF"))
Beta.estimates1$Position <- Beta.estimates1$Position[ , drop=TRUE]
Beta.estimates1 <- merge(Beta.estimates1, Master)
A stripchart is used to graph the peak ages of the players against the
fielding position. (See Figure 8.8.) Since some of the peak age estimates are
not reasonable values, the limits on the horizontal axis are set to 20 and 40.
stripchart(Peak.Age ~ Position, data=Beta.estimates1,
xlim=c(20, 40), method="jitter", pch=1)
special <- with(Beta.estimates1, identify(Peak.Age, Position,
n=5, labels=nameLast))
Generally, for all fielding positions, the peak ages for these 1990 players tend
to fall between 27 and 32. The variability in the peak age estimates reflects
the fact that hitters have different career trajectory shapes. There are three
outfielders and two shortstops who seem to stand out by having a high peak
age estimate. Using the identify function, the mouse is used to point out
these unusual values and the row numbers for these players are stored in
the variable special. The five players are Eric Davis, Gary DiSarcina, Jim
Eisenreich, Alvaro Espinoza, and Tony Phillips.
A new data frame dnew is formed containing the hitting statistics for these
Davis
OF Phillips
Eisenreich
C
DiSarcina
SS
Espinoza
3B
2B
1B
20 25 30 35 40
Peak.Age
FIGURE 8.8
Peak age estimates for players with mid-career 1985–1995 graphed against
fielding position.
five players with large peak age estimates, and this information is merged with
the Master data frame. The ggplot function is used to graph the data together
with the quadratic fits for these five players. (See Figure 8.9.) The geom point
function adds the points, the stat smooth function adds the quadratic curves,
and the facet wrap function creates separate panels for the five players.
dnew <- subset(Batting.2000, playerID %in% Beta1$playerID[special])
dnew <- merge(dnew, Master)
ggplot(dnew, aes(Age, OPS)) + geom_point(size=4) +
facet_wrap(~ nameLast, ncol=2) + ylim(0.4, 1.05) +
stat_smooth(method="lm", se=FALSE, size=1.5,
formula=y ~ poly(x, 2, raw=TRUE)) + theme_bw()
Looking at Figure 8.9, Davis had an unusual trajectory where he appeared to

have a hitting slump in the middle of his career. DiSarcina’s fitted increasing
trajectory is likely caused by the one large OPS value towards the end of his
career. The general impression is that Eisenreich, Espinoza, and Phillips had
pretty consistent level OPS values through their careers.
Davis DiSarcina
1.0 ● ● ●
● ● ●
● ●
● ●
0.8 ●● ●
●
● ● ●
● ● ● ●
0.6 ● ●● ●
● ●
0.4
Eisenreich Espinoza
1.0
●
0.8 ●● ● ●●●
OPS
● ●● ● ●
● ●
0.6
● ●
● ●●
● ● ●
● ●
0.4 ●
Phillips
1.0
●●
●
0.8 ● ● ●● ●
● ●
● ●● ●●
● ●
0.6 ●
0.4
25 30 35
Age
FIGURE 8.9
OPS trajectory graphs for five players with unusually large peak age estimates.
8.6 Further Reading

James (1982) wrote an essay on “Looking For the Prime”; based on a statis-
tical study, he came to the conclusion that batters tend to peak at age 27.
Berry et al. (1999) give a general discussion of career trajectories of athletes
from hockey, baseball, and golf. Chapter 11 of Albert and Bennett (2003)
considers the career trajectories of the home run rates of nine great histor-
ical sluggers. Albert (2002, 2009) discuss general patterns of trajectories of
hitters and pitchers in baseball history, and Fair (2008) performs an extensive
analysis of baseball career trajectories based on quadratic models. Albert and
Rizzo (2012), Chapter 7, give illustrations of regression modeling using R.
8.7 Exercises
1. (Career Trajectory of Willie Mays)
(a) Use the gets.stats function to extract the hitting data for Willie
Mays for all of his seasons in his career.
(b) Construct a scatterplot of Mays’ OPS season values against his age.
(c) Fit a quadratic function to Mays’ career trajectory. Based on this
model, estimate Mays’ peak age and his estimated largest OPS value
based on the fit.
2. (Comparing Trajectories)
(a) Using James’ similarity score measure (function similar), find the
five hitters with hitting statistics most similar to Willie Mays.
(b) Fit quadratic functions to the (Age, OPS) data for Mays and the five
similar hitters. Display the six fitted trajectories on a single panel.
(c) Based on your graph, describe the differences between the six player
trajectories. Which player had the smallest peak age?
3. (Comparing Trajectories of the Career Hits Leaders)
(a) Find the batters who have had at least 3200 career hits.
(b) Fit the quadratic functions to the (Age, AVG) data for this group of
hitters, where AVG is the batting average. Display the fitted trajec-
tories on a single panel.
(c) On the basis of your work, which player was the most consistent hitter
on average? Explain how you measured consistency on the basis of
the fitted trajectory.
4. (Comparing Trajectories of Home Run Hitters)
(a) Find the ten players in baseball history who have had the most career
home runs.
(b) Fit the quadratic functions to the home run rates of the ten players,
where HR.RAT E = HR/AB. Display the fitted trajectories on a
single panel.
(c) On the basis of your work, which player had the highest estimated
home run rate at his peak? Which player among the ten had the
smallest peak home run rate?
(d) Do any of the players have unusual career trajectory shapes? Is there
any possible explanation for these unusual shapes?
5. (Peak Ages in the History of Baseball)
(a) Find all the players who entered baseball between 1940 and 1945 with
at least 2000 career at-bats.
(b) Find all the players who entered baseball between 1970 and 1975 with
at least 2000 career at-bats.
(c) By fitting quadratic functions to the (Age, OPS) data, estimate the
peak ages for all players in parts (a) and (b).
(d) By comparing the peak ages of the 1940s players with the peak ages
of the 1970s players, can you make any conclusions about how the
peak ages have changed in this 30-year period?
9
Simulation
CONTENTS
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
9.2 Simulating a Half Inning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
9.2.1 Markov chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
9.2.2 Review of work in runs expectancy . . . . . . . . . . . . . . . . . . . . . 213
9.2.3 Computing the transition probabilities . . . . . . . . . . . . . . . . . . 215
9.2.4 Simulating the Markov chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
9.2.5 Beyond runs expectancy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
9.2.6 Transition probabilities for individual teams . . . . . . . . . . . . 220
9.3 Simulating a Baseball Season . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
9.3.1 The Bradley-Terry model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
9.3.2 Making up a schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
9.3.3 Simulating talents and computing win probabilities . . . . 225
9.3.4 Simulating the regular season . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
9.3.5 Simulating the post-season . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
9.3.6 Function to simulate one season . . . . . . . . . . . . . . . . . . . . . . . . . 227
9.3.7 Simulating many seasons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
9.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
9.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
9.1 Introduction
A baseball season consists of a collection of games between teams, where each
game consists of nine innings, and a half-inning consists of a sequence of plate
appearances. Because of this clean structure, the sport can be represented
by relatively simple probability models. Simulations from these models are
helpful in understanding different characteristics of the game.
One attractive aspect of the R system is its ability to simulate from a
wide variety of probability distributions. In this chapter, we illustrate the
use of R functions to simulate a game consisting of a large number of plate
appearances. Also, R is used to simulate the game-to-game team competition
of teams during an entire season.
Section 9.2 focuses on simulating the events in a baseball half-inning us-
211
ing a special probability model called a Markov chain. The runners on base
and the number of outs define a state and this probability model describes
movements between states until one reaches three outs. The movement or
transition probabilities are found using actual data from the 2011 season. By
simulating many half-innings using this model, one gets a basic understanding
of the pattern of run scoring.
Section 9.3 describes a simulation of an entire baseball season using the
Bradley-Terry probability model. Teams are assigned talents from a bell-
shaped (normal) distribution and a season of baseball games is played us-
ing win probabilities based on the talents. By simulating many seasons, one
learns about the relationship between a team’s talent and its performance in
a 162-game season. We describe simulating the post-season series and assess
the probability that the “best” team, that is, the team with the best ability
actually wins the World Series.
9.2 Simulating a Half Inning

9.2.1 Markov chains
A Markov chain is a special type of probability model useful for describing
movement between locations, called states. In the baseball context, a state
is viewed as a description of the runners on base and the number of outs in
an inning. Each of the three bases can be occupied by a runner or not, and
so there are 2 × 2 × 2 = 8 possible runner situations. Since there are three
possible numbers of outs (0, 1, or 2), there are 8 × 3 = 24 possible runner
and outs states. If we include the 3 outs state, there are a total of 25 possible
states during a half-inning of baseball.
In a Markov chain, a matrix of transition probabilities is used to describe
how one moves between the different states. For example, suppose that there
are currently runners on first and second with one out. Based on the outcome
of the plate appearance, the state can change. For example, the batter may
hit a single; the runner on second scores and the runner on first moves to
third. In this case, the new state is runners on first and third with one out.
Or maybe the batter will strike out, and the new state is runners on first and
second with two outs. By looking at a specific row in the transition probability
matrix, one learns about the probability of moving to first and third with one
out, or moving to first and second with two outs, or any other possible state.
In a Markov chain, there are two types of states – transition states and
absorbing states. Once one moves into an absorbing state, one remains there
and can’t return to other transition states. In a half-inning of baseball, since
the inning is over when there are 3 outs, this 3-outs state acts as an absorbing
state.
Simulation 213
There are some special assumptions in a Markov chain model. We assume

that the probability of moving to a new state only depends on the current
state. So any baseball events that happened before the current runners and
outs situation are not relevant in finding the probabilities. In other words, this
model assumes there is not a momentum effect in batting through an inning.
Also we are assuming that the probabilities of these movements are the same
for all teams, against all pitchers, and for all innings during a game. Clearly,
this assumption that all teams are average is not realistic, but we will address
this issue in one of the sections of this chapter.
There are several attractive aspects of using a Markov chain to model
a half-inning of baseball. First, the construction of the transition probabil-
ity matrix is easily done with 2011 season data using computations from the
runs expectancy chapter. One can use the model to play many half-innings of
baseball and the run scoring patterns that are found resemble the actual run
scoring of actual MLB baseball. Last, there are special properties of Markov
chains that simplify some interesting calculations, such as the number of play-
ers who come to bat during an inning.
9.2.2 Review of work in runs expectancy

To construct the transition matrix for the Markov chain, one needs to know
the frequencies of transitions from the different runners/outs states to other
possible runners/outs states. One obtains these frequencies using the Ret-
rosheet play-by-play data from a particular season, and we review the work
from Chapter 5.
We begin by reading in play-by-play data for the 2011 season, creating the
data frame data2011.
The variable HALF.INNING is created which creates a unique identification

for each half-inning in each baseball game. The new variable RUNS.SCORED
gives the number of runs scored in each play.
data2011$HALF.INNING <- with(data2011,
paste(GAME_ID, INN_CT, BAT_HOME_ID))
data2011$RUNS.SCORED <- with(data2011, (BAT_DEST_ID > 3) +
(RUN1_DEST_ID > 3) + (RUN2_DEST_ID > 3) + (RUN3_DEST_ID > 3))
The function get.state defines a state variable based on the runners

on the three bases (runner1, runner2, runner3) and the number of outs
(outs). Using this function, the new variable STATE is computed which gives
the runner locations and the number of outs at the beginning of each play.
The variable NEW.STATE contains the runners and outs information at the
conclusion of the play.
get.state <- function(runner1, runner2, runner3, outs){

runners <- paste(runner1, runner2, runner3, sep="")
paste(runners, outs)
}
RUNNER1 <- ifelse(as.character(data2011[,"BASE1_RUN_ID"])=="", 0, 1)

data2011$STATE <- get.state(RUNNER1, RUNNER2, RUNNER3,
data2011$OUTS_CT)
NRUNNER1 <- with(data2011, as.numeric(RUN1_DEST_ID==1 |

BAT_DEST_ID==1))
RUN2_DEST_ID==2 | BAT_DEST_ID==2))
RUN2_DEST_ID==3 | RUN3_DEST_ID==3 | BAT_DEST_ID==3))
NOUTS <- with(data2011, OUTS_CT + EVENT_OUTS_CT)
data2011$NEW.STATE <- get.state(NRUNNER1, NRUNNER2, NRUNNER3, NOUTS)
By using the subset function, we focus on plays where there is a change in

the state or in the number of runs scored. By another application of subset,
we restrict attention to complete innings where there are three outs – the new
data frame is named data2011C. Last, by use of the BAT EVENT FL variable, we
only consider plays where there is a batting event. So non-batting plays such
as steals, caught stealing, wild pitches, and passed balls are ignored. There
is obviously some consequence of removing these non-batting plays from the
viewpoint of run production, and this issue is discussed later in this chapter.
data2011 <- subset(data2011, (STATE != NEW.STATE) | (RUNS.SCORED > 0))
library(plyr)
data.outs <- ddply(data2011, .(HALF.INNING), summarize,
Outs.Inning=sum(EVENT_OUTS_CT))
data2011 <- merge(data2011, data.outs)
data2011C <- subset(data2011, Outs.Inning == 3)
data2011C <- subset(data2011, BAT_EVENT_FL == TRUE)
In our definition of the NEW.STATE variable, we recorded the runner loca-

tions when there were three outs. The runner locations don’t matter, so we
recode NEW.STATE to the value “3” when the number of outs is equal to 3.
library(car)
data2011C$NEW.STATE <- recode(data2011C$NEW.STATE,
"c(’000 3’, ’100 3’, ’010 3’, ’001 3’,
’110 3’, ’101 3’, ’011 3’, ’111 3’)=’3’")
Simulation 215
9.2.3 Computing the transition probabilities

Now that the STATE and NEW.STATE variables are defined, one can compute
the frequencies of all possible transitions between states by use of the table
function – the matrix of counts is T.matrix. There are 24 possible values of the
beginning state STATE, and 25 values of the final state NEW.STATE including
the 3-outs state.
T.matrix <- with(data2011C, table(STATE, NEW.STATE))
This matrix can be converted to a probability matrix by use of the

prop.table function; the resulting matrix is denoted by P.matrix.
P.matrix <- prop.table(T.matrix, 1)
We add a row to this transition probability matrix corresponding to transitions

from the 3-out state. When the inning reaches 3-outs, then it stays at 3-outs,
so the probability of staying in this state is 1.
P.matrix <- rbind(P.matrix, c(rep(0, 24), 1))
To better understand this transition matrix, the transition probabilities

starting at the “000 0” state, no runners and no outs are displayed. (Only
the positive probabilities are shown and the data.frame function is used to
display the probabilities vertically.) The most likely transitions are to the “no
runners, one out” state with probability 0.677 and to the “runner on first, no
outs” state with probability 0.240. The probability of moving from the “000
0” state to the “000 0” state is 0.027; in other words, the chance of a home
run with no runners on with no outs is 0.027.
P1 <- round(P.matrix["000 0", ], 3)
data.frame(Prob=P1[P1 > 0])
Prob
000 0 0.027
000 1 0.677
001 0 0.006
010 0 0.050
100 0 0.240
Let’s contrast this with the possible transitions starting from the “010 2”
state, runner on second with two outs. The most likely transitions are “3 outs”
(probability 0.640), “runners on first and second with two outs” (probability
0.163), and “runner on first with 2 outs” (probability 0.083).
P2 <- round(P.matrix["010 2", ], 3)
data.frame(Prob=P2[P2 > 0])
Prob
000 2 0.020
001 2 0.006
010 2 0.055
100 2 0.083
101 2 0.034
110 2 0.163
3 0.640
9.2.4 Simulating the Markov chain

One can simulate this Markov chain model a large number of times to obtain
the distribution of runs scored in a half-inning of 2011 baseball. The first
step is to construct a matrix giving the runs scored in all possible transitions
between states. Let Nrunners denote the number of runners in a state and O
denote the number of outs. For a batting play, the number of runs scored is
equal to
(b)
RU N S = (Nrunners + O(b) + 1) − (Nrunners
(a)
+ O(a) ).
In other words, the runs scored is the sum of runners and outs before (b)
the play minus the sum of runners and outs after (a) the play plus one. For
example, suppose there are runners on first and second with one outs, and
after the play, there is a runner on second with two outs. The number of runs
scored is equal to
RU N S = (2 + 1 + 1) − (1 + 2) = 1.
A new function count.runners.outs is defined which takes a state as in-
put and returns the sum of the number of runners and outs. This function is
applied across all the possible states (using the sapply function) and the cor-
responding sums are stored in the vector runners.outs. The outer function
with the “-” operation performs the RUNS calculation for all possible pairs
of states and the resulting matrix is stored in the variable R. If one inspects
the matrix R, one will notice some negative values and some strange large
positive values. But this is not a concern since the corresponding transitions,
for example a movement between a “000 0” state and a “000 2” state in one
batting play, are not possible. An additional column of zeros is added to this
run matrix by use of the cbind function.
count.runners.outs <- function(s)
sum(as.numeric(strsplit(s,"")[[1]]), na.rm=TRUE)
runners.outs <- sapply(dimnames(T.matrix)[[1]], count.runners.outs)[-25]
R <- outer(runners.outs + 1, runners.outs, FUN="-")
dimnames(R)[[1]] <- dimnames(T.matrix)[[1]][-25]
dimnames(R)[[2]] <- dimnames(T.matrix)[[1]][-25]
R <- cbind(R, rep(0, 24))
We are now ready to simulate a half-inning of baseball using a new function

simulate.half.inning. The inputs are the probability transition matrix P,
the run matrix R, and the starting state s (an integer between 1 and 24). The
output is the number of runs scored in the half-inning.
Simulation 217
simulate.half.inning <- function(P, R, start=1){

s <- start; path <- NULL; runs <- 0
while(s < 25){
s.new <- sample(1:25, 1, prob=P[s, ])
path <- c(path, s.new)
runs <- runs + R[s, s.new]
s <- s.new
}
runs
}
There are two key statements in this simulation. If the current state is s, the
function sample will simulate a new state using the s row in the transition
matrix P – the new state is denoted s.new. The total number of runs scored
in the inning is updated using the value in the s row and the s.new column
of the runs matrix R.
Using the replicate function, one can simulate a large number of half-
innings of baseball. In the below code, we simulate 10,000 half-innings starting
with no runners and no outs (state 1), collecting the runs scored in the vector
RUNS.
RUNS <- replicate(10000, simulate.half.inning(T.matrix, R))
To find the possible runs scored in a half-inning, the table function is used
to tabulate the values in RUNS.
table(RUNS)
RUNS
0 1 2 3 4 5 6 7 8 9 10
7483 1334 659 312 133 44 22 8 1 3 1
In our 10,000 simulations, five or more runs scored in 44 + 22 + 8 + 1 + 3 +

1 = 79 half-innings, so the chance of scoring five or more runs would be 79 /
10,000 = 0.0079. This calculation can be checked using the sum function.
sum(RUNS[RUNS >= 5]) / 10000
[1] 0.0079
The mean number of runs scored is computed by applying the mean function
on RUNS.
mean(RUNS)
[1] 0.4584
Over the 10,000 half-innings, an average of 0.4584 runs were scored.

To understand the runs potential of different runners and outs situations,
one can repeat this simulation procedure for other starting states. A function
RUNS.j is written to compute the mean number of runs scored starting with
state j. By use of the sapply function, we apply the function RUNS.j over all
of the possible starting states 1 through 24. The output is a vector of mean
runs scored stored in the variable Runs.Expectancy.
RUNS.j <- function(j){

mean(replicate(10000, simulate.half.inning(T.matrix, R, j)))
}
Runs.Expectancy <- sapply(1:24, RUNS.j)
Runs.Expectancy <- t(round(matrix(Runs.Expectancy, 3, 8), 2))
dimnames(Runs.Expectancy)[[2]] <- c("0 outs", "1 out", "2 outs")
dimnames(Runs.Expectancy)[[1]] <- c("000", "001", "010", "011", "100",
"101", "110", "111")
Runs.Expectancy
0 outs 1 out 2 outs
000 0.45 0.26 0.09
001 1.36 0.91 0.32
010 1.08 0.63 0.30
011 1.90 1.33 0.50
100 0.82 0.49 0.21
101 1.73 1.13 0.47
110 1.43 0.86 0.40
111 2.23 1.52 0.69
Recall that our simulation model is based only on batting plays. To under-
stand the effect of non-batting plays (stealing, caught stealing, wild pitches,
etc.) on run scoring, we compare this runs expectancy matrix with the one
found in Chapter 5 using all batting and non-batting plays. A new matrix
Runs is created with the earlier matrix and we compute the difference Runs
- Runs.Expectancy – this is the contribution of non-batting plays to the
average number of runs scored.
Runs <- matrix(
c(0.47, 0.25, 0.10, 1.45, 0.94, 0.32,
1.06, 0.65, 0.31, 1.93, 1.34, 0.54,
0.84, 0.50, 0.22, 1.75, 1.15, 0.49,
1.41, 0.87, 0.42, 2.17, 1.47, 0.76),
8, 3, byrow=TRUE)
Runs - Runs.Expectancy
0 outs 1 out 2 outs
000 0.02 -0.01 0.01
001 0.09 0.03 0.00
010 -0.02 0.02 0.01
011 0.03 0.01 0.04
100 0.02 0.01 0.01
101 0.02 0.02 0.02
110 -0.02 0.01 0.02
111 -0.06 -0.05 0.07
Note that most of the values of the difference are positive, indicating that
these non-batting plays generally do create runs. The three largest values are
0.09, 0.07, and 0.04 corresponding to the “001, 0 outs”, “111, 2 outs”, and
“011, 2 outs” situations. These positive values make sense since these are all
Simulation 219
situations with a runner on third who can score with a wild pitch or passed
ball.
9.2.5 Beyond runs expectancy

By using properties of Markov chains, it is straightforward to use the transi-
tion probability matrix to learn more about the movement through the run-
ners/outs states.
By multiplying the probability matrix P.matrix by itself three times, we
can learn about the state of the inning after three plate appearances. In R,
matrix multiplication is indicated by the %*% symbol. The result is stored in
the matrix P.matrix.3.
P.matrix.3 <- P.matrix %*% P.matrix %*% P.matrix
The first row of P.matrix.3 gives the probabilities of being in each of the 25
states after three hitters starting at the “000 0” state. We round these values
to three decimal places, sort from largest to smallest, and display the largest
values.
sorted.P <- sort(round(P.matrix.3["000 0", ], 3), decreasing=TRUE)
head(data.frame(Prob=sorted.P))
Prob
3 0.369
100 2 0.241
110 1 0.086
010 2 0.083
000 2 0.045
001 2 0.030
After three PAs, the most likely outcomes are three outs (probability 0.369),
runner on first with 2 outs (probability 0.241), and runners on first and second
with one out (probability 0.086).
It is also easy to learn about the number of visits to all runner-outs states.
Define the matrix Q to be the 24-by-24 submatrix found from the transition
matrix by removing the last row and column (the three outs state). By sub-
tracting the matrix Q from the identity matrix and taking the inverse of the
result, we obtain the fundamental matrix N of an absorbing Markov chain.
(The diag function is used to construct the identity matrix and the function
solve takes the matrix inverse.)
Q <- P.matrix[-25, -25]
N <- solve(diag(rep(1, 24)) - Q)
To understand the fundamental matrix, the beginning entries of the first

row of the matrix are displayed.
N.0000 <- round(N["000 0", ], 2)
head(data.frame(N=N.0000))
N
000 0 1.04
000 1 0.74
000 2 0.58
001 0 0.01
001 1 0.03
001 2 0.05
Starting at the beginning of the inning (the “000 0” state), the average number
of times the inning will be in the “000 0” state is 1.04, the average number
of times in the “000 1” state is 0.74, the average number of times in the “000
2” state is 0.58, and so on. By using the sum function, we find the average
number of states that are visited.
sum(N.0000)
[1] 4.28
In other words, the average number of plate appearances in a half-inning

(before three outs) is 4.28.
We can compute the average number of batting plays until three outs for
all starting states by multiplying the fundamental matrix N by a column vector
of ones. The vector of average number of plays is stored in the variable Length
and eight values of this vector are displayed.
Length <- round(t(N %*% rep(1, 24)), 2)
data.frame(L=Length[1, 1:8])
L
000 0 4.28
000 1 2.88
000 2 1.47
001 0 4.37
001 1 2.97
001 2 1.51
010 0 4.33
010 1 2.95
This tells us the length of the remainder of the inning, on average, starting
with each possible state. For example, starting at the bases empty, one out
state, we expect on average to have 2.88 more batters. In contrast, with a
runner on third with two outs, we expect to have 1.51 more batters.
9.2.6 Transition probabilities for individual teams

The transition probability matrix describes movements between states for an
average team. Certainly, these probabilities will vary for teams of different
batting abilities, and the probabilities will also vary against teams of different
pitching abilities. We focus on different batting teams and discuss how to
obtain good estimates of the transition probabilities for all teams.
To get the relevant data, a new variable BATTING.TEAM needs to be defined
Simulation 221
that gives the batting team in each half-inning. By use of the substr function,
we define the home team variable HOME TEAM ID, and an ifelse function is
used to define the batting team.
data2011C$HOME_TEAM_ID <- with(data2011C, substr(GAME_ID, 1, 3))
data2011C$BATTING.TEAM <- with(data2011C,
ifelse(BAT_HOME_ID == 0,
as.character(AWAY_TEAM_ID),
as.character(HOME_TEAM_ID)))
By use of the table function, a three-way table Team.T is constructed

giving the counts of each team in the transitions from the current to new
states.
Team.T <- with(data2011C, table(BATTING.TEAM, STATE, NEW.STATE))
For example, the matrix Team.T[’ANA’, , ] gives the transition counts for
Anaheim in the 2011 season.
If one is interested in comparing run productions for different batting
teams, it is necessary to make some adjustments to the team transition prob-
ability matrices to get realistic predictions of performance. To illustrate the
problem, we focus on transitions from the “100 2” state. The transition counts
are stored in the variable Team.T.S and a few rows of this table are displayed.
for six of the teams.
d.state <- subset(data2011C, STATE == ’100 2’)
Team.T.S <- with(d.state, table(BATTING.TEAM, NEW.STATE))
Team.T.S
NEW.STATE
BATTING.TEAM 000 2 001 2 010 2 011 2 100 2 101 2 110 2 3
ANA 11 3 7 8 0 16 56 253
ARI 11 4 13 2 0 15 73 240
ATL 7 2 4 7 0 23 68 273
...
TEX 12 5 16 6 1 20 67 268
TOR 7 3 9 10 1 18 51 269
WAS 9 1 5 10 0 25 61 243
For some of the less common transitions, there is much variability in the
counts across teams and this causes the corresponding team transition prob-
abilities to be unreliable. If pT EAM represents the team’s transition probabil-
ities for a particular team, and pALL are the average transition probabilities,
then a better estimate at the team’s probabilities has the form
n K
pEST = pT EAM + pALL ,
n+K n+K
where n is the number of transitions for the team and K is a smoothing count.
The description of the methodology is beyond the level of this book, but in this
case a smoothing count of K = 1274 leads to a good estimate at the team’s

transition probabilities. (The choice of K depends on the starting state.)
This method is illustrated for Washington’s transition counts starting from
the “100 2” state. The transition counts are stored in WAS.Trans, the total
number of transitions are in WAS.n, and the team transition proportions in
P.WAS. The transition counts for all teams are stored in ALL.Trans, and the
overall transition proportions are in P.ALL. The smoothing count of 1274 is
stored in K.
WAS.Trans <- Team.T.S["WAS", ]
WAS.n <- sum(WAS.Trans)
P.WAS <- WAS.Trans / WAS.n
ALL.Trans <- with(subset(data2011C, STATE == ’100 2’),

table(NEW.STATE))
P.ALL <- ALL.Trans / sum(ALL.Trans)
K <- 1274
The improved estimate at Washington’s transition proportions is computed

using the formula and stored in P.EST. The three sets of proportions (Wash-
ington, overall, and improved) are displayed in a data frame.
P.EST <- WAS.n / (K + WAS.n) * P.WAS + K / (K + WAS.n) * P.ALL
data.frame(WAS=round(P.WAS, 4),
ALL=round(c(P.ALL), 4),
EST=round(c(P.EST), 4))
WAS ALL EST
000 2 0.0254 0.0254 0.0254
001 2 0.0028 0.0073 0.0063
010 2 0.0141 0.0242 0.0220
011 2 0.0282 0.0209 0.0225
100 2 0.0000 0.0005 0.0004
101 2 0.0706 0.0504 0.0548
110 2 0.1723 0.1885 0.1850
3 0.6864 0.6827 0.6835
Note that the improved transition proportions are a compromise between the
team’s proportions and the overall values. For example, for a transition from
the state “100 2” to “010 2”, the Washington value is 0.0141, the overall value
is 0.0242, and the improved value 0.0220 falls between the Washington and
overall values. This method is especially helpful for transitions such as “100
2” to “100 2” which did not occur for Washington in this season but we know
there is a positive chance of these transitions happening in the future.
This smoothing method can be applied for all teams and all rows of the
transition matrix to obtain improved estimates at teams’ probability transi-
tion matrices. With the team transition matrices computed in this way, one
can explore the run-scoring behavior of individual batting teams.
Simulation 223
9.3 Simulating a Baseball Season

9.3.1 The Bradley-Terry model
An attractive method of modeling paired comparison data such as baseball
games is the Bradley-Terry model. This modeling and simulation is illustrated
for the 1968 baseball season where the regular season and playoff system had a
relatively simple structure. It is straightforward to adapt these methods to the
present baseball season with a more complicated schedule and playoff system.
In 1968, there were 20 teams, 10 in the National League and 10 in the
American League. Suppose each team has a talent or ability to win a game.
The talents for the 20 teams are represented by the values T1 , ..., T20 . We
assume that the talents are distributed from a normal curve model with mean
0 and standard deviation sT . A team of average ability would have a talent
value close to zero, “good” teams would have positive talents, and bad teams
would have negative talents. Suppose team A plays team B in a single game.
By the Bradley-Terry model, the probability team A wins the game is given
by the logistic function
exp(TA )
P (A wins) = .
exp(TA ) + exp(TB )
This model is closely related to the log5 method developed by Bill James
in his Baseball Abstract books in the 1980s. If PA and PB are the winning
percentages of teams A and B, then James’ formula is given by
PA /(1 − PA )
P (A wins) = .
PA /(1 − PA ) + PB /(1 − PB )
Comparing the two formulas, one sees that the log5 method is a special case
of the Bradley-Terry model where a team’s talent T is set equal to the log
odds of winning log(P/(1 − P )). A team with a talent T = 0 will win (in the
long run) half of its games (P = 0.5). In contrast, a team with talent T = 0.2
will win (using the log 5 values) approximately 55% of its games and a team
with talent T = −0.2 will win 45% of its games.
Using this model, one can simulate a baseball season as follows.
1. Construct the 1968 baseball schedule. In this season, each of the 10 teams
in each league play each other team in the same league 18 games, where 9
games are played in each team’s ballpark. (There was no interleague play
in 1968.)
2. Simulate 20 talents from a normal distribution with mean 0 and standard
deviation sT . The value of sT is chosen so that the simulated season win-
ning percentages from this model resemble the actual winning percentages
during this season.
3. Using the probability formula and the talent values, one computes the
probabilities that the home team wins all games. By a series of coin flips
with these probabilities, one determines the winners of all games.
4. Determine the winner of each league (ties need to be broken by some
random mechanism) and play a best-of-seven World Series using winning
probabilities computed using the Bradley-Terry formula and the two talent
numbers.
9.3.2 Making up a schedule

The first step in the simulation is to construct the schedule of games. A short
function make.schedule is written to help with this task. The inputs are the
vector of team names teams and the number of games k that will be played
between two teams in the first team’s home park. The output is a data frame
where each row corresponds to a game and Home and Visitor give the names
of the home and visiting teams. The gl function is helpful for generating a
factor which repeats each team label a particular number of replications, and
the rep function generates repeated copies of a vector.
make.schedule <- function(teams, k){
n.teams <- length(teams)
Home <- rep(gl(n.teams, n.teams, length=n.teams ^ 2,
labels=teams), k)
Visitor <- rep(gl(n.teams, 1, length=n.teams ^ 2, labels=teams), k)
schedule <- data.frame(Home=Home, Visitor=Visitor)
subset(schedule, Home != Visitor)
}
This function is used to construct the schedule for the 1968 season. Two
vectors NL and AL are constructed containing abbreviations for the National
League and American League teams. We apply the function make.schedule
twice, once for each league, using k = 9 since one team hosts another team
nine games. The rbind is used to paste together the NL and AL schedules,
creating the data frame schedule.
NL <- c("ATL", "CHN", "CIN", "HOU", "LAN", "NYN", "PHI",
"PIT", "SFN", "SLN")
AL <- c("BAL", "BOS", "CAL", "CHA", "CLE", "DET", "MIN",
"NYA", "OAK", "WS2")
teams <- c(NL, AL)
league <- c(rep(1, 10), rep(2, 10))
schedule <- rbind(make.schedule(NL, 9),
make.schedule(AL, 9))
Simulation 225
9.3.3 Simulating talents and computing win probabilities

The next step to compute the win probabilities for all of the games in the
season schedule. The team talents are assumed to come from a normal dis-
tribution with mean 0 and standard deviation s.talent, which we assign
s.talent = 0.20. (Recall that this value of the standard deviation is chosen
so that the season team win percentages generated from the model resemble
the actual team win percentages. ) The talents are simulated using the func-
tion rnorm and a data frame TAL is created that assigns the talents to the 20
teams. By use of several applications of the merge function, we add the team
talents to the schedule data frame – the new data frame is called SCH. Last,
since we have the talents for the home and visiting teams for all games, the
Bradley-Terry model is applied to compute home team winning probabilities
for all games – these probabilities are stored in the variable prob.Home.
s.talent <- 0.20
talents <- rnorm(20, 0, s.talent)
TAL <- data.frame(Team=teams, League=league, Talent=talents)
SCH <- merge(schedule, TAL, by.x="Home", by.y="Team")
names(SCH)[4] <- "Talent.Home"
SCH <- merge(SCH, TAL, by.x="Visitor", by.y="Team")
names(SCH)[6] <- "Talent.Visitor"
SCH$prob.Home <- with(SCH,
exp(Talent.Home) / (exp(Talent.Home) + exp(Talent.Visitor)))
The first six rows of the data frame SCH are displayed, where one sees the
games scheduled, the talents of the home and away teams, and the probability
that the home team wins the matchup.
head(SCH)
Visitor Home League.x Talent.Home League.y Talent.Visitor prob.Home
1 ATL PHI 1 -0.02757542 1 -0.06858703 0.5102515
2 ATL PIT 1 0.14765145 1 -0.06858703 0.5538500
3 ATL SFN 1 -0.07264467 1 -0.06858703 0.4989856
4 ATL NYN 1 -0.20925903 1 -0.06858703 0.4648899
5 ATL HOU 1 -0.19522186 1 -0.06858703 0.4683835
6 ATL SLN 1 -0.09780258 1 -0.06858703 0.4926966
9.3.4 Simulating the regular season

To simulate an entire season of games, we are doing a series of coin flips, where
the probability the home team wins depends on the winning probability. The
function rbinom performs the coin flips for the 1620 scheduled games – the
outcomes are a sequence of 0s and 1s. By use of the ifelse function, we define
the winner variable to be the Home team if the outcome is 1, and the Visitor
otherwise.
SCH$outcome <- with(SCH, rbinom(nrow(SCH), 1, prob.Home))

SCH$winner <- with(SCH, ifelse(outcome, as.character(Home),
as.character(Visitor)))
The teams, home win probabilities, and outcomes of the first six games are
displayed.
head(SCH[, c("Visitor", "Home", "prob.Home", "outcome","winner")])
Visitor Home prob.Home outcome winner
1 ATL PHI 0.5102515 0 ATL
2 ATL PIT 0.5538500 1 PIT
3 ATL SFN 0.4989856 1 SFN
4 ATL NYN 0.4648899 1 NYN
5 ATL HOU 0.4683835 1 HOU
6 ATL SLN 0.4926966 1 SLN
How did the teams perform during this particular simulated season? Using
the table function, we find the number of wins for all teams. This information
is collected together with the team names in the data frame WIN, and using
the merge function the season results are combined with the team talents to
create the data frame RESULTS.
wins <- table(SCH$winner)
WIN <- data.frame(Team=names(wins), Wins=as.numeric(wins))
RESULTS <- merge(TAL, WIN)
9.3.5 Simulating the post-season

After the regular season, one simulates the post-season series. A function
win.league is written that simulates a league championship – the inputs are
the data frame RR of teams and win totals and the league number (1 for
National and 2 for American). The function first identifies the teams that
have the largest number of wins. If one team has the maximum number, then
an indicator variable Winner.Lg is created which is 1 for that particular team.
If there is a tie in win totals for two or more teams, then we randomly choose
one of the teams to be the pennant winner (using the random multinomial
function rmultinom) where the winning probabilities are proportional to the
exponentials of the talents.
win.league <- function(RR, league){
wins <- RR$Wins * (RR$League == league)
MAX <- max(wins)
if(sum(wins == MAX) > 1){
prob <- exp(RR$Talent) * (wins == MAX)
outcome <- c(rmultinom(1, 1, prob))
RR$Winner.Lg <- RR$Winner.Lg + outcome
}
if(sum(wins == MAX) == 1){
Simulation 227
RR$Winner.Lg <- RR$Winner.Lg + as.numeric(wins == MAX)}

RR
}
To simulate the post-season, we initialize two new variables Winner.Lg

and Winner.WS – these are indicators for the league champions and the World
Series winners. By two applications of win.league, we find the winners of each
league. The World Series is simulated by flipping a coin seven times, where the
win probabilities are proportional to exp(Talent). The vector winner indicates
the team winning a majority of the games.
RESULTS$Winner.Lg <- 0
RESULTS$Winner.WS <- 0;
for(j in 1:2)
RESULTS <- win.league(RESULTS, j)
teams <- (1:20)[RESULTS$Winner.Lg == 1]
outcome <- c(rmultinom(1, 7, exp(RESULTS$Talent)[teams]))
winner <- teams[1] * (diff(outcome) < 0) + teams[2] * (diff(outcome) > 0)
RESULTS$Winner.WS[winner] <- 1
9.3.6 Function to simulate one season

It is convenient to place all of these commands including the functions
make.schedule and win.league in a single function one.simulation.68.
The only input is the standard deviation s.talent that describes the spread
of the normal talent distribution. The output is a data frame containing the
teams, talents, number of season wins, and success in the post-season. We
illustrate simulating one season and display the data frame RESULTS that is
returned.
RESULTS <- one.simulation.68(0.20)
RESULTS
Team League Talent Wins Winner.Lg Winner.WS
1 ATL 1 0.080997187 93 0 0
2 BAL 2 -0.219407302 70 0 0
3 BOS 2 0.136735454 93 0 0
4 CAL 2 -0.410259781 58 0 0
5 CHA 2 -0.142939718 72 0 0
6 CHN 1 -0.007529692 78 0 0
7 CIN 1 0.130597992 84 0 0
8 CLE 2 -0.261658757 79 0 0
9 DET 2 -0.006537641 77 0 0
10 HOU 1 -0.332069747 63 0 0
11 LAN 1 0.004409475 79 0 0
12 MIN 2 -0.180103045 81 0 0
13 NYA 2 0.105354274 86 0 0
14 NYN 1 -0.190907202 61 0 0
15 OAK 2 0.184825855 91 0 0
16 PHI 1 -0.057648893 79 0 0
17 PIT 1 0.678827397 111 1 1
18 SFN 1 -0.191052425 85 0 0
19 SLN 1 -0.230921646 77 0 0
20 WS2 2 0.451958661 103 1 0
A new function display.standings is written to put the season wins in a

more familiar standings format. The inputs to this function are the RESULTS
data frame and the league indicator.
display.standings <- function(RESULTS, league){
Standings <- subset(RESULTS, League == league)[, c("Team", "Wins")]
Standings$Losses <- 162 - Standings$Wins
Standings[order(Standings$Wins, decreasing=TRUE), ]
}
This function is applied twice (once for each league) and the cbind function
combines the two standings into a single data frame. The league champions
and the World Series winner are also displayed.
cbind(display.standings(RESULTS, 1), display.standings(RESULTS, 2))
Team Wins Losses Team Wins Losses
17 PIT 111 51 WS2 103 59
1 ATL 93 69 BOS 93 69
18 SFN 85 77 OAK 91 71
7 CIN 84 78 NYA 86 76
11 LAN 79 83 MIN 81 81
16 PHI 79 83 CLE 79 83
6 CHN 78 84 DET 77 85
19 SLN 77 85 CHA 72 90
10 HOU 63 99 BAL 70 92
14 NYN 61 101 CAL 58 104
with(RESULTS, as.character(Team[Winner.Lg == 1]))
[1] "PIT" "WS2"
with(RESULTS, as.character(Team[Winner.WS == 1]))
[1] "PIT"
In this particular simulated season, Pittsburgh (PIT) won the National

League with 111 wins and Washington (WS2) won the American League with
103 wins. Pittsburgh defeated Washington in the World Series. The team with
the best talent in this season was Pittsburgh (talent equal to 0.6788) and they
won the World Series. In other words the “best team in baseball” was also the
most successful during this simulated season. We will shortly see if the best
team typically wins the World Series.
9.3.7 Simulating many seasons

One can learn about the relationship between a team’s ability and its sea-
son performance by simulating many seasons of baseball. We initially set a
Simulation 229
Many.Results data frame to NULL and use a for loop to repeat the simulation
for 1000 seasons, storing the output in Many.Results.
Many.Results <- NULL
for(j in 1:1000)
Many.Results <- rbind(Many.Results, one.simulation.68(0.20))
The data frame Many.Results contains the talent number and number of
wins for 1000 × 20 = 20,000 teams. The smoothScatter function is used to
construct a smoothed scatterplot of Talent and Wins and Figure 9.1 shows
the result. (Here the plot function would have resulted in an overly cluttered
scatterplot.)
with(Many.Results, smoothScatter(Talent, Wins))
120
100
Wins
80
60
40
−0.5 0.0 0.5

Talent
FIGURE 9.1
Smoothed scatterplot of talent and number of season wins for teams in 1000
simulated seasons.
As expected, there is a positive trend in the graph, indicating that better teams
tend to win more games. But there is much vertical spread in the scatterplot
which says that the relationship between talent and wins is not strong.
To reinforce the last point, suppose we focus on “average” teams that have
a talent number between −0.05 and 0.05. Using the subset function, a new
data frame Results.avg is created containing the talent and wins data for
these average teams. A histogram is constructed of the season wins for these
teams. (See Figure 9.2.)
Results.avg <- subset(Many.Results, Talent > -0.05 & Talent < 0.05)
hist(Results.avg$Wins)
Histogram of Results.avg$Wins
1000
800
600
Frequency
400
200
0
50 60 70 80 90 100 110
Results.avg$Wins
FIGURE 9.2
Histogram of the number of season wins for “average” teams in the 1000
simulated seasons.
One expects these average teams to win about 80 games. But what is surprising
is the variability in the win totals – average teams can regularly have win totals
between 70 and 90, and it is possible (but not likely) to have a win total close
to 100.
What is the relationship between a team’s talent and its post-season suc-
cess? Consider first the relationship between a team’s talent (variable Talent)
and winning the league (the variable Winner.Lg). Since Winner.Lg is a bi-
nary (0 or 1) variable, a common approach for representing this relationship
is a logistic model – this is a generalization of the usual regression model
where the response variable is binary instead of continuous. The glm func-
tion with the family=binomial argument is used to fit a logistic model – the
Simulation 231
output is stored in the variable fit1. In a similar fashion, a logistic model

is used to model the relationship between winning the World Series (variable
Winner.WS) and the talent – the output is in the variable fit2.
fit1 <- glm(Winner.WS ~ Talent, data = Many.Results, family=binomial)
fit2 <- glm(Winner.Lg ~ Talent, data = Many.Results, family=binomial)
A logistic model has the form
exp(a + bT )
p= ,
1 + exp(a + bT )
where T is a team’s talent, (a, b) are the regression coefficients, and p is the
probability of the event. In the following code, the regression coefficients of
the “win pennant” logistic fit are stored in the variable b1. By use of the
curve function, the fitted probability of winning the pennant is graphed as
a function of the talent. A second application of curve is used to overlay
the fitted probability of winning the World Series. The completed graph is
displayed in Figure 9.3.
b1 <- coef(fit1)
curve(exp(b1[1] + b1[2] * x) / (1 + exp(b1[1] + b1[2] * x)),
-0.4, 0.4, xlab="Talent", ylab="Probability", lwd=2,
ylim=c(0, 1))
b2 <- coef(fit2)
curve(exp(b2[1] + b2[2] * x) / (1 + exp(b2[1] + b2[2] * x)),
add=TRUE, lwd=2, lty=2)
legend(-0.2, 0.8, legend=c("Win Pennant", "Win World Series"),
lwd=2, lty=c(1, 2))
As expected, the chance of a team winning the pennant (solid line) increases
as a function of the talent. An average team with T = 0 has only a small
chance of winning the pennant; an excellent team with a talent close to 0.4
has about a 60% chance of winning the pennant. The probabilities of winning
the World Series (represented by a dashed line) are substantially smaller than
the chances of winning the pennant. For example, this excellent (T = 0.4)
team has only about a 35% chance of winning the World Series. In fact, it can
be demonstrated that the team winning the World Series is likely not to be
the team with the best talent (largest value of T ).
9.4 Further Reading

A general description of the Markov chain probability model is contained in
Kemeny and Snell (1976). Pankin (1987) and Bukiet et al. (1997) illustrate
the use of Markov chains to model baseball. Chapter 9 of Albert (2003) gives
1.0
0.8
Win Pennant
Win World Series
0.6
Probability
0.4
0.2
0.0
−0.4 −0.2 0.0 0.2 0.4

Talent
FIGURE 9.3
Probability of winning the league and the World Series for teams of different
talents.
an introductory description of Markov chains and illustrates the construction

and use of the transition matrix using 1987 season data. The Bradley-Terry
model (Bradley and Terry, 1952) is a popular statistical model for paired
comparisons. Chapter 9 of Albert and Bennett (2003) describes the application
of the Bradley-Terry model for baseball team competition. The use of R in
simulation is introduced in Chapter 11 of Albert and Rizzo (2012).
9.5 Exercises
1. (A Simple Markov Chain)
Suppose one is interested only in the number of outs in an inning. There
are four possible states in an inning (0 outs, 1 out, 2 outs, and 3 outs)
and you move between these states in each plate appearance. Suppose at
each PA, the chance of not increasing the number of outs is 0.3, and the
Simulation 233
probability of increasing the outs by one is 0.7. The following R code puts
the transition probabilities of this Markov chain in a matrix P.
P <- matrix(c(.3, .7, 0, 0,
0, .3, .7, 0,
0, 0, .3, .7,
0, 0, 0, 1), 4, 4, byrow=TRUE)
(a) If one multiplies the matrix P by itself P to obtain the matrix P2:
P2 <- P %*% P
The first row of P2 gives the probabilities of moving from 0 outs
to each of the four states after two plate appearances. Compute P2.
Based on this computation, find the probability of moving from 0
outs to 1 out after two plate appearances.
(b) The fundamental matrix N is computed as
N <- solve(diag(c(1, 1, 1)) - P[-4, -4])
The first row gives the average number of PAs at 0 out, 1 out, and 2
outs in an inning. Compute N and find the average number of PAs in
one inning in this model.
2. (A Simple Markov Chain, Continued)

The following function simulate.half.inning will simulate the number
of plate appearances in a single half-inning of the Markov chain model
described in Exercise 1 where the input P is the transition probability
matrix.
simulate.half.inning <- function(P){
s <- 1; path <- NULL;
while(s < 4){
s.new <- sample(1 : 4, 1, prob=P[s, ])
path <- c(path, s.new)
s <- s.new
}
length(path)
}
(a) Use the replicate function to simulate 1000 half-innings of this

Markov chain and store the lengths of these simulated innings in
the vector lengths.
(b) Using this simulated output, find the probability that a half-inning
contains exactly four plate appearances.
(c) Use the simulated output to find the average number of PAs in a
half-innings. Compare your answer with the exact answer in Exercise
1, part (b).
3. (Simulating a Half Inning)

In Section 9.2.4, the expected number of runs as calculated for each one of
the 24 possible runners-outs situations using data from the 2011 season.
To see how these values can change across seasons, download play-by-
play data from Retrosheet for the 1968 season, construct the probability
transition matrix, simulate 10,000 half-innings from each of the 24 situa-
tions, and compute the runs expectancy matrix. Compare this 1968 runs
expectancy matrix compares with the one computed using 2011 data.
4. (Simulating the 1950 Season)
Suppose you are interested in simulating the 1950 regular season for the
American League. In this season, the team abbreviations were “PHI,”
“BRO,” “NYG,” “BSN,” “STL,” “CIN,” “CHC,” and “PIT” and each
team played every other team 22 games (11 games at each park).
(a) Using the function make.schedule, construct the schedule of games

for this AL season.
(b) Suppose the team talents follow a normal distribution with mean 0
and standard deviation 0.25. Using the Bradley-Terry model, assign
home win probabilities for all games on the schedule.
(c) Use the rbinom function to simulate the outcomes of all 616 games
of the AL 1950 season.
(d) Compute the number of season wins for all teams in your simulation.
5. (Simulating the 1950 Season, Continued)
(a) Write a function to perform the simulation scheme described in Exer-

cise 4. Have the function return the team with the largest talent and
the team with the most wins. (If there is a tie for the league pennant,
have the function return one of the best teams at random.)
(b) Repeat this simulation for 1000 seasons, collecting the most talented
team and the most successful team for all seasons.
(c) Based on your simulations, what is the chance that the most talented
team wins the pennant?
6. (Simulating the World Series)
(a) Write a function to simulate a World Series. The input is the the
probability p the AL team will defeat the NL team in a single game.
(b) Suppose an AL team with talent 0.40 plays a NL team with talent
0.25. Using the Bradley-Terry model, determine the probability p
that the AL wins a game.
(c) Using the value of p determined in part (b), simulate 1000 World
Series and find the probability the AL team wins the World Series.
Simulation 235
(d) Repeat parts (b) and (c) for AL and NL teams who have the same
talents.
10
Exploring Streaky Performances
CONTENTS
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
10.2 The Great Streak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
10.2.1 Finding game hitting streaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
10.2.2 Moving batting averages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
10.3 Streaks in Individual At-Bats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
10.3.1 Streaks of hits and outs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
10.3.2 Moving batting averages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
10.3.3 Finding hitting slumps for all players . . . . . . . . . . . . . . . . . . . 243
10.3.4 Were Suzuki and Ibanez unusually streaky? . . . . . . . . . . . . 246
10.4 Local Patterns of Weighted On-Base Average . . . . . . . . . . . . . . . . . . . 249
10.5 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
10.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
10.1 Introduction
Some of the most interesting phenomena in baseball are streaky or hot/cold
performances by hitters and pitchers. During particular periods in the season,
a particular player will hit for a high batting average, and in other periods,
the player will be in a “cold streak” and all batted balls appear to be fielded
for outs. In this chapter, we’ll use R to explore streaky hitting performances.
One of the great hitting accomplishments in baseball history is Joe DiMag-
gio’s 56-game hitting streak and Section 10.2 explores DiMaggio’s game-to-
game hitting for the 1941 season. An R function is used to find all of DiMag-
gio’s hitting streaks, and a moving average function is used to explore DiMag-
gio’s batting average over short time intervals. Retrosheet play-by-play data
records batters’ performances in all plate appearances and we use this data in
Section 10.3 to explore hitting streaks in individual at-bats. Suppose a hitter
is going through an “0 for 20” hitting slump – should we be surprised? One
way of answering this question is to find the longest hitting slumps for all
hitters in a particular baseball season. A second way to understand the size of
this hitting slump is to contrast this hitting with pattern of slumps under a
random model. A method for simulating a random pattern of hits and outs is
237
described and this method is used to assess if a particular player exhibits more
streakiness in his hitting sequence than what one would expect by chance.
This discussion of streakiness focuses on patterns of hits and outs, and
certainly the quality of an at-bat depends on more than just getting a hit.
Section 10.4 discusses patterns of streakiness using the players’ weighted on-
base percentage (wOBA) where positive outcomes of a plate appearance are
weighted by their run values. We look at players’ wOBA over groups of five
games during a season. A way to describe streaky hitting behavior is to look
at the variability of the five-game wOBA values. Using this measure of streak-
iness, we identify the streaky hitters during the 2011 season.
10.2 The Great Streak

10.2.1 Finding game hitting streaks
Whenever there is a discussion of great streaky performances in baseball, one
has to talk about the “Great Streak” where Joe DiMaggio hit in 56 consecutive
games during the 1941 season. Many people think that this particular hitting
accomplishment is one of the few baseball records that will not be broken in
our lifetime. We use DiMaggio’s game-to-game hitting data to motivate how
we can use R to explore streaky performances.
Play-by-play hitting records are currently not available in Retrosheet for
the 1941 season, but baseball-reference.com gives a game-to-game hitting log
for DiMaggio for this season. By cutting and pasting the data table from the
web page to a text file, we create a new file “dimaggio.1941.csv” that contains
this hitting log saved in csv format. By use of the read.csv function, we load
this file into R, creating a new data frame joe.
joe <- read.csv("dimaggio.1941.csv")
For each game during the season, the data frame records AB, the number of
at-bats, and H, the number of hits. As a quick check that the data has been
entered correctly, we compute DiMaggio’s season batting average by summing
the game hit totals and dividing by the total at-bats.
sum(joe$H) / sum(joe$AB)
[1] 0.3567468
The result agrees with DiMaggio’s 1941 batting average of .357. (Actually,
although this was a high average, it was overshadowed by Ted Williams’ .406
average during the 1941 season.)
A hitting streak is commonly defined as the number of consecutive games in
which a player gets at least one base hit. Suppose we’re interested in computing
all of DiMaggio’s hitting streaks for the 1941 season. Towards this goal, using
Exploring Streaky Performances 239
the ifelse function, we create a new variable HIT for each game that is either
1 or 0 depending on whether DiMaggio recorded at least one hit in the game.
joe$HIT <- ifelse(joe$H >= 1, 1, 0)
We display the values of HIT that visually shows DiMaggio’s streaky hitting
performance.
joe$HIT
[1] 1 1 1 1 1 1 1 1 0 0 0 1 1 1 0 1 1 0 0 1 0 0 0 1 1 1 0 0 1 1 1 1
[33] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
[65] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1
[97] 1 1 1 1 1 0 0 1 1 1 1 0 0 1 1 0 0 0 1 1 1 1 0 0 0 1 1 1 1 1 1 1
[129] 0 1 0 1 1 1 1 1 0 1 1
We see that DiMaggio started the season with an eight-game hitting streak,
then had three games with no hits, a hitting streak of three games, and so on.
Suppose we wish to compute all hitting streaks for a particular player.
This is conveniently done using the following user-defined function streaks.
The input to this function is a vector y of 0s and 1s corresponding to game
results where the player was hitless (0) or received at least one hit (1). The
output will be a vector containing the lengths of all hitting streaks.
streaks <- function(y){
n <- length(y)
where <- c(0, y, 0) == 0
location.zeros <- (0 : (n+1))[where]
streak.lengths <- diff(location.zeros) - 1
streak.lengths[streak.lengths > 0]
}
In the function, we define n to the number of values in the vector y. A zero

is appended to the beginning and end of y, and a vector location.zeros is
defined giving the indices of the vector locations containing 0s. The lengths of
all streaks of 1s are found using the function diff to compute the differences
in the locations with 0s.
The function streaks is read into R and this function is applied to DiMag-
gio’s game hit/no-hit sequence stored in the variable joe$HIT.
streaks(joe$HIT)
[1] 8 3 2 1 3 56 16 4 2 4 7 1 5 2
This function picks up DiMaggio’s famous 56-game hitting streak. It is in-

teresting to note that Joe followed his 56-game streak immediately with a
16-game hitting streak.
The media is also fascinated with streaks of no-hit games. One can find
DiMaggio’s streaks of hitless games by creating a new variable NO.HITS that
records 1 or 0 if the player has respectively no hits or hits, and then applying
the streaks function to the new variable.
joe$NO.HIT <- 1 - joe$HIT

streaks(joe$NO.HIT)
[1] 3 1 2 3 2 1 2 2 3 3 1 1 1
It is interesting that the length of the longest streak of no-hit games was only
three for DiMaggio’s 1941 season.
10.2.2 Moving batting averages

An alternative way of looking at streaky hitting performances uses batting av-
erages computed over short time intervals. One may be interested in exploring
DiMaggio’s batting average in this manner. He must have been a hot hitter
during his 56-game hitting streak, and perhaps DiMaggio was somewhat cold
in other periods during the season.
In general, suppose we are interested in computing a player’s batting av-
erage over a width or window of 10 games. We want to compute the batting
average over games 1 to 10, over games 2 to 11, over games 3 to 12, and so on.
These batting averages would be the sum of hits divided by the sum of at-bats
over the 10-game periods. These short-term batting averages are commonly
called moving averages.
The function moving.average below computes these moving averages. The
arguments to the function are the vector of hits H, the vector of at-bats AB,
and the window of games width.
moving.average <- function(H, AB, width){
require(TTR)
game <- 1 : length(H)
P <- data.frame(Game=SMA(game, width),
Average=SMA(H, width) / SMA(AB, width))
P[complete.cases(P), ]
}
A vector game is defined to be the integers 1 through n, where n is the number
of games. The function SMA from the package TTR computes the simple moving
averages of game, H and AB, and those quantities are used to compute the
moving batting average AVG. The output of the function is a data frame with
two variables Game and Average. The variable Game gives the game number
value in the middle of the window, and Average is the corresponding batting
average over the game window.
After the function moving.average is read into R, it is easy to compute
DiMaggio’s batting average over short time intervals. Suppose we consider
a window of 10 games. In the following code, we use moving.average to
compute the moving batting averages and the output is passed to plot to
construct a line graph of these averages (see Figure 10.1). A horizontal line
is added at DiMaggio’s season batting average so one can easily see when
Joe was relatively hot and cold during the season. To relate this display with
DiMaggio’s hitting streaks, we use the rug function to display the games where
Joe had at least one hit on the horizontal axis.
0.5
0.4
Average
0.3
0.2
20 40 60 80 100 120
Game
FIGURE 10.1
Moving average plot of DiMaggio’s batting average for the 1941 season using
a window of 10 games. The horizontal line shows DiMaggio’s season batting
average. The games where DiMaggio had at least one base hit are displayed
on the horizontal axis.
plot(moving.average(joe$H, joe$AB, 10), type="l")

abline(h=sum(joe$H) / sum(joe$AB))
game.hits <- (1:nrow(joe))[as.logical(joe$HIT)]
rug(game.hits)
This figure dramatically shows that DiMaggio’s hitting performance climbed

steadily during his 56-game hitting streak and he actually had a short-term 10-
game batting average over .500 during the streak. DiMaggio had a noticeable
hitting slump in the second half of the season and he hit bottom about game
110. In practice, the appearance of this graph may depend on the choice
of time interval (argument width in the function moving.average) and one
should experiment with several width choices to get a better understanding
of a hitter’s short-term batting performance.
10.3 Streaks in Individual At-Bats

The previous section considered hitting streaks at a game-to-game level. Since
records of individual plate appearances are available in the Retrosheet play-
by-play files, it is straightforward to explore hitting streaks at this finer level.
Ichiro Suzuki is one of the most exciting hitters in baseball, especially for his
ability to hit singles, many of the infield variety. The streakiness patterns in
Suzuki’s play-by-play hitting data is explored for the 2011 season.
We begin by reading the Retrosheet play-by-play file for the 2011 season,
storing the file in the data frame data2011.
The subset function is used to define a new data frame ichiro.AB; records
are chosen where the batting id is “suzui001” (Suzuki’s code id) and the at-bat
flag is TRUE. (In this exploration, only Suzuki’s official at-bats are considered.)
ichiro.AB <- subset(data2011, BAT_ID == "suzui001" & AB_FL == TRUE)
10.3.1 Streaks of hits and outs

We record at each at-bat if a hit occurred. There is a current variable H FL
in the data2011 data frame recording the number of bases for a hit. Using
the ifelse function, a new variable HIT is defined that is 1 if a hit occurs
and 0 otherwise. To make sure that these at-bats are correctly ordered in time
during the season, a variable DATE (extracted from the GAME ID variable using
the substr function) is defined, and the order function sorts the data frame
ichiro.AB by date.
ichiro.AB$HIT <- ifelse(ichiro.AB$H_FL > 0, 1, 0)
ichiro.AB$DATE <- substr(ichiro.AB$GAME_ID, 4, 12)
ichiro.AB <- ichiro.AB[order(ichiro.AB$DATE), ]
From the variable HIT, the lengths of all hitting streaks are identified,
where a streak refers to a sequence of consecutive base hits. Using the streaks
function defined in Section 10.2, the streak lengths are obtained for Suzuki in
the 2011 season.
streaks(ichiro.AB$HIT)
[1] 1 1 1 2 1 1 1 1 1 1 1 2 1 1 1 1 2 2 1 1 1 4 1 1 1 1 2 3 1 1
[31] 1 1 1 1 1 2 2 1 1 1 1 1 2 2 1 2 2 1 1 1 2 1 2 1 1 2 1 3 1 3
[61] 1 1 1 1 1 1 1 1 1 2 1 1 2 1 2 1 1 1 3 1 1 1 2 1 4 1 1 2 3 1
[91] 1 1 1 1 1 1 2 1 2 1 1 1 2 1 2 1 1 1 1 1 1 1 1 1 1 1 2 1 1 1
[121] 1 1 4 1 1 1 1 2 2 1 1 3 1 1 1 2
As expected, most of the hitting streaks lengths are 1, although several times
Suzuki had four consecutive hits.
It may be more interesting to explore the lengths of the gaps between hits.
By the operation 1 - HIT, the roles of 0 and 1 are reversed in the sequence,
and the function streaks is applied to find the lengths of all of the gaps
between hits that are 1 or larger.
streaks(1 - ichiro.AB$HIT)
[1] 1 2 2 4 8 1 3 6 2 3 6 1 2 8 4 1 1 1 5 2
[21] 1 1 4 4 1 1 4 1 5 6 2 3 1 2 4 4 5 4 3 2
[41] 10 2 2 12 11 5 16 2 2 15 2 1 1 2 1 1 2 2 4 2
[61] 3 3 9 1 3 2 1 3 10 2 1 1 10 3 15 6 4 3 1 2
[81] 3 3 3 5 4 11 5 1 2 4 3 5 1 3 4 5 10 3 3 3
[101] 1 1 3 2 1 2 3 2 2 2 1 1 1 4 6 1 5 3 1 8
[121] 4 1 1 7 1 7 6 5 1 2 6 5 6 2 3 6
This output is more interesting; a frequency table of this output is constructed

using the table function.
table(streaks(1 - ichiro.AB$HIT))
1 2 3 4 5 6 7 8 9 10 11 12 15 16
35 28 22 15 11 9 2 3 1 4 2 1 2 1
It is seen that Suzuki had a streak of 12 outs once, a streak of 15 outs twice,
and a streak of 16 outs once.
10.3.2 Moving batting averages

Another way to view Suzuki’s streaky batting performance is to consider his
batting average over short time intervals, analogous to what we did for DiMag-
gio for his game-to-game hitting data. Using the moving.average function, a
moving average plot is constructed of Ichiro’s batting average using a window
of 30 at-bats (see Figure 10.2). Using the rug function, we display the at-bats
where Ichiro had hits. The long streaks of outs are visible as gaps in the rug
plot. During the middle of the season, Ichiro had a 30 at-bat batting average
exceeding 0.500, while during other periods, his 30 at-bat average was as low
as 0.100.
ichiro.AB$AB <- 1
plot(moving.average(ichiro.AB$HIT, ichiro.AB$AB, 30), type="l", xlab="AB")
abline(h=mean(ichiro.AB$HIT))
rug((1:nrow(ichiro))[ichiro.AB$HIT == TRUE])
10.3.3 Finding hitting slumps for all players

In our exploration of Suzuki’s batting performance, we saw that he had a
“0 for 16” hitting performance during the season. Should we be surprised by
0.5
0.4
Average
0.3
0.2
0.1
0 100 200 300 400 500 600
AB
FIGURE 10.2
Moving average plot of Ichiro Suzuki’s batting average for the 2011 season
using a window of 30 at-bats. The horizontal line shows Suzuki’s season batting
average. The at-bats where Suzuki had at least one base hit are shown on the
horizontal axis.
a hitting slump of length 16? Let’s compare Suzuki’s long slump with the
longest slumps for all regular players during the 2011 season.
First a new function longest.ofer is written that computes the length of
the longest hitting slump for a given batter. (An “ofer” is a slang word for a
hitless streak in baseball.) The inputs to this function are the batter id code
batter and the batting data frame data. The output of the function is the
length of the longest slump.
longest.ofer <- function(batter, data){
d.AB <- subset(data, BAT_ID == batter & AB_FL == TRUE)
d.AB$HIT <- ifelse(d.AB$H_FL > 0, 1, 0)
d.AB$DATE <- substr(d.AB$GAME_ID, 4, 12)
d.AB <- d.AB[order(d.AB$DATE), ]
source("streaks.R")
max(streaks(1 - d.AB$HIT))
}
After reading this function into R, we confirm that it works by finding the
longest hitting slump for Suzuki.
longest.ofer("suzui001", data2011)
[1] 16
Suppose we want to compute the length of the longest hitting slump for all
players in this season with at least 400 at-bats. Using the aggregate function,
the number of at-bats is computed for all players, and players.400 is defined
to be the vector of the id codes of all players with 400 or more at-bats. By use
of the sapply function together with the longest.ofer function, the length
of the longest slump is computed for all regular hitters. Using the data.frame
function, the vector S is converted to a data frame, adding the variable Player.
(The rownames(S) <- NULL command is used to remove the player ids from
the row names.)
A <- aggregate(data2011$AB_FL, list(Player=data2011$BAT_ID), sum)
players.400 <- A$Player[A$x >= 400]
S <- sapply(players.400, longest.ofer, data2011)
S <- data.frame(Player=names(S), Streak=S)
rownames(S) <- NULL
To decipher the player ids, it is helpful to merge the data frame of the
longest hitting slumps S with the player roster information contained in the
Retrosheet file “roster2011.csv.” This roster file is read into R, saving it in
the data frame roster2011. Then the merge function is applied, merging
data frames S and roster2011, matching on the variables Player (in S) and
Player.ID (in roster2011).
roster2011 <- read.csv("roster2011.csv")
S1 <- merge(S, roster2011, by.x="Player", by.y="Player.ID")
The slump lengths are ordered in decreasing order using the function order
with the decreasing=TRUE argument and the rows of the data frame S1 are
reordered using this ordering. The top six slump lengths are displayed by the
head function.
S.ordered <- S1[order(S1$Streak, decreasing=TRUE), ]
head(S.ordered)
Player Streak X Last.Name First.Name Bats Pitches Team V7
80 ibanr001 35 941 Ibanez Raul L R PHI OF
6 aybae001 30 3 Aybar Erick B R ANA SS
113 mcgec001 27 726 McGehee Casey R R MIL 3B
122 olivm001 27 1095 Olivo Miguel R R SEA C
152 ruizc001 26 958 Ruiz Carlos R R PHI C
48 ellim001 25 896 Ellis Mark R R OAK SS
The six longest hitting slumps during the 2011 season were by Raul Ibanez
(35), Erick Aybar (30), Casey McGehee (27), Miguel Olivo (27), Carlos Ruiz
(26), and Mark Ellis (25). Relative to these long hitting slumps, Suzuki’s
hitting slump of 16 at-bats looks short.
10.3.4 Were Suzuki and Ibanez unusually streaky?

In the previous section, patterns of streakiness of hit/out data were compared
for all players in the 2011 season. An alternative way to look at the streakiness
of a player is to contrast his streaky pattern of hitting with streaky patterns
under a “random” model.
To illustrate this method, consider a hypothetical player who bats 13 times
with the outcomes
0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 1, 1, 1.
A measure of streakiness is defined based on this sequence of hits and outs.
One good measure of streakiness or clumpiness in the sequence is the sum of
squares of the gaps between successive hits. In this example, the gaps between
hits are 1, 2, and 4, and the sum of squares of the gaps is S = 12 +22 +42 = 21.
Is the value of streakiness statistic S = 21 large enough to conclude that
this player’s pattern of hitting is non-random? This question is answered by
a simple simulation experiment. If the player sequence of hit/out outcomes is
truly random, then all possible arrangements of the sequence of 6 hits and 7
outs are equally likely. We randomly arrange the sequence 0, 1, 0, 0, 1, 1, 0,
0, 0, 0, 1, 1, 1, find the gaps, and compute the streakiness measure S. This
randomization procedure is repeated many times, collecting, say, 1000 values
of the streakiness measure S. A histogram of the values of S is constructed
– this histogram represents the distribution of S under a random model. If
the observed value of S = 21 is in the middle of the histogram, then the
player’s pattern of streakiness is consistent with a random model. On the
other hand, if the value S = 21 is in the right tail of this histogram, then
the observed streaky pattern is not consistent with “random” streakiness and
there is evidence that the player’s pattern of hits and outs is non-random.
This method is first illustrated for Ichiro Suzuki’s 2011 hitting data. As
before, Suzuki’s hitting data is read into the data frame ichiro.AB – the
variable HIT is a sequence of 0s and 1s, where 0 corresponds to an out and 1
corresponds to a hit.
ichiro.AB <- subset(data2011, BAT_ID == "suzui001" & AB_FL == TRUE)
ichiro.AB$HIT <- ifelse(ichiro.AB$H_FL > 0, 1, 0)
ichiro.AB$DATE <- substr(ichiro.AB$GAME_ID, 4, 12)
ichiro.AB <- ichiro.AB[order(ichiro.AB$DATE), ]
The clumpiness or streakiness is measured by the sum of squares of all

gaps between hits. The function streaks is used to find the gaps which are
stored in the vector st. Each of the gap values is squared and the sum function
computes the sum.
source("streaks.R")
st <- streaks(1 - ichiro.AB$HIT)
sum(st ^ 2)
[1] 3047
The value of Suzuki’s streakiness statistic is S = 3047.

Next, a function random.mix is written to perform one iteration of the
simulation experiment where the input y is a vector of 0s and 1s. The sample
function finds a random arrangement of y, the streaks function finds the
gaps between hits, and the function clump.stat finds the sum of squares of
the gaps.
random.mix <- function(y){
source("streaks.R")
clump.stat <- function(sp) sum(sp ^ 2)
mixed <- sample(y)
clump.stat(streaks(1 - mixed))
}
By use of the replicate function, this simulation experiment is repeated

1000 times, storing the values of the streakiness statistic in the vector ST.
ST <- replicate(1000, random.mix(ichiro.AB$HIT))
A histogram of the values of ST is constructed (using the truehist function in

the package MASS) and the abline function is used to overlay the clumpiness
value (3047) for Suzuki. (See Figure 10.3.)
library(MASS)
truehist(ST)
abline(v=3047, lwd=3)
text(3250, 0.0016, "Suzuki", cex=1.5)
Since the value of 3047 is in the middle of the histogram distribution, the
streakiness pattern in Suzuki’s hitting is consistent with a random model.
There is insufficient evidence that Suzuki is truly streaky.
Suzuki
0.0015
0.0010
0.0005
0.0000
2500 3000 3500 4000
ST
FIGURE 10.3
Histogram of one thousand values of the clumpiness statistic assuming all
arrangements of hits and outs for Suzuki are equally likely. The observed
value of the clumpiness statistic for Suzuki is shown using a vertical line.
This method can be used to check if the streaky patterns of any hitter
are non-random. A new function clump.test is constructed using the R code
previous discussed. The input is the player id code playerid and the season
batting data frame data. One thousand values of the clumpiness measure are
computed by 1000 replications of the simulation procedure. A histogram of
the clumpiness measures is constructed and the observed clumpiness statistic
is shown as a vertical line.
clump.test <- function(playerid, data){
player.AB <- subset(data, BAT_ID == playerid & AB_FL == TRUE)
player.AB$HIT <- ifelse(player.AB$H_FL > 0, 1, 0)
player.AB$DATE <- substr(player.AB$GAME_ID, 4, 12)
player.AB <- player.AB[order(player.AB$DATE), ]
ST <- replicate(1000, random.mix(player.AB$HIT))
truehist(ST, xlab="Clumpiness Statistic")
stat <- sum((streaks(1 - player.AB$HIT)) ^ 2)
abline(v=stat, lwd=3)
text(stat * 1.05, 0.0010, "OBSERVED", cex=1.2)}
To investigate the non-randomness of Raul Ibanez’s sequence of hit/out

data, the function clump.test is run using Ibanez’s player id cole “ibanr001”
and the resulting histogram display is shown in Figure 10.4.
clump.test("ibanr001", data2011)
Note that Ibanez’s clumpiness measure is in the right tail of this distribution,
indicating that Ibanez clearly displayed more streakiness than one would ex-
pect by chance.
10.4 Local Patterns of Weighted On-Base Average

In our discussion of hitting slumps and streaks, the focus is on either getting a
hit or an out in an official at-bat. A successful plate appearance goes beyond
simply getting a hit or an out, and we would like to explore patterns of slumps
and streaks using a better measure of hitting performance.
One popular measure of hitting performance, the weighted on-base per-
centage or wOBA, is based on giving weights to each positive outcome of a
plate appearance where the weights are proportional to the run values of the
outcomes. This measure is obtained by summing these weights and dividing
the sum by the number of plate appearances.
In one definition, the non-intentional walks (N IBB), hit-by-pitches
(HBP ), singles (1B), reached base by error (RBOE), doubles (2B), triples
(3B), and home runs (HR) are given the respective weights 0.72, 0.75, 0.90,
0.92, 1.24, 1.56, and 1.95, and wOBA is given by the sum of the weights
0.0015
0.0010
OBSERVED
0.0005
0.0000
2500 3000 3500 4000

Clumpiness Statistic
FIGURE 10.4
Histogram of one thousand values of the clumpiness statistic assuming all
arrangements of hits and outs for Raul Ibanez are equally likely. The observed
value of the clumpiness statistic for Ibanez is shown using a vertical line.
divided by the plate appearances (P A). A new function sum.woba is writ-

ten that computes the sum of plate appearances and sum of wOBA weights
for a vector event of codes of plate appearance events. This function uses
the same codes used in the Retrosheet variable “EVENT CD”. (The code 14
corresponds to a walk, the code 16 corresponds to a hit-by-pitch, and so on.)
sum.woba <- function(event){
s <- 0.72 * (event == 14) + 0.75 * (event == 16) +
0.90 * (event == 20) + 0.92 * (event == 18) +
1.24 * (event == 21) + 1.56 * (event == 22) +
1.95 * (event == 23)
c(sum(s), length(s))
}
Since we wish to focus our exploration to plate appearances, the subset

function is used to create a new data frame data2011a containing only of
batting events (BAT EVENT FL == TRUE). The aggregate function is used to-
gether with the sum.woba function to find the sum of weights and number of
PAs for each player for each game in the 2011 season. Looking at a few lines
of the new data frame S, we see Bobby Abreu (abbreviation “abreb001”) in
the first game had a total weight of 1.44 in 4 plate appearances – the value of
wOBA for this game would be 1.44/4 = 0.360.
data2011a <- subset(data2011, BAT_EVENT_FL == TRUE)
S <- with(data2011a,
aggregate(EVENT_CD, list(Player=BAT_ID, Game=GAME_ID), sum.woba))
head(S)
Player Game x.1 x.2
1 abreb001 ANA201104080 1.44 4.00
2 bautj002 ANA201104080 0.90 4.00
3 bourp001 ANA201104080 0.00 4.00
4 calla001 ANA201104080 0.72 4.00
5 congh001 ANA201104080 0.90 4.00
6 davir003 ANA201104080 0.00 4.00
Say we are interested in looking at a player’s wOBA over groups of five

games – games 1-5, games 6-10, games 11-15, and so on. A player’s batting
performance for all games in a season is represented as a matrix, where rows
correspond to games, the first column corresponds to the wOBA weights for
all games, and the second column corresponds to the game PAs. The new
function regroup collapses a player’s batting performance matrix into groups
of size g, where a particular row will correspond to the sum of weights and
sum of PAs in a particular group. (In our exploration, groups of size g = 5
will be used.)
regroup <- function(d, g){
n <- dim(d)[1]
n.g <- floor(n / g); n.r <- n - n.g * g
gp <- rep(1 : n.g, each=g, length.out=n)
if(n.r > 0) {
i <- (n.g * g + 1) : n
gp[i] <- rep(n.g, length(i))
}
aggregate(d, list(gp), sum)[, -1]
}
To illustrate this grouping operation, the game-to-game hitting data for

Suzuki is collected in the data frame suzuki. As before, to make sure the
data is chronologically ordered, a new variable Date is created and the rows
are ordered by increasing values of Date. Showing the first few rows of suzuki,
we see that the variable x (labeled by x.1 and x.2) is the matrix of hitting
data. The regroup function is applied on the matrix suzuki$x. The output
is a matrix with two columns: V1 contains the sum of weights for each group
of five games and V2 is the number of PAs in each group. (Not all rows of the
matrix are shown.)
suzuki <- subset(S, Player == "suzui001")
suzuki$Date <- with(suzuki, substr(Game, 4, 12))
suzuki <- suzuki[order(suzuki$Date), ]
head(suzuki)
Player Game x.1 x.2 Date
32293 suzui001 OAK201104010 3.44 5.00 201104010
32314 suzui001 OAK201104020 1.80 5.00 201104020
32335 suzui001 OAK201104030 1.62 4.00 201104030
46061 suzui001 TEX201104040 1.24 5.00 201104040
46079 suzui001 TEX201104050 0.92 4.00 201104050
46098 suzui001 TEX201104060 1.44 4.00 201104060
regroup(suzuki$x, 5)
V1 V2
1 9.02 23
2 6.66 23
3 5.18 23
4 7.36 22
...
30 4.32 23
31 8.56 23
32 5.22 25
The process of finding the five-game hitting data has been illustrated for
Suzuki. When we look at the sequence of five-game wOBAs for an arbitrary
player, the wOBAs for a consistent player will have small variation, and the
wOBA values for a streaky player will have high variability. A common mea-
sure of variability is the standard deviation, the average size of the deviations
from the mean.
A new function is written to compute the mean and standard deviation of
the group wOBAs for a given player. This function get.streak.data performs
this operation for a given player with id code playerid, the weight matrix
for all plays S, and a grouping of g games (by default g = 5). The output is
a vector with the sum of plate appearances N, the mean of the group wOBAs
Mean and the standard deviation of the group wOBAs SD.
get.streak.data <- function(playerid, S, g=5){
S.player <- subset(S, Player == playerid)
S.player$Date <- with(S.player, substr(Game, 4, 12))
S.player <- S.player[order(S.player$Date), ]
S.player.gp <- regroup(S.player$x, g)
s.woba.avg <- with(S.player.gp, V1 / V2)
c(N=sum(S.player.gp$V2),
Mean=mean(s.woba.avg), SD=sd(s.woba.avg))
}
To illustrate the use of this function, we apply it to Suzuki’s hitting data.

get.streak.data("suzui001", S, 5)
N Mean SD
721.00000000 0.28498259 0.09212598
Suzuki had 721 plate appearances, the mean of his five-game wOBAs was
0.285 and the standard deviation of his five-game wOBAs was 0.092.
The function get.streak.data is applied to all players in the 2011 season.
The vector player.list is defined to be the vector of all unique player ids
and the sapply function is used to apply get.streak.data to all players in
player.list. We want to focus on “regular” players and the subset function
is used to collect the streakiness data only for players where the number of
plate appearances (variable N) is 500 or greater.
player.list <- unique(S$Player)
Results <- data.frame(Player=player.list,
t(sapply(unique(S$Player),
get.streak.data, S)))
Results.500 <- subset(Results, N >= 500)
A scatterplot is constructed of the mean and standard deviation of wOBAs

of all regular players in Figure 10.5. Using the identify function, we locate
the points corresponding to the largest and smallest standard deviations and
place the last names of these players on the graph (using information from
the Master data frame).
Master <- read.csv("Master.csv")
with(Results.500, plot(Mean, SD))
ids <- with(Results.500, identify(Mean, SD, n=2, Player, plot=FALSE))
pids <- as.character(Results.500[ids, "Player"])
with(Results.500, text(Mean[ids], SD[ids],
subset(Master, retroID %in% pids)$nameLast, pos=2))
The streakiest hitter during the 2011 season using this standard deviation
measure was Justin Upton. Likewise, the most consistent player, Ichiro Suzuki,
is identified as the one with the smallest standard deviation of the period
Upton
0.16
0.14
0.12
SD
0.10
0.08
Suzuki
0.30 0.35 0.40

Mean
FIGURE 10.5
Scatterplot of mean and standard deviation of five-game wOBAs for all players
in the 2011 with at least 500 PA. Two players are identified, Suzuki and Upton,
who had small and large standard deviations, respectively.
wOBAs. These two players can be compared graphically by plotting their

five-game wOBA values against the period number.
A new function get.streak.data2 is created to compute the vector of
five-game wOBA values for a particular player. This function is a simple mod-
ification of the function get.streak.data where the last line is replaced by
the line:
data.frame(Period=1:length(s.woba.avg), wOBA=s.woba.avg)
The function returns the period number Period and the corresponding
weighted on-base percentage wOBA.
Using this new function, a data frame d1 is created with Suzuki’s data and
similar data frame d2 is created for Upton’s data and the two data frames
are merged by use of the rbind function. The graphics functions ggplot,
geom line, and facet grid in the ggplot2 package are used to create the
line graphs. One nice feature of ggplot2 graphics is that it automatically
uses the same vertical scale for the two panels and shows the player names on
the right of the graph.
d1 <- get.streak.data2("suzuk001", S)
d2 <- get.streak.data2("uptoj001", S)
d <- rbind(data.frame(Player="Suzuki", d1),
data.frame(Player="Upton", d2))
library(ggplot2)
ggplot(d, aes(Period, wOBA)) +
geom_line(size=1) + facet_grid(Player ~ . )
Note that, as expected, Suzuki and Upton have dramatically different patterns
of five-game wOBAs. Most of Suzuki’s five-game wOBAs fall between 0.200
and 0.400. In contrast, Upton had a change in wOBA of 0.000 to 0.740 in two
periods; he was a remarkably streaky hitter during the 2011 season.

There is much interest in streaky performances of baseball players in the
literature. Gould (1989), Berry (1990), and Seidel (2002) discuss the signif-
icance of DiMaggio’s hitting streak in the 1941 season. Albert and Bennett
(2003), Chapter 5, describe the difference between observed streakiness and
true streakiness and give an overview of different ways of detecting streakiness
of hitters. Albert (2008) and McCotter (2009) discuss the use of randomiza-
tion methods to detect if there is more streakiness in hitting data than one
would expect by chance.
0.6
0.4
Suzuki
0.2
0.0
wOBA
0.6
0.4
Upton
0.2
0.0
0 10 20 30
Period
FIGURE 10.6
Line plots of five-game wOBAs against period number for Ichiro Suzuki and
Justin Upton for the 2011 season. Suzuki had a very consistent pattern of
wOBAs, while Upton’s pattern of wOBAs is very volatile.
10.6 Exercises
1. (Ted Williams)
The data file “williams.1941.csv” contains Ted Williams game-to-game
hitting data for the 1941 season. This season was notable in that Williams
had a season batting average of .406 (the most recent season batting av-
erage exceeding .400). Read this dataset into R.
(a) Using the R function streaks, find the lengths of all of Williams’
hitting streaks during this season. Compare the lengths of this hitting
streaks with those of Joe DiMaggio during this same season.
(b) Use the function streaks to find the lengths of all hitless streaks of
Williams during the 1941 season. Compare these lengths with those
of DiMaggio during the 1941 season.
2. (Ted Williams, Continued)
(a) Use the R function moving.average to find the moving batting av-
erages of Williams for the 1941 season using a window of 5 games.
Graph these moving averages and describe any hot and cold patterns
in Williams hitting during this season.
(b) Compute and graph moving batting averages of Williams using sev-
eral alternative choices for the window of games.
3. (Streakiness of the 2008 Lance Berkman)
Lance Berkman had a remarkable hot period of hitting during the 2008
season.
(a) Download the Retrosheet play-by-play data for the 2008 season, and
extract the hitting data for Berkman.
(b) Using the function streaks, find the lengths of all hitting streaks of
Berkman. What was the length of his longest streak of consecutive
hits?
(c) Use the streaks function to find the lengths of all streaks of consec-
utive outs. What was Berkman’s longest “oh-for” during this season?
(d) Construct a moving batting average plot using a window of 20 at-
bats. Comment on the patterns in this graph – was there a period
when Berkman was unusually hot?
4. (Streakiness of the 2008 Lance Berkman, Continued)
(a) Use the method described in Section 10.3.4 to see if Berkman’s
streaky patterns of hits and outs are consistent with patterns from a
random model.
(b) The method of Section 10.3.4 used the sum of squares of the gaps as
a measure of streakiness. Suppose one uses the longest streak of con-
secutive outs as an alternative measure. Rerun the method with this
new measure and see if Berkman’s longest streak of outs is consistent
with the random model.
5. (Streakiness of All Players During the 2008 Season)
(a) Using the 2008 Retrosheet play-by-play data, extract the hitting data
for all players with at least 400 at-bats.
(b) For each player, find the length of the longest streak of consecutive
outs. Find the hitters with the longest streaks and the hitters with
shortest streaks. How does Berkman’s longest “oh-for” compare in
the group of longest streaks?
6. (Streakiness of All Players During the 2008 Season, Continued)

(a) For each player and each game during the 2008 season, compute the
sum of wOBA weights and the number of plate appearances P A.
(See Section 10.4.)
(b) For each player with at least 500 PA, compute the wOBA over groups
of five games (games 1-5, games 6-10, etc.) For each player, find the
standard deviation of these five-game wOBA, and find the ten most
streaky players using this measure.
7. (The Great Streak)
The Retrosheet website recently added play-by-play data for the 1941
season when Joe DiMaggio achieved his 56-game hitting streak.
(a) Download the 1941 play by play data from the Retrosheet website.
(b) Confirm that DiMaggio had three “0 for 12” streaks during the 1941
season.
(c) Use the method described in Section 10.3.4 to see if DiMaggio’s
streaky patterns of hits and outs in individual at-bats are consistent
with patterns from a random model.
(d) DiMaggio is perceived to be very streaky due to his game-to-game
hitting accomplishment during the 1941 season. Based on your work,
is DiMaggio’s pattern of hitting also very streaky on individual at-
bats?
11
Learning About Park Effects by Database
Management Tools
CONTENTS
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
11.2 Installing MySQL and Creating a Database . . . . . . . . . . . . . . . . . . . . . 260
11.3 Connecting R to MySQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
11.3.1 Connecting using package RMySQL . . . . . . . . . . . . . . . . . . . . . . . 262
11.3.2 Connecting using Package RODBC . . . . . . . . . . . . . . . . . . . . . . . . 263
11.4 Filling a MySQL Game Log Database from R . . . . . . . . . . . . . . . . . . 264
11.4.1 From Retrosheet to R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
11.4.2 From R to MySQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
11.5 Querying Data from R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
11.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
11.5.2 Coors Field and run scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
11.6 Baseball Data as MySQL Dumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
11.6.1 Lahman’s database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
11.6.2 Retrosheet database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
11.6.3 PITCHf/x database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
11.7 Calculating Basic Park Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
11.7.1 Loading the data in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
11.7.2 Home run park factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
11.7.3 Assumptions of the proposed approach . . . . . . . . . . . . . . . . . 277
11.7.4 Applying park factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
11.8 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
11.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
11.1 Introduction
In this book, analyses were performed entirely from baseball datasets loaded
into R. That was possible because we were dealing with datasets with a rel-
atively small number of rows. However, when one wants to work on multiple
seasons of play-by-play (or pitch-by-pitch) data, it become more difficult to
259
manage all of the data inside R.1 While Retrosheet gamelogs consist of ap-
proximately 250,000 records, there are approximately 10 million Retrosheet
play-by-play events, and MLBAM provides data on roughly 800,000 pitches
per year for MLB games.
A solution to this “big data” problem is to store the data using a Database
Management System (DBMS) and, by communicating the system with R,
access only the data needed for the particular analysis. In this chapter some
guidance is provided on this task. Our choice for the DBMS is MySQL, likely
the most popular open-source DBMS. However, readers familiar with other
software can find similar solutions for their DBMS of choice.
The use of MySQL and the R interface are used to gain some understanding
of park effects in baseball. Unlike most of the other team sports, baseball
ballparks vary greatly in shape and dimensions. The left-field wall in Fenway
Park, home of the Boston Red Sox, is listed at 310 feet from home plate, while
the left-field fence in Wrigley Field (home of the Chicago Cubs) is 355 feet
away. The left-field wall in Boston, commonly known as The Green Monster,
is 37 feet high, while the left-field fence Dodger Stadium in Los Angeles is
only four feet high. Such differences in ballpark shapes and dimensions and
the prevalent weather conditions have a profound effect on the game and the
associated player measures of performance.
We first show how to obtain and set up MySQL, and then illustrate con-
necting R to a MySQL database for the purpose of retrieving data in R and
appending data to MySQL tables. This interface is used to present evidence
of the effect of Coors Field (home of the Colorado Rockies) on run scoring.
The reader is directed to online resources providing baseball data (of the sea-
sonal to pitch-by-pitch type) ready for import into MySQL. The chapter is
concluded by providing the readers with a basic approach for calculating park
factors and using these factors to make suitable adjustments to players’ stats.
11.2 Installing MySQL and Creating a Database

Since this is a book focusing on R, we emphasize the use of MySQL to-
gether with R. A user can install MySQL on its own from www.mysql.com/
downloads. Alternatively, one can obtain this database software by the in-
stallation of XAMPP, a distribution containing an Apache Web server, plus
MySQL, PHP, and Perl, freely available for Linux, Windows, Mac OS X, and
Solaris. On the XAMPP Web page2 one finds information on how to install
the distribution for the desired platform. Below we show examples related to
1 R by default reads data into memory (RAM) thus imposing limits on the size of datasets
it can read.
2 www.apachefriends.org/en/xampp.html
Learning About Park Effects by Database Management Tools 261
the Windows distribution, but readers using other operating systems should
find a similar installation procedure.
If one has not modified the defaults during the installation, XAMPP can
be launched from the Programs option under the Start Menu. In the box
that appears (Figure 11.1) one clicks on the Start button. beside Apache and
MySQL to start those services. By clicking on Admin... on the right side of
MySQL the default browser is opened and one is taken to the localhost/
phpmyadmin page (Figure 11.2).
FIGURE 11.1
XAMPP Control Panel Application.
Let’s now create a new database named RBaseball. First, click on the
Databases tab, then type RBaseball on the text box under Create new
database and click the Create button. After a few seconds one should see
the RBaseball database listed on the left frame.3 If one clicks on the newly
created database, a message will appear saying there are no tables in the
database.
3 If that is not the case, simply clicking the Reload navigation frame green arrow will
make it appear.
FIGURE 11.2
Browser window displaying the localhost/phpmyadmin/ page.
11.3 Connecting R to MySQL

The RMySQL package provides R users with functions to connect to a MySQL
database. Unfortunately the installation of such a package is usually not as
straightforward as the typical installation of an R package (as shown in Chap-
ter 2), especially under Windows. For this reason we provide an alternate way
to connect to MySQL, by using the RODBC package.
Readers who plan to make extensive use of MySQL connections are en-
couraged to make the necessary efforts for installing RMySQL,4 since that will
have a much better performance. However, in the sections that follow, we
provide alternate code for those working with RODBC.
11.3.1 Connecting using package RMySQL

Once RMySQL is installed, the R function dbConnect creates the connection.
The user and pwd arguments indicate the user name and the password for
4 Some details on how to get RMySQL to work under Windows can be found at tinyurl.
com/RMySQL1 and tinyurl.com/RMySQL2.

accessing the MySQL database (if they have been specified when the database
was created), while dbname indicates the default database to which R will be
connected (in our case RBaseball, created in Section 11.2). Note that the
connection is assigned to an R object (conn here), as it will be required as an
argument by several functions.
library(RMySQL)
conn <- dbConnect(MySQL(), user=’username’, pwd=’password’
, dbname=’RBaseball’)
To remove a connection one uses dbDisconnect(conn=conn), where the ar-

gument conn requires an open connection.
11.3.2 Connecting using Package RODBC

The package RODBC provides an alternate way for connecting to a MySQL
database. By making use of Open Database Connectivity (ODBC), RODBC
provides a single solution to access databases from several DBMS.5
The first step is to set up an ODBC connection. ODBC is DBMS-
independent because it uses drivers as “translation layers” between the appli-
cation and the DBMS. To enable ODBC connection to MySQL (or any other
DBMS) an appropriate driver is required. On the MySQL website the nec-
essary driver can be downloaded at dev.mysql.com/downloads/connector/
odbc.
Once the appropriate driver has been downloaded and installed, the sub-
sequent required step is the configuration of a connector. The MySQL website
provides detailed instructions on how to perform the task according to one’s
operating system: the relevant URL is dev.mysql.com/doc/refman/5.0/en/
connector-odbc-configuration.html.
Here we illustrate the process for the configuration on a Windows XP
machine, creating a connector to the RBaseball database created in Section
11.2.
1. From the Windows Start menu click on the Control Panel.

2. If the Control Panel is displayed in Classic View, click on Administra-
tive Tools. If the Control Panel is displayed in Category View click on
Performance and Maintenance and then on Administrative Tools.
3. Click on Data Sources (ODBC).
4. In the ODBC Data Source Administrator box, click on Add... to add a
new data source.
5. In the Create New Data Source box select the MySQL ODBC driver.
5 A number of packages similar to RMySQL exist for connections to other DBMS systems:
RPostgreSQL, ROracle, and RSQLite are a few of them.

6. Set the parameters in the MySQL Connector/ODBC Data Source Config-

uration box as follows:
(a) Enter a Data Source Name (DSN): in our case we choose RBaseball.
(b) Optionally add a Description.
(c) Enter localhost in the Server text box and leave the default Port
value.
(d) Enter a User and a Password for the connection.
(e) From the Database popup menu, select the RBaseball database.
7. Click OK and the DSN is saved.6
Once the ODBC connector has been configured, the database can be ac-
cessed from R by use of the function odbcConnect. The argument dsn is the
name of the data source used when creating the ODBC connection (point
6a in the previous subsection), while uid and pwd need to be provided if a
user name and a password have been specified (6d) when setting the ODBC
connection.
library(RODBC)
conn <- odbcConnect(dsn="RBaseball", uid="user"
, pwd="password")
To close the connection in R, one types

odbcClose(channel=conn)
11.4 Filling a MySQL Game Log Database from R

The game log data files are currently available at the Retrosheet Web page
www.retrosheet.org/gamelogs/index.html. By clicking on a single year one
obtains a compressed (zip) file containing a single text file of the season’s
game logs. Here we first create a function for loading a season of game logs
into R, then show how to append the data into a MySQL table. Finally code
is provided which loops through several seasons, downloading the game logs
from Retrosheet and appending them to the MySQL table.
6 The Test button can be used to make sure the connection has been successfully config-
ured.
11.4.1 From Retrosheet to R

The load.gamelog function, shown below, inputs the season number and
performs the following operations:
1. Downloads the zip file of the season from the Retrosheet game log Web
page.
2. Extracts the text file contained in the downloaded zip file.
3. Reads the text file into R.
4. Removes both the compressed and the extracted files.
load.gamelog <- function(season){

download.file(
url=paste("http://www.retrosheet.org/gamelogs/gl", season
, ".zip", sep="")
, destfile=paste("gl", season, ".zip", sep="")
)
unzip(paste("gl", season, ".zip", sep=""))
gamelog <- read.table(paste("gl", season, ".txt", sep="")
, sep=",", stringsAsFactors=F)
file.remove(paste("gl", season, ".zip", sep=""))
file.remove(paste("gl", season, ".txt", sep=""))
gamelog
}
After the function has been read into R, one season of game logs (for
example, the year 2012) is inputted into R by typing the command:
gl2012 <- load.gamelog(2012)
Since game log files downloaded from Retrosheet do not have column head-
ers, the resulting gl2012 data frame has column names assigned by de-
fault by R. They can be replaced with meaningful names stored in the
game log header.csv file as done elsewhere in this book.
glheaders <- read.csv("retrosheet/game_log_header.csv")
names(gl2012) <- names(glheaders)
11.4.2 From R to MySQL

Using a ODBC connection, the gl2012 data frame currently loaded in R can
be transferred to the RBaseball MySQL database. In the lines that follow,
the RODBC package is loaded, a connection is set to the RBaseball database,
and finally the sqlSave function is used to append the data to a table in the
MySQL database.
library(RODBC)
conn <- odbcConnect("RBaseball")
sqlSave(channel=conn, dat=gl2012, tablename="gamelogs"
, append=TRUE, rownames=FALSE)
Here are some notes on the arguments in the sqlSave function:
• The channel argument requires an open connection; here the one that
was previously set (conn) is specified.
• The dat argument requires the name of the R data frame to be appended
to the table in the MySQL database.
• The tablename argument requires a string indicating the name of the
table (in the database) where the data are to be appended.
• Setting append to TRUE indicates that, should a table by the name
"gamelogs" already exist, data from gl2012 will be appended to the ta-
ble. If append is set to FALSE the table "gamelogs" (if it exists) will be
overwritten.
• The rownames=FALSE argument indicates the row names will not be
appended.
The R code using RMySQL is similar to the RODBC code, but requires signif-
icantly less time for appending data. The following code is preferable if one
has successfully installed the RMySQL package.
library(RMySQL)
conn <- dbConnect(MySQL(), user=’username’
dbWriteTable(conn, name="gamelogs", value=gl2012
, append=TRUE, row.names=F)
In the previous section, code was provided for appending one season of
game logs into a MySQL table. However, we have demonstrated in previous
chapters that it is straightforward to use R to work with a single season of
game logs. To fully appreciate the advantages of storing data in a DBMS, a
MySQL table will be populated with game logs going back through baseball
history. With a historical database and an R connection, we demonstrate the
use of R to perform analysis over multiple seasons.
A new function appendGameLogs is written which loops through a specified
set of years (potentially from 1871 to the present), downloads the files from
Retrosheet,7 and appends the data to the gamelogs table in the RBaseball
7 The downloading of data from Retrosheet is performed by the previously presented
load.gamelog function, thus the reader has to make sure said function is loaded for the
code in this section to work.
MySQL database.8 The whole process may take several minutes. If one is not
interested in downloading files dating back to 1871, seasons from 1995 are
sufficient for reproducing the example of the next section.
The function appendGameLogs takes the following parameters as inputs:
• start and end indicate the seasons one wants to download from Ret-
rosheet and append to the MySQL database. By default the function will
work on seasons from 1871 to 2012.
• connPackage provides the user with the option of selecting whether to
use the RODBC or the RMySQL package for performing the work. Note that
RMySQL works faster and is the preferred choice.
• headersFile points to the full path where the file containing the game
log headers is stored.
• dbTableName specifies the name of the table in the MySQL database
where the data are to be uploaded.
• ther parameter inputs as required by either the odbcConnect or the

dbConnect functions, such as the DSN name, the user name, and the
password to access the database.9
appendGameLogs <- function(

start=1871, end=2012, connPackage="RODBC"
, headersFile="retrosheet/game_log_header.csv"
, dbTableName="gamelogs", ...
){
require(package=connPackage, character.only=TRUE)
glheaders <- read.csv(headersFile)
if(connPackage == "RMySQL"){
conn <- dbConnect(MySQL(), ...)
} else {
conn <- odbcConnect(...)
}
for(season in start:end){
print(paste(Sys.time(), "working on season:", season))
flush.console()
gamelogs <- load.gamelog(season)
glheaders <- read.csv(headersFile)
names(gamelogs) <- names(glheaders)
gamelogs$GAME_ID <- paste(gamelogs$HomeTeam, gamelogs$Date
8 The proposed code actually allows the reader to specify both a database and a table of
choice.
9 The three-dots (...) construct is used here for allowing the user to specify additional
arguments to the appendGameLogs function as needed by the functions called inside it (either
dbConnect or odbcConnect).
, gamelogs$DoubleHeader, sep="")
gamelogs$YEAR_ID <- substr(gamelogs$Date, 1, 4)
if(connPackage == "RMySQL"){
dbWriteTable(conn, name=dbTableName, value=gamelogs
, append=TRUE, row.names=F)
} else {
sqlSave(conn, dat=gamelogs, tablename=dbTableName
, append=TRUE, rownames=FALSE)
}
}
}
The following lines of code demonstrate the use of the function

appendGameLogs both with RODBC and RMySQL.
appendGameLogs(start=1871, end=1872, connPackage="RODBC"
, dsn="rbaseball", uid="user", pwd="password")
appendGameLogs(start=1873, end=1874, connPackage="RMySQL"

, dbname="rbaseball", user="user", password="password")
11.5 Querying Data from R

11.5.1 Introduction
Both RMySQL and RODBC provide functions for querying data stored in a
MySQL database. The sqlFetch function from RODBC, for example, allows
the import of a full MySQL table into an R data frame. However, the purpose
of having data stored in a MySQL database is to selectively import data into
R for particular analysis; one selectively imports data by querying one or more
tables of the database.
For example, suppose one is interested in comparing the attendance of the
two Chicago teams by day of week since the 2000 season. The following code
retrieves the raw data in R.
library(RODBC)
chiAttendance <- sqlQuery(conn, "
select date, hometeam, dayofweek
, attendence
from gamelogs
where date > 20000101
and hometeam in (’CHN’, ’CHA’)
")
head(chiAttendance)
hometeam dayofweek attendance
1 CHN Thu 55000
2 CHN Mon 38655
3 CHN Wed 26838
4 CHN Thu 20152
5 CHN Fri 21324
6 CHA Fri 38912
The sqlQuery function provides the database query. Its arguments in this
function are the connection handle established in the previous line (conn)
and a string consisting of a valid SQL statement. Readers familiar with SQL
will have no problem in understanding the meaning of the query. For those
unfamiliar with SQL, we present here a brief explanation of the purpose of
the query, inviting anyone who is interested in learning about the language to
look for the numerous resources devoted to the subject.
The first row in the SQL statement indicates the columns of the table that
are to be select-ed (in this case date, hometeam, dayofweek, and attendance).
The second line states from which table they have to be retrieved (gamelogs).
Finally, the where clause specifies conditions for the rows that are to be re-
trieved: the date has to be greater than 20000101 and the value of hometeam
has to be one of CHN and CHA.
The same task can be performed with the following code if one wants to
use RMySQL rather than RODBC.
library(RMySQL)
chiAttendance <- dbGetQuery(conn, "
select date, hometeam, dayofweek
, attendence
from gamelogs
where date > 20000101
and hometeam in (’CHN’, ’CHA’)
").
The comparison between attendance at the two Chicago ballparks is dis-

played in Figure 11.3. The plot has been obtained by first setting as NA the
games reporting attendance values of zero,10 then transforming the dayofweek
column into a factor ordered from Sun to Sat and calculating the average
attendance by day of week. Finally the xyplot function from the lattice
package has been used to actually draw the plot.
10 In case of single admission doubleheaders (i.e., when two games are played on the same
day and a single ticket is required for attending both) the attendance is reported only for
the second game, while it is set at zero for the first.
chiAttendance$attendence <- ifelse(chiAttendance$attendence == 0

, NA, chiAttendance$attendence)
chiAttendance$dayofweek <- factor(chiAttendance$dayofweek
, levels=c("Sun", "Mon", "Tue", "Wed", "Thu", "Fri", "Sat"))
avgAtt <- aggregate(attendence ~ hometeam + dayofweek
, data=chiAttendance, FUN=mean)
library(lattice)
xyplot(attendence ~ dayofweek, data=avgAtt
, groups=hometeam
, pch=c("S", "C"), cex=2, col= "black"
, xlab="day of week"
, ylab="attendance"
)
40000
C
C
C
C
C
C C
35000
attendance
S
30000
S
S
S
25000
S
S S
Sun Mon Tue Wed Thu Fri Sat
day of week
FIGURE 11.3
Comparison of attendance by day of the week on games played at home by
the Cubs (C) and the White Sox (S).
11.5.2 Coors Field and run scoring

As an example of accessing multiple years of data, we explore the effect of
Coors Field (home of the Colorado Rockies in Denver) on run scoring through
the years. Coors Field is a peculiar ballpark because it is located at an alti-
tude of about one mile over the sea level. Coors Field has thinner air than
other stadiums, allowing both batted balls to travel farther and making the
curveballs delivered by pitchers “flatter.”
As a first step, we use the function dbConnect to connect R to the MySQL
database containing the game logs. In the use of this function, remember to
adjust the argument values wherever necessary.
library(RMySQL)
, dbname=’retrosheet’, host=’localhost’)
Then, by using SQL language in the function dbGetQuery,11 data is re-

trieved for the games played by the Rockies, either at home or on the road,
since 1995, the year they moved to Coors Field.
rockies.games <- dbGetQuery(conn, "select year_id, date, parkid
, visitingteam, hometeam
, visitorrunsscored as awR
, homerunsscore as hmR
from gamelogs
where (hometeam=’COL’
or visitingteam=’COL’)
and year_id > 1994")
The game data is conveniently stored in the rockies.games data frame and
can be explored with R commands. The sum of runs scored in each game is
computed by adding the runs scored by the home team and the visiting team.
A new column coors is added indicating whether the game was played at
Coors Field.12
rockies.games$runs <- rockies.games$awR + rockies.games$hmR
rockies.games$coors <- (rockies.games$parkid == "DEN02")
The offensive output by the Rockies and their opponents is compared at

Coors and other ballparks by a graph constructed using the ggplot2 package
introduced in Chapter 6. The resulting visualization is displayed in Figure
11.4.
library(ggplot2)
ggplot(aes(x=year_id, y=runs, linetype=coors), data=rockies.games) +
11 The keyword as in SQL has the purpose of assigning different names to columns. Thus
visitorrunsscored as awR tells SQL that, in the results returned by the query, the column
visitorrunsscored will be named awR.
12 Retrosheet code for Coors Field is DEN02. A list of all ballparks codes is available at
www.retrosheet.org/parkcode.txt.
stat_summary(fun.data="mean_cl_boot") +
xlab("season") +
ylab("runs per game (both teams combined)") +
scale_linetype_discrete(name="location"
, labels=c("other parks", "Coors Field"))
The stats summary layer is used in ggplot2 to summarize the y values at
every unique value of x. The fun.data argument lets the user specify a sum-
marizing function; in this case "mean cl boot" implements a nonparametric
bootstrap procedure for obtaining confidence bands for the population mean.
The output resulting from this layer are the vertical bars appearing for each
data point. The scale linetype discrete layer is used for labeling the series
(name argument) and assigning a name to the legend (labels).
16
●
●
●
runs per game (both teams combined)
14
●
●
●
● ●
●
● location
12 ●
● other parks
● ● Coors Field
● ● ●
●
●
●
10
●
●
●
● ● ●
●
●
● ●
●
●
●
●
8 ●
● ●
●
1995 2000 2005 2010

season
FIGURE 11.4
Comparison of runs scored by the Rockies and their opponents at Coors Field
and in other ballparks.
From Figure 11.4 one notices how Coors Field has been an offense-friendly
park, boosting run scoring by as high as six runs per game. However the effect
of the Colorado ballpark has somewhat decreased in the new millennium,
displaying minimal differences in the 2006-2008 period. One reason for Coors
becoming less of an extreme park is the installation of a humidor. Since the
2002 season, baseballs have been stored, prior to each game, in a room at
a higher humidity, with the intent of compensating for the unusual natural
atmospheric conditions.13
11.6 Baseball Data as MySQL Dumps

Section 11.4.1 illustrated populating a MySQL database from within R using
an example of creating a table of Retrosheet game logs. Several so-called SQL
dumps are available online for creating and filling databases with baseball
data. SQL dumps are simple text files (featuring a .sql extension) containing
SQL instructions for creating and filling SQL tables.
11.6.1 Lahman’s database

Sean Lahman provides his historical database of seasonal stats in several for-
mats. So far in this book we have used the comma-delimited version, con-
sisting of several .csv files. However, the database is also available as a SQL
dump, which can be downloaded from seanlahman.com/baseball-archive/
statistics/ (look for the SQL version). The downloaded file is a zip archive,
which needs to be decompressed. The resulting extracted files include one file
dedicated to instructions (readme 2012.txt in the version downloaded at the
time of this writing) and the dump file (lahman2012.sql.)
There are several options for having the dump file perform its task, includ-
ing typing commands in the OS shell or importing the file via the phpAdmin
panel. It is beyond the scope of this book to illustrate said processes, thus
here we provide a command to obtain the desired result from within R by use
of the shell function which invokes the Command Prompt (in Windows). In
order for the following code to work, the MySQL service should be running14
(see Section 11.2 and Figure 11.1).
shell("mysql < C:\\Baseball\\Datadumps\\lahman2012.sql -u username
-p password")
Note that the SQL dump file creates a new database named lahman.
13 For a detailed analysis of the humidor effects, see this article by Alan Nathan,
Professor Emeritus of Physics at the University of Illinois at Urbana-Champaign: www.

baseballprospectus.com/article.php?articleid=13057.
14 Also, make sure to change the directory containing the .sql file and the user name and
the password as appropriate.

11.6.2 Retrosheet database

In Appendix A we provide R code to download Retrosheet files and transform
them in formats easily readable by R. By slightly adapting the code pro-
vided in Section 11.4.2, one can also append them to a MySQL database.
Jeff Zimmerman provides SQL dumps of the Retrosheet database at the
Baseball Heat Maps website. At the URL www.baseballheatmaps.com/
retrosheet-database-download/ one has several download options, includ-
ing the full Retrosheet database and downloads by decade.
Once one has downloaded the desired file and extracted the content from
the zip archive, a command similar to the one previously shown for the Lah-
man’s database will create the tables in MySQL. Note that the SQL dump
provided by Zimmerman does not contain code to create (and use) a new
database, thus one has to specify the name of the database where operations
are to be performed in the following code15 we have indicated the RBaseball
database after the password parameter.
shell("mysql < C:\\Baseball\\Datadumps\\2010s_Retrosheet.sql
-u username -p password RBaseball")
Zimmerman’s SQL dump creates a number of lookup tables, useful for decod-
ing information such as the type of batted ball, the base situation, and so on.
Details on coded values are provided in Appendix A.
11.6.3 PITCHf/x database

Zimmerman also provides downloads, updated daily during the season, of
PITCHf/x data at www.baseballheatmaps.com/pitch-fx-download/. Sim-
ilarly to what was shown in the previous section, after downloading the zip
archive and extracting its content to a folder of choice, the following command
in R will provide the reader with an up-to-date MySQL PITCHf/x database.
shell("mysql < C:\\Baseball\\Datadumps\\pbp2.sql

-u username -p password RBaseball")
Zimmerman’s data dump comes with a handful of tables. The bulk of

PITCHf/x data is contained in the pitches table, which features most of the
information the reader was presented with in Chapter 6, when the verlander
and cabrera data frames were introduced. A more detailed explanation on
the columns featured in the pitches table can be found in Appendix B,
particularly in Section B.4.2.
Similarly, some details on the contents of the atbats table are provided
in Sections B.4.1 and B.4.3. Whenever ones needs to combine information
contained in pitches with information contained in atbats, the ab id is the
common column to be used for merging purposes.
15 For the code sample we have downloaded Retrosheet data for the 2010-present period.
The SQL dump also contains a games table, featuring general information
on the games, which can be linked to atbats with the common game id col-
umn. Finally a few reference tables provide details on players, umpires, teams,
and pitch types.
11.7 Calculating Basic Park Factors

Park factors (usually abbreviated as PF) have been used for decades by base-
ball analysts as a tool for mitigating the effect of the ballpark when assessing
the value of players. Park factors have been calculated in several ways and in
this section we illustrate a very basic approach, focusing on year 1996, one of
the most extreme seasons for Coors Field, as displayed in Figure 11.4.
In the explanations that will follow, we presume the reader has Retrosheet
data for the 1990s in his database.16
11.7.1 Loading the data in R

As a first step, we connect to the MySQL database (in this example via an
ODBC connection) and retrieve the desired data: by using SQL language we
select the columns containing the home and away teams and the event code
from the events table, keeping only the rows where the year is 1996 and the
event code corresponds to one indicating a batted ball (see Appendix A). The
results of the query are stored in the hrPF R data frame.
# Connect to MySQL database

options(stringsAsFactors=F)
library(RODBC)
# get data
hrPF <- sqlQuery(conn, "
select away_team_id, home_team_id, event_cd
from events
where year_id=1996
and event_cd in (2, 18, 19, 20, 21, 22, 23)
")
16 Refer to Section 11.6.2 for performing the necessary steps to get the data into a MySQL
database.
11.7.2 Home run park factor

A ballpark can have different effects on the various players performance statis-
tics. The unique configuration of Fenway Park in Boston, for example, en-
hances the likelihood of a batted ball to become a double, especially flyballs
to left, which often carom off the Green Monster. On the other hand, home
runs are rare on the right side of Fenway Park due to the unusually long
distance of the right-field fence from home plate.
In this example we explore the stadium effect on home runs in 1996, cal-
culating park factors for home runs. To begin, a new column event fl is
created which indicates, for every row in the hrPF data frame, the occurrence
of a home run.
hrPF$event_fl <- ifelse(hrPF$event_cd == 23, 1, 0)
Next the frequency of home runs per batted ball is computed for all MLB
teams both at home and on the road. The two resulting data frames are
merged and the first column of the newly obtained data frame evCompare is
renamed to team id.
evAway <- aggregate(event_fl ~ away_team_id, data=hrPF
, FUN=mean)
evHome <- aggregate(event_fl ~ home_team_id, data=hrPF
, FUN=mean)
evCompare <- merge(evAway, evHome, by.x="away_team_id"
, by.y="home_team_id", suffixes=c("_A", "_H"))
names(evCompare)[1] <- "team_id"
head(evCompare)
team_id event_fl_A event_fl_H
1 ANA 0.03808699 0.03053270
2 ARI 0.03013345 0.03935524
3 ATL 0.04524887 0.04268827
4 BAL 0.03729904 0.04293456
5 BOS 0.04663212 0.03374167
6 CHA 0.04045377 0.05190166
Park factors are typically calculated so that the value 100 indicates a neutral
ballpark (one which has no effect on the particular statistic) while values over
100 indicate playing fields that increase the likelihood of the event (home
run in this case) and values under 100 indicate ballparks that decrease the
likelihood of the event.
The 1996 home run park factors are obtained with the first line of the
following code. The resulting data frame evCompare is ordered in descending
order by PF. The head function is used to display the most HR-friendly parks
and the tail function displays the least-friendly home run parks. Coors Field
is at the top of the HR-friendly list, displaying an extreme value of 158 – this
park boosted home run frequency by over 50% in 1996. At the other end of
the spectrum, in 1996 was Dodgers Stadium in Los Angeles, featuring a home
run park factor of 71.
evCompare$PF <- round(100 * evCompare$event_fl_H
/ evCompare$event_fl_A)
evCompare <- evCompare[order(-evCompare$PF),]
head(evCompare)
team_id event_fl_A event_fl_H PF
9 COL 0.03405430 0.05383393 158
4 CAL 0.03870425 0.04831843 125
1 ATL 0.03228680 0.03717311 115
3 BOS 0.03849444 0.04425145 115
10 DET 0.04571550 0.05058280 111
6 CHN 0.03741794 0.04066623 109
tail(evCompare)
team_id event_fl_A event_fl_H PF
13 KCA 0.03380102 0.02905167 86
8 CLE 0.04402386 0.03722397 85
5 CHA 0.04240652 0.03486662 82
12 HOU 0.03441600 0.02723649 79
19 NYN 0.03634809 0.02888087 79
14 LAN 0.03595703 0.02561144 71
11.7.3 Assumptions of the proposed approach

The proposed approach to calculating park factors makes several simplifying
assumption. The first assumption is that the home team and the ballpark
are considered interchangeable. While that is true for most games, sometimes
alternate ballparks have been used for particular games. For example, during
the present season, the Oakland Athletics played their first six home games in
Cashman Field (Las Vegas, NV) while renovations at the Oakland-Alameda
County Coliseum were being completed, and the San Diego Padres had a series
of three games against the New York Mets at Estadio de Beisbol Monterrey
in Mexico as a marketing move by MLB.17
Another assumption of the proposed approach is that a single park factor
is appropriate for all players, without considering how ballparks might affect
some categories of players differently. Asymmetric outfield configurations, in
fact, cause playing fields to have unequal effects on right-handed and left-
handed players. For example, the aforementioned Green Monster in Boston,
being situated in left field, comes into play more frequently when right-handed
batters are at the plate; and the recent new version of the Yankee Stadium
17 A list of games played in alternate sites is displayed on the Retrosheet website at the
url www.retrosheet.org/neutral.htm.
has seen left-handed batters take advantage of the short distance of the right
field fence.
Finally, the proposed park factors (as well as most published versions of
park factors) essentially ignore the players involved in each event (in this case
the batter and the pitcher). As teams rely more on the analysis of play-by-play
data, they typically adapt their strategies to accommodate the peculiarities
of ballparks. For example, while the diminished effect of Coors Field on run
scoring displayed in Figure 11.4 is mostly attributable to the humidor, part of
the effect is certainly due to teams employing different strategies when playing
in this park. For example, teams can use pitchers who induce a high number
of groundballs that are less impacted by the effect of the rarified air.
11.7.4 Applying park factors

In the 1996 season, four Rockies players hit 30 or more home runs – Andres
Galarraga led the team with 47, followed by Vinnie Castilla and Ellis Burks
(tied at 40) and Dante Bichette (31). Behind them Larry Walker had just
18 home runs, but with a very limited playing time due to injuries. In fact,
Walker’s HR/AB ratio was second to only Galarraga’s. Their offensive output
was certainly boosted by playing 81 of their games in Coors Field. Using the
previously calculated park factors, one can estimate the number of home runs
Galarraga would have hit in a neutral park environment.
We begin with retrieving from the MySQL database every Galarraga’s
1996 plate appearance ending with a batted ball, and a column event fl is
defined indicating a value of 1 for home runs and 0 otherwise, as was done in
Section 11.7.2.
# Connect to MySQL database
options(stringsAsFactors=F)
library(RODBC)
conn <- odbcConnect("rbaseball")
# Get Galarraga’s plate appearances

andres <- sqlQuery(conn, "
select away_team_id, home_team_id, event_cd
from events
where year_id=1996
and event_cd in (2, 18, 19, 20, 21, 22, 23)
and bat_id=’galaa001’
")
# identify HRs
andres$event_fl <- ifelse(andres$event_cd == 23, 1, 0)
Then the previously calculated park factors are added to the andres data
frame. This is done by merging data frames andres and evCompare, using
the columns home team id and team id as the merging columns. Using the
merged data frame adresPF, we calculate the mean park factor for Galarraga’s
plate appearances.
andres <- merge(andres, evCompare[,c("team_id", "PF")]
, by.x="home_team_id"
, by.y="team_id")
andresPF <- mean(andres$PF)
andresPF
[1] 129.1384
The compounded park factor for Galarraga, derived from the 252 batted balls
he had at home and the 225 he had on the road (ranging from 9 in Dodgers
Stadium in Los Angeles to 23 at the Astrodome in Houston), indicates Andres
had his home run frequency increased by an estimated 29% relative to a
neutral environment. In order to get the estimate of home runs in a neutral
environment, we divide Galarraga’s home runs by his average home run park
factor multiplied by 100.
47 / andresPF * 100
[1] 36.39507
According to our estimates, Galarraga’s benefit from the ballparks he played
in (particularly his home Coors Field) amounted to roughly 47 - 36 = 11 home
runs in the 1996 season.

Chapter 2 of Adler (2006) has detailed instructions on how to obtain and
install MySQL, and on how to set up an historical baseball database with
Retrosheet data. Hack #56 (Chapter 5 of the same book) provides SQL code
for computing and applying Park Factors.
MySQL reference manuals are available in several formats at dev.mysql.
com/doc/ on MySQL website. The HTML online version features a search box
which allows users to quickly retrieve pages pertaining to specific functions.
11.9 Exercises
1. (Runs Scored at the Astrodome)
(a) Using either the sqlQuery function from the RODBC package or the
dbGetQuery function from the RMySQL package, select games featuring

the Astros (as either the home or visiting team) during the years when
the Astrodome was their home park (i.e., from 1965 to 1999).
(b) Draw a plot to visually compare through the years the runs scored
(both teams combined) in games played at the Astrodome and in
other ballparks.
2. (Astrodome Home Run Park Factor)
(a) Select data from one season between 1965 and 1999. Keep the
columns indicating the visiting team identifier, the home team iden-
tifier and the event code, and the rows identifying ball-in-play events.
Create a new column which identifies whether an home run has oc-
curred.
(b) Prepare a data frame containing the team identifier in the first col-
umn, the frequency of home runs per batted ball when the teams
plays on the road in the second column, and the same frequency
when the team plays at home in the third column.
(c) Compute home run park factors for all MLB teams and check how
the domed stadium in Houston affected home run hitting.
3. (Applying Park Factors to “Adjust” Numbers)
(a) Using the same season selected for the previous exercise, obtain data
from plate appearances (ending with the ball being hit into play)
featuring one Astros player of choice. The exercise can either be
performed on plate appearances featuring an Astros pitcher on the
mound or an Astros batter at the plate. For example, if the selected
season is 1988, one might be interested in discovering how the As-
trodome affected the number of home runs surrendered by veteran
pitcher Nolan Ryan (Retrosheet id: ryann001) or the number of home
runs hit by rookie catcher Craig Biggio (id: biggc001).
(b) As shown in Section 11.7.4 merge the selected player’s data with the
Park Factors previously calculated and compute the player’s individ-
ual Park Factor (which is affected by the different playing time the
player had in the various ballparks) and use it to estimate a “fair”
number of home runs hit (or surrendered if a pitcher was chosen).
4. (Park Factors for Other Events)
(a) Park Factors can be estimated for events other than Home Runs.
SeamHeads.com Ballpark Database18 for example features Park Fac-
tors for seven different events, plus it offers split factors accord-
ing to batters’ handedness. See for example the page for the As-
trodome: www.seamheads.com/ballparks/ballpark.php?parkID=
HOU02&tab=pf1.
18 www.seamheads.com/ballparks/index.php
(b) Choose an event (even different from the seven shown at SeamHeads)
and calculate how ballparks affect its frequency. As a suggestion, the
reader may want to look at seasons in the ’80s, when artificial turf
was installed in close to 40% of MLB fields, and verify whether parks
with concrete/synthetic grass surfaces featured a higher frequency of
batted balls (home runs excluded) converted into outs.19
19 SeamHeads provides information on the surface of play in each stadium’s page. For
example, in the previously mentioned page relative to the Astrodome, if one hovers the
mouse over the ballpark name in a given season, a pop-up will appear providing information
both on the ballpark cover and its playing field surface. SeamHeads is currently providing
its ballpark database as a zip archive containing comma-separated-value (.csv) files that
can be easily read by R: the link for downloading it is found at the bottom of each page in
the ballpark database section.
12
Exploring Fielding Metrics with Contributed
R Packages
CONTENTS
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
12.2 A Motivating Example: Comparing Fielding Metrics . . . . . . . . . . . 284
12.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
12.2.2 The fielding metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
12.2.3 Reading an Excel spreadsheet (XLConnect) . . . . . . . . . . . . . 286
12.2.4 Summarizing multiple columns (doBy) . . . . . . . . . . . . . . . . . . 287
12.2.5 Finding the most similar string (stringdist) . . . . . . . . . . 288
12.2.6 Applying a function on multiple columns (plyr) . . . . . . . 291
12.2.7 Weighted correlations (weights) . . . . . . . . . . . . . . . . . . . . . . . . 291
12.2.8 Displaying correlation matrices (ellipse) . . . . . . . . . . . . . . 292
12.2.9 Evaluating the fielding metrics (psych) . . . . . . . . . . . . . . . . . 293
12.3 Comparing Two Shortstops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
12.3.1 Reshaping the data (reshape2) . . . . . . . . . . . . . . . . . . . . . . . . . 296
12.3.2 Plotting the data (ggplot2 and directlabels) . . . . . . . . 296
12.4 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
12.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
12.1 Introduction
R packages have been introduced in Chapter 2 of this book. Here the evalua-
tion of fielding is used as a motivating example to further illustrate the capa-
bilities of a number of R packages. To use any of the packages described in this
chapter, the package needs to be installed in R before loading the package with
the library function. One installs a package by use of the install.packages
function or the Install Packages button in RStudio’s Packages tab.
283
12.2 A Motivating Example: Comparing Fielding Met-

rics
12.2.1 Introduction
The evaluation of defensive contribution is one of sabermetrics’ holy grails.
Fielding measures have vastly improved since the days when fielding percent-
age1 was the only available metric for defensive evaluation. In evaluating field-
ing fifty years ago, one had to choose between using a very imperfect tool or
ignoring fielding altogether. Big strides in evaluating defense have been possi-
ble due to the increasing availability of data. Retrosheet has made decades of
play-by-play information accessible to the public, and companies like STATS,
Baseball Info Solutions and Sportvision have been tracking, at increasing lev-
els of detail and precision, the ball trajectories and the positioning of the
fielders.
The quantification of defensive contribution is a subject which could fill the
pages of an entire book. Recently, tax lawyer Michael A. Humphreys published
a book Wizardry - Baseball’s All-Time Greatest Fielders Revealed in which
he provides an overview of the publicly known defensive metrics and proposes
his own measure which allows the assessment of defensive value throughout
the history of baseball.
In this chapter we introduce a handful of defensive metrics with the goal of
using these measures to evaluate the fielding of shortstops during the 2009 sea-
son. Since the ratings for one of the metrics are available as a Microsoft Excel
spreadsheet, we introduce the XLConnect package that provides functions for
working with Excel files within R. A few shortstops played for different teams
during the 2009 season, but we are interested in their defensive ratings for the
full season. The package doBy allows for summarizations of data frames over
multiple columns and this package is helpful for describing defensive ratings
split by teams.
It can be problematic combining data from two different sources with no
common identifier and different spellings of the players’ names. In this work,
it is helpful to use the stringdist package and its functions devised for the
comparison of strings. Then the function colwise function from the plyr
package is shown helpful in applying the same function over multiple columns
of a data frame.
Several useful functions for working with correlations will be demonstrated.
After correlations are computed using the cor function from the stats pack-
age (which comes installed with the basic version of R), a special function from
the weights package computes correlations that allow for the assignment of
1 Fielding percentage is defined as the ratio between a fielder’s successful chances (putouts
+ assists) and his total chances (putouts + assists + errors).

Exploring Fielding Metrics with Contributed R Packages 285
different weights to the observations in the data frame. Correlations can be

attractively displayed by use of functions in the ellipse package.
12.2.2 The fielding metrics

The majority of defensive metrics are derived through basically the same pro-
cess. Each measure compares the number of successful plays of a fielder with
the expected number an average fielder would make with the same opportu-
nities and translates the extra (fewer) plays made into runs saved (cost) to
the team. Given this common framework, defensive metrics differ on the data
used, the computation of the expected number of plays, and the method in
which plays are converted to runs. Six fielding metrics are compared in this
chapter; we provide a brief introduction on the history of each measure and
how it is defined.
• Ultimate Zone Rating (UZR), developed by Mitchel Lichtman, is

based on proprietary data collected by Baseball Info Solutions (BIS). For
each batted ball, the system estimates the probability of getting an out
given several parameters such as the speed and trajectory of the batted
ball, the handedness of the batter, the base/out situation, and the ballpark.
• Defensive Runs Saved (DRS) was introduced in The Fielding Bible

by John Dewan, the owner of Baseball Info Solutions. This system esti-
mates the number of extra (fewer) plays made by a fielder given a set of
parameters defining the batted ball, including the direction, the batted
ball type (groundball, flyball, line drive, pop up), and the type of contact
(soft, medium, hard). DRS uses proprietary BIS data, obtained with the
work of so-called videoscouts–i.e., people entering information in databases
while watching recorded broadcasts of games.
• Total Zone (TZL), developed by Sean Smith, makes use of the freely
available Retrosheet data. It determines the fielder responsible for a batted
ball using (whenever available) the batted ball type and location. Since
that data is not available in every season covered by Retrosheet, Smith
uses information about batter handedness, pitcher’s ground ball/fly ball
tendency, and the fielder ultimately collecting the batted ball as proxies
to designate the responsible fielder.
• Revised Zone Rating (RZR) divides the field of play into several
zones and assigns the responsibility of fielding a batted ball to one fielder
depending on the zone where the ball is hit. In its basic version, Zone
Rating (ZR) is simply the ratio of balls in a fielder’s zone converted into
outs. In its revised form, a bonus is added to account for the plays made
by fielders on balls outside of their zones.
• Fan Scouting Report (FSR) is a subjective defensive evaluation sys-

tem. Its numbers are based on the results of a yearly survey among baseball
fans conducted by analyst and blogger Tom Tango.
• Wizardry ’s Runs (Runs) are taken from Humphreys’ book and are
derived from seasonal data available in the Lahman’s database. One ad-
vantage of this system is that it is possible to calculate fielding ratings
throughout baseball history.
As a rule of thumb for the counting metrics (all of the above, except
RZR), one can identify the players who are credited with saving 15 or more
runs as extremely good ones and, conversely, as awful defenders those who are
estimated costing 15 or more runs to their teams. For RZR, where an average
fielder usually posts a rate around .835, the best players are found at the .940
mark, while the worst ones record values close to .700.
12.2.3 Reading an Excel spreadsheet (XLConnect)

The appendices of Humphreys (2011) are available online on the book’s
website.2 Appendix C of Humphreys (2011) consists of nine Excel spread-
sheets containing historical defensive contributions calculated by the author’s
method. In the examples that follow, we will make use of the shortstops’
spreadsheet measures, thus readers who are following this code on R should
download this spreadsheet and place it in a folder of choice.
A few contributed packages are helpful for reading Excel files into R includ-
ing the XLConnect package. In the following R script, the package is loaded
with the library function, and the loadWorkbook function is used to read in
the Excel file of spreadsheet measures; the workbook in R is named xlwzr. The
readWorksheet function is used to read the first sheet of xlwzr into a data
frame with name wzr, starting from row number seven (with the startRow
argument) to skip the header lines. The final lines of the code use the subset
function to create a new data frame restricting Year to 2009, and a new vari-
able Name is created which combines the first and last names of the players.
library(XLConnect)
xlwzr <- loadWorkbook("Appendix_C_Shortstop.xls")
wzr <- readWorksheet(xlwzr, sheet=1, startRow=7)
wzr <- subset(wzr, Year == 2009)
wzr$Name <- paste(wzr$First, wzr$Last)
head(wzr)
Year L T.mcs First Last Pos IP Runs
236 2009 N FLA Alfredo Amezaga SS 42.0000 1.0780581
275 2009 A BAL Robert Andino SS 478.3333 -0.7244126
284 2009 A TEX Elvis Andrus SS 1238.0000 5.0929674
417 2009 A KC Mike Aviles SS 269.3333 5.0588753
2 www.oup.com/us/companion.websites/9780195397765/appendices/?view=usa.
425 2009 A LA Erick Aybar SS 1189.3333 11.1041201

571 2009 N STL Brian Barden SS 27.0000 0.5520377
v.Tm Name
236 1.514699 Alfredo Amezaga
275 -3.394066 Robert Andino
284 -4.307379 Elvis Andrus
417 3.166190 Mike Aviles
425 7.698325 Erick Aybar
571 0.180907 Brian Barden
The data frame wzr contains, in order, the year, the league, and the team
for which the shortstop played, the player’s first and last name, his position,
the number of innings played, his defensive runs, the difference between his
defensive runs and those of the rest of his team at the same position, and the
newly created variable with the full player’s name.
12.2.4 Summarizing multiple columns (doBy)

Some of the players in the wzr data frame have multiple entries, such as
Orlando Cabrera who played shortstop for multiple teams in 2009.
subset(wzr, Name == "Orlando Cabrera")
Year L T.mcs First Last Pos IP Runs
1650 2009 A OAK Orlando Cabrera SS 887.6667 -12.7156058
1651 2009 A MIN Orlando Cabrera SS 501.0000 -0.7264558
v.Tm Name
1650 -4.554030 Orlando Cabrera
1651 -2.043441 Orlando Cabrera
To obtain the cumulative seasonal data for all the players, one needs to sum
the IP, the Runs, and the v.Tm columns for players who played in multiple
teams. The aggregate function allows one to perform this task for one column
at a time; with three applications of aggregate, the result is three separate
data frames that need to be merged using the merge function.
An alternative approach to aggregate uses the doBy package which fea-
tures functions for groupwise calculations. One particularly useful function
is summaryBy which works similarly to aggregate, but allows for the use of
multiple columns on the left side of the aggregating formula. In the following
R code, the doBy package is loaded and the summaryBy is used to sum3 on
the three columns as desired, grouping the data by the Name column. The
keep.names arguments is set to TRUE in order to maintain the original names
for the columns.
3 The FUN argument in summaryBy also accepts a vector, thus multiple functions can be
passed, and thus performed, at once. For example with FUN=c(mean, sd) one simultaneously
computes the mean and the standard deviation on the variables specified on the left side
of the formula. When more than one function is passed to FUN, it is not possible to set
keep.names as TRUE.
library(doBy)
wzr <- summaryBy(IP + Runs + v.Tm ~ Name, data=wzr, FUN=sum,
keep.names=TRUE)
To confirm that this function is operating correctly, the first few lines of the
wzr data frame are displayed.
head(wzr)
Name IP Runs v.Tm
1 A. Hernandez 289.6667 -3.2744424 -0.6064142
2 Aaron Miles 44.0000 0.0488766 -0.3746494
3 Adam Everett 942.6667 2.7007132 -2.3298411
4 Adam Rosales 33.0000 0.8441731 0.7099337
5 Alberto Callaspo 2.0000 -0.2630282 -0.2770827
6 Alberto Gonzalez 279.3333 -4.9981058 -4.6320992
12.2.5 Finding the most similar string (stringdist)

Data for several other defensive metrics can be downloaded from the Fan-
Graphs website.4 Pointing the mouse on the “Leaders” tab at the top of the
page and clicking on the year 2009 under “Batting Leaders,” one finds a page
with several choices. Click on the “Fielding” tab, then select shortstops (“SS”
tab), and in the drop-down menu labeled “Min Inn” choose zero (so that every
player who played shortstop is included). If not selected by default, click on
the “Advanced” tab.5 The website now displays a table featuring defensive
values for 2009 shortstops according to several fielding metrics. On the top-
right angle of the table there is an “Export Data” tag, which allows to save
the table in a comma-delimited (csv) format.
Once the csv file has been saved, it can easily be read into R. The first
column has some unusual characters in the header of the first column and we
rename the first column to “Name” to remove these characters.
fg <- read.csv("FanGraphs Leaderboard.csv")
names(fg)[1] <- "Name"
head(fg)
Name Team Pos Inn rSB rGDP rARM rGFP rPM DRS BIZ
1 Jack Wilson - - - SS 917.1 NA 2 NA 1 25 28 271
2 Elvis Andrus TEX SS 1238.0 NA 2 NA 2 11 15 362
3 Paul Janish CIN SS 592.1 NA 0 NA 1 10 11 194
4 Cesar Izturis BAL SS 934.2 NA 0 NA 3 11 14 312
5 Alex Gonzalez - - - SS 948.1 NA 0 NA 1 -6 -5 250
6 Brendan Ryan STL SS 830.2 NA 2 NA 1 19 22 318
Plays RZR OOZ CPP RPP TZL FSR ARM DPR RngR ErrR UZR
4 www.fangraphs.com.
5 The table can be reached directly at tinyurl.com/fgSS09.
1 223 0.823 48 NA NA 8.4 10 NA 1.4 13.9 -0.2 15.1

2 307 0.848 45 NA NA 6.1 14 NA 0.6 14.1 -2.6 12.1
3 167 0.861 20 NA NA 11.9 6 NA 0.8 6.5 3.4 10.7
4 259 0.830 39 NA NA 16.1 8 NA -0.9 6.5 3.5 9.1
5 204 0.816 31 NA NA -3.3 -1 NA 1.5 2.4 3.8 7.7
6 266 0.837 54 NA NA 12.4 11 NA 2.6 1.4 3.3 7.2
UZR.150 playerid
1 22.7 1017
2 13.5 8709
3 22.7 7412
4 12.1 656
5 12.2 520
6 9.6 6073
The merge function could be used to combine the Wizardry’s and the
FanGraphs’ data frames if they contained a common column. Unfortunately,
wzr does not contain any player identifiers, thus the most likely candidate
columns for the joining of the two datasets are those labeled Name. In the
following R code, we use the setdiff function (which performs asymmetric
difference on two sets) to look for unmatched names between the two data
frames. The fg.mismatches vector contains elements of the column Name in
the fg data frame not found among the elements of the column Name in the
wzr data frame. The wzr.mismatches vector will contain differences in the
opposite direction.
fg.mismatches <- setdiff(fg$Name, wzr$Name)
wzr.mismatches <- setdiff(wzr$Name, fg$Name)
fg.mismatches
[1] "Brent Lillibridge" "Anderson Hernandez"
[3] "Willie Bloomquist" "Yuniesky Betancourt"
wzr.mismatches
[1] "A. Hernandez" "B. Lillibridge" "W. Bloomquist"
[4] "Y. Betancourt"
We have learned that Humphreys has elected to use the initial of the first
name for some players with a long last name. While in this case, with only
four elements on each side, matching by hand would be feasible, but it is
helpful to demonstrate a way for identifying similar strings, which could be
useful in case of longer lists of elements to match. The stringdist pack-
age provides functions for computing so-called string distances. The function
stringdistmatrix compares the elements of the fg.mismatches and the
wzr.mismatches vectors, yielding the distance matrix shown in the follow-
ing code.
library(stringdist)
dm <- stringdistmatrix(fg.mismatches, wzr.mismatches)

dm
[,1] [,2] [,3] [,4]
[1,] 15 4 15 16
[2,] 7 16 17 16
[3,] 16 15 5 13
[4,] 16 19 16 7
To interpret the above results, the first row of the matrix displays the
string distances between the first element of fg.mismatches (i.e., “Brent Lil-
libridge”) and every element of war.mismatches. The smallest distance has a
value of four and is found on the second column of the matrix, corresponding
to the second element of wzr.mismatches (which is “B. Lillibridge”). Having
not specified a value for the method argument, the default value "osa", corre-
sponding to the Optimal string alignment, has been used. The method, also
known as the Damerau-Levenshtein distance, computes the distance between
two strings as the minimum number of operations needed to transform one
string into the other. The operations can be either the insertion of a char-
acter, the deletion of a character, the substitution of a character, and the
transposition of two adjacent characters.
The next step consists of creating a data frame that maps the mismatched
names from FanGraphs with the mismatched names from Wizardry. We first
write a function index.min that inputs a vector and returns the index of
the element featuring the minimum value. Then the apply function is used to
execute index.min on the first margin (i.e., the rows) of the dm matrix. Finally
the names.mapped data frame is built, matching the mismatched FanGraphs
names to the appropriate mismatched Wizardry names.
index.min <- function(v) which(v == min(v))
idx <- apply(dm, MARGIN=1, FUN=index.min)
names.mapped <- data.frame(fgName=fg.mismatches,
wzrName=wzr.mismatches[idx])
The last step before merging the two data frames consists of adding a col-
umn in one data frame (we choose wzr), mapping each name (both those
matched and those initially unmatched) to the names of the other data
frame. The first line of the following code merges the wzr data frame to the
names.mapped just created. By setting the all.x argument to TRUE, we re-
quest to keep every element of the first data frame passed to merge (i.e., wzr);
the elements of wzr that do not have a match in names.mapped will feature
empty values in the columns of wzr. The second line fills those values by copy-
ing from the Name column whenever the fgName column has a missing (NA)
value.
wzr <- merge(wzr, names.mapped, by.x="Name", by.y="wzrName"
, all.x=TRUE)
wzr$fgName <- ifelse(is.na(wzr$fgName), wzr$Name, wzr$fgName)
12.2.6 Applying a function on multiple columns (plyr)

After the mapping of names in the previous section, we are now ready to merge
the data coming from the two different sources. This is accomplished with the
merge function, keeping only the columns needed for the further analysis.
defense <- merge(fg, wzr[,-1], by.x="Name", by.y="fgName")

defense <- defense[,c("Name", "Inn", "Plays", "UZR", "DRS"
, "TZL", "RZR", "FSR", "Runs")]
Some of the players have played very little at the shortstop position in 2009
and thus may not have been rated by one or more of the selected defensive
metrics.6 One can choose to ignore them when computing pairwise correla-
tions between two metrics.7 Another option is to assume those players with
limited playing time to have performed at an average rate, thus substituting
the missing values with zeros for the defensive metrics expressed in runs.8
A new function coalesce is written that takes a value x as input and
returns another value dflt (zero if not otherwise specified) whenever x is
NA. In the lines of code that follow we apply coalesce to the RZR column,
substituting NAs with the average value of .801.
coalesce <- function(x, dflt=0) ifelse(is.na(x), dflt, x)
defense$RZR <- coalesce(defense$RZR, dflt=.801)
The coalesce function is then applied to every column of defense. One can
avoid calling this function multiple times thanks to the colwise function in
the plyr package which takes a vector-input function as the argument and
returns another function which works columnwise on a data frame.
library(plyr)
coalesceColumns <- colwise(coalesce)
defense <- coalesceColumns(defense)
12.2.7 Weighted correlations (weights)

The function cor computes pairwise correlations between the various defensive
metrics contained in the data frame defense. The [,-1:-3] indexing excludes
the first three columns (the player’s name, the innings he played and the plays
he made) from the computations. The round function is used to display only
three decimal digits.
6 Try typing the command subset(defense, Inn < 5) in the R console for displaying
some of them.
7 Such exclusion would be obtained by specifying the value "pairwise.complete.obs" to
the use argument of the cor function.

8 Since RZR is expressed as a rate of succesful plays, we will substitute NAs with the
average success rate, which was .801 for shortstops in 2009.

round(cor(defense[,-1:-3]),3)
UZR DRS TZL RZR FSR Runs
UZR 1.000 0.782 0.727 0.142 0.489 0.305
DRS 0.782 1.000 0.724 0.090 0.596 0.387
TZL 0.727 0.724 1.000 0.124 0.495 0.322
RZR 0.142 0.090 0.124 1.000 0.050 0.059
FSR 0.489 0.596 0.495 0.050 1.000 0.089
Runs 0.305 0.387 0.322 0.059 0.089 1.000
Every observation (i.e., every row in the defense data frame) counts the
same when calculating the correlations. This may be inappropriate – a short-
stop who has more opportunities perhaps should have greater weight than
a part-time shortstop in the computation of the correlations. The function
wtd.cor in the weights package can compute a correlation allowing for dif-
ferent weights for the observations. The function wtd.cor has two inputs: the
matrix of observations and a vector of weights which in this case corresponds
to the number of innings played. The function returns three matrices: a ma-
trix of correlation coefficients, and matrices of t-values and p-values. In the
code that follows we limit the output to the correlation matrix, which can be
compared to the one produced by cor.
library(weights)
round(wtd.cor(defense[,-1:-3], weight=defense$Inn)$correlation, 3)
UZR DRS TZL RZR FSR Runs
UZR 1.000 0.794 0.731 0.618 0.502 0.241
DRS 0.794 1.000 0.745 0.433 0.598 0.364
TZL 0.731 0.745 1.000 0.565 0.487 0.288
RZR 0.618 0.433 0.565 1.000 0.275 0.171
FSR 0.502 0.598 0.487 0.275 1.000 0.018
Runs 0.241 0.364 0.288 0.171 0.018 1.000
The matrix of weighted correlations is slightly different than the raw (un-
weighted) correlations. In particular, it appears the very low correlation be-
tween the Revised Zone Rating (RZR) and the other metrics was highly affected
by the numbers posted by the occasional players.
12.2.8 Displaying correlation matrices (ellipse)

A correlation matrix can be displayed graphically by the use of ellipse-shaped
glyphs (Murdoch and Chow (1996)). The ellipses will have a bottom-left to
top-right orientation for positive correlations and a bottom-right to top-left
orientation for negative correlations. They will appear narrower for higher lev-
els of correlations and will look more like circles as the correlation coefficients
approximate zero.
The package ellipse provides the function plotcorr to produce such
visualizations. The results of the following code can be seen in Figure 12.1
library(ellipse)
Dcor <- wtd.cor(defense[,-1:-3], weight=defense$Inn)$correlation
plotcorr(Dcor)
Runs
DRS
UZR
RZR
FSR
TZL
UZR
DRS
TZL
RZR
FSR
Runs
FIGURE 12.1
Visualization of the correlation matrix for defensive metrics.
12.2.9 Evaluating the fielding metrics (psych)

The ellipses in Figure 12.1 indicate positive correlations between any pair of
defensive metrics, but the strengths of the associations among the various field-
ing metrics are variable, as shown graphically by the thickness of the ellipses
and numerically in the correlation matrix in Section 12.2.7. Not surprisingly,
UZR and DRS have the highest correlation (0.79) – this is understandable
since both measures are based on Baseball Info Solution data. The measure
FSR, based solely on subjective evaluations, shows a fair degree of association
with some of the other metrics. On the other hand, the method outlined in
Wizardry, labeled as Runs, displays the lowest correlations across the board.
By following steps similar to the ones outlined in Sections from 12.2.3 to
12.2.5, we created a file named teamD.csv, which contains team cumulative
defensive ratings for the seasons 2007 to 2009. In addition to the six met-
rics previously presented, this file features an additional rating, the Defensive
Efficiency Record, labeled DER. This fielding measure, first proposed by Bill
James, is calculated by dividing the outs recorded on batted balls by the
number of balls playable by the defense (every batted ball except home runs).
DER provides a fair, albeit not perfect, estimation of defensive value at the
team level.9 In the remainder of this section the six previously shown metrics
will be compared with DER.
The teamD.csv file is read into R and the weighted correlations of the six
metrics are computed, with the help of the wtd.cor function in the weights
package, as done in Section 12.2.7.
teamD <- read.csv("teamD.csv")
Dcor <- wtd.cor(teamD[,c("DER", "UZR", "DRS", "TZL", "RZR"
, "FSR", "Runs")], weight=teamD$IP)
A new visualization, produced by the function cor.plot in the psych, is
used to display the resulting correlation matrix. In our example, the function
mat.sort is first used to sort the correlation matrix so that similar items
are grouped together. Then the function cor.plot displays the correlation
matrix where darker boxes correspond to larger correlation values. By setting
the numbers argument to TRUE), the values of the correlations, on a 0 to 100
scale, are also displayed.
library(psych)
sortedCor <- mat.sort(Dcor$correlation)
cor.plot(sortedCor, numbers=TRUE)
According to Figure 12.2, TZL has the highest correlation with DER for the
three-year period.
12.3 Comparing Two Shortstops

In this section, the long careers10 of two great shortstops are used to illustrate
functions available in several contributed R packages. Specifically, suppose one
is interested in visualizing fielding ratings of the careers of Yankee captain
Derek Jeter and long-time Phillies player Jimmy Rollins according to the
metrics UZR and DSR. The FanGraphs website allows registered users11 to
9 The main shortcoming of DER consists in valuing all unsuccessful defensive plays the
same, thus not accounting for the different outcomes of batted balls (singles, doubles, and
so on).
10 We will actually be using part of the career of Rollins and Jeter, particularly the seasons
covered by BIS data, thus having UZR and DRS values available.
11 Registration is free of charge.
DER 100 76 78 63 57 69 30
Runs 76 100 66 62 66 59 20
TZL 78 66 100 58 56 58 40
UZR 63 62 58 100 63 50 45
DRS 57 66 56 63 100 50 63
RZR 69 59 58 50 50 100 44
FSR 30 20 40 45 63 44 100
DER
Runs
TZL
UZR
DRS
RZR
FSR
FIGURE 12.2
Correlation plot for the comparison of team defensive metrics.
create custom leaderboards, by selecting desired stats and a subset of players.

For this particular example, we request a table displaying Jeter’s and Rollins’
year-by-year UZR and DRS measures and the resulting table is exported in
the jeter rollins.csv file. In the following R code, the file is loaded into
the jetrol data frame and its first few lines are displayed.
jetrol <- read.csv("jeter_rollins.csv")
head(jetrol)
Season Name Team DRS UZR playerid
1 2008 Jimmy Rollins PHI 18 12.9 971
5 2009 Derek Jeter NYY 3 6.4 826
12.3.1 Reshaping the data (reshape2)

Data in the jetrol data frame require a little reshaping in order to be easily
processed through one of the powerful graphics packages (either lattice or
ggplot2). The function melt in the reshape2 package is used to give the
jetrol data frame the desired shape.
The arguments of the function melt are the data frame, id.vars,
measure.vars, variable.name, and value.names. In this example, we indi-
cate, by the id.vars argument, that the new data frame will have the first two
columns featuring the season and the name of the player. The measure.vars
argument indicates that the measured variables DRS and UZR will be the ones
undergoing the reshaping – two new columns will appear in the data frame.
The variable.name argument indicates that fieldingMetric is the name of
the defensive metric, and the value.name argument indicates that runs is the
name of the defensive rating.
library(reshape2)
jetrol2 <- melt(jetrol
, id.vars=c("Season", "Name")
, measure.vars=c("DRS", "UZR")
, variable.name="fieldingMetric"
, value.name="runs")
head(jetrol2)
Season Name fieldingMetric runs
1 2008 Jimmy Rollins DRS 18
5 2009 Derek Jeter DRS 3
12.3.2 Plotting the data (ggplot2 and directlabels)

The jetrol2 data frame is now ready to be plotted using the ggplot2
package, extensively introduced in Chapter 6. The ggplot object is initial-
ized by mapping the Season to the x-axis, the runs to the y-axis, and the
fieldingMetrics to the color aesthetic. By adding the geom line object, we
indicate we want a line graph, and the facet grid function is used to place
the players (Name) in separate panels. The scale color manual layer is used
for manually selecting the line colors (the label appearing as the legend title
is specified here as well), while in the scale x continuous layer the position
of the tick marks is indicated on the x-axis. Finally a dotted reference line is
drawn at zero runs saved. Since the whole ggplot object has been assigned
to the variable p, one needs to type p in the R console to actually display the
plot.
library(ggplot2)
p <- ggplot(jetrol2, aes(x=Season, y=runs
, col=fieldingMetric)) +
geom_line() +
facet_grid(. ~ Name) +
scale_color_manual(name="Fielding\nmetric"
, values=c("black", "grey70")) +
scale_x_continuous(breaks=seq(2004, 2012, 4)) +
geom_hline(yintercept=0, lty=3)
The package directlabels provides a quick-to-use function direct.label
which adds labels directly to the plot, thus removing the color legend, as shown
in Figure 12.3.12
library(directlabels)
direct.label(p)
The plots in Figure 12.3 show that Jeter has been constantly rated as a
below-average fielder (despite both UZR and DRS indicating an unusually
good season for him in 2009), while Rollins has mostly been an above-average
fielding shortstop. UZR and DRS numbers follow more or less the same path
for the Yankee captain, but their portrait of the Philadelphia shortstop convey
different information: where DRS seems to indicate a steep decline in Jimmy’s
defensive value, he appears to have maintained his ability according to UZR.

Rickey (1954) is an article that appeared on the August 2, 1954, issue of Life
magazine, in which Branch Rickey, considered one of the greatest baseball
executives ever, illustrates what are among the earliest known sabermetric-
like analyses. On page 83 of the magazine, he admits that fielding value as
was measured at the time (with fielding percentage) is “utterly worthless.”
Humphreys (2011) outlines how several of the fielding metrics developed in
the past 15 years of work, and illustrates a method of his own to estimate runs
saved by fielders using data available throughout most of baseball history. John
Dewan (2006, 2009, 2012) periodically publishes a volume of The Fielding
Bible, containing essays and fielding leaderboard tables based on the Baseball
Info Solution proprietary system.
12 Note that in most cases, adding labels to the area where data are displayed unnecessarily
clutters the plot.

Derek Jeter Jimmy Rollins

20
DRS
10
UZR
0 UZR
runs
−10
−20
DRS
2004 2008 2012 2004 2008 2012
Season
FIGURE 12.3
Derek Jeter’s and Jimmy Rollins’ defensive value through the years as mea-
sured by UZR and DRS. Labels added directly on the plot area.
12.5 Exercises
1. (Data Reshaping: Exploring Team Plate Discipline)
(a) On the FanGraphs website (www.fangraphs.com/) generate a report
containing team batting plate discipline stats from 2008 to the latest
completed season. Export the data as a csv file and load it into R.
Rename the columns if necessary.
(b) Calculate the correlation between percentage of swings outside the
zone (O-Swing%) and the percentage of pitches in the strike zone
(Zone%)). Are teams with a high propensity toward swinging at bad
pitches fed with a lower number of pitches in the strike zone?
(c) Reshape the loaded data so that the new data frame will consist of
four columns: season, team, the name of the discipline statistic, and
the value of the discipline statistic.
(d) Subset the data frame so that it only contains O-Swing% and Zone%
rows.
(e) Draw a plot where teams are displayed in different panels, each fea-
turing the year-to-year trend of both O-Swing% and Zone%.
2. (Applying a Function on Multiple Columns: Exploring the Avail-
ability of Statistics Through Baseball History)
Not every baseball statistic has been tracked since the early days of base-
ball. Stolen Bases and Caught Stealing are examples of numbers which are
not available in some earlier seasons.
(a) Load the batting statistics from Lahman’s database and create a new
data frame containing only the following information: year, league,
stolen bases, and runners caught stealing.
(b) Write a function that takes a vector as its only argument and returns
the percentage of missing values (NAs) in the vector. Note that is.na
is a function for identifying NAs and that the function length returns
the number of elements in a vector.
(c) Apply the newly written function to the columns of the previously
created data frame.
(d) Use the ddply function (from the plyr package) to apply the function
by year and by league. Was the recording of SBs and CSs introduced
at the same time in the National and the American League?
3. (Comparing Defensive Ratings and Stealing Ability in Center-
fielders)
As speed is an important trait of a centerfielder skill set, good base stealers
are often found among players manning the middle outfielder position.
(a) Download seasonal fielding ratings for centerfielders from Wizardry’s
online resources. Read the Excel file into R.
(b) Using tables from Lahman’s database, prepare a data frame contain-
ing seasonal data on offensive stolen bases and caught stealing for
players who have appearances (in the given season) at centerfield.
(c) Estimates like those performed in Chapter 5 indicate that successful
stolen bases are worth roughly 0.2 runs to the offensive team, while
a caught stealing cost it 0.5 runs. Using these values compute stolen
base runs for every player-season.
(d) Merge defensive ratings with stolen base rating data. Note that on
the previous point you may want to gather the players’ first and
last names from the Master table. Try to perform string matching
as shown in Section 12.2.5 to match the highest number of players-
seasons.
(e) Compute the correlation between Stolen Base Runs and Defensive
Runs (as reported in Wizardry).
A
Retrosheet Files Reference
CONTENTS
A.1 Downloading Play-by-Play Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
A.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
A.1.2 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
A.1.3 Using a special function for a particular season . . . . . . . . . 302
A.1.4 Reading the files into R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
A.1.5 The function parse.retrosheet.pbp . . . . . . . . . . . . . . . . . . . 302
A.2 Retrosheet Event Files: a Short Reference . . . . . . . . . . . . . . . . . . . . . . . 304
A.2.1 Game and event identifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
A.2.2 The state of the game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
A.3 Parsing Retrosheet Pitch Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
A.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
A.3.2 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
A.3.3 Evaluating every count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
A.1 Downloading Play-by-Play Files

A.1.1 Introduction
The play-by-play data files are currently available at the Retrosheet Web
page www.retrosheet.org/game.htm. By clicking on a single year, say 1950,
one obtains a compressed (zip) file containing a collection of files, one set of
files containing information on the plays for the home games for all teams,
and another set of files giving the rosters of the players for each team. This
appendix describes the use of a single function parse.retrosheet.pbp to
create a single file, in csv format, of all plays for games played in a particular
season, and also create a csv file of the rosters of all players. We assume
that one is working in a Windows session, although similar operations can be
performed on other systems.
301
A.1.2 Setup
In the current R working directory, create a new folder “download.folder”
and in this new folder, create two folders “zipped” and “unzipped.” There
are special software tools for working with the Retrosheet files that should be
downloaded from sourceforge.net/projects/chadwick/files/. When the
compressed file is expanded, the file “cwevent.exe” should be placed in the
“unzipped” folder.
A.1.3 Using a special function for a particular season

Open R and make sure that you are in the working directory containing the
folder “download.folder.” Read in the function “parse.retrosheet.pbp.R” by
use of the source function.
source("parse.retrosheet.pbp.R")
The function parse.retrosheet.pbp has a single argument season. To ob-

tain the play-by-play files for the 1950 season, the relevant Retrosheet files
are obtained by typing
parse.retrosheet.pbp(1950)
If you look in the “download.folder/unzipped” folder, you will see two new
files: “all1950.csv” contains the play for play records for the 1950 season, and
“roster1950.csv” contains roster information for all players of that season.
A.1.4 Reading the files into R

To read the data files into R, one should first make sure that the
files “all1950.csv” and “roster1950.csv” are in the current working direc-
tory. In addition, the file “field.csv” containing the names of all of the
fields in the play-by-play record file should be placed in this directory.
This file was created by the authors from the information in the page
chadwick.sourceforge.net/doc/cwevent.html. The following R script will
read in the three files by three applications of the read.csv function. In ad-
dition, the names function is used to assign names to the play-by-play data
frame using the Header variable from the fields data frame.
data <- read.csv("all1950.csv", header=FALSE)
roster <- read.csv("roster1950.csv")
names(data) <- fields[, "Header"]
A.1.5 The function parse.retrosheet.pbp

The R function parse.retrosheet.pbp listed below contains five functions.
We briefly describe what is done in each of these functions.
Retrosheet Files Reference 303
The function download.retrosheet downloads the Retrosheet zip file

from the particular season and saves the zip file in the “zipped” folder.
The function unzip.retrosheet unzips the file and puts the individ-
ual event files and roster files in the “unzipped” folder. The function
create.csv.file(season) reads the individual team record files into R,
merges the files, and writes a new single combined file in csv format in the
“unzipped” folder. The function create.csv.roster function does a similar
thing with the individual roster files. Last, the cleanup function removes the
individual record and roster files that are no longer needed.
parse.retrosheet.pbp <- function(season){
download.retrosheet <- function(season){
download.file(
url=paste("http://www.retrosheet.org/events/", season, "eve.zip", sep="")
, destfile=paste("download.folder", "/zipped/", season, "eve.zip", sep="")
)
}
unzip.retrosheet <- function(season){
unzip(paste("download.folder", "/zipped/", season, "eve.zip", sep=""),
exdir=paste("download.folder", "/unzipped", sep=""))
}
create.csv.file <- function(year){
wd <- getwd()
setwd("download.folder/unzipped")
shell(paste(paste("cwevent -y", year, "-f 0-96"),
paste(year, "*.EV*", sep=""),
paste("> all", year, ".csv", sep="")))
setwd(wd)
}
create.csv.roster <- function(year){
filenames <- list.files(path = "download.folder/unzipped/")
filenames.roster <-
subset(filenames, substr(filenames, 4, 11) == paste(year,".ROS",sep=""))
read.csv2 <- function(file)
read.csv(paste("download.folder/unzipped/", file, sep=""), header=FALSE)
R <- do.call("rbind", lapply(filenames.roster, read.csv2))
names(R)[1 : 6] <- c("Player.ID", "Last.Name", "First.Name",
"Bats", "Pitches", "Team")
wd <- getwd()
write.csv(R, file=paste("roster", year, ".csv", sep=""))
setwd(wd)
}
cleanup <- function(){
wd <- getwd()
shell("del *.EVN")
shell("del *.EVA")
shell("del *.ROS")
shell("del TEAM*")
setwd(wd)
}
download.retrosheet(season)
unzip.retrosheet(season)
create.csv.file(season)
create.csv.roster(season)
cleanup()
}
A.2 Retrosheet Event Files: a Short Reference

As was mentioned in Chapter 1, Retrosheet event files come in a format ex-
pressly devised for them, and require the use of some software tools for con-
verting them in a format suitable for data analysis. Retrosheet provides such
software tools (www.retrosheet.org/tools.htm) and a step-by-step example
(www.retrosheet.org/stepex.txt) for performing the conversion.
Another website1 (whose name is a tribute to baseball pioneer Henry Chad-
wick credited for devising the box score) provides similar tools for parsing
Retrosheet event files,2 that have been used for creating the play-by-play files
used in this book.
Chadwick tools generate a line for each play in the Retrosheet event files,
consisting of 97 “regular” columns (the same that are obtained using the tools
provided by Retrosheet) plus 63 “extended” fields, allowing to easily access
all of the information contained in the Retrosheet event files. Going through
every one of the more than 150 columns generated by the Chadwick tools is
beyond the scope of this book, thus we point to the documentation on the
Chadwick website for the full list.3 In this section we present the main fields
describing an event and the state of the game when it happens.
A.2.1 Game and event identifiers

The games are identified in Retrosheet event files by 12-character strings (the
GAME ID column): the first three characters identify the home team, the fol-
lowing eight characters indicate the date when the game took place (in the
YYYYMMDD format), and the last character is used to distinguish games of
1 Chadwick:Software Tools for Scoring Baseball Games chadwick.sourceforge.net/
doc/index.html.
2 Download link: sourceforge.net/projects/chadwick/files.
3 The documentation for all the software tools is available at chadwick.sourceforge.
net/doc/cwtools.html. In particular, the tool for processing the event files (cwevent) is
documented at chadwick.sourceforge.net/doc/cwevent.html#cwtools-cwevent.
doubleheaders (thus “1” indicates the first game, “2” the second game, and
“0” means only one game was played on the day).
Events are progressively numerated in each game (column EVENT ID), thus
every single action in the Retrosheet database can be uniquely identified by
the combination of the game identifier and the event identifier.
A.2.2 The state of the game

Several fields are helpful for defining the state of the game when a particular
event happened. The inning and the team on offense variables are stored
in the INN CT and BAT HOME ID fields respectively. The latter field can assume
values “0” (away team batting, i.e., top of the inning) or “1” (home team bat-
ting, bottom of the inning). The visitor score and the home score variables
are recorded in the AWAY SCORE CT and HOME SCORE CT.
The number of outs before the play is indicated in the OUTS CT column,
while the situation of runners on base is coded in the field START BASES CD,
using numbers from 1 to 7 as shown in Table A.1.4
TABLE A.1
Retrosheet coding for the situation of runners on base.
Code Bases occupancy

0 Empty
1 1B only
2 2B only
3 1B & 2B
4 3B only
5 1B & 3B
6 2B & 3B
7 Loaded
The actual description of the event resides in the EVENT TX column,

consisting in a string describing the outcome of the play (e.g., strikeout, single,
etc.), some additional details (e.g. the type and location of the batted ball),
and the advancement of any runner on base. Several columns are generated
by decoding the EVENT TX string:
• EVENT CD is a numeric code reflecting the basic event; Table A.2 dis-
plays the codes for the possible plays coded in this column.
• BAT EVENT FL is a flag indicating whether an event is a batting event,
in which case it is labeled as T. Non-batting events include, for example,
4 An analogous column named END BASES CD contains the base state at the end of the
play, coded in the same way.

stolen bases, wild pitches and, generally, any event that does not mark the
end of a plate appearance.
• H CD is a numeric code indicating the base hit type, going from 1 for
a single to 4 for a home run.
• BATTEDBALL CD is a single character code denoting the batted ball
type. It can assume one of the following values: G (ground ball), L (line
drive), F (fly ball), P (pop-up). Note that for most of the seasons in the
Retrosheet database, the batted ball type is reported only for plate appear-
ances ending with the batter making an out, while they are not available
on base hits.
• BATTEDBALL LOC TX is a string indicating the batted ball loca-
tion, coded according to the diagram shown at www.retrosheet.org/
location.htm. Note that this information is available for a limited num-
ber of seasons.
• FLD CD is a numeric code denoting the fielder first touching a batted
ball, coded with the conventional baseball fielding notation going from 1
(the pitcher) to 9 (the right fielder).
The sequence of pitches is recorded in the PITCH SEQ TX and has been
addressed in Chapter 7, where Table 7.1 displays how the different pitch out-
comes are coded. Several columns are generated from this one, indicating
counts of the various types of pitch outcomes, as displayed in Table A.3.
A.3 Parsing Retrosheet Pitch Sequences

A.3.1 Introduction
Chapter 7 showed how to compute, by the use of regular expressions, whether
a plate appearance went through either a 1-0 or a 0-1 count. Here the code
is provided to retrieve the same information for every possible balls/strikes
count.
A.3.2 Setup
We first load Retrosheet data for the 2011 season.
pbp2011 <- read.csv("retrosheet/all2011.csv")
headers <- read.csv("retrosheet/fields.csv")
names(pbp2011) <- headers$Header
TABLE A.2
Retrosheet coding for the type of event.
Code Event type

2 Generic Out
3 Strikeout
4 Stolen Base
5 Defensive Indifferen
6 Caught Stealing
8 Pickoff
9 Wild Pitch
10 Passed Ball
11 Balk
12 Other Advance
13 Foul Error
14 Nonintentional Walk
15 Intentional Walk
16 Hit By Pitch
17 Interference
18 Error
19 Fielder Choice
20 Single
21 Double
22 Triple
23 Homerun
Then a new column sequence is created in which the pitch sequence is re-
ported, stripped by any character not indicating an actual pitch to the batter.5
pbp2011$sequence <- gsub("[.>123+*N]", "", pbp2011$PITCH_SEQ_TX)
A.3.3 Evaluating every count

Every plate appearance starts with a 0-0 count. The code for both the 1-0
and 0-1 counts was described in Chapter 7.
pbp2011$c00 <- TRUE
pbp2011$c10 <- grepl("^[BIPV]", pbp2011$sequence)
pbp2011$c01 <- grepl("^[CFKLMOQRST]", pbp2011$sequence)
A number inside the square brackets indicates the minimum number of
times the preceding expression has to be repeated in the string to be parsed.
The following lines look for plate appearances going through the counts 2-0,
3-0, and 0-2.
5 See Table 7.1 in Chapter 7 for reference.
TABLE A.3
Columns reporting counts of various pitch types.
Column name Column description

PA BALL CT No. of balls in plate appearance
PA CALLED BALL CT No. of called balls in plate appearance
PA INTENT BALL CT No. of intentional balls in plate
appearance
PA PITCHOUT BALL CT No. of pitchouts in plate appearance
PA HITBATTER BALL CT No. of pitches hitting batter in plate
appearance
PA OTHER BALL CT No. of other balls in plate appearance
PA STRIKE CT No. of strikes in plate appearance
PA CALLED STRIKE CT No. of called strikes in plate appearance
PA SWINGMISS STRIKE CT No. of swinging strikes in plate
appearance
PA FOUL STRIKE CT No. of foul balls in plate appearance
PA INPLAY STRIKE CT No. of balls in play in plate appearance
PA OTHER STRIKE CT No. of other strikes in plate appearance
pbp2011$c20 <- grepl("^[BIPV]{2}", pbp2011$sequence)

pbp2011$c30 <- grepl("^[BIPV]{3}", pbp2011$sequence)
pbp2011$c02 <- grepl("^[CFKLMOQRST]{2}", pbp2011$sequence)
The | (vertical bar) character is used to separate alternatives. The follow-
ing lines parse the sequence string looking for the different sequences that
can lead to 1-1, 2-1, and 3-1 counts.
pbp2011$c11 <- grepl("^([CFKLMOQRST][BIPV]|[BIPV][CFKLMOQRST])"
, pbp2011$sequence)
pbp2011$c21 <- grepl("^([CFKLMOQRST][BIPV][BIPV]
|[BIPV][CFKLMOQRST][BIPV]
|[BIPV][BIPV][CFKLMOQRST])", pbp2011$sequence)
pbp2011$c31 <- grepl("^([CFKLMOQRST][BIPV][BIPV][BIPV]
|[BIPV][CFKLMOQRST][BIPV][BIPV]|[BIPV][BIPV][CFKLMOQRST][BIPV]
|[BIPV][BIPV][BIPV][CFKLMOQRST])", pbp2011$sequence)
On two-strike counts, batters can indefinitely foul pitches off without af-
fecting the count. In the lines below, sequences reaching two strikes before
reaching the desired number of balls feature the [FR]* expression, denoting a
foulball6 happening any number of times, including zero, as indicated by the
asterisk.
pbp2011$c12 <- grepl("^([CFKLMOQRST][CFKLMOQRST][FR]*[BIPV]

6F encodes a foulball, R a foulball on a pitchout. See Table 7.1.
|[BIPV][CFKLMOQRST][CFKLMOQRST]
|[CFKLMOQRST][BIPV][CFKLMOQRST])", pbp2011$sequence)
pbp2011$c22 <- grepl("^(
[CFKLMOQRST][CFKLMOQRST][FR]*[BIPV][FR]*[BIPV]
|[BIPV][BIPV][CFKLMOQRST][CFKLMOQRST]
|[BIPV][CFKLMOQRST][BIPV][CFKLMOQRST]
|[BIPV][CFKLMOQRST][CFKLMOQRST][FR]*[BIPV]
|[CFKLMOQRST][BIPV][CFKLMOQRST][FR]*[BIPV]
|[CFKLMOQRST][BIPV][BIPV][CFKLMOQRST]
)", pbp2011$sequence)
pbp2011$c32 <- grepl("^[CFKLMOQRST]*[BIPV][CFKLMOQRST]*
[BIPV][CFKLMOQRST]*[BIPV]", pbp2011$sequence)
& grepl("^[BIPV]*[CFKLMOQRST][BIPV]*[CFKLMOQRST]"
, pbp2011$sequence)
B
Accessing and Using MLBAM Gameday and
PITCHf/x Data
CONTENTS
B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
B.2 Where are the Data Stored? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
B.3 Suitable Formats for PITCHf/x Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
B.3.1 Obtaining data from on-line resources . . . . . . . . . . . . . . . . . . 314
B.3.2 Parsing in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
B.3.2.1 A wrapper function . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
B.4 Details on the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
B.4.1 atbat attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
B.4.2 pitch attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
B.4.3 hip attributes (hit locations data) . . . . . . . . . . . . . . . . . . . . . . 318
B.5 Special Notes About the Gameday and PITCHf/x Data . . . . . . . . 319
B.6 Miscellanea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
B.6.1 Calculating the pitch trajectory . . . . . . . . . . . . . . . . . . . . . . . . . 320
data: pitchRx . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
B.6.3 Cross-referencing with other data sources . . . . . . . . . . . . . . . 323
B.6.4 Online resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
B.1 Introduction
This section provides further details on Gameday and PITCHf/x data. It
will be shown where those data are stored, how they can be retrieved us-
ing resources available on the Web, and how they can be parsed using R. A
description of the most important fields will be provided, together with an
overview of the issues a researcher should be aware of when analyzing these
data.
311
B.2 Where are the Data Stored?

Data relevant to games of one particular day can be found following a link such
as gd2.mlb.com/components/game/mlb/year_2012/month_06/day_13/. By
simply modifying the year, the month, and the day, one can reach any MLB
game of the Gameday era.1 Figure B.1 displays the page featuring the games
played on June 13, 2012.
FIGURE B.1
MLB directory of games played on June 13, 2012.
To access data relative to a specific game, one follows one of the

links beginning with the characters gid . For example, clicking on the
gid 2012 06 13 houmlb sfnmlb 1/ link, one reaches the page for the Houston
Astros @ San Francisco Giants game, in which Matt Cain pitched a perfect
game. Once the gd2.mlb.com/components/game/mlb/year_2012/month_06/
day_13/gid_2012_06_13_houmlb_sfnmlb_1/ page is reached, all of the play-
by-play data, including PITCHf/x and hit locations, are available by follow-
ing the inning/ link. PITCHf/x data for the whole game are stored in the
1 Actually, by modifying the mlb part of the URL, one can even access minor league
games.
Accessing and Using MLBAM Gameday and PITCHf/x Data 313
inning all.xml XML page and they appear on the Web page as in Figure
B.2. Similarly, data on hit locations are stored in the inning hit.xml page.
FIGURE B.2
Pitch-by-pitch data of the Astros – Giants game played on June 13, 2012.
The form of XML documents is a tree structure. The inning all.xml files
have the following structure.
<game>
<inning>
<top>
<atbat>
<pitch>
</atbat>
</top>
<bottom>
<atbat>
<pitch>
</atbat>
</bottom>
</inning>
</game>
The <game> element is said to be the root element of the tree – this indicates
that this document is a game. The <inning> element is a child of the root,
and <top> and <bottom> are in turn children of <inning>. Both <top> and
<bottom> elements have <atbat> children, which then have <pitch> children.
Note that all elements are delimited by an <element> tag at the beginning
and an </element> tag at the end.
In the opening tag (the one without the slash) of elements, attributes can
be found. For example, the <inning> element usually has the format
<inning num="1" away_team="hou" home_team="sfn" next="Y">

The <inning> element has four attributes: the inning number (num which has
value "1" in this case), the teams involved (away team and home team, set as
"hou" for Houston and "sfn" for San Francisco), and whether another inning
has been played (next, which has value "Y" in this case).
The attributes of the <atbat> and the <pitch> elements are the ones con-
taining the play-by-play and pitch-by-pitch information, including PITCHf/x
data, and will be described in Sections B.4.1 and B.4.2.
B.3 Suitable Formats for PITCHf/x Data

To analyze pitch-by-pitch data, the XML information needs to be converted
in a more suitable format. In this section a few online resources for converting
PITCHf/x data to a user-ready form will be described and some details on
how to parse XML data in R will be provided.
B.3.1 Obtaining data from on-line resources

Starting from Joseph Adler’s publication of Baseball Hacks (Adler, 2006), sev-
eral people have shared their codes to download and parse Gameday data.
Kyle Wilkomm developed his parser, named BBOS (short for Base-
ball On a Stick), and made it available at sourceforge.net/projects/
baseballonastic/. BBOS uses Python code to create and populate a MySQL
database of pitch-by-pitch data and can be run by users not familiar with
Python. Jeff Zimmerman has further simplified life for aspiring baseball
analysts, as he makes his MySQL pitch-by-pitch database (updated daily
during the season) available for download at www.baseballheatmaps.com/
pitch-fx-download/ (see Section 11.6.2 for details).
B.3.2 Parsing in R
The XML package provides XML parsing functionality in R. The first step in
the parsing process reads the XML file into R using the xmlParse function.
library(XML)
gameUrl <- "gd2.mlb.com/components/game/mlb/year_2012/mo
nth_06/day_13/gid_2012_06_13_houmlb_sfnmlb_1/inning/inning_al
l.xml"
xmlGame <- xmlParse(gameUrl)
Several functions are available to access specific nodes of the tree. For
example, one can obtain the inning nodes by using the getNodeSet function.
xmlInnings <- getNodeSet(xmlGame, "//inning")
length(xmlInnings)
[1] 9
By using the xmlAttrs function, one can retrieve the attributes of a node.
The following code obtains the attributes for the first inning.
xmlAttrs(xmlInnings[[1]])
num away_team home_team next

"1" "hou" "sfn" "Y"
To obtain the attributes for all the nine innings, one can use the function
ldply from the plyr package analogous to the function ddply introduced in
Chapter 2.2
library(plyr)
ldply(xmlInnings, xmlAttrs)
num away_team home_team next

1 1 hou sfn Y
2 2 hou sfn Y
3 3 hou sfn Y
4 4 hou sfn Y
5 5 hou sfn Y
6 6 hou sfn Y
7 7 hou sfn Y
8 8 hou sfn Y
9 9 hou sfn N
B.3.2.1 A wrapper function

The few lines presented in the previous section are most of what is needed to
parse XML data into R data frames. However, since the corresponding XML
elements may not always have the same number of columns (for example, if one
pitch did not get tracked by the PITCHf/x system, it will lack all the relevant
2 ldply differs from ddply only for the input data it accepts: where ddply takes a data
frame (hence the ”d” at the beginning of its name), ldply requires a list (hence the ”l”).
attributes), the ldply function might encounter problems in generating the

data frame due to the mismatch of attributes. The following wrapper function
grabXML takes care of this issue and can be used for the parsing of most XML
documents.
grabXML <- function(XML.parsed, field){
parse.field <- getNodeSet(XML.parsed, paste("//", field, sep=""))
results <- t(sapply(parse.field, function(x) xmlAttrs(x)))
if(typeof(results)=="list"){
do.call(rbind.fill, lapply(lapply(results, t), data.frame
, stringsAsFactors=F))
} else {
as.data.frame(results, stringsAsFactors=F)
}
}
We illustrate the use of the grabXML function on the <pitch> nodes. This
function requires a parsed XML document as the first argument (the pre-
viously created xmlGame is used here) and a character string indicating the
nodes on which the process is applied ("pitch" here).
pitchesData <- grabXML(xmlGame, "pitch")
dim(pitchesData)
[1] 279 39
For the June 13, 2012 game between Houston and San Francisco, data on 279
pitches are parsed into the pitchesData data frame which consists of 279
rows and 39 columns. Functions specifically devised for retrieving PITCHf/x
data are available in the pitchRx package, introduced in Chapter 12.
B.4 Details on the Data

In this section we describe the atbat and pitch attributes most frequently
used in baseball analysis. Good collections of descriptions of the PITCHf/x
attributes are provided by Mike Fast at fastballs.wordpress.com/2007/
08/02/glossary-of-the-gameday-pitch-fields/ and Alan Nathan at
webusers.npl.illinois.edu/~a-nathan/pob/tracking.htm, and our de-
scriptions that follow heavily draw from those sources.
B.4.1 atbat attributes

• batter and pitcher: MLBAM six-digit codes identifying the batter and
the pitcher.
• stand and p throws: handedness of the batter and the pitcher for the
at-bat.
• des and event: detailed and short descriptions of the at-bat outcome.
B.4.2 pitch attributes

• des: Short description of the pitch outcome.
• type: One-character code of the pitch outcome (one among B=ball,
S=strike, X=ball in play).
• x and y: Horizontal and vertical coordinates of the pitch as it crosses
home plate, manually recorded. Note: these are not the coordinates recorded
by the PITCHf/x system.
• sv id: Timestamp denoting the second when PITCHf/x detected the
pitched ball, formatted as YYMMDD hhmmss.
• start speed and end speed: Speed in miles per hour at the release
point and when the ball crosses the front of home plate.
• sz top and sz bot: Vertical coordinates for the top and the bottom
of the strike zone of the batter currently at the plate. Both variables are
expressed as feet from the ground and they are manually recorded at the
beginning of every at-bat.
• pfx x and pfx z: Horizontal and vertical movement of the pitch com-
pared to a theoretical pitch of the same speed with no spin-induced move-
ment. Both variables are measured in inches.
• px and pz: Horizontal and vertical location of the pitch, measured when
the pitch crosses the front of home plate as recorded by the PITCHf/x
system. The coordinate system is centered on the middle of home plate
and at ground level and viewed from the catcher/umpire point of view,
thus a positive value of px indicates the pitch crosses the plate to the
right of its middle and a negative value to the left. A negative value of pz
indicates a pitch that bounced before reaching home plate. Both px and
pz variables are measured in feet.
• x0, y0, z0: Coordinates indicating the calculated position of the ball
at the release point. The y0 parameter indicates the distance from home
plate and is generally set at 50 feet from home plate3 ; researchers have
found 55 feet as a distance that better approximates the true release point
of the pitch and it is thus advisable to recalculate the coordinates at the
55ft mark,4 as illustrated in Section B.6.1. x0, y0 and z0 are the left and
3 Inthe first years of PITCHf/x this value varied from 40 to 55 feet.
4 PITCHf/x data provided in the data frames of this book have undergone such trans-
formation.
right position and the height of the release point in the same coordinate
system as px and pz.
• vx0, vy0, and vz0: Components of the pitch velocity in three dimen-
sions, measured at release in feet per second.
• ax, ay, and az: Components of the pitch acceleration in three dimen-
sions, measured at release in f t/s2 .
• break y, break angle, and pitch type: Quantities defining the

“break” of a pitch by comparing its trajectory to that of a straight
line going from release point to location at home plate. The variable
break length is the maximum deviation from the straight line in inches,
break y is the distance from home plate where such deviation occurs (in
inches), and break angle is the direction of the deviation, with positive
values indicating a break toward a right-handed hitter.5
• pitch type: Two-character abbreviation of the type of pitch as classi-
fied by the MLBAM algorithm. See Table B.4.2 for the decoding of the
abbreviations.
• spin dir: Direction of the spin of the ball, where 0◦ indicates a perfect
top spin and 180◦ indicates a perfect bottom spin.
• spin rate: Spin rate of the ball in revolutions per minute.
B.4.3 hip attributes (hit locations data)

While the attributes of both the atbat and pitch elements are located in the
inning all.xml file, hit location data are stored in the inning hit.xml file,
and the elements whose attributes are to be parsed are named hip.
• des: Short description of the batted ball outcome.

• x and y: Horizontal and vertical coordinates of the batted ball, manually
recorded when the ball is first touched by a fielder (at the landing spot for
home runs). The Gameday stringer marks the spot on a 250 × 250 pixel
diagram of the ballpark; the position of home plate and the pixels-to-feet
ratio are different from park to park.
5 Fora visual explanation of the three break attributes, see the figure displayed at the
end of Mike Fast’s glossary.
TABLE B.1
Key for the pitch types abbreviations used by MLBAM.
abbreviation description
AB Automatic Ball
CH Change-up
CU Curveball
EP Eephus pitch
FA Fastball (unspecified)
FC Cut-fastball (cutter)
FF Four-seam fastball
FO Forkball
FS Split-fingered fastball (splitter)
FT Two-seam fastball
IN Intentional ball
KC Knuckle curve
KN Knuckleball
PO Pitchout
SC Screwball
SI Sinker
SL Slider
UN Unknown pitch type
B.5 Special Notes About the Gameday and PITCHf/x

Data
Gameday and PITCHf/x have provided baseball analysts with an incredible
wealth of data, but such amount of information should be used keeping in
mind some issues that are tied to it, some of which are illustrated in Mike
Fast’s article The Internet cried a little when you wrote that on it.6
• Release point estimate: the release point is not directly tracked by

PITCHf/x, but inferred by calculating the full trajectory of the pitch.
Pitchers actually release the ball at different distances from home plate.
That fact combined with the different trajectory of pitch types adds sys-
tematic errors in the release point values which are estimated a fixed dis-
tance from home plate. This issue should make analysts very careful when
presenting evidence of release point changes for a particular pitcher, espe-
cially if comparisons come from different ballparks.
• Pitch classification: The pitch type associated to each tracked pitch

(see Table B.4.2) is obtained by an algorithm developed by MLBAM which
6 www.hardballtimes.com/main/article/the-internet-cried-a-little-when-you-wrote-that-
on-it/.
makes use of the PITCHf/x data. The algorithm has been altered through
the years. While the modifications have generally improved the accuracy
of the classification of pitches, the year-to-year comparison of pitchers’
repertoires require extra caution, as differences might be a product of the
changes in the classifying algorithm.
• Batted ball locations: Batted ball location data are recorded man-
ually and suffer from several problems. First, coordinates systems vary
from ballpark to ballpark, as stringers mark spots on 250 × 250 pixel field
diagrams with inconsistent home plate positioning and pixel-to-feet ratio.
Second, researchers have shown that biases exist due to both the position
(i.e. height) the stringer is assigned at the ballpark and the outcome of the
batted ball. Finally, the stringers are instructed to mark the place where
the ball is collected by a fielder and in case of deflections or caroms off the
walls, it is impossible to infer the original angle of the batted ball.
B.6 Miscellanea
B.6.1 Calculating the pitch trajectory
As seen in the previous sections, PITCHf/x tracks data on location, velocity,
and acceleration of a pitch. Using the kinematics equation for constant accel-
eration, the position of the ball at a given time t can be determined by the
following equations:
1
x = x0 + xv0 t + axt (B.1)
2
1
y = y0 + yv0 t + ayt (B.2)
2
1
z = z0 + zv0 t + azt (B.3)
2
The previous equations are translated to R with use of the following func-
tion pitchloc.7
pitchloc <- function(t, x0, ax, vx0, y0, ay, vy0, z0, az, vz0) {
x <- x0 + vx0 * t + 0.5 * ax * I(t ^ 2)
y <- y0 + vy0 * t + 0.5 * ay * I(t ^ 2)
z <- z0 + vz0 * t + 0.5 * az * I(t ^ 2)
if(length(t) == 1) {
loc<-c(x, y, z)
7 The code in this section has been slightly adapted from code.google.com/p/
r-pitchfx/.
} else {
loc <- cbind(x, y, z)
}
return(loc)
}
The function pitch.trajectory calculates the trajectory of a pitch from
release point to home plate at specified time intervals (the default choice of
the argument interval is 0.01 seconds).
pitch.trajectory <- function(x0, ax, vx0, y0, ay, vy0, z0, az,
vz0, interval = .01) {
cross.plate <- (-1 * vy0 - sqrt(I(vy0 ^ 2) - 2 * y0 * ay)) / ay
tracking <- t(sapply(seq(0, cross.plate, interval), pitchloc,
x0 = x0, ax = ax, vx0 = vx0, y0 = y0, ay = ay, vy0 = vy0,
z0 = z0, az = az, vz0 = vz0))
colnames(tracking) <- c("x", "y", "z")
tracking <- data.frame(tracking)
return(tracking)
}

data: pitchRx
The pitchRx package, contributed by Carson Sievert, provides functions for
obtaining PITCHf/x data from MLBAM and for producing advanced visual-
izations from this data. The scrapeFX is a convenient function for downloading
the data from MLBAM and storing them in R objects; by specifying a start-
ing and an ending date, one obtains PITCHf/x data for every game played
between the given dates.8
The following code uses scrapeFX to download PITCHf/x data for the
games played between May 31, 2012 and June 1, 2012, and save the results in
the object dat. This object consists of a list of two data frames, one containing
information at the plate appearance level (atbat), and the second at the pitch-
by-pitch detail (pitch) including pitch speed, movement, location, and type.
library(pitchRx)
dat <- scrapeFX(start="2012-05-31", end="2012-06-01")
The pitchRx package contains an example dataset pitches, containing
data on every four-seamer and cutter thrown by either Mariano Rivera or
Phil Hughes during the 2011 season. The strikeFX function in this package
is used for creating a heatmap of the pitch locations, split by pitcher and
opponent’s handedness; the result is displayed in Figure B.3.
8 Downloading data requires a few seconds per game, thus massive downloading of data
requires plenty of time.

strikeFX(pitches, geom="tile"
, layer=facet_grid(pitcher_name ~ stand))
L R
5
Mariano Rivera
3
2
density
Height from Ground
0.3
1
0 0.2
5
0.1
4
Phil Hughes
3
0
−2 −1 0 1 2 −2 −1 0 1 2
Horizontal Pitch Location
FIGURE B.3
Locations of four-seamers and cutters delivered by Mariano Rivera and Phil
Hughes in 2011, by batter handedness.
The animateFX function produces several plots that, if displayed sequen-

tially, create animations of the pitch trajectories from the pitcher’s release
point to home plate. With the following code one obtains an animation of
Rivera’s and Hughes’ pitches, which is not displayed.
animateFX(pitches, layer=facet_grid(pitcher_name ~ stand))

The reader is encouraged to read Carson’s Personal Blog 9 and particularly
the pitchRx demo page10 to fully appreciate the potential of the pitchRx
package.
9 cpsievert.wordpress.com/
10 cpsievert.github.io/pitchRx/demo/
B.6.3 Cross-referencing with other data sources

The master table in Lahman’s database is a useful resource for cross-
referencing players across several data sources such as the Baseball-Reference
website and the Retrosheet files. Unfortunately it currently does not con-
tain a column for the MLBAM player identifier, thus the master table is
not useful for merging PITCHf/x data to information coming from other
sources. Baseball Prospectus currently provides a comprehensive and up-to-
date list of players featuring both the Retrosheet and the MLBAM identi-
fiers at www.baseballprospectus.com/sortable/playerid_list.php. The
csv file linked at the top of the page can be read directly into R using the
read.csv function.
players <- read.csv("www.baseballprospectus.com/sortable/
playerids/playerid_list.csv")
Recently another very comprehensive source of players’ identifiers has
been made available online: The Register at Chadwick Baseball Bureau. At
chadwick-bureau.com/the-register/ one finds a link for the download of
a zip file containing a register of players, managers, and umpires at any pro-
fessional level (including, other than the Major Leagues, the Minor and Inde-
pendent Leagues, Winter Leagues, Japanese and Korean top levels, and the
Negro Leagues).
B.6.4 Online resources

The following PITCHf/x resources are currently available on the World Wide
Web. Note that, due to site maintainers being hired by MLB front offices or
exclusive licensing contracts, these resources are subject to being removed or
moved.
• Brooks Baseball (www.brooksbaseball.net/): Created and main-
tained by Dan Brooks, its main features are the Player Cards, consisting of
tables and charts for every pitcher who has ever played in a ballpark with
the PITCHf/x system installed. Tables and charts report information on
characteristics of pitches, their usage (including sequencing), and the out-
comes they produce. The classification of pitches used at Brooks Baseball
is not the MLBAM one, as pitches are classified by Pitch Info LLC. An-
other useful resource of Brooks Baseball is the PitchFX Tool, which allows
site visitors to select one pitcher for one game and obtain a pitch-by-pitch
table.
• Baseball Prospectus (www.baseballprospectus.com/): In its Statis-
tics section, Baseball Prospectus offers PITCHf/x Hitters Profiles,
PITCHf/x Pitchers Profiles, PITCHf/x Leaderboards, and PITCHf/x
Matchups. The building blocks of these resources come from the previ-
ously mentioned Brooks Baseball.
• FanGraphs (www.fangraphs.com/): FanGraphs has PITCHf/x ta-

bles and charts for individual players. For example, pitcher James
Shields’s PITCHf/x page is available at www.fangraphs.com/pitchfx.
aspx?playerid=7059&position=P.
• F/X by Texas Leaguers (pitchfx.texasleaguers.com/): Allows
one to set a time frame and find PITCHf/x pitching or batting data for one
particular player. This site includes charts on trajectory and movement,
tables on pitch characteristics, and outcomes and pitcher/batter match-
ups.
• Prof. Alan Nathan’s The Physics of Baseball (webusers.npl.
illinois.edu/~a-nathan/pob/index.html): Contains research on base-
ball physics and has a section dedicated to pitch tracking using video tech-
nology at webusers.npl.illinois.edu/~a-nathan/pob/pitchtracker.
html.
• Katron’s MLB Gameday BIP Location (katron.org/projects/
baseball/hit-location/): Allows to transpose hit location data of a
given ballpark in another ballpark of choice. Keeping in mind all the
caveats previously illustrated for batted ball data, it can be used to explore
the effect moving to a new team can have on a player’s batting.
• Sportvision (www.sportvision.com/baseball): Sportvision is the
company which has devised the PITCHf/x system. On its website, videos
for other products they exclusively license to MLB teams (such as HITf/x,
FIELDf/x and COMMANDf/x) are featured.
Bibliography
[1] Adler, J. (2006), Baseball Hacks: Tips & Tools for Analyzing and Win-
ning with Statistics, O’Reilly Media.
[2] Albert, J. (2002), “Smoothing career trajectories of baseball hitters,”

manuscript, Bowling Green State University.
[3] Albert, J. and Bennett, J. (2003), Curve Ball: Baseball, Statistics, and
the Role of Chance in the Game, Springer, New York.
[4] Albert, J. (2003), Teaching Statistics Using Baseball, Mathematical As-

sociation of America.
[5] Albert, J. (2008), “Streaky hitting in baseball,” Journal of Quantitative
Analysis in Sports, 4, 1.
[6] Albert, J. (2009), “Is Roger Clemens whip trajectory unusual?” Chance,
22, 2, 9-20.
[7] Albert, J. and Rizzo, M. (2012), R by Example, Springer, New York.
[8] Allen, D. (2009a), “Deconstructing the Non-Fastball Run Maps”,
Baseball Analysts website, http://baseballanalysts.com/archives/
2009/03/deconstructing_1.php.
[9] Allen, D. (2009b), “Platoon Splits for Three Types of Fastballs”, Baseball
Analysts website, http://baseballanalysts.com/archives/2009/05/
platoon_splits.php.
[10] Berry, S., Reese, S., and Larkey, P. (1999), “Bridging different eras in
sports,” Journal of the American Statistical Association, 94, 661-676.
[11] Berry, S. (1990), “The summer of ’41: a probabilistic analysis of DiMag-
gio’s streak and Williams’s average of .406,” Chance 4, 4, 8-11.
[12] Bradley, R. and Terry, M. (1952), “Rank analysis of incomplete block
designs: I. The method of paired comparisons,” Biometrika, 39, 324-345.
[13] Bukiet, B., Elliotte R., and Palacios, J. (1997), “A Markov chain ap-
proach to baseball,” Operations Research, 45, 14-23.
325
326 Bibliography
[14] Cleveland, W. (1979), “Robust locally weighted regression and smooth-

ing scatterplots,” Journal of the American Statistical Association, 74,
829–83.
[15] Cleveland, W. (1994), Elements of Graphing Data, Hobart Press.
[16] Caola, R. (2003), “Using calculus to relate runs to wins: Part i,” By the
Numbers, 13, 9–16.
[17] Davenport, C. and Woolner K. (1999), “Revisiting the Pythagorean

Thorem,” Baseball Prospectus website, www.baseballprospectus.com/
article.php?articleid=342.
[18] Dewan, J. (2006), The Fielding Bible. ACTA Publications.
[19] Dewan, J. (2009), The Fielding Bible Volume II. ACTA Publications.
[20] Dewan, J. and Jedlovec, B. (2012), The Fielding Bible Volume III. ACTA
Publications.
[21] Dolphin, A., Lichtman, M. and Tango, T. (2007), The Book: Playing the
Percentages in Baseball, Potomac Books Inc.
[22] Fair, R. (2008), “Estimated Age Effects in Baseball,” Journal of Quan-

titative Analysis of Sports, 4, 1.
[23] Fast, M. (2009), “What the heck is pitchf/x?” In The Hardball Times
Baseball Annual 2010. ACTA Publications.
[24] Gould, S. (1989), “The streak of streaks,” Chance, 2, 2, 10-16.

[25] Heipp, B. (2003), “W% estimators,” Buckeyes and Sabermetrics website,
gosu02.tripod.com/id69.html.
[26] Humphreys, M. (2011), Wizardry: Baseball’s All-Time Greatest Fielders
Revealed. Oxford University Press.
[27] James, B. (1980), Bill James Baseball Abstract, self-published, Lawrence,
KS.
[28] James, B. (1982), Bill James Baseball Abstract, Ballantine Books.
[29] Kabacoff, R. (2011), R in Action, Manning Publications.

[30] Kemeny, J. and Snell, L. (1976), Finite Markov Chains, Vol. 210,
Springer-Verlag, New York.
[31] Keri, J. and Baseball Prospectus (2007), Baseball Between the Numbers:
Why Everything You Know about the Game Is Wrong, Basic Books.
Bibliography 327
[32] Lahman, S. (2012), “Lahman’s Baseball Database, 1871-2012, v.

2012, Comma-delimited version,” seanlahman.com/files/database/
lahman2012-csv.zip.
[33] Lindsey, G. (1963), “An investigation of strategies in baseball,” Opera-
tions Research, 11, 4, 477-501.
[34] Marchi, M. (2010), “Platoon splits 2.0”, Hardball Times website, http:
//www.hardballtimes.com/main/article/platoon-splits-2.0/.
[35] McCotter, T. (2009), “Hitting streaks don’t obey your rules: Evidence
that hitting streaks aren’t just by-products of random variations,” The
Baseball Research Journal, 37, 62-70.
[36] Murdoch, D.J. and Chow, E.D. (1996), “A graphical display of large
correlation matrices,” The American Statistician, 50, 178-180.
[37] Murrell, P. (2006), R Graphics, Chapman and Hall, Boca Raton, Florida.
[38] Nathan, Alan M. (2013), The Physics of Baseball website, baseball.
physics.illinois.edu.
[39] Palmer P. (1983), “Balls and Strikes,” In Baseball Analyst, Is-

sue 5, February 1983. Available at http://sabr.org/research/
baseball-analyst-archives.
[40] Pankin, M. (1987), “Baseball as a Markov chain,” In The Great American
Baseball Stat Book, 520-524.
[41] R Development Core Team (2013), “R: A language and environment for
statistical computing,” R Foundation for Statistical Computing, Vienna,
Austria, www.R-project.org.
[42] Rickey B. (1954), “Goodby to some old baseball ideas,” In LIFE, August
2, 1954 issue. Available at goo.gl/mZiG5.
[43] RStudio (2013). RStudio: Integrated development environment for R
(Version 0.97.336) [Computer software]. Boston, MA. Retrieved June,
22, 2013. Available from www.rstudio.org.
[44] Sarkar, D. (2008), Lattice: Multivariate Data Visualization with R (Use
R!), Springer, New York.
[45] Schwarz, A. (2005), The Numbers Game: Baseball’s Lifelong Fascination
with Statistics, St. Martin’s Griffin.
[46] Seidel, M. (2002), Streak: Joe DiMaggio and the summer of ’41, Bison
Books.
328 Bibliography
[47] Star, J. (2011), “The Road to October: Sept. 29, 2011,” MLB
website, http://mlb.mlb.com/news/article.jsp?content_id=
25380714&vkey=roadtooctober2011&ymd=20110929.
[48] Triumph Books (2012), 2012 Official Rules of Major League Baseball,
Triumph Books.
[49] Venables, W. N., Smith, D. M., and the R Development Core Team
(2011). “An Introduction to R,” Version 2.13.0 (2011-04-13).
[50] Walsh, J. (2008), “Searching for the games best pitch,” Hard-
ball Times website, http://www.hardballtimes.com/main/article/
searching-for-the-games-best-pitch/.
[51] Walsh, J. (2010), “The Compassionate Umpire,” Hardball
Times website, http://www.hardballtimes.com/main/article/
the-compassionate-umpire/.
[52] Wickham, H. (2009), ggplot2: Elegant Graphics for Data Analysis,
Springer, New York.
[53] Wilkinson, L. (2005), The Grammar of Graphics, second edition,

Springer, New York.
Statistics
Analyzing Baseball Data with R provides an introduction to R for

sabermetricians, baseball enthusiasts, and students interested in
exploring the rich sources of baseball data. It equips you with the
necessary skills and software tools to perform all of the analysis steps,
from gathering the datasets and entering them in a convenient format
to visualizing the data via graphs to performing a statistical analysis.
The authors first present an overview of publicly available baseball
datasets and a gentle introduction to the type of data structures and
exploratory and data management capabilities of R. They also cover
the traditional graphics functions in the base package and introduce
more sophisticated graphical displays available through the lattice and
ggplot2 packages. Much of the book illustrates the use of R through
popular sabermetrics topics, including the Pythagorean formula, runs
expectancy, career trajectories, simulation of games and seasons,
patterns of streaky behavior of players, and fielding measures. All of
the datasets and R code used in the text are available online.
This book helps you answer questions about baseball teams, players,
and strategy using large, publicly available datasets. It offers detailed
instructions on downloading the datasets and putting them into
formats that simplify data exploration and analysis. Through the
book’s various examples, you will learn about modern sabermetrics
and be able to conduct your own baseball analyses.
Max Marchi is a statistician at the Emilia-Romagna Regional Health
Agency. He has been a regular contributor to The Hardball Times and
Baseball Prospectus websites and has consulted for MLB clubs.
Jim Albert is a professor of statistics at Bowling Green State University.
He has authored or coauthored several books and is the editor of the
Journal of Quantitative Analysis of Sports.
K16473

Analyzing Baseball Data With R-Chapman and Hall - CRC (2013)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Analyzing Baseball Data With R-Chapman and Hall - CRC (2013)

Uploaded by

Copyright:

Available Formats

The R Series

Duncan Temple Lang Hadley Wickham

Aims and Scope

Analyzing Baseball Data with R, Max Marchi and Jim Albert

Dynamic Documents with R and knitr, Yihui Xie

Event History Analysis with R, Göran Broström

Programming Graphical User Interfaces with R, Michael F. Lawrence and

R Graphics, Second Edition, Paul Murrell

Reproducible Research with R and RStudio, Christopher Gandrud

Statistical Computing in C++ and R, Randall L. Eubank and Ana Kupresanin

No claim to original U.S. Government works

International Standard Book Number-13: 978-1-4665-7023-8 (eBook - PDF)

1 The Baseball Datasets 1

3.7.1 Parallel stripcharts . . . . . . . . . . . . . . . . . . . . 74

4 The Relation Between Runs and Wins 87

5 Value of Plays Using Run Expectancy 105

5.1 The Runs Expectancy Matrix . . . . . . . . . . . . . . . . . 105

6 Advanced Graphics 129

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

6.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 130

7 Balls and Strikes Effects 161

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

8 Career Trajectories 187

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

10 Exploring Streaky Performances 237

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

10.3 Streaks in Individual At-Bats . . . . . . . . . . . . . . . . . . 242

11 Learning About Park Effects by Database Management

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

12 Exploring Fielding Metrics with Contributed R Packages 283

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

12.2.7 Weighted correlations (weights) . . . . . . . . . . . . 291

A Retrosheet Files Reference 301

A.1 Downloading Play-by-Play Files . . . . . . . . . . . . . . . . 301

B Accessing and Using MLBAM Gameday and PITCHf/x

B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

• “We use: R, MySQL / Oracle, perl, php.”

1.2 The Lahman Database: Season-by-Season Data

1.2.2 Obtaining the database

4. Extract the content of the downloaded zipped file.

1.2.3 The Master table

1.2.4 The Batting table

field name value

1.2.5 The Pitching table

3 ruthba01 1916 1 BOS AL 67 67 136 18 37 5 3 3 15 0 10 23 0 4 67

1.2.6 The Fielding table

Clearly, Ruth’s range as an outfielder deteriorated towards the end of his

1.2.7 The Teams table

1.2.8 Baseball questions

1.3 Retrosheet Game-by-Game Data

1.3.3 Game logs

1.3.4 Obtaining the game logs from Retrosheet

1.3.5 Game log example

1.3.6 Baseball questions

Q In which months are home runs more likely to occur? What

1.6, are reported in Game log files.

1.4 Retrosheet Play-by-Play Data

1.4.2 Event example

Spencer’s throw and flipped it backhand to Posada in time to nail Giambi at

1.4.3 Baseball questions

fielding plays of all time.

1.5 Pitch-by-Pitch Data